Plant disease epidemiology: facing challenges of the 21st Century

Plant disease epidemiology: facing challenges of the

21st Century

Under the aegis of an International Plant Disease Epidemiology Workshop

held at Landernau, France, 10–15th April, 2005

Edited by
S. Savary and B.M. Cooke

Reprinted from European Journal of Plant Pathology, Volume 115 Issue 1, 2006

123
A C.I.P catalogue record for this book is available from the library of Congress

ISBN 1-4020-5019-4 (HB)

ISBN 1-4020-5020-8 (e-book)

Published by Springer,
P.O. Box 17, 3300 AA, Dordrecht, The Netherlands

Printed on acid-free paper

Cover photos:
Patterns of change in multiple pathosystems over space: spatial distribution of four diseases in a groundnut
plot, Côte d’Ivoire, France.
A- Groundnut rust, Puccinia arachidis;
B- Early leafspot, Cercospora arachidicola;
C- Late leafspot, Cercosporidium personatum (Phaeoisariopsis personata),
D- Web blight, Rhizoctonia solani.
Disease assessments were made at 90 days after sowing. Rust, early leaf spot, and late leaf spot: severity
(% diseased leaf area) scales; web blight: incidence (% diseased plants) scale. From Lannou and Savary,
1991, modified.

Springeronline.com
All Rights Reserved

2006 Springer
No part of the material protected by this copyright notice may be reproduced or
utilized in any form or by any means, electric or mechanical,
including photocopying, recording or by any information storage and
retrieval system, without written permission from the copyright owner.

Printed in the Netherlands

Contents

Foreword 1–2
Botanical epidemiology: some key advances and its continuing role in disease 3–23
management
L.V. Madden
Framework development in plant disease risk assessment and its application 25–34
X.B. Yang
Ecological genomics and epidemiology 35–51
K.A. Garrett, S.H. Hulbert, J.E. Leach and S.E. Travers
The practical considerations of scale in plant pathology 53–59
W.W. Turechek
Trends in theoretical plant epidemiology 61–73
H. Scherm, H.K. Ngugi and P.S. Ojiambo
Establishing priorities for plant science research and developing world food security 75–93
R.W. Herdt
Disease assessment concepts and the advancements made in improving the accuracy and 95–103
precision of plant disease data
F.W. Nutter Jr., P.D. Esker and R.A. Coelho Netto
Relation between soil health, wave-like fluctuations in microbial populations, 105–122
and soil-borne plant disease management
A.H.C. van Bruggen, A.M. Semenov, A.D. van Diepeningen, O.J. de Vos and W.J. Blok
Patterns and management of crop multiple pathosystems 123–138
S. Savary, B. Mille, B. Rolland and P. Lucas
European Journal of Plant Pathology (2006) 115:1–2
Springer 2006
DOI 10.1007/s10658-005-5946-6

Foreword

Plant Disease Epidemiology: Facing Challenges of 20th century, and will continue to expand in the
the 21st Century next century. Exchanges of plant materials at very
Plant disease epidemiology deals with diseases in different scales, local to global, have profound ef-
plant populations. During the past century, it has fects on plant diseases. Plant disease epidemiolo-
become a vibrant field of science, achieving signif- gists have become experts in assessing the risk of
icant conceptual innovations with important im- irruption of novel pathogens in plant communities,
pact on the management of plant diseases. Plant the consequences it may have on ecosystems, and
disease epidemiology mobilises concepts and ways to manage such perturbations. The concepts
methods from ecology, genetics, environmental related to biological invasions or population dis-
physics, botany, and mathematics. It deals with placements certainly are not new to plant pathol-
cultivated and non-cultivated plants in environ- ogists: the epidemiological community in fact
ments where human activities have had large, or contributed to craft them in the past century. New
lesser, impact. As in many other fields of science, threats may now also exist, whereby exotic or
plant disease epidemiology faces important, some- novel plant pathogens would intentionally be
times new, questions. By and large, many of these introduced: these threats must be dealt with. The
questions emerge from changes in human societies consequences of plant pathogen transport are
and changes in the status of the planet on which we many: on local performances of spontaneous
live. ecosystems and agricultural ecosystems; on farm-
Global climate is changing at a rapid rate: will it ers’ livelihoods; on local, national, and regional
render plant diseases more, or less, harmful to man- economies; and perhaps more importantly, they
made and spontaneous ecosystems? There is much can have adverse consequences on trade regula-
debate on this issue, because global climate has tion. Will plant disease epidemiologists provide
varying, sometimes very large effects on the local answers to such pressing questions?
environment of growing plant canopies, and be- Biodiversity, a buzzword of the past century, is
cause the physical micro-environment and its var- also of global concern. The decline in global bio-
iation strongly influence plant diseases and their diversity that is currently taking place has been
consequences on ecosystem functioning and per- referred to as the sixth great extinction process our
formance; in addition, changes in global climate planet has experienced during its history, but this
trigger many profound changes in the way ecosys- time, it is man-made. Generations of plant
tems, cultivated or not, are managed. Interest- pathologists, and especially of plant disease epi-
ingly, much of the early literature on botanical demiologists, have been dealing with biodiversity.
epidemiology dealt with climate-disease or climate- The huge diversity of life that resides in the rhi-
pathogen relationships – in fact these kinds of zosphere and the phyllosphere are causes both of
relationships have long been perceived as the bulk diseases in plants, and of their suppression. Much
of epidemiological research by many. Plant disease current research is addressing ways of harnessing
epidemiologists thus have a strong scientific tradi- such biodiversity not as enemies – of which
tion in studying climate-pathogen-disease rela- pathogens are an inherent part – but rather as
tionships. Can such an asset be mobilised by the important biological allies to control disease epi-
epidemiological community to answer questions demics. The diversity of plants is another facet of
about the effect of climate change on plant diseases? global biodiversity, and there are concerns about
Global trade, and thus, trade of plant products, the decline in the genetic diversity of crop plants. It
have increased at an unprecedented rate during the is from this diversity that possibly the most potent
2

instrument for disease management has been Is it so that globalised exchanges, novel biological
developed by plant pathologists: genetic host plant technologies, and the self-regulating mechanisms
resistance. Will we run short of resistance genes of trade, will be sufficient to fulfil the needs of
against major plant pathogens? Host plant diver- future generations? Will these not have negative
sity, and the disease resistance genes it harbours, side-effects, and will they truly prevent the current
can be deployed over time and space, according to over exploitation of natural resources, water and
epidemiological principles. In-depth knowledge of land in particular?
the characteristics of individual pathogens causing Sustainable production and crop protection
specific diseases that must be controlled has been systems need to be devised, which could exploit
mobilised to develop appropriate strategies at the scarcer resources sparingly, and if possible en-
plant population, field, landscape, and sub-re- hance the resource base. Can these production and
gional levels. Major successes have been achieved protection systems be designed so that they gen-
using such strategies, and the end of the past erate healthy, high-quality products that would
century has seen their recognition by the scientific find niche markets both locally and globally, and
community. Will epidemiologists succeed in the so provide farmers with the income they require,
future in fully sharing these technologies with the and offer consumers products that suit their needs
farmer so that they are more fully utilised? and their incomes? Plant disease epidemiologists
Food security was a central concern of the alone cannot provide answers to such questions,
global agricultural research community in the but certainly could significantly contribute to these
middle of the 20th century, but apparently, not new strategies.
anymore. However, the world population still The five-day International Plant Disease Epi-
increases, and is expected to do so for several demiology Workshop (held 10–15th April, 2005,
decades. One out of six human beings living on in Landernau, France, the ninth of a series) re-
earth today suffers from lack of food. Many of ported in this special issue of the European
today’s poor live in cities, with no access to land Journal of Plant Pathology, obviously could not
and agriculture, and most of the projected in- address all of these issues, and others, with all the
crease in the world population will take place in depth good science demands. However it provided
the world’s largest cities. Pests, including plant a unique opportunity for scientists interested in
pathogens, cause losses in pre-harvest yield in the this field to meet and face challenging questions,
range of 20–40%; estimates of post-harvest losses contribute to animated debates, and reflect on the
are inadequate, but it is a fair assumption that future development of the science of plant disease
they are often higher than 10 or 20%. Why are epidemiology.
our estimates – the raison d’être of plant pathol-
ogy – still so vague today? Seldom do economists SERGE SAVARY
currently address the issue of food security – why? MIKE COOKE
European Journal of Plant Pathology (2006) 115:3–23
Springer 2006
DOI 10.1007/s10658-005-1229-5

Botanical epidemiology: some key advances and its continuing role in disease
management

Laurence V. Madden
Department of Plant Pathology, The Ohio State University, Wooster, OH, 44691-4096, USA
(E-mail: [email protected])

Accepted 13 July 2005

Key words: basic reproduction number, Bayesian decision theory, epidemic, mixed models, model, pop-
ulation dynamics, receiver operating characteristic (ROC) curve

Abstract

Epidemiology involves the study of the temporal, spatial, and spatio-temporal dynamics of disease in
populations, and the utilization of results of experiments and surveys to describe, understand, compare, and
predict epidemics. Such understanding and description of epidemics can lead directly to the development
and evaluation of efficient control strategies and tactics. Mathematical and statistical models are key tools
of the epidemiologist. Recent advances in statistics, including linear and nonlinear mixed models, are
allowing a more appropriate matching of data type and experimental (or survey) design to the statistical
model used for analysis, in order to meet the objectives of the investigator. Coupled ordinary and partial
differential equations, as well as simpler growth-curve equations, are especially useful deterministic models
for representing plant disease development in fields in time and space over single seasons or many years,
and their use can lead to appraisal of control strategies through metrics such as the basic reproduction
number, a summary parameter that may be calculated for many general epidemic scenarios. Recently,
compelling arguments have been made for the use of Bayesian decision theory in developing and evaluating
real-time disease prediction rules, based on measured disease or weather conditions and either empirical or
mechanistic models for disease or control intervention. Through some simple calculations of predictor
accuracy and (prior) probability of an epidemic (or the need for control), the success of any predictor can
be quantified in terms of the estimated probability of random observations being epidemics when predicted
to be epidemics or not epidemics. Overall, despite the many contributions in epidemiology over the past
four decades, more effort is still needed to convince those outside of epidemiology to more fully use
epidemiological results and insights into the development and evaluation of disease controls.

Introduction growth in the discipline within plant pathology

during the 1960s, 1970s, and 1980s (e.g., Campbell
In 1963, van der Plank made a most compelling and Madden, 1990; Kranz, 1990; Jones, 1998;
case for the importance of botanical epidemiology, Zadoks, 2001) has been eclipsed by growth in the
both for understanding plant diseases at the pop- larger field of molecular biology over the last two
ulation scale and for determining disease man- decades. Nevertheless, more than 40 years after
agement strategies (van der Plank, 1963). He also van der Plank’s book (1963), botanical epidemi-
made the bold statement at the time that ‘epide- ology is still here, and still of utmost importance in
miology is here to stay.’ Individual disciplines en- giving a sound theoretical and practical basis for
joy ‘ups and downs’ of popularity, of course, and disease management. This view may not always be
epidemiology is no exception. The tremendous held outside of the discipline, however, and it
4

remains a challenge for epidemiologists to con- ida, karnal bunt in Arizona, and plum pox in
tinue to make the compelling case that epidemi- Pennsylvania, U.S., are three examples of disease
ology matters. invasions (Gildow et al., 2004; Gottwald et al.,
Until molecular biology or more traditional 2001; Rush et al., 2005).
breeding results in durable resistance to all plant New pathogens (or pathogens new to a given
pathogens on all crops, coupled with the accep- crop) continue to be discovered, as well as strains,
tance of the new cultivars by growers and the races, or biotypes of previously known pathogens.
public, there will be plant disease epidemics, and The new very aggressive biotype of African cas-
many of these will result in substantial reductions sava mosaic virus in Africa is an example of a
in yield. There is certainly increasing use of crop newly evolved isolate (Legg, 1999; Strange and
GMOs around the world (James, 2003), but cul- Scott, 2005) that is proving very difficult to con-
tivars with very broad-acting and durable resis- trol. Sudden oak death, caused by Phytophthora
tance have yet to be developed. Moreover, the ramorum, is a newly identified disease of oak and
public opinion against their use remains strong in several other plant species, which is spreading
many regions; thus, it would be naı̈ve to expect naturally and (unfortunately) with the assistance
‘super resistant’ cultivars in the foreseeable future. of man, in the U.S. and elsewhere (Rizzo et al.,
Use of fungicides and other chemicals in a pro- 2002).
tectant or curative manner is only practical for For diseases such as sudden oak death or Asian
some crops and some diseases, and there is soybean rust (newly introduced into the U.S.),
increasing societal pressure to (drastically) reduce there is a great need to know the extent of spread
the use of these chemicals in many regions. Thus, a from current locations (e.g., from the point of
scientific basis for applying or not applying introduction) to other locations. For any disease
chemicals is needed, and the decision clearly in- that is locally concentrated (e.g., around the point
volves knowledge (or prediction) of the disease of a new introduction), or does not yet exist in a
dynamics under different environmental condi- country or region, ethically one cannot deliber-
tions. The development of resistance to fungicides ately introduce the pathogen where it does not
and antibiotics continues, and new cultivars have a occur in order to study spore movement and
finite lifetime. resulting disease intensity. Thus, modelling is a key
No control tactics are known that will totally research tool for understanding risks based on key
eliminate epidemics in crops and forests where the epidemiological characteristics or traits of a dis-
pathogen is present over large areas. Biological ease (Madden and van den Bosch, 2002; Madden
and cultural controls may be very beneficial, and Wheelis, 2003). Epidemiology as a discipline
depending on the pathosystem (Maloy, 1993), but depends heavily on the tools of mathematical and
variability of control efficacy may be high with the statistical modelling (Campbell and Madden,
former, and grower acceptance may be low with 1990), so epidemiologists are, in general, quite
the latter (e.g., unwillingness to rotate crops). prepared to tackle the problem of disease spread
The public and the scientific community have through modelling. Model parameters for these
been definitely reminded of the importance of types of situations can be obtained from observa-
epidemiology, and the research tools that epi- tions where the disease of interest does occur
demiologists can bring to a problem, in recent naturally.
years. A few examples are given. With increasing Most practicing epidemiologists would strongly
world trade of agricultural commodities as well as support van der Plank’s (1963) statement that
international travel, the risk of pathogen invasion epidemiology sets the strategy for disease control,
of new countries or regions is well recognized and numerous examples can be given where epi-
(NRC, 2002), and predictions of the risk of inva- demic knowledge leads to better control (Zadoks
sion involve many epidemiological characteristics and Schein, 1979; Fry, 1982; Maloy, 1993). Fur-
of pathogens, such as survival probabilities and thermore, epidemiological principles and results
reproductive potential (Madden and Wheelis, can also lead to specific control recommendations,
2003). Moreover, the decision to attempt to erad- through the process of disease forecasting or risk
icate or not also involves knowledge of disease prediction (Hardwick, 1998; Hughes et al., 1999),
epidemiology. The cases of citrus canker in Flor- as demonstrated 45 years ago (Waggoner, 1960).
 5

However, as pointed out recently by Jeger (2004), with the same model (depending on a realized
many controls are utilized and evaluated without value of a shape parameter). A feature of these
explicit consideration of disease dynamics in fields. models is that they are all based on a single re-
Although there is great danger in basing conclu- sponse variable (disease intensity, y) in relation to
sions on disease intensity measured at one time in continuous time, which can be obtained as a
an epidemic (especially for polycyclic diseases; see solution for the rate of change of y with time, dy/dt
Campbell and Madden, 1990), this unfortunately [e.g., dy/dt=rLy(1)y) for the logistic model]. In
happens too often. Thus, epidemiologists still need some cases, the solution can be expressed as a
to be pro-active in working with others in devel- linear model, e.g., logit(y)=a + rLt, where a is a
oping and evaluating disease control measures. transformation of disease intensity at time 0, rL is
In the remainder of this article, I discuss a few the per capita rate parameter, and logit(y) is a
developments that I consider to be very important linearizing transformation of y.
in the development of plant disease epidemiology. A good fit of an empirical model, or even a
Many more topics could have been covered. I have perfect fit, to data collected over time, is not proof
two major themes. One deals with the advance- of any mechanism for population growth (Camp-
ment in our theoretical understanding of the bell and Madden, 1990; Zadoks, 2001). But a good
population-dynamic processes of disease spread in fit of a particular model for several disease pro-
space and increase in time, coupled with the gress curves could lead one to hypothesize about
improvements in relating certain models (or their mechanisms, and then test the hypothesis with
parameters) to empirical results (i.e., model fit- additional data or experiments. Moreover, using a
ting). The other theme deals with the prediction of model that provides a (reasonably) good fit to data
plant disease on a real-time basis, or prediction of is extremely important to accurately compare
the need to impose a control measure, based on epidemics; among other things, using an inappro-
principles from probability theory. Citations are priate model will lead to biased estimates of the
deliberately sparse, and are mainly to major re- rate parameter and its standard error (Neter et al.,
views of topics rather than to all the (many) 1983).
important original papers published over the last One clear trend in botanical epidemiology is the
few decades. I assume throughout that modelling dramatically increasing complexity of statistical
and statistical data analysis are methodological models and methods that have been applied to all
foundations for understanding epidemics and uti- epidemiological data over the last few decades (e.g.,
lizing any gained knowledge in disease control. Gilligan, 2002; van Maanen and Xu, 2003). This is
a natural development given the fact that epide-
miology is a science of populations, and popula-
Temporal and spatial dynamics of disease tions can only be adequately characterized and
compared using the methodology of statistics. Al-
Growth curve modelling and analysis though I am sure there are some who feel that the
emphasis on mathematics and statistics obscures
Van der Plank (1963) used the monomolecular and the understanding of the biology of epidemics, I
logistic equations as heuristic models of monocy- would make the opposite claim, and declare
clic (simple interest) and polycyclic (compound emphatically that mathematical and statistical
interest) disease epidemics. These models continue modelling are foundations for understanding epi-
to be the benchmarks for quantification of epi- demics. I further believe that, with some excep-
demics, especially over single growing seasons. tions, the use of statistical analysis is actually still
However, plant pathologists discovered in the inadequate in most of epidemiological research,
1960s and 1970s that these two models did not and certainly in most of plant pathology research!
necessarily provide an adequate description (based Many investigators still only: measure disease at a
on statistical principles of model fitting) for many single time, do not match the chosen form of data
disease progress curves (Campbell and Madden, analysis to the type of disease intensity variable
1990). Several alternative models were proposed or (discrete for incidence, continuous but unequal
developed, some of them flexible in the sense that variance for severity, ordinal for many disease
different degrees of skewness could be represented rating scales); do not base their analysis on the
6

chosen experimental design; or perform inefficient progress model parameters for each epidemic, with
(and sometimes uninformative) analyses. An a follow-up analysis of variance, through mixed
example of the latter is the still common practice of models one can simultaneously estimate the dis-
performing a separate data analysis for each ease progress parameters and their appropriate
assessment time during an epidemic rather than standard errors based on the explicit features of
simultaneously analyzing treatments (between- the design. The former approach (e.g., estimated
subject factors) and time (within-subject factors), slope for each plot, and then an ANOVA of these
and their interactions. Garrett et al. (2004) and slopes), still common with researchers, is known to
citations therein can lead the reader to some of the be the least powerful approach to detect differ-
important recent advances in statistical data anal- ences in treatments (Wolfinger, 1996). Through
ysis of relevance in plant pathology. these mixed-model methods, random effects (such
It has been known for many years (Madden, as locations, blocks, and possibly genotypes), and
1986) that disease values collected over time in the their interactions with fixed effects (treatments)
same experimental or sampling unit (e.g., plot) are can be appropriately estimated and realistic infer-
serially correlated and that the variation in disease ences made.
over time within plots is different from the varia- Many population dynamic processes can be
tion between plots. This may be in part due to the expressed only in nonlinear form (e.g., y=f(t; a,b),
cumulative nature of disease progress curves (see where f(d) is a nonlinear function). The recent
pp. 521–522 in Schabenberger and Pierce, 2002, advances in nonlinear mixed models (Garrett
for general discussion of cumulative processes over et al., 2004) can be applied to these situations, but
time). Serial correlations, sometimes called tem- the range of experimental designs is much more
poral autocorrelations, are especially troublesome limited (currently), and considerably larger data
in the comparison of treatments. My recent studies sets are required to estimate and compare param-
now show, however, that fitting of appropriate eters. Nevertheless, statistically savvy and moti-
population-growth models to disease progress vated epidemiologists can make considerable
data often reduces the correlation of residuals to progress here.
near zero for individual disease progress curves,
reducing the need to directly utilize cumbersome Mechanistic modelling (linked differential
adjustments to standard errors for calculated rates equations)
(unpublished). However, in the larger setting of
multiple disease progress curves, corresponding to Van der Plank (1963) clearly realized that models
multiple treatment factors and blocks, there will such as the logistic were inadequate for a biologi-
always be non-zero correlations of observations cally meaningful characterization of disease pro-
within the plots by the nature of the experimental gress in time. His approach was to use a so-called
design (Schabenberger and Pierce, 2002). How- differential-delay equation in order to represent
ever, the structure of the correlations and vari- polycyclic disease development. This model relates
ances may be quite complex, due to the cumulative dy/dt to the infectious disease intensity rather than
process of disease development, but simple vari- to total disease intensity, with infectious disease
ance-covariance models can adjust for this prop- estimated based on assumed fixed-duration latent
erty. For disease progress models that can be and infectious periods. Although the use of dif-
expressed in linear form through the use of a ferential-delay equations serve as a good founda-
transformation of y [e.g., logit(y)], linear mixed tion for developing computer simulation models
models provide a tremendous (and still underuti- with fixed time steps, such equations are extremely
lized) tool for a thorough analysis of the epidemics cumbersome for mathematical analysis, making it
(Garrett et al., 2004). Most plant pathologists difficult to explore implications of different bio-
(including epidemiologists) are not aware of the logical properties of hosts and pathogens, or of
major advances made in mixed model analysis in different control strategies, on long-term disease
statistics, a field that encompasses classical ANO- development. Eventually, plant pathologists dis-
VA and regression, and many other topics in a covered the mathematical elegance of linked or
unified manner (Schabenberger and Pierce, 2002; coupled differential equations for characterizing
Garrett et al., 2004). Instead of estimating disease disease progress (Jeger, 1986a, b; van Maanen and
 7

Xu, 2003). The approach – which was utilized as mean latent period, 1/l is the mean infectious
long ago as 1911 for representing malaria epi- period, and b is the per capita transmission rate
demics (Ross, 1911) – is to use two-to-several (new diseased individuals per diseased individual
differential equations, with some variables of per healthy individual per unit time). For fungal
interest and parameters appearing in more than (or oomycetes) diseases, b is the product of spore
one of the equations. The beauty of this approach production per time unit per infectious individual,
is that new terms can be easily added, as needed, to the probability that a spore comes in contact with
meet the objectives of the investigator and the a healthy individual, and the probability that a
details of the pathosystem, and asymptotic and spore in contact with a healthy host individual
steady state results (such as disease persistence) causes an infection. Total disease at any time is
can be explored quantitatively. Furthermore, even determined as Y=L+I+R, and disease intensity
though analytical solutions cannot generally be as a proportion is given by y=Y/(H+L+I+R). If
obtained (i.e., one cannot write out y as a function initial disease intensity is very low, then at t=0,
of parameters and time without the use of the initial total host density is virtually the same as
integral symbol), numerical solutions are now easy initial healthy host density, H0. The product bH0 is
to obtain with many mathematical programmes analogous to van der Plank’s (1963) corrected
such as MATHCAD and MATHEMATICA. basic infection rate (new diseased individuals per
Statistical software such as PROC MODEL of the diseased individual per unit time).
SAS/EST system allows direct parameter estimation A fundamental result with this model is that
of one or more parameters for these types of disease will increase (i.e., an epidemic will occur)
models (Madden et al., 1987). The approach is only if bH0/l>1. The expression to the left of the
iterative and computationally intensive, but read- inequality is known as the basic reproduction
ily accomplished by those who have a good number, R0 (Diekmann and Heesterbeek, 2000).
understanding of nonlinear models and statistics. This composite parameter also indicates the final
However, unlike the case for models with analyt- intensity of disease (after a long time) and the
ical solutions (linear or nonlinear; see previous initial exponential rate of increase (see Segarra
sub-section above), one cannot easily incorporate et al., 2001, for details). An example realization of
the features of the experimental design (e.g., split the model in equation 1 is shown in Figure 1 for
plot, etc.) into the model fitting. Rather, one gen- the situation with R0=2.5. Final disease is less
erally needs to estimate parameters for each indi- than 100%, and is estimated by iteratively solving
vidual epidemic (e.g., each field or plot) and then y¥=1-exp()R0y¥). Control strategies are devel-
perform t-tests or analysis of variance on the oped or evaluated by finding combinations of b,
estimated parameters (depending on the experi- x, and l that give R0 < 1; specific control tactics
mental design). (e.g., host resistance, protectant fungicide, cura-
A relatively simple coupled differential equation tive fungicide) can then be directed at reducing b,
model for a polycyclic disease with no plant mor- etc.
tality is given by: An advantage of the equation 1 formulation is
the easy expansion for other situations. For in-
dH
¼
bHI stance, a simple-interest disease component
dt (infections from resident inoculum throughout the
dL epidemic, rather than just at the start) can be
¼ bHI
xL
dt incorporated by using the pxH term, where x is the
ð1Þ
dI density of inoculum and p is a simple-interest rate
¼ xL
lI parameter. One can consider x to be constant or to
dt
change (typically, decline over time), so that dx/
dR
¼ lI dt=Jx. When x does not change, then px is
dt
equivalent to the monocyclic rate parameter (rM)
where H, L, I and R are the densities of disease- of the monomolecular model. The pxH term is
free (healthy), latently infected, infectious, and subtracted from dH/dt and added to dL/dt in
post-infectious (removed) individuals (e.g., plants, equation 1. A pure simple-interest epidemic results
leaves, roots, or even sites on leaves), 1/x is the if b=0; otherwise, a composite of polycyclic and
8

1.00 dH
H ¼
bHI
pxH þ gðHmax
HÞ
Y dt
Disease intensity

0.80
dL
¼ bHI þ pxH
xL
gL
0.60 R dt
dI
0.40 ¼ xL
lI
gI ð2Þ
Fixed host dt
0.20 I dR
¼ lI
gR
L dt
0.00
0 50 100 150 200 250 dx
¼
#x
Time (t) dt

1.00 Note that in this example, total host size

H Dynamic host
(H+L+I+R) does not change, even though there
Disease intensity

0.80
Y is continuous loss and addition of the host indi-
0.60 viduals (with a balance between the additions and
R losses). This can be seen by noting that
0.40
Hmax=H+L+I+R and adding the rates: dH/
0.20 dt+dL/dt+dI/dt+dR/dt=0. The model can be
I written in different ways to unlink the growth and
L
0.00 mortality, to incorporate more complicated link-
0 60 120 180 240 300
Time (t) ages, and to account for more than one disease or
more than one host genotype at a time, but the
Figure 1. Density of healthy (H), latently infected (L), infec- example is useful to show one model formulation.
tious (I), and post-infectious (R) individuals (on a proportion When p=0 (no simple interest component), an R0
scale), together with total disease (Y=L+I+R), based on
equation 1 (upper frame) and equation 2 (lower frame). Mean
can be defined for many host-growth/mortality
latent period (1/x) was 7, and mean infectious period (1/l) model situations. For instance, with p=0 (no
was 10 time units. Upper frame: bH0=0.25 per time unit. simple-interest component), R0=[bHmax/
Lower frame: bH0=0.35 per time unit, g=0.02, and p=0 (no (l+g)]Æ[x/(x+g)]. An example realization of this
simple-interest component). Because of proportion scale, y model is shown in the lower frame of Figure 1.
and Y are the same here.
Note that Y (=L+I+R) and H oscillate a little
before settling down to the steady states. The
monocyclic processes occurs over time, very steady-state level of disease at a given R0 is lower
typical for root diseases (Gilligan, 2002). Host for the dynamic host than the fixed-host situation
mortality can be incorporated by using a death- (equation 1); without the simple-interest compo-
rate parameter g. Then gH, gL, gI, and gR are nent, the steady state Y is 1)(1/R0).
subtracted from the right hand sides of the equa- This approach of using a dynamic (but fixed
tions for dH/dt, dL/dt, dI/dt, and dR/dt, respec- total) host population size has been used in plant
tively. Host growth can be incorporated in various disease epidemiology (e.g., Madden et al., 2000),
ways. One approach is to consider just a single per and even more so in medical epidemiology
capita growth rate (X) for disease-free individuals, (Anderson and May, 1991) to determine whether
and add the term X to the right hand side of the or not an epidemic can occur (i.e., a disease inva-
dH/dt equation. Suppose, further, that host size sion) as well as the persistence (or not) of disease
(e.g., number of citrus trees in a region) is fixed long term. With primary infections occurring
(say, at Hmax), and that new trees are only planted throughout the epidemic (p>0), the concepts be-
if others die. Then, the growth rate is also the come a little more complicated, but there may still
mortality rate, and new host individuals can be be a threshold (combination of parameters) that
expressed as X=gHmax; the combined growth/ must be met for disease to persist (see review in
mortality for H can then be written as g(Hmax)H). Gilligan, 2002, and references cited therein).
A more general epidemic model can be writ- Many other biological features can be incorpo-
ten as rated in the model of equation 2. For instance,
 9

since most plant viruses are transmitted by dimensions, we need to use partial derivatives
arthropod vectors, the rate of change in H and L rather than ordinary derivatives. Expanding
does not directly depend on infectious plant indi- equation 1, we can write the spatio-temporal
viduals (I) but on infective vectors per plant (Z). model as:
Thus, the contact rate term, bHI in the first two Z1
equations of the model must be replaced by bHZ, @Hðt;sÞ
¼
bHðt;sÞ Iðt;nÞDðs
nÞdn
where Z is the density of infective vectors per @t
plant. Other components would be unchanged.
1

There is also a need to add equations for the @Lðt;sÞ

dynamics of the vector population, including ¼ bHðt;sÞ
@t
virus-free and infective vectors. Details are given in Z1
Madden et al. (2000) and Jeger et al. (2004). Other
 Iðt;nÞDðs
nÞdn
xLðt;sÞ
expansions can incorporate disruptions caused by
harvesting and/or planting for a multi-season time
1

scale, as well as host responses to infection (e.g., @Iðt;sÞ

¼ xLðt;sÞ
lIðt;sÞ ð3Þ
Gilligan, 2002; Madden and van den Bosch, 2002). @t
The models shown so far are all deterministic.
@Rðt;sÞ
These can all be expressed in stochastic form, ¼ lIðt;sÞ
which is useful if one is specifically interested in @t
heterogeneity of epidemics, small population sizes, where all parameters are as defined before, and
or the epidemic outcome for individual plants or D(s)n) is the contact distribution, which is simply
plant units. Gilligan (2002) and Gibson et al. a scaled version of a disease gradient. Example
(1999) provide more details. The mathematics contact distributions include the exponential, Pa-
definitely becomes more difficult with stochastic reto, Cauchy, and normal. Unlike with the simpler
models. purely temporal model(s), the rate of decline in
healthy host individuals at location s (and the rate
Some spatial aspects of epidemics of increase in latently infected host individuals at
s) is explicitly based on the integration of the
There are two different threads to the character- contributions of infectious individuals at all loca-
ization of the spatial component of plant disease tions (all n values). The specific contribution at n
epidemics. One thread deals with dispersal and to disease at s is the product of magnitude of
resulting disease gradients, and the use of observed infectious individuals at n multiplied by the
gradients to elucidate the form of the contact probability that a unit of inoculum (say, spore) at
distribution (Campbell and Madden, 1990), the n reaches location s (based on the contact
probability of a unit of inoculum at one location distribution).
(n) coming in contact with a host individual at Both so-called wave-like and non-wave-like
location s. This approach has been especially disease expansion is documented, where the
valuable for determining the rate of disease velocity of disease expansion into new areas is
expansion from a focus, both within fields and constant or increases with time, and supported by
higher spatial scales (e.g., continents) (van den the theory summarized in equation 3. The velocity
Bosch et al., 1999). The contributions of van den of expansion (or the acceleration of expansion) is
Bosch and Zadoks (see Zadoks, 2001), Ferrandino generally proportional to ln(R0), so that there is no
(1993), and Aylor (1999) are especially noteworthy spread if R0 £ 1. The form of the contact distri-
for aerial pathogens, and of Gilligan and col- bution makes the difference in type of expansion.
leagues (2002) for root diseases. An example realization is shown in Figure 2 for
One of the advantages of the coupled differential non-wave-like expansion. The linkage of temporal
equation approach of the previous section is that it population dynamics of disease and focus expan-
can be directly expanded to account for disease at sion rates is of fundamental importance because it
any location as well as any time. With two physical shows (qualitatively and quantitatively) how
dimensions, it is now necessary to be explicit in reproduction (infection) and contact probabilities
notation about time t and location s. With two (dispersal) fully determine spatio-temporal
10

1000 studying epidemics because there is no single

spatial starting point. With many original starting
Disease intensity

800 points (foci with disease at time 0), numerical

600 solutions to equation 3 – or solutions to stochastic
analogues of equation 3 (Xu and Ridout, 1998;
400 van Maanen and Xu, 2003), – can be used to de-
scribe epidemics and explore implications of bio-
200
logical and physical features on disease progress,
0 but it is more difficult to develop general principles
-4 0 -2 4 -8 8 24 40 or characterize expansion rates. Moreover, fitting
Distance (s) a model such as equation 3 to data is generally
impractical with standard statistical programmes.
1000 Thus, in epidemiology – as in ecology (Pielou,
Disease intensity (log scale)

1977) for that matter – more statistical (rather

100 than mathematical) approaches have been gener-
ally followed to study spatial aspects of epidemics
10 (Madden and Hughes, 1995, 2002; Hughes et al.,
1997). This is the second thread of spatial char-
1 acterization of epidemics. Concepts of clustering,
aggregation, and regularity are utilized in terms of
0. 1 many different (but interrelated) statistical meth-
-4 0 -2 4 -8 8 24 40 ods such as indices of dispersion, correlation, semi-
Distance (s) variograms, and distance statistics. This concep-
tual approach goes back to Cochran (1936) and
Figure 2. Density of diseased individuals (Y=L+I+R) vs. Bald (1937) in plant pathology. A further advan-
distance from a line source at 10-day time increments based
on the numerical solution of equation 3. H0=1000. Mean la-
tage of the statistical approaches is that results (or
tent period (1/x) was 7, and mean infectious period (1/l) was concepts) are often directly useful for developing
10 time units. bH0=0.4 per time unit. A Pareto distribution sampling plans, for either estimating disease
was used for the contact distribution. The horizontal distance intensity or making a decision regarding a control
between pairs of successive curves at a single Y value (e.g., intervention (Madden and Hughes, 1999; Hughes
0.1), divided by 10 gives the velocity of disease expansion.
et al., 2002).
The interrelationships between spatial aggrega-
outcomes, given a set of initial conditions. Control tion of disease and temporal dynamics is gradually
strategies are based, once again, on reducing R0 to becoming more apparent. Using stochastic simu-
below 1, as well as reducing the scale of the contact lation, Xu and Ridout (1998) nicely showed how
distribution (spread parameter of the dispersal initial conditions, reproduction, and spatial con-
gradient) to a low value. tact distribution affect disease dynamics. A more
Equation 3 can be expanded for host growth, theoretical approach has been to incorporate
simple-interest dynamics, and so on, just as equa- spatial properties of epidemics without explicitly
tion 1 was expanded to equation 2. It is (much) using a spatial dimension (i.e., using models simi-
more difficult to work with partial differential lar to equation 1). Models of this type are some-
equations than with ordinary ones, and finding times called spatially implicit, in contrast to the
numerical solutions can even be tedious. When the spatially explicit ones such as equation 3. The
epidemic starts with a single focus (say, at the edge approach generally involves using a nonlinear
or centre of a region), then mathematical progress function of I and/or H in the contact term, where
can be made, usually with additional assumptions the function depends on degree of aggregation
(van den Bosch et al., 1990). (Zhang et al., 2000).
When there are several initial foci of infections, In recent years there has been considerable
or unknown number and locations of initial inoc- progress in bringing the two threads together
ulum, spatio-temporal differential-equation mod- (Gibson, 1997; Keeling et al., 2004), through the
els, such as equation 3, are much less useful for ingenious use of stochastic models and parameter
 11

estimation. The results are primarily for the situ- of parameters and variables compared with large
ation where individual plants are spatially refer- multi-variable and multi-parameter systems mod-
enced, and disease intensity is measured as a els is relevant here.
binary variable (diseased or healthy). There is still Thus, for many objectives, relatively simple
more work to do in this area, both in terms of models – such as the logistic, exponential, and
testing the new approaches and for expanding monomolecular disease progress models, and
approaches for other spatial situations (e.g., spa- empirical regression equations – will continue to
tial referencing of just sampling units, not indi- be indispensable tools for the epidemiologist
vidual plants) and other measures of disease (Jeger, 2004). Although these models clearly are
intensity (e.g., severity). approximations, so are more complicated models.
As stated by Bertrand Russell, ‘Although this may
General thoughts on spatio-temporal disease seem a paradox, all exact science is dominated by
dynamics the idea of approximation’ (Auden and Kronen-
berger, 1966). Whether or not a model is too much
There is no doubt that through the expansion of of an approximation will always depend on the
models such as equations 1–3, as much detail as needs of the investigator.
desired can be incorporated into models of plant
disease epidemics (van Maanen and Xu, 2003).
Such expansions require both knowledge of the Decision making in epidemiology
pathosystem and knowledge of mathematics to
realistically link model structure and parameteri- The case for disease forecasting
zation to meaningful population-dynamic proper-
ties of the disease. Even with a model with just a As stated by Gilligan (1985), ‘Of the potential
few parameters, such as equation 3, mathematical benefits of mathematical modelling to improving
insight may no longer be feasible unless starting the efficiency of control of crop disease, prediction
conditions are restricted (e.g., one initial focus). stands foremost.’ Sometimes predictions or fore-
Once other complicating factors are introduced, casts can be based explicitly on the rate parameter
such as incorporation of environmental effects on of a model such as the logistic or exponential (or
the parameters (i.e., turning parameters into new even more complicated mechanistic model), as
variables that are functions of environment and done with EPIPRE (Hardwick, 1998). That is, one
new parameters), results will be limited to inter- can either use rL to predict disease intensity some
pretation of numerical simulations with the time period into the future based on either: (1)
model(s). calculated rL from previous estimates of disease in
Although a model can be made indefinitely the current epidemic; or (2) predicted rL based on
complex to represent an indefinitely complex bio- environment (etc.), where the equation was devel-
logical system (such as plant pathosystem), such a oped in other studies. However, predictions need
model would violate the important principle of not necessarily be tied to population growth
parsimony – keeping the model as simple as pos- models in an explicit manner. A wide range of
sible for the objectives of the investigator. Models, empirical models (often derived from regression or
by definition, are simplifications of reality, which discriminant analysis) are utilized to simply iden-
are useful for many purposes, including descrip- tify conditions leading to a disease outbreak or a
tions, comparisons, statistical inference, predic- large reduction in yield (Madden and Ellis, 1988;
tion, and developing understanding. Constructing Campbell and Madden, 1990). In fact, the pre-
models that are more complex than needed to meet diction model (risk algorithm) may actually be
the needs of the investigator – whether or not the derived without any formal statistical analysis; a
basic biological knowledge is available for the good example of this is the collection of early
construction of the model – is inefficient and can prediction models for late blight of potato
lead to faulty conclusions because of unrecognized (Hardwick, 1998).
(possibly erroneous) properties of the complicated Epidemiologists continue to develop new pre-
model. The conclusions of Jeger (1986a) regarding diction systems for plant diseases, usually used for
the value of models with (relatively) small numbers scheduling fungicide applications, that is, for
12

decision-making in real time regarding the need for Fusarium head blight of wheat in north America.
a control intervention (e.g., spray or not spray) The model, which is really a prediction rule in this
(Madden and Ellis, 1988). A major development in scenario, was developed based on an analysis of 50
this area over the last decade has been the appli- location-years for the disease in several parts of the
cation of formal Bayesian decision theory to either U.S. The predictive system has evolved in several
the construction or evaluation of the prediction ways since the 2003 publication (L.V. Madden,
systems. Growers and others (e.g., crop consul- unpublished), with many more observations ana-
tants) make numerous decisions before, during, lyzed as well as the development of new models, but
and after growing seasons, such as when and I restrict the discussion here to the data and results
where to plant, which cultivar to grow, whether to of the published paper. Eighteen of the location-
treat seeds with fungicide, and whether or not to years were considered to be major epidemics (i.e.,
apply a pesticide at any given time. Each decision requiring control, if available), simply called epi-
can be correct or incorrect, and Bayesian decision demics for convenience. One can thus consider the
theory provides a framework for making decisions so-called prior probability of an epidemic (E+) to
objectively (e.g., spraying a crop) and for evalu- be estimated or predicted as Prob(E+)=18/
ating decisions that have been made. 50=0.36. Of course, the data set for analysis here is
The key contributions in plant pathology have not necessarily representative of all locations for an
been by J. Yuen, G. Hughes, and some of their indefinite period of time, but we use this calculated
colleagues (Yuen et al., 1996; Hughes et al., 1999; prior probability for now since other information
Yuen and Hughes, 2002; Yuen, 2003). A very re- was not available. Yuen (2003) discusses the use of
cent and thorough example is Turechek and Wil- location-dependent prior probabilities. With Fusa-
cox (2005). The approach outlined below is rium head blight of wheat, there is a little more than
explicitly used in the disease predictive system for a one in three chance overall that an epidemic will
Sclerotina stem rot of oilseed rape (Twengström occur in a given location and year in the U.S. With
et al., 1998), and qualitative aspects of the ap- no other information, such as measured weather
proach are used implicitly by most investigators variables or inoculum levels in the atmosphere or on
developing and using predictive systems. I would crop debris, one would predict no epidemic – that is,
argue, however, that a fuller utilization of the one would bet against an epidemic at a given loca-
quantitative aspects of the approach will lead to tion and year, (even though one would sometimes
better predictive systems and more efficient use of lose the bet). This idea could be applied not just at a
the ones already developed. The Bayesian-decision yearly time scale. For apple scab, one could deter-
method centres on the determination of the prob- mine the proportion of days (or weeks, for instance)
ability of a disease outbreak (or need for a control where an infection period occurs in the spring. This
intervention) before and after using the predictor. can be considered the estimated prior probability of
This approach has much in common with medical the need to apply a fungicide, independent of any
diagnostic research, where the prediction of a other information (e.g., ignoring weather data).
disease epidemic here is analogous to the diagnosis Returning to Fusarium head blight, the proba-
of an individual for a given disease condition. Both bility of no epidemic (E)) in the De Wolf et al. data
areas involve decisions (predictions of disease in a set is given by Prob(E))=32/50=0.64=1–0.36.
field or region, or the prediction of a disease status For ease of calculations, it is convenient to deter-
of an individual) that can only be made with some mine the odds from the probability. In general, if A is
error. Plant disease prediction for crops has an some event, then odds(A)=Prob(A)/[1-Prob(A)]. If
additional level of uncertainty compared with the odds are known, the probability is obtained
medical diagnosis, since the decision is made for an from Prob(A)=odds(A)/[1+odds(A)]. With the
entire population (e.g., a field of a given crop, or example, the odds are: odds(E+)=0.36/[1–
even a region where the crop is grown) rather than 0.36]=0.563, and odds(E))=0.64/[1–0.64]=1.778.
just for the individual. Note that the odds(A)=1 when Prob(A)=0.5.
Decision theory for disease prediction in plant Probabilities above ½ give odds above 1, and
pathology can be explained best with a detailed probabilities less than ½ give odds below 1. The
example. De Wolf et al. (2003) developed a model main symbols used in this part of the article are
(their Model B) to predict major epidemics of summarized in Table 1, for convenience.
 13

Table 1. Some of the notation used regarding decision theory for disease prediction

Symbol Description

Prob(E+) Prior probability of an epidemic (or major epidemic, or for the need for a
control intervention).
Prob(E)) Prior probability of no epidemic (or for the lack of need for a control
intervention)
odds(E+) Prior odds of an epidemic: Prob(E+)/[1)Prob(E+)]
odds(E)) Prior odds of no epidemic: Prob(E))/[1)Prob(E))]
Prob(P+|E+) Probability of an actual epidemic being correctly predicted using some
specified disease forecasting or predictive system; the conditional probability
of a prediction of an epidemic (i.e., Z > threshold) given that an epidemic has
occurred.
Prob(P)|E)) Probability of an actual non-epidemic being correctly predicted using some
specified disease forecasting or predictive system; the conditional probability
of a prediction of a non-epidemic (i.e., Z < threshold) given that an epidemic
has not occurred.
Prob(P+|E)) Probability of an actual non-epidemic being incorrectly predicted to be an
epidemic; [=1)Prob(P)|E))]; the conditional probability of a prediction of
an epidemic given that an epidemic has not occurred.
Prob(P)|E+) Probability of an actual epidemic being incorrectly predicted to be a non-
epidemic [=1)Prob(P+|E+)]; the conditional probability of a prediction of
a non-epidemic given that an epidemic has occurred.
Z Indicator of the risk of an epidemic (or the need for a control intervention).
Can be derived with statistical or non-statistical methods.
TPP True positive proportion (sensitivity); proportion of epidemics correctly
predicted. An estimate of Prob(P+|E+)
TNP True negative proportion (specificity); proportion of non-epidemics correctly
predicted. An estimate of Prob(P)|E))
FPP False positive proportion; proportion of non-epidemics incorrectly predicted
to be epidemics [=1)TNP]. An estimate of Prob(P+|E)).
FNP False negative proportion; proportion of epidemics incorrectly predicted to be
non-epidemics [=1)TPP]. An estimate of Prob(P)|E+)
J A measure of accuracy [= TPP + TNP – 1=TPP – FPP], known as
Youden’s index.
Prob(E+|P+) Posterior probability of an epidemic given that one is predicted. Also known
as the positive predictive value, PV(+).
Prob(E)|P)) Posterior probability of no epidemic given that one is not predicted. Also
known as negative predictive value, PV()).
Prob(E-|P+) Posterior probability of no epidemic given that one is predicted [=1)Pro-
b(E+|P+)].
Prob(E+|P)) Posterior probability of an epidemic given that one is not predicted
[1)Prob(E)|P))].
LR(+) Likelihood ratio of a positive prediction (i.e., prediction of an epidemic), TPP/
FPP.
LR()) Likelihood ratio of a negative prediction (i.e., prediction of a non-epidemic),
FNP/TNP.
ROC Receiver operating characteristic curve, a plot of TPP vs. FPP. Can be written
mathematically as TPP=f(FPP).
odds(E+|P+) Posterior odds of an epidemic given that one is predicted (see equation 4).
odds(E)|P)) Posterior odds of a non-epidemic given that one is not predicted (see
equation 5).
CR Cost ratio, approximately equal to the cost of a false positive (CFP) divided by
the cost of a false negative (CFN)
LR*(+) Instantaneous likelihood ratio; the slope of the tangent to the ROC curve at
any (FPP, TPP) point. Also given as first derivative f¢(FPP), of the model for
the ROC curve (see equation 8 for example).
14

The question that arises in the context of disease the fraction of epidemics is fairly far from ½. A
forecasting or predictive systems is: can one sub- better overall measure of accuracy is given by
stantially change the predicted probability of an Youden’s index, J=TPP + TNP – 1=TPP – FPP;
epidemic (or the odds) based on other informa- J equals 1 for a perfect predictor. For the example,
tion? De Wolf et al. (2003) used logistic regression J=0.677.
to develop a risk algorithm (an equation in this The TPP is often called the sensitivity of a pre-
case) for predicting an epidemic. The following dictor or sensitivity of a model, and is an estimate
predictor was obtained: of the probability of an actual epidemic being
correctly predicted, Prob(P+|E+). Likewise,
Z ¼
3:725 þ 10:5ðX1 X2 Þ
TNP is often called the specificity of a predictor (or
where X1 is number of hours that temperature is of a model), and is an estimate of the probability
between 15 and 30 C for the 7 days prior to wheat that a non-epidemic is correctly predicted,
flowering, and X2 is the number of hours that Prob(P)|E)). FPP is the estimate of the proba-
temperature is between 15 and 30 C and relative bility that an actual non-epidemic is incorrectly
humidity is at least 90% for the 10 days starting at predicted to be an epidemic, Prob(P+|E)); FNP is
flowering, and Z is the predicted logit of the the estimate of the probability that an actual epi-
probability of an epidemic given the two weather demic is incorrectly predicted to be an non-epi-
variables. For other pathosystems, Z could rep- demic, Prob(P)|E+). The following table
resent a direct measurement of, for instance, hours summarizes the metrics and the estimates for the
of wetness, rather than being a function derived example.
from other variables. Z could also represent an
estimate of disease intensity (e.g., measured disease
early in a growing season) that is used to predict Predicted fi P+ P-
final disease or crop yield (see Turechek and Wil- Actualfl
cox, 2005; Yuen and Hughes, 2002).
E+ TPP FNP (= 1)TPP)
It turns out that the chosen threshold to use for Prob(P)|E+)
Prob(P+|E+)
predicting an epidemic with this data set is 0.833 0.167
Z=)0.40 (recall that Z is a logit); for a given E) FPP (1)TNP) TNP
location-year, if Z is above the threshold, then Prob(P+|E)) Prob(P)|E))
0.156 0.844
predict an epidemic (and label this P+), otherwise
predict no epidemic (and label this P)). Using this
rule, 15 of the 18 known epidemics were correctly
predicted, giving a true positive proportion of Although TPP and TNP are very similar here,
TPP=15/18=0.833. Also, 27 out of the 32 known this is not necessarily the case.
non-epidemics were correctly predicted, giving a The effectiveness of a predictor can be expressed
true negative proportion of TNP=27/32=0.844. in another way, which is extremely useful for some
One can also calculate the proportion of known calculations below. The likelihood ratio of a po-
non-epidemics predicted to be epidemics, which is sitive prediction (i.e., prediction of an epidemic) is
the false positive proportion, FPP=5/32=0.156. estimated by: LR(+)=TPP/(1-TNP)=TPP/FPP.
Finally, the proportion of known epidemics pre- Furthermore, the likelihood ratio of a negative
dicted to be non-epidemics is the false negative prediction (i.e., prediction of a non-epidemic) is
proportion, FNP=3/18=0.167. It can be shown estimated by: LR())=(1-TPP)/TNP=FNP/TNP.
that FPP=1-TNP, and that FNP=1-TPP. All of For the example, one obtains LR(+)=5.34 and
these calculations are based on the known (or as- LR())=0.20. An accurate predictor has, in gen-
sumed) status of each observation in the data set. eral, large LR(+) (above 1) and small LR())
Overall accuracy could be reported as (15+27)/ (close to 0).
50=0.840. However, this metric depends on the The use of a threshold of )0.4 for Z gives an
TPP and TNP values as well as the fraction of overall high accuracy (high J), but this is not the
observations in each category, and can thus be a only possibility. This can be seen by the TPP and
misleading indicator of model (predictor) success if TNP values over the full range of possible Z
 15

Hughes et al., 1999), which is a plot of TPP vs.

True positive proportion (TPP)

1.00 FPP (see Figure 3), that is, a plot of sensitivity vs.
Correct decisions
0.80
1-specificity. The curve goes from (0,0) at the lower
1.00 TPP
0.80
TNP left corner to (1,1) at the upper right corner. The
0.60 0.60 upper corner represents the lowest threshold tested
0.40
0.20
(i.e., smallest Z value), corresponding to maximum
0.40
0.00 sensitivity (high TPP) but minimum specificity
-5 -3 -1 1 3 5
0.20 (low TNP). The lower left corner represents the
Predictor threshold
highest threshold tested (i.e., largest Z value),
0.00 corresponding to minimum sensitivity and maxi-
0.00 0.20 0.40 0.60 0.80 1.00
mum specificity. If the predictor is of no value, the
False positive proportion (FPP = 1-TNP)
ROC curve will give a straight line through these
Figure 3. Main graph: An ROC curve, that is, the true posi- two extremes, with a slope of 1. An ideal predictor
tive proportion (TPP) vs. the false positive proportion (FPP), will give a curve that goes vary rapidly from (0,0)
for the predictor model in De Wolf et al. (2003). Inset graph: up to a TPP value of 1 at an FPP barely above 0
TPP and true negative proportion (TNP=1)FPP) vs. a full (i.e., 0+). The maximum J over all possible
range of decision thresholds.
thresholds of Z for accuracy occurs at the point on
the ROC curve that is closest to the upper left
values, as shown by the insert of Figure 3. If one corner. The area under the ROC curve is an
chose a low threshold (say, )3.5), then all the ac- overall measure of the prediction accuracy, with a
tual epidemics would be correctly predicted maximum of 1; for the example, the area is 0.9.
(TPP=1), since, essentially, all observations are One can think of the ROC curve as representing
then predicted to be epidemics (all observations the model TPP=f(FPP). It can be shown that the
have Zs larger than )3.5). There is a major con- first derivative of this model [f¢(FPP)] is the
sequence of a low threshold, however, in that the instantaneous value of LR(+) at any point FPP,
actual non-epidemics are also predicted to be epi- that is, the tangent to the TPP:FPP curve at any
demics (TNP=0 or FPP=1). The J value would FPP (Hughes and Madden, 2003). I call this like-
then be 0, a completely undesirable result. As the lihood ratio LR*(+). The more common calcu-
threshold increases above )3.5 for a predicted lation of LR(+), and the only one possible when
epidemic, TPP goes down, since it is now harder to the ROC curve is not available, is the straight-line
predict an epidemic, and increasing numbers of the slope over the interval from the point (0,0) to the
actual epidemics are predicted to be non-epidemics point (FPP,TPP), which equals LR(+)=TPP/
(i.e., some actual epidemics have Z values less than FPP (as indicated above).
the threshold). However, TNP goes up with the
increasing threshold, as higher numbers of non- Predictors in practice
epidemics are being correctly predicted (i.e., many
of the actual non-epidemics have Z values below All of the statistics shown so far deal with the
the threshold, as desired). Ultimately, with a very success of the predictor for known epidemics and
high threshold, all actual non-epidemics are cor- non-epidemics (i.e., for known status of the
rectly predicted (TNP=1 or FPP=0), since the observations). To assess the predictor in practice,
threshold is higher than all the observations. one must determine the probability that a random
However, this means that all the actual epidemics observation of unknown status (a particular loca-
are also predicted to be non-epidemics (TPP=0, tion-year in the example) is an epidemic, given that
FNP=1), since it is then impossibly hard to pre- the predictor score is positive (Z > the threshold),
dict an epidemic (i.e., all actual epidemics have Z written as Prob(E+|P+), or the probability that a
values less than the high threshold). In between random observation of unknown status is not an
these extremes, there are thresholds around )1 to 0 epidemic, given that the predictor score is negative,
where both TPP and TNP are high. written as Prob(E)|P)). Note that the conditional
The overall performance of any predictor can be probabilities have been turned around from that
summarized with a receiver operating characteris- used in developing the predictor, where the
tic (ROC) curve (Metz, 1978; Linnet, 1988; epidemic status was known; now, the prediction
16

status is known and not the actual status of an epidemic – one is no more certain of the epidemic
observation. To determine these and related status of a random observation when the predictor
probabilities for the population of interest (i.e., all is used compared to when it is not used.
location-years where the predictor is being used), One can also determine the posterior probability
one invokes Bayes’ Theorem (Yuen and Hughes, that an observation is not an epidemic when an
2002; Hughes and Madden, 2003), which can be epidemic is actually predicted, Prob(E)|P+), by
most easily written (and interpreted) in terms of first calculating the posterior odds: odds(E)|-
odds rather than probabilities. P+)=odds(E))/LR(+). Note that Prob(E)|P+)
The odds of an epidemic, given one is predicted is also given by 1-Prob(E+|P+). For the example,
[odds(E+|P+)] depends on the accuracy of the Prob(E)|P+)=0.25. Thus, there is still a reason-
predictor, expressed as LR(+), and the prior odds able probability (¼) that an observation is not an
that an observation is an epidemic [odds(E+)]. epidemic even when one is predicted, with the gi-
This new odds value can be written as: ven accuracy of the model. Another valuable term
is the posterior probability that an observation is
odds(E þ jPþÞ ¼ LRðþÞðodds(EþÞ ð4Þ
not an epidemic, given that a non-epidemic is
which is a simple direct product of the two terms. predicted, Prob(E)|P)). This can be readily
The left-hand side of this equation is known as the determined from:
posterior odds of an epidemic (or a disease out-
odds(E
jP
Þ ¼ odds(E
Þ=LRð
Þ ð5Þ
break, or the need for a control intervention, etc.),
given that one is predicted. The posterior odds will With the example, the posterior odds is estimated
be high (relatively speaking) if the prior odds is by: odds(E)|P))=1.778/0.20=8.98. The posterior
high (e.g., a relatively high proportion of location- probability is thus: Prob(E)|P))=8.98/
years for Fusarium head blight are epidemics) or if (1+8.98)=0.90. In other words, the probability
LR(+) is high (i.e., high accuracy). For the that there will not be an epidemic when an epi-
example, the posterior odds is given by: odd- demic is not expected (= 0.9) is increased com-
s(E+|P+)=5.34Æ0.563=3.0. The posterior prob- pared to the prior probability of a non-epidemic
ability can be determined by the transformation of (0.64) when no other information is available. Fi-
the odds (see above), which produces Pro- nally, the posterior probability of an epidemic gi-
b(E+|P+)=3.0/(1+3.0)=0.75. The posterior ven that a non-epidemic is predicted,
probability can be calculated directly, without use Prob(E+|P)), can be determined from 1-Pro-
of prior and posterior odds (Yuen and Hughes, b(E)|P)), or by first calculating the posterior odds
2002), but the formula is cumbersome and less as: odds(E+|P))=odds(E+)LR()). For the
intuitive. example, one obtains Prob(E+|P))=0.10, mean-
In typical usage, we would only predict an epi- ing that there is only a small probability that a
demic if the posterior odds is above 1 [or Pro- random unknown observation is actually an epi-
b(E+|P+) > ½]. In the example, the predicted demic when a non-epidemic is predicted. Because
odds of an epidemic occurring when the model of the properties of the predictor model and the
predicts one is more than five times the average (or prior probability of an epidemic, in the example
overall) odds of an epidemic when the predictor- one would have somewhat more confidence in
variable information is not known (or not used). predictions of non-epidemics than in predictions of
The predicted posterior probability of an epidemic epidemics.
when one is predicted (0.75) is a little more than The predictor is clearly of value based on the
double the prior probability (0.36) when no achieved likelihood ratios and the prior odds (or
information is known (or used). Note that a pre- prior probabilities). The point of interest here is
dictor is only valuable if LR(+) is larger than 1 in how the success of the predictor depends on the
this simple situation. When LR(+)=1, the prior prior odds of an epidemic (which comes directly
and posterior odds are the same, as well as the from the prevalence of epidemics). For instance,
prior and posterior probabilities, which means that consider what would happen if epidemics were
using the predictor is not giving the user any much less common than assumed originally, but
additional information about the chance of an one still wanted to use the developed predictor.
 17

I call this scenario B (and the original diseases, unless one has a predictor with an ex-
scenario above as A). Assume the prior tremely high overall accuracy (very large LR(+)).
probability of an epidemic is Prob(E+)=0.05, With an imperfect predictor (i.e., TPP < 1,
which gives: odds(E+)=0.053, Prob(E))=0.95, TNP < 1), there is uncertainty in any predictions
odds(E))=19.0. For Fusarium head blight this is of epidemics. Given that epidemics do not occur
unrealistically low, but the value is used for dem- that often (hypothetically, when
onstration purposes. The likelihood ratio is un- Prob(E+)=0.05), the evidence must be stronger
changed since this is a property of the predictor, than that obtainable from the use of the predictor
not the prior probability. The posterior odds of an to conclude (at least with more than a 0.50 prob-
epidemic then is calculated from equation 4 as ability, or more than an odds of 1) that an epi-
odds(E+|P+)=5.34Æ0.053=0.28, resulting in a demic will occur when predicted. However, if
posterior probability of an epidemic when one is LR(+) was 20 (i.e., a much more accurate pre-
predicted of Prob(E+|P+)=0.28/(1+0.28)=0.22 dictor), then the posterior odds would be 1.06
(Table 2). The posterior probability of an obser- (when prior odds of an epidemic was 0.053), and
vation not being an epidemic when one is pre- the posterior probability would be 0.51. Under
dicted is very high, equal to Prob(E)|P+)=1– these circumstances, the use of the predictor would
0.22=0.78. The other posterior probabilities are: be of greater value in management. However,
Prob(E)|P))=0.99, and Prob(E+|P))=0.01 (i.e., finding such accurate predictors in plant pathology
there is a very low probability that a random may be very difficult.
observation is an epidemic when one is not pre- There is an alternative to improving the overall
dicted). Using environmental data in the form of prediction accuracy for rare diseases. One can use
the predictor (Z), the probability that a random a different threshold of Z for an epidemic, which
observation (a location-year) is an epidemic clearly can be demonstrated with the Fusarium head
increases when one is predicted (from the prior blight results. As shown in Figure 3, TPP declines,
probability of 0.05 to the posterior probability of and TNP increases, as the threshold increases. If
0.22). However, there is still considerably less than one used a threshold (on the logit scale) of +2
a 50% chance that an observation is an epidemic (instead of )0.4), one would obtain TPP=0.39,
when one is predicted, and there is nearly an 80% TNP=0.99, FPP=1–0.99=0.01, and FNP=1–
chance (posterior probability of 0.78) that an 0.39=0.61; the likelihood ratios would be
observation is a non-epidemic when one is pre- LR(+)=39.0 and LR())=0.62. I call this sce-
dicted. In other words, use of the current predictor nario C (see Table 2). Using a higher threshold
(with the selected threshold of Z for a positive means moving down the ROC curve (Figure 3)
prediction) would be of little value in disease towards the lower left corner. By using a high
management when epidemics are rare – most threshold, almost all the known non-epidemics are
control interventions would be wasted since most correctly predicted (more specifically, almost all
of the predicted epidemics would turn out to be the known non-epidemics have Z values less than
non-epidemics. This shows in general that disease the new higher threshold; TNP  1), but only 40%
forecasting may not be of direct value for rare of the known epidemics are correctly predicted

Table 2. Evaluation of disease predictor for Fusarium head blight of wheat (see De Wolf et al., 2003) under various scenarios of
prior probability of an epidemic and threshold used for predicting an epidemica

Scenario Prob(E+) Threshold TPP TNP LR(+) LR()) Prob(E+|P+) Prob(E)|P))

A b
0.36 )0.4 0.833 0.844 5.34 0.20 0.75 0.90
B 0.05 )0.4 0.833 0.844 5.34 0.20 0.22 0.99
C 0.05 +2.0 0.39 0.99 39.0 0.62 0.67 0.97
D 0.85 )0.4 0.833 0.844 5.34 0.20 0.97 0.47
E 0.85 )1.7 0.944 0.656 2.74 0.085 0.94 0.67
a
See text and Table 1 for explanation of symbols and notation, as well as for terms not given in table.
b
Scenario A is the nominal (or standard) use of the predictor as described in the article.
18

(TPP  0.4) because many of these cases have Z combined higher LR(+) and lower LR())], one could
values below the new threshold. In other words, it move the threshold to a lower value (see Figure 3),
is now more difficult to correctly predict a known which corresponds to a higher TPP and lower TNP
epidemic. Although this may seem to be undesir- (the opposite direction than used when epidemics
able, the low prior probability of an epidemic were rare). A lower threshold means moving up the
(0.05) means that the posterior probability of an ROC curve towards the upper right corner.
epidemic when one is predicted to occur is actually With a threshold of )1.7, one obtains:
improved (i.e., predicted epidemics are more likely TPP=0.944, TNP=0.656, FPP=1–0.656=0.344,
to actually be epidemics). Using the numbers in and FNP=1–0.944=0.056; the likelihood ratios
this paragraph (including a prior probability of an would be LR(+)=2.74 and LR())=0.085. This
epidemic of 0.05), Prob(E+|P+)=0.67 (substan- is scenario E (Table 2). By making it easier to
tially higher than the 0.22 posterior probability for predict known epidemics (i.e., lowering the
scenario B), which means that in only about one threshold of Z for deciding in favour of an
third of the time, on average, would a predicted epidemic) when epidemics are common, one does
epidemic actually correspond to a non-epidemic not change the predictions of epidemics very
[i.e., Prob(E)|P+)=1–0.67=0.33]. When one re- much; that is, because of high prevalence of
quires stronger evidence for an epidemic [a higher epidemics, Prob(E+|P+)=0.94, only slightly
threshold, giving a larger LR(+)], there is less of a less than under scenario D]. However, it is much
chance that the prediction of an epidemic is wrong. more likely that an observation predicted to be a
There is a slight cost here to using the higher non-epidemic is, in fact, a non-epidemic. That is,
threshold – the posterior probability of an obser- Prob(E)|P))=0.67, compared to 0.47 for sce-
vation being a non-epidemic when predicted to not nario D. But there is still a fairly high proba-
be an epidemic is reduced to Prob(E)|P))=0.97, bility that a random observation is an epidemic
compared with 0.99 with the nominal predictor when a non-epidemic is predicted [Pro-
threshold (scenario B). This is due to the increase in b(E+|P))=0.33].
LR()) compared to the original choice of thresh- The above evaluation was totally presented in
old. Here, very little was lost in identifying non- terms of commonness of epidemics (or of the need
epidemics by changing the threshold for a positive to control, in general), measured by estimated
prediction (since it is, relatively speaking, easy to prior odds, and the accuracy of the predictor for
predict non-epidemics when epidemics are rare). known cases, measured by likelihood ratios. The
One can also consider the implication of much entire evaluation can be recast in terms of costs for
more common occurrence of epidemics. For in- each of the four possible decisions (true positive,
stance, if the prior probability of an epidemic is true negative, false positive, and false negative), or
Prob(E+)=0.85 (much higher than realistic for more simply, the costs of the two incorrect deci-
Fusarium head blight), one obtains: odds(E+)=5.67, sions (false positives and false negatives) (Linnet,
Prob(E))=0.15, odds(E))=0.176. Consider the 1988). Hughes and Madden (2003) give a detailed
predictor used at the nominal threshold ()0.4), account of this for regulatory problems (invasive
which gives, once again, LR(+)=5.34 and organism risk analysis) rather than for disease
LR())=0.2. I call this scenario D (Table 2). One forecasting. In brief, if CFP and CFN are the costs
obtains the following posterior probabilities: Pro- of a false positive and a false negative prediction,
b(E+|P+)=0.97, Prob(E)|P+)=1–0.97=0.03, respectively, then define CR as the ratio of these:
Prob(E)|P))=0.47, and Prob(E+|P))=0.53. CR  CFP/CFN (see Table 1). CR actually depends
Here, the predictor works very well for predicting also on the costs of true positives and true nega-
epidemics (there is only a 3% chance that an tives, but relative to the costs of the errors, it is
observation is a non-epidemic when one is pre- quite practical to consider these other costs as nil.
dicted), but works less well for predicting the non- Then, the decision rule that minimizes the average
epidemics. Based on Prob(E+|P)), about half the cost of using the predictor can be shown to be:
observations predicted to be non-epidemics are, on
odds(E þ jPþÞ > CR, then predict an epidemic;
average, actually epidemics. In the absence of a
more accurate predictor model [that would give a odds(E þ jPþÞ<CR, then predict a non-epidemic:
 19

The posterior odds is calculated from equation 4,

True positive proportion, TPP = f(FPP)

based on the prior odds and accuracy of the pre- 1.00 10
dictor (the likelihood ratio). The lower the CR, the

Derivative, f '(FPP)=LR*(+)
lower the posterior odds needed to predict an 0.80
epidemic. A low CR would occur when the costs of 1
false positives (such as the cost of spraying a crop 0.60
where the fungicide application is not needed) is
relatively low compared to the costs of false neg- 0.40
atives (such as the yield loss due to the disease that LR*(+) 0.1
Observed
occurs because a needed fungicide spray was not 0.20
used). As shown by Hughes and Madden (2003),
Predicted
the optimum threshold of Z to use for minimizing 0.00 0.01
costs of using a predictor can be determined based 0.00 0.20 0.40 0.60 0.80 1.00
on CR. In particular, one can write: False positive proportion (FPP = 1-TNP)

LR ðþÞopt ¼ CR/odds(EþÞ ð6Þ Figure 4. Left-hand axis: An ROC curve (true positive pro-
portion, TPP, vs. the false positive proportion, FPP) for the
where the left-hand side is the optimum LR*(+) in predictor model in De Wolf et al. (2003), together with pre-
which to make epidemic predictions (a function of dicted TPP from equation. Right-hand axis: derivative of
equations 7 and 8 vs. FPP, the instantaneous likelihood ratio
TPP and FPP, which are properties of the pre-
[LR*(+)].
dictor). Equation 6 is easy to calculate. The right
hand side is based on commonness of epidemics
and the relative costs of predictor errors (simply as
a ratio, so that absolute values of the costs are not which is based on equation 5 in Hughes and
needed), but does not involve the accuracy of the Madden (2003). As required because of the shape
predictor. It should be re-emphasized that the of the ROC curve, f ¢ (FPP)=LR*(+) declines
predicted LR*(+) here is the ‘instantaneous’ with increasing FPP (Figure 4; right-hand axis).
change in TPP with change in FPP (slope of the Mathematically, one can solve the f ¢ (FPP) equa-
tangent to the TPP:FPP curve at FPP), given as f ¢ tion for FPP, and then obtain TPP based on
(FPP). To translate equation 6 into an exact pre- f(FPP) (equation 7). From this combination of
dicted combination of TPP and TNP (or FPP =1- TPP and FPP (which gives TNP and FNP), one
TNP), one must first have a specific model for the can determine the standard LR(+) (= TPP/FPP)
ROC, TPP=f(FPP), in order to obtain f ¢ (FPP) at and use equation 4 for risk assessment.
the optimum point (Hughes and Madden, 2003). The use of equations 6 and 7 can be demon-
An example is equation 4 in Hughes and Madden strated with the Fusarium head blight data. Pre-
(2003). The fit of this model to the ROC curve in viously, an epidemic was predicted in practice if
Figure 3 using nonlinear least squares results in odds(E+|P+) was greater than 1. This is equiva-
the following equation: lent to specifying that the cost ratio, CR, equals 1,
where both types of errors are equally costly.



1=2:37
TPP ¼ 1 þ e
4:42 FPP
2:37
1 Using odds(E+)=0.563, as stated previously, one
finds from equation 6 that LR*(+)opt = 1.78
ð7Þ
when CR=1. Graphically, one finds this value of
As shown in Figure 4, the model provides a good f ¢ (FPP) in Figure 4, and then determines the
fit to the TPP values. The first derivative of corresponding FPP and TPP at this value. One can
equation 7 is given as: see that FPP  0.18 and TPP  0.81 at f ¢
(FPP)=1.78 in the graph, which are similar to the


1=2:37
4:42
1 þ e
4:42 FPP
2:37
1 e FPP
2:37
f 0 ðFPPÞ ¼


2:37
 ð8Þ
FPP 1 þ e
4:42 FPP
1
20

values used in the nominal situation described in want a low FNP). Conversely, as CR increases
Table 1 (with a Z threshold of )0.4; scenario A). (e.g., false positives are more costly than false
(There will be some discrepancy because negatives), one moves down the ROC curve to-
equation 7 is not a perfect fit to the ROC curve.) wards the left-hand corner (lower TPP and FPP;
With the TPP and FPP values here, LR(+)  4.5, higher TNP and FNP), which means that a higher
and the posterior odds of an epidemic when one Z threshold is used for predictions of epidemics.
is predicted is 4.5Æ0.563=2.5, giving Pro- Loosely speaking, with high cost of false positives,
b(E+|P+)=0.71 (close to the value found at one would not want to make too many of these
slightly different TPP and TNP values in Table 1). errors (i.e., one would want a low FPP). The ap-
It is important to emphasize that the posterior proach outlined here can also be coupled with
odds of an epidemic (or non-epidemic) are actually consideration of different prior probabilities of
calculated with LR(+), not with the instantaneous epidemics, as presented previously in this section.
rate LR*(+). It is quite possible for a given pathosystem that
Now consider the situation in which a false there are combinations of prior probabilities and
negative decision is four times as costly as a false costs of false predictions that one would always
positive decision (CR=CFP/CFN=1/4=0.25). assume that an epidemic will occur or always as-
With the nominal prior odds of an epidemic, one sume that an epidemic will not occur. The deci-
finds that LR*(+)opt=0.25/0.563=0.444. From sion-theory approach provides the formal
Figure 4, one can see that this derivative occurs at mechanism for evaluating these scenarios.
FPP  0.32 (or TNP  0.68) and TPP  0.94. An The analyses discussed here are just the begin-
increased TPP and decreased TNP compared to nings of the possibilities for applying risk assess-
the nominal situation (with CR=1) is higher up ment to disease prediction (Yuen and Hughes,
the ROC curve (towards the right-hand corner), 2002; Yuen, 2003). In fact, only the initial aspects
which means that an epidemic is predicted to occur of this approach have been formally applied to
at a lower Z threshold (Figure 3). That is, there is Fusarium head blight forecasting at this stage;
a less stringent criterion to predict an epidemic. costs of decisions and consideration of prior
Using the listed sensitivities and specificities here, probabilities of epidemics for this pathosystem will
one obtains LR(+)=0.94/0.32=2.94, and be addressed more formally after a more accurate
LR())=(1–0.94)/0.68=0.088. The posterior odds prediction system is developed for known epi-
of an epidemic when one is predicted then is demics and non-epidemics. Other areas requiring
odds(E+|P+)=2.94Æ0.563=1.66, giving a pos- research for plant diseases in general include:
terior probability of Prob(E+|P+)=1.66/ having more than a dichotomy of decisions (such
(1+1.66)=0.62. It can be shown that the posterior as spray, do not spray, and wait-and-see what
probability of a non-epidemic when a non-epi- happens); dealing with more than a dichotomy of
demic is predicted is Prob(E)|P))=0.95. One is predictions (such as predicting the degree of ex-
more certain about the true epidemic status of a pected damage from a disease, predicting spraying
random observation when non-epidemics are pre- once or weekly for the rest of the season); dealing
dicted compared to when epidemics are predicted. with more than a dichotomy of measured out-
Also, Prob(E+|P+) is lower here than when comes (such as intensity of disease on a continuum
CR=1, but this reduction in certainty of epidem- for different predictions, for either binary predic-
ics is required to minimize the average cost of tions or a continuum of predictions); and dealing
making predictions. with multiple diseases or pests.
In general, as demonstrated in the previous
paragraph, as CR declines at a given prior prob-
ability of an epidemic, one moves up the ROC Discussion
curve towards the right-hand corner (higher TPP
and FPP; lower TNP and FNP), which means that Botanical epidemiology has advanced on many
a lower Z threshold is used for predictions of fronts in the years since van der Plank’s first book
epidemics. Loosely speaking, with high cost of was published in 1963, and the discipline continues
false negative decisions, one would not want to to be a foundation for understanding and pre-
make too many of these errors (i.e., one would dicting diseases at the population scale. I have
 21

chosen to outline just two out of many possible pathologists, including epidemiologists, clearly are
broad topics where substantial advances have not yet thinking formally or explicitly in terms of
been, and continue to be, made. Many of the prior and posterior odds, and likelihood ratios, the
speakers and poster presentations at the 9th concepts follow directly from intuitive under-
International Workshop on Plant Disease Epide- standings of how prediction rules are applied when
miology reported in this special issue of the there is some inaccuracy in the predictors and
European Journal of Plant Pathology have dealt when epidemics and non-epidemics (or the need to
with other valuable topics. intervene or not with a control method) are not
The use of growth-curve and mechanistic pop- equally common in a given area. The next gener-
ulation dynamic models, especially the coupled ation of advances in this area will deal more with
(ordinary and partial) differential equations out- the costs of decisions (correct or incorrect), and in
lined in this article, provide a flexible and powerful addressing some of the biological and environ-
methodology for representing the temporal, spa- mental interactions in more complicated patho-
tial, and spatio-temporal dynamics of diseases, systems, possibly involving multiple diseases (and
and provides the framework to elucidate general crops simultaneously) (McRoberts et al., 2003).
thresholds for epidemic occurrences (disease Whether one is working with elaborate popu-
invasion), long-term persistence of disease, veloc- lation-dynamic models or testing prediction rules
ity at which disease expands from foci, and the for decision making, I accept as an axiom that
initial rates of disease increase over time. Many of appropriate statistical methods be used to fit
these qualitative and quantitative properties of models to data, compare results within and be-
epidemics can be summarized by the basic repro- tween studies, and test hypotheses. Developments
duction number (R0). in linear and nonlinear mixed models, for instance,
The coupled differential equations (or other are drastically improving the matching of the sta-
model formulations, such as stochastic difference tistical methods to the data and experimental de-
equations) can be made extremely complicated, sign, and intended inferences of the investigator
and care should be taken to keep the principle of (Schabenberger and Pierce, 2002). There is also
model parsimony in mind when modelling epi- growing evidence that Bayesian methods are also
demics! Strategies for disease control can be useful for developing prediction equations (Mila
readily explored by finding the combination of and Carriquiry, 2004), and not just in evaluating
disease properties (e.g., latent and infectious peri- the performance of predictors (as demonstrated in
ods, transmission rate) that result in a R0 less than this article). Except for the most quantitative of
1. This approach is less useful for real-time pre- the botanical epidemiologists, more effort is still
diction of epidemics, or for determining the need needed to encourage plant pathologists to keep
to intervene with a control measure, because pre- abreast of developments in statistics and utilize the
cise estimates of parameters under specific envi- appropriate old and new techniques (Garrett
ronmental conditions are often not available. et al., 2004).
Usually, simpler prediction equations (such as
regression equations, discriminant functions, or
ad hoc rules) are used to actually make predic- Acknowledgements
tions. This use of more descriptive equations can
be further justified by the fact that even very Salaries and research support were provided by
complicated mechanistic models often result in state and federal funds to the Ohio Agricultural
simple exponential-type population increase when Research and Development Center, Ohio State
disease intensity is not high (see Segarra et al., University, USA.
2001).
The incorporation of probabilistic decision the-
ory into disease predictions (whether these come
References
from empirical rules or equations, or even popu-
lation-dynamic models) is the second topic I cov- Anderson RM and May RM (1991) Infectious Diseases of
ered where key advances have been made recently Humans; Dynamics and Control, Oxford Science Publica-
(Yuen and Hughes, 2002). Although most plant tion, Oxford.
22

Auden WH and Kronenberger L (1966) The Viking Book of Hughes G, Gottwald TR and Yamamura K (2002) Survey
Aphorisms, Viking, New York. methods for assessment of citrus tristeza virus incidence in
Aylor DE (1999) Biophysical scaling and the passive dispersal urban citrus populations. Plant Disease 86: 367–372
of fungus spores: relationship to integrated pest manage- James C (2003) Global Status of Commercialized Transgenic
ment strategies. Agricultural and Forest Meteorology 97: Crops: 2003, International Service for the Acquisition of
275–292. Agri-Biotech Applications (ISAAA), Manila
Bald JG (1937) Investigations on ‘spotted wilt’ of tomatoes III: Jeger MJ (1986a). The potential of analytic compared with
infection in field plots. Bulletin 106. Melbourne, Australia: simulation approaches to modeling in plant disease epide-
Council for Scientific and Industrial Research. miology. In: Leonard KJ and Fry WE (eds) Plant Disease
Campbell CL and Madden LV (1990) Introduction to Plant Epidemiology, Vol. 1: Population Dynamics and Manage-
Disease Epidemiology, John Wiley & Sons, New York ment (pp 255–281) Macmillan Publishing Company, NY
Cochran WG (1936) The statistical analysis of field counts of Jeger MJ (1986b) Asymptotic behaviour and threshold criteria
diseased plants. Supplement to Journal of Royal Statistical in model plant disease epidemics. Plant Pathology 35: 355–
Society 3: 49–67. 361
DeWolf ED, Madden LV and Lipps PE (2003) Risk assessment Jeger MJ (2004) Analysis of disease progress as a basis for
models for wheat Fusarium head blight epidemics based on evaluating disease management practices. Annual Review
within-season weather data. Phytopathology 93: 428–435 of Phytopathology 42: 61–82
Diekmann O and Heesterbeek JAP (2000) Mathematical Jeger MJ, Holt J, vanden Bosch F and Madden LV (2004)
Epidemiology of Infectious Diseases: Model Building, Epidemiology of insect-transmitted plant viruses: modeling
Analysis, and Interpretation, John Wiley and Son, Ltd., disease dynamics and control interventions. Physiological
Chichester, U.K, 303 Entomology 29: 1–14
Ferrandino FJ (1993) epidemic waves: I. Focus expansion Jones DG(.) (1998) The Epidemiology of Plant Diseases,
within a linear planting. Phytopathology 83: 795–802 Kluwer Academic Publishers, Dordrecht, 460
Fry WE (1982) Principles of Plant Disease Management, Keeling MJ, Brooks SP and Gilligan CA (2004) Using
Academic Press, New York conservation of pattern to estimate spatial parameters from
Garrett KA, Madden LV, Hughes G and Pfender WF (2004) a single snapshot. Proceedings National Academy of
New applications of statistical tools in plant pathology. Science USA 101: 9155–9160
Phytopathology 94: 999–1003 Kranz J(.) (1990) Epidemics of Plant Diseases: Mathematical
Gibson GJ (1997) Investigating mechanisms of spatiotemporal Analysis and Modeling, 2nd edn. Springer-Verlag, Berlin
epidemic spread using stochastic models. Phytopathology Legg JP (1999) Emergence, spread and strategies for controlling
87: 139–146 the pandemic of cassava mosaic virus disease in east and
Gibson GJ, Gilligan CA and Kleczkowski A. (1999) Predicting central Africa. Crop Protection 18: 627–637
variability in biological control of a plant-pathogen system Linnet K (1988) A review of the methodology for assessing
using stochastic models. Proceedings Royal Society London diagnostic tests. Clinical Chemistry 34: 1379–1386
B 266: 1743–1753 Madden LV (1986). Statistical analysis and comparison of
Gildow F, Damsteegt B, Stone A, Schneider W, Luster D and disease progress curves. Pages 55–84. In: Leonard K and
Levy L (2004) Plum pox in North America: identification of Fry WE (eds) Plant Disease Epidemiology (pp 372)
aphid vectors and a potential role for fruit in virus spread. Macmillan Publishing Co, New York
Phytopathology 94: 868–874 Madden LV and Hughes G (1995) Plant disease incidence:
Gilligan CA (1985). Introduction. In: Gilligan CA ) Advances distributions, heterogeneity, and temporal analysis. Annual
in Plant Pathology, Vol. 3; Mathematical Modelling of Review of Phytopathology 33: 529–564
Crop Disease (pp 1–10) Academic Press, London Madden LV and Hughes G (1999) Sampling for plant disease
Gilligan CA (2002) An epidemiological framework for disease incidence. Phytopathology 89: 1088–1103
management. Advances in Botanical Research 38: 1–64 Madden LV and Hughes G (2002). Plant epidemics, models,
Gottwald TR, Hughes G, Graham JH, Sun X and Riley T and analysis. In: El-Shaarawi A and Piegorsch W (eds)
(2001) The citrus canker epidemic in Florida: the scientific Encyclopedia of Environmetrics (pp 1557–15623) John
basis of regulatory eradication policy for an invasive Wiley & Sons
species. Phytopathology 91: 30–34 Madden LV and Ellis MA (1988). How to develop disease
Hardwick NV (1998). Disease forecasting. In: Jones DG (eds) forecasters. In: Kranz J and Rotem J (eds) Experimental
The Epidemiology of Plant Diseases (pp 207–230) Kluwer Techniques in Plant Disease Epidemiology (pp 191–208)
Academic Publishers, Dordrecht Springer-Verlag Publishers
Hughes G and Madden LV (2003) Evaluating predictive models Madden LV and vanden Bosch F (2002) A population-
with application in regulatory policy for invasive weeds. dynamics approach to assess the threat of plant pathogens
Agricultural Systems 76: 755–774 as biological weapons against annual crops. BioScience 52:
Hughes G, McRoberts N, Madden LV and Nelson SC (1997) 65–74
Validating mathematical models of plant disease progress in Madden LV and Wheelis M (2003) The threat of plant
space and time. IMA Journal of Mathematics Applied in pathogens as weapons against U.S. crops. Annual Review
Medicine and Biology 14: 85–112 of Phytopathology 41: 155–176
Hughes G, McRoberts N and Burnett FJ (1999) Decision- Madden LV, Pirone TP and Raccah B (1987) Temporal
making and diagnosis in disease management. Plant analysis of two viruses increasing in the same tobacco
Pathology 48: 147–153 fields. Phytopathology 77: 974–980
 23

Madden LV, Jeger MJ and vanden Bosch F (2000) A Turechek WW and Wilcox WW (2005) Evaluating predictors of
theoretical assessment of the effects of vector-virus trans- apple scab with receiver operating characteristic (ROC)
mission mechanism on plant virus disease epidemics. curve analysis. Phytopathology 95: 679–691
Phytopathology 90: 576–594 Twengström E, Sigvald R, Svensson C and Yuen J (1998)
Maloy OC (1993) Plant Disease Control, John Wiley & Sons, Forecasting sclerotinia stem rot in spring sown oilseed rape.
New York Crop Protection 17: 405–411
McRoberts N, Hughes G and Savary S (2003) Integrated vanden Bosch F, Metz JAJ and Diekmann O (1990) The
approaches to understanding and control of diseases and velocity of population expansion. Journal of Mathematical
pests in field crops. Australasian Plant Pathology 32: 167– Biology 28: 529–565
180 vanden Bosch F, Metz JAJ and Zadoks JC (1999) Pandemics of
Metz CE (1978) Basic principles of ROC analysis. Seminars in focal plant disease, a model. Phytopathology 89: 495–505
Nuclear Medicine 8: 283–298 vanMaanen A and Xu X-M (2003) Modelling plant disease
Mila AL and Carriquiry AL (2004) Bayesan analysis in plant epidemics. European Journal of Plant Pathology 109: 669–682
pathology. Phytopathology 94: 1027–1030 vander Plank JE (1963) Plant Diseases: Epidemics and Control,
Neter J, Wasserman W and Kutner MH (1983) Applied Linear Academic Press, NY, 349
Regression Models, Richard, D. Irwin, Inc, Homewood, Waggoner PE (1960). Forecasting epidemics. In: Horsfall JG
IL, 547 and Dimond AE (eds) Plant Pathology, AN Advanced
NRC (2002). Predicting Invasions of Non-indigenous Plants Treatise (pp 291–313III) Academic Press, New York
and Plant Pests. National Research Council, National Wolfinger RD (1996) Heterogeneous variance–covariance
Academy of Sciences, National Academy Press, Washing- structures for repeated measures. Journal of Agricultural,
ton, DC Biological, and Environmental Statistics 1: 205–230
Pielou EC (1977) Mathematical Ecology, John Wiley & Sons, Xu X-M and Ridout MS (1998) Effects of initial epidemic
New York conditions, sporulation rate, and spore dispersal gradient
Rizzo DM, Garbelotto M, Davidson JM, Slaughter GW and on the spatio-temporal dynamics of plant disease epidemics.
Koike ST (2002) Phytophthora ramorum as the cause of Phytopathology 88: 1000–1012
extensive mortality of Quercus spp. and Lithocarpus Yuen J (2003) Bayesian approaches to plant disease forecasting.
densiflorus in California. Plant Disease 86: 205–214 Online. Plant Health Progress doi:10.1094/PHP-2003-1113-
Ross R (1911) The Prevention of Malaria, 2nd edn. Murray, 06-RV. [http://www.plantmanagementnetwork.org/php]
London Yuen JE and Hughes G (2002) Bayesian analysis of plant
Rush CM, Stein JM, Bowden RL, Riemenschneider R, disease prediction. Plant Pathology 51: 407–412
Boratynski T and Royer MH (2005) Status of Karnal bunt Yuen JE, Twengström E and Sigvald R (1996) Calibration and
of wheat in the United States 1996 to 2004. Plant Disease verification of risk algorithms using logistic regression.
89: 212–223 European Journal of Plant Pathology 102: 847–854
Schabenberger O and Pierce FJ (2002) Contemporary Statisti- Zadoks JC (2001) Plant disease epidemiology in the twentieth
cal Models for the Plant and Soil Sciences, CRC Press, century: a picture by means of selected controversies. Plant
Boca Raton, Florida Disease 85: 808–816
Segarra J, Jeger MJ and vanden Bosch F (2001) Epidemic Zadoks JC and Schein RD (1979) Epidemiology and Plant
dynamics and patterns of plant diseases. Phytopathology Disease Management, Oxford University Press, Oxford, 427
91: 1001–1010 Zhang X-S, Holt J and Colvin J (2000) A general model of
Strange RN and Scott PR (2005) Plant disease: a threat to global plant-virus disease infection incorporating vector aggrega-
food security. Annual Review of Phytopathology 43: (in press) tion. Plant Pathology 49: 435–444
This page intentionally blank
European Journal of Plant Pathology (2006) 115:25–34
Springer 2006
DOI 10.1007/s10658-005-4513-5

Framework development in plant disease risk assessment and its application

X.B. Yang
Department of Plant Pathology, Iowa State University, Ames, Iowa, 50011, USA (E-mail: xbyang@
iastate.edu)

Accepted 20 October 2005

Abstract

This article reviews recent developments in plant disease risk assessment. The role of risk assessment as an
application area in macrophytopathology and its contribution to the development of macroscale disease
study are discussed. This article also discusses the concepts and components of risk assessment for different
end points and the assessment framework of different potential ranges of a new pathogen: establishment
range, suitability range, damage range, and dispersal range. Different end points generate risk information
suitable for decision makers at different levels. New insights gained from selected major diseases, especially
from risk assessment due to the recent global movement of soybean rust, are presented. The role of
pathologists in presenting risk information has extended beyond the professional research domain and has
become critical in influencing decision-making, evident by soybean rust in both South and North America.
The bias components of risk communication are defined, and different levels of receivers for risk infor-
mation are identified based on their interpretation capability of risk information, bias potential, and
utilization of risk information. Lack of predictability of dispersal potential contributes to uncertainty of
risk assessment for airborne diseases. Potential research areas in disease risk assessment are discussed.

Introduction Definition of risk assessment

Risk assessment is part of botanical epidemiology. In a narrow sense, risk assessment involves deter-
Some principles of disease risk assessment can be mining the potential epidemiological and eco-
found in disease epidemic forecasting in the early nomic impact of emerging or new diseases
development stage of epidemiology. In the 1970s, (domestic or foreign). The information is critical to
threat analysis was used in regulatory plant decision makers at higher levels, for example, in
pathology to determine quarantine subjects. Now, deciding policies to deal with risk mitigation and
different terms such as threat analysis, risk analy- risk preparation at regional or national levels. In a
sis, or risk assessment are used for studies to sense, risk assessment is the application of epide-
determine the epidemic potential of exotic, new, miology to regulatory plant pathology or to dis-
and emerging diseases. Over the last ten years, ease management. A study of risk assessment is a
advances in computer computation have made macroscale, long-term disease prediction that
quantitative analysis with large sets of climatic encompasses assessment of establishment poten-
data possible, along with risk assessment of exotic tial, entry potential (when the range of a disease
diseases (Yang et al., 1991). The most systematic expands beyond a political border), and epidemic
study is on soybean rust, Phykoposora pachyrhizi. potential (epidemic frequency and epidemic
Since the introduction of Asian soybean rust into severity or extent of the disease), and the potential
the New World, tremendous resources and effort losses in a region or country once an epidemic
have gone toward assessing the risk of soybean occurs. Risk assessment is an epidemiological
rust. This article reviews recent developments in study to predict future occurrence of a disease
risk assessment of exotic and emerging new dis- by using non-experimental approaches, often
eases with special reference to soybean rust. involving computer modelling. The modelling
26

Table 1. Comparison of purpose and scales of risk information among different users of risk information for Sclerotinia stem rot
of soybean, caused by Sclerotinia sclerotiorum, in the North Central region of the United States (from Yang, 2004)

User Purpose Temporal scale Spatial scale

Farmers Chemical control In-season Fields/farm

Variety selection Coming season Fields/farm
Tillage Coming season Fields/farm
Extension In-season Fields/area
Agronomists Advice Coming season Fields/area
Seed/chemical Marketing strategies Next year Regional
Companies Breeding decisions Next few years Regional
Product development Years/decade(s) Regional
Government Funding decision Years Regional

results are not repeatable with extrapolation from been attributed to climate change (Rosenzweig
limited field studies or published data. et al., 2001). Such changes can result in the emer-
In a broad sense, risk prediction or risk assess- gence of new threats from minor diseases or to
ment has been extended to predicting the seasonal range expansion of a disease to production regions
occurrence of endemic diseases, especially in hor- where the disease previously was not a concern.
ticultural and high-value crops (Luo et al., 2001). Expansion of the damage range of soybean
Research in this area involves predicting the bean pod mottle virus into the US North Central
upcoming season’s disease occurrence in a defined Region and sudden death syndrome (SDS) into the
production area. Use of the term ‘disease risk’ for northern United States are recent examples. The
disease forecasting reflects advances in disease risk increase of SDS prevalence in US north central
communication at the farm-level. Yang (Yang, region has been associated with increased planting
2003) recently outlined a conceptual example of of soybean in early spring, a production measure
risk prediction according to a temporal and spatial associated with warmer springs. The third trend is
scale that uses risk information (Table 1). In this the change in farming practices, exemplified by the
article, I exclude the seasonal prediction of the expansion of no-tillage systems.
outbreak risk of an endemic disease in a region for As an applied research area, disease risk
disease management practices. assessment has recently gained importance in the
political arena of biosecurity (Madden and Van
Need for risk assessment den Bosch, 2002;Madden and Wheelis, 2003).
Because biosecurity-related risk assessment is new
The rapid development of risk assessment reflects and is highly relevant to political issues, risk
the adaptation of plant pathology to new global assessment in this area will likely embed bias
agriculture trends, the consequence of which leads potential, and contributions of research from this
to an increased risk of new diseases. The first trend area to risk assessment are needed.
is frequent seed and plant material movement by
international companies. For example, some soy-
bean and corn seeds planted in the US Midwest Risk assessment and macrophytopathology
are now produced in South America. The move-
ment of large amounts of germplasm is thus Macrophytopathology is the study of disease
unavoidable and could facilitate the movement of occurrence patterns and disease management at
plant pathogens and consequently the introduc- the macroscale (Zeng, 2003). A narrower defini-
tion of new diseases. Assessments are made for tion is the study of statistical patterns of disease
regulatory decision-making, and several studies distribution, disease range, and epidemic frequen-
are underway in this area. The second trend is cies on large spatial and temporal scales. Ecol-
climate change (IPCC, 2001). Climate change has ogically, macrophytopathology addresses the
been shown to be a driving force in the long-term questions when, where, and why a new disease
dynamics of plant diseases (Yang and Scherm, emerges and becomes a major production threat.
1997); new diseases and re-emerged diseases have Such a study can be used to predict the damage
 27

potential of a disease (Yang, 2003). A theoretical two-dimensional distribution of the occurrence of a

framework of macrophytopathology has not yet disease over a geographic area. When an exotic
been developed. The concept of risk assessment disease is introduced into a new geographic region
predates macrophytopathology. Risk assessment is or a new disease emerges, assessment of the poten-
the application of epidemiological methodology to tial range of the disease is to determine the geo-
predict the long-term risk of new or emerging pathosystem range, which is directly associated with
diseases. With such information, disease risk can impact assessment. Epidemiologically, a disease
be mitigated by managing the movement and dis- range is related to the following four other ranges
tribution of a new disease on a macroscale (Mag- important to the ultimate impact of a disease.
arey, personal communication). Regulatory
1) Establishment range, or year-round survival
measures are effective approaches for risk mitiga-
range, in which a pathogen can sustain itself
tion. Strategically, a sound assessment helps make
from one growing season to the next by
decisions in developing resistance programmes,
completing disease cycles. For a soil-borne
such as screening for resistance germplasm.
disease, the establishment range is the same
Development of resistant varieties is an expensive,
as the disease range. For an airborne disease,
long-term investment and would not be initiated
the establishment range is smaller than the
until the entry of the disease.
disease range, and the establishment region
It has been suggested that macrophytopatholo-
serves as the source area of inoculum for
gy is basically the same as geophytopathology as
other regions during a season. For example,
initially proposed by Weltzien (Weltzien, 1972).
with wheat rust in the United States, the
Conceptually, macrophytopathology is different
establishment range is the overwintering
from geophytopathology in two respects. In his
range of wheat rust fungi in southern Texas
review, Weltzien used the idea of geophytopa-
and Louisiana.
thology in which ‘documentation, analysis, and
2) Suitability range defines a geographic area
prognosis of plant epidemics seem to be an
where the conditions are suitable for disease
appropriate theme for maps as basic contribu-
to occur, which is important for airborne dis-
tions.’ For geophytopathology, the study began in
ease risk assessment. A range suitable for a
advanced stages of epidemiology and uses quan-
disease to occur does not necessarily mean
titative epidemiological approaches to study dis-
the disease will occur in all areas defined by
ease occurrence patterns on a large scale. At the
the range because of the uncertainty of inoc-
time geophytopathology was proposed, botanical
ulum availability. The airborne inoculum
epidemiology as a discipline or field was in its
may never reach certain areas defined in the
cinfancy. In macrophytopathology, new informa-
region. Assessment based on this range rep-
tion technology is integrated with sophisticated
resents the maximum risk.
modelling techniques to handle vast climatological
3) Dispersal range defines the geographic area
data. It also includes disease management on a
into which airborne diseases can spread sea-
macroscale. The core area of macroscale study,
sonally from overwintering areas defined by
risk assessment, is the application of theories and
the establishment range. Dispersal range of a
methods of epidemiology. In the context of mac-
disease does not mean range of inoculum
rophytopathology, risk assessment is equivalent to
spread, which is far larger than seasonal dis-
disease forecasting in the conventional scale of
ease dispersal range. Recent studies show
plant disease study. Weltzien (1972) did foresee
that air currents can, within a season, carry
the potential of geophytopathology for disease
the spores of soybean rust from Brazil to al-
management, but it has yet to be demonstrated in
most anywhere in the Western Hemisphere.
macrophytopathology.
4) Damage range is a region where the disease
can cause significant economic yield loss in a
Disease range concepts and risk assessment frequency that warrants implementation of
production measures. Sometimes, it can be
The concept of disease range was proposed by Yang referred to as the endemic region of a dis-
and Feng (Yang and Feng, 2001) to describe the ease. Geographically, the physical sizes of
28

these ranges for an airborne disease have the distance during the growing season, the pathogen
following order: establishment range < dam- should not be considered a threat. For this
age range < dispersal range < suitability assessment, the key aspects are to determine the
range. For soil-borne diseases, the establish- availability of alternate hosts of the pathogen and
ment range should be equal to the suitability its overwintering potential in a non-host growing
range and no smaller or larger than the dam- season. Quantification of the source strength early
age range (establishment range = suitability in spring may be important to mid-term season
range > damage range). disease forecasting for airborne diseases.

Suitability assessment
Components in risk assessment
This assessment determines the suitability of the
climate in the studied geographic region or coun-
The fundamental parameters of potential risk of a
try. Suitability assessment is almost always the first
new or exotic disease to crop production in a
assessment to address in risk analysis. If the
geographical region are disease range, frequency
climate is unsuitable for the occurrence of the dis-
of potential epidemics, and intensity of epidemics
ease, no further assessment is needed. The normal
in terms of economic losses. Epidemiologically,
approach for this assessment is to use a disease
risk assessment for a new disease consists of the
model together with climatic data to assess the
following exercises, each of which can be inde-
epidemic potential in the region. Sometimes, the
pendent. Negative results of each assessment
epidemic potential is further fed into a yield-loss
indicate the non-threatening status of a new dis-
model to determine the maximum yield loss po-
ease. These components are establishment poten-
tential of a disease. When long-term climatic data
tial, suitability of the environment to disease
are available, determination of potential epidemic
occurrence, dispersal potential, and yield-loss
frequencies and severity, a higher level of assess-
potential. In risk assessment, almost all assessment
ment, are useful to policy makers. If severe epi-
starts with a suitability assessment, and the order
demics of a disease have a frequency one-in-eight
of assessment procedures is determined by the
or ten years, the disease may not be a major pro-
availability of techniques and data at the time
duction concern. However, suitability assessment
when the assessment is made. Because effects of
is almost always made with an assumption that
temperature and dew are known key factors in
initial inoculum is available early in a growing
determining the infectivity of a plant pathogen,
season. Therefore, the estimated risk would be
these factors are investigated first in epidemiolog-
greater than the real loss if the disease is airborne
ical studies and therefore are available for the
because initial inoculum of an airborne disease is
environment suitability assessment of risk assess-
not always available. Risk estimated from such an
ment. For soybean rust, there are three critical
assessment represents maximum risk.
components of uncertainty: 1) suitability of cli-
matic conditions for rust epidemics in soybean
Dispersal potential assessment
production areas, 2) likelihood of establishment of
the fungus in North America, and 3) the seasonal
This assessment applies to a disease caused by
dispersal potential of the pathogen from over-
an airborne pathogen that establishes regionally in
wintering regions to major soybean production
a country but poses a threat to the rest of the
regions.
production area. The damage level of the disease
depends on the yearly reintroduction potential of
Establishment assessment inoculum. It is not a concern for soilborne dis-
eases, however, once the establishment assessment
This assessment addresses the question: once an is completed. For soybean rust, the pathogen
exotic pathogen is introduced, can it survive from overwinters in southern Florida and Texas, which
season to season in a country or geographic is far from the major US soybean production
region, and if so, where? If the disease cannot be region. For risk assessment, this is the last com-
established in the region or in an area of reachable ponent to study because epidemiology does not
 29

the type of assessment. Information from the

establishment assessment is significant for quar-
antine purposes. The risk from the suitability
assessment is the maximum risk useful for policy
decision-making in which accountability is a con-
cern. The maximum risk could be far from the real
losses because of a lack of inoculum in a season or
missing a dispersal component in the study. For
epidemiologists, the challenging part is to deter-
mine the most likely losses, which is information
useful for industry, whose concern is on the impact
of a new disease on its profitability. The arrow in
Figure. 2 indicates the most likely losses assessed
with comprehensive epidemiological information.
Figure 1. Flow chart showing the process of a risk assessment Depending on the availability of information/data
study. To have a completed risk study, the last step of risk
and skills provided to an assessment, the assessed
interpretation and communication by the modellers is essen-
tial. risk may be higher or lower than the most likely
risk. Over time, the assessed value from a risk
assessment should approach the most likely risk.
provide methodology for such an assessment. For
soybean rust, this uncertainty has been a factor in
decision-making in Argentina where the disease Examples of assessment for a soilborne disease
can overwinter only in the northern production
regions. Figure 1 Risk assessment of soybean sudden death syn-
Conceptually, risk assessment for exotic, new, or drome (SDS) caused by Fusarium solani f. sp.
emerging diseases could be generalized according glycines is an example of an emerging disease. This
to the types of assessment (Figure 2). The outcome soilborne disease was first reported in Arkansas in
of each assessment and its usefulness depend on the early 1970s and caused endemic production

Figure 2. Conceptual framework of risk assessment for exotic, new, or emerging diseases, showing the types of assessment, out-
come of each type, and users of each outcome. The arrow indicates the most likely losses assessed with comprehensive epidemio-
logical information.
30

problems in the southern states (Roy et al., 1997). based on research in a containment facility at Fort
It was initially domestic and no establishment Detrick in Federick, MD, and in fields in Asia,
assessment was needed. Because it is soilborne, the which provided baseline information for disease
dispersal assessment was also not applicable. The modelling; and 2) development of computer mod-
need for a potential impact assessment was not elling to quantitatively assess the potential effects
realized until 1993 when the disease was found in of rust on soybean yield in the United States.
Iowa, a leading soybean production state. The Research under controlled conditions focused
soybean industry needed to know the level of the on determining the importance of each epidemic
threat from this disease to the North Central Re- component or subcomponent in the soybean rust
gion, which produces 78% of US soybean, so that disease cycle and on quantifying its response to
the funding agencies could prioritize investment of host and environmental variation. The compo-
research funds. A risk assessment for SDS was nents—spore germination, infection, latent period,
conducted using Climex, computer software lesion expansion, sporulation, and senescence of
developed by CSIRO (Sutherst and Maywald, uredia—were studied by several researchers at Ft.
1991), with disease parameters generated from Detrick and elsewhere (Keogh, 1974; Marchetti
experiments conducted under controlled condi- et al., 1976; Bromfield et al., 1980; Meching et al.,
tions (Scherm and Yang, 1999). The assessment 1989); Patil et al., 1997; Hundekar and Hiremath,
predicted that the disease would cause more losses 2001). The effects of dew duration and tempera-
in the north central region than in the southern ture on infection have been quantified as a two-
United States where the disease originated. The dimensional relationship from which an infection
assessment has proved correct. By 2002, SDS had model was developed to estimate infection
spread into Canada and Minnesota and was (Marchetti et al., 1976). These studies provided
ranked the number one damaging fungal disease in critical background information for building epi-
the north central US soybean production. When demiological models for risk assessment. The field
data were presented in 1995 at a regional soybean experiments were conducted at the Asian Vegeta-
conference, SDS immediately gained the attention ble Research and Development Centre in southern
of the soybean industry. Breeding for resistance to Taiwan. Soybean can grow year-round in this
SDS had started before the disease became a area, and disease occurs most of the year, except
production problem in the US North Central Re- during the winter. These data allowed analysis of
gion. Now, many seed companies have resistant the seasonal variation in rust epidemics.
varieties available to growers. Without this risk From the data compiled from field and green-
assessment, which promoted resistance breeding house studies, a computer simulation model,
and management research, current disease preva- SOYRUST was developed. This simple disease
lence levels and the frequency of epidemics would model includes most weather variables that influ-
probably be higher. ence disease epidemics. The model was validated
with data from Taiwan, and predictions matched
observations. SOYRUST was integrated, as a
Example of assessment for an airborne disease subroutine, into the soybean growth model SOY-
GRO (Wilkerson et al., 1985), developed at the
Suitability assessment University of Florida to simulate disease progress
during the production season and to predict yield.
Risk assessment of soybean rust, caused by the With the assumption that spores are available
fungus Phakopsora pachyrhizi, to US soybean early in the growing season, the simulation results
production is the best example for an airborne showed that considerable yield loss could occur in
disease. The assessment started in the early 1980s some areas of the United States. In recent years,
and was one of the earliest risk assessment pro- USDA–APHIS generated a US risk map by using
grammes. Risk assessment of this disease has con- continuous moisture measurements and dew days
tributed to the development of general concepts data. The general consensus is that the environ-
and quantitative methodology. For suitability mental conditions in the US soybean production
analysis, research efforts were divided into two regions are suitable for the occurrence of this dis-
phases: 1) understanding infection components ease (Bromfield, 1984).
 31

Establishment assessment southern Texas and Florida. In the United States,

the fungus is likely to overwinter in areas where
Predicting the year-round survival of the soybean climatic conditions in winter are similar to those in
rust fungus is important for determining avail- southern China. During mild winters, the coastal
ability of spores in the spring and for determining region of the Gulf of Mexico is also in the P.
potential dissemination into major soybean pro- pachyrhizi year-round survival zone (Figure. 3).
duction regions during a growing season from an The reported occurrence of soybean rust in kudzu
overwintering area. Models were used to predict plants near Tampa, Florida, in February of 2005
where climatological conditions are suitable for validated the assessment.
the year-round persistence of P. pachyrhizi
worldwide. Long-term meteorological averages Dispersal assessment
were used to assess stress by using the CLIMEX
software developed by Sutherst and May- Several means existed for long distance spread of
wald(1985). Integration of stresses was used to airborne diseases, such as the tobacco blue model
predict the likelihood of survival of P. pachyrhizi to assess spore movement from the Caribbean to
within a location (Pivonia and Yang, 2004). The the southeastern US. For soybean rust, there are
assessment shows that areas presumed suitable for several means for northward movement. Clima-
year-round survival of P. pachyrhizi in the Western tological models have been integrated with epide-
Hemisphere extend from southern Brazil to miological models for prediction of soybean rust

Figure 3. Illustration of the potential range of soybean rust, P. pachyrhizi, in North America with a focus on the United States
(Yang, 2003). The establishment range is indicated by the probability of overwintering, which is the US coastal area. The suitabil-
ity range is estimated to extend to from Gulf Coast to northern border of the US in east of Rocky Mountain. The dispersal range
and the damage range are yet to determined.
32

movement. Such integration could be a potential farms. The capability of digesting risk information
research area to generate new directions for epi- varies, and the purpose of taking risk information
demiological research. The MM 5 model is a glo- differs. Therefore, interpretations of the risk on the
bal circulation model for air particle movement receiver sides are different. Receivers of risk
and high split model for rust prediction. The information can be grouped by three levels based
model has been used to correctly predict the 2004 on their knowledge and interpretation of disease
season rust spore movement to Argentina and risk.
Colombia. In August 2004, it also predicted the Level 1. This level consists of pathologists or
spore movement from Colombia to the southern groups of professionals who have in-depth
US before the disease was found in Louisiana. knowledge of plant pathology. Detailed outcomes
on risk assumption can be explained and the
uncertainty of the information is fully understood
Risk communication by receivers. This level includes industry experts in
chemical and seed companies whose annual prof-
To plant pathologists, risk communication is rela- itability relies on decision with a measured risk.
tively new in the framework of risk study compared The setting for risk information delivery includes
with plant disease risk analysis. Elements of risk professional meetings through presentations by
communication consist of receivers of risk infor- authorities on specific topics, and this level has no
mation, interpretation of the uncertainty, and bias in interpretation of disease risk.
delivery of the information with an approach or Level 2. This includes decision makers at the
method according to information receivers and the policy-making level at the public domain, such as
capability to digest risk information of receiver government officials or society leaders who have
aspects. To the risk communicators who are often limited understanding of assumptions or who
plant pathologists, it is crucial to understand the overlook the assumptions made for assessment for
epidemiological principals used in the assessments. political liability and address the maximum risk.
Most models are built with certain assumptions Bias interpretations of the risk assessment are not
that are critical to risk interpretation. Without uncommon. To make a better argument for mak-
stating the underlining assumption while discussing ing policies, the maximum risk is used most fre-
the risk, the risk level can be overestimated, which quently for accountability reasons. The example is
often occurs when risk information is disseminated the frequent use of soybean rust economic assess-
by non-pathologists. For soybean rust, when cita- ment of US $ 7.2 billion (Kuchler et al., 1984) in
tions are made for the study of USDA economic decision-making, although a later much smaller
assessment or yield loss (Yang et al., 1991), the figure (US $ 1–2 billion) has been made in a new
writers often reported figures of potential losses risk assessment.
from a risk assessment without providing the For high-level policy makers, decision-making is
assumptions addressed by authors in their studies. based on political rational and accountability,
which has a tendency towards self-protection and
Receivers of risk information therefore naturally embeds bias of select use of risk
information. For soybean rust, decisions at higher
After the risk assessment is made, the information levels were likely made using the worse-case sce-
is disseminated to the public. There are several nario. Economic calculation is less essential com-
levels of receivers: policy makers, scientists in pared with the producers or profit-driven
chemical or seed companies for product develop- businesses. Finally, risk communication varies
ment, managers of funding agencies, producers for from culture to culture. Some cultures are more
day-to-day farm operation, and crop advisers in sensitive to disease risk than others. For soybean
private and public domains. Based on the decision rust, the response of industry to soybean rust in
levels, the risk information is generated differently the United States has been much greater than in
in terms of temporal and spatial scales (Table 1) South American countries.
(Yang, 2003). For decision-making, the temporal Level 3. This includes laypersons or producers
scale ranges from decades to weeks, and the spatial who have a limited knowledge of plant pathology.
scale ranges from the entire country to individual Some highly competitive producers, however, have
 33

good plant pathology knowledge and should be maximum damage of a soilborne disease. How-
considered in level 1. For level 3, communication is ever, for airborne diseases, the establishment range
made through indirect approaches. Information is of a disease would be no larger than damage range
delivered mostly through media, agricultural mag- and smaller than the suitability range, depending
azines, or radio, where normally the maximum risk on the dispersal potential of the disease. To have a
of a disease is presented without further explana- more realistic assessment, the long-distance conti-
tion. Risk is often selectively stated by media to nental dispersal potential of the airborne disease
achieve sensational effects. In the dissemination of needs to be predicted.
soybean rust information, only maximum yield Dispersal potential is a key uncertainty in pre-
losses of assessments (loss of 80%) were used, dicting the risk of an exotic airborne disease,
without providing the dispersal assumptions used which is important in determining the entry risk
in the assessment. Unfortunately, producers are and damage risk after establishment. For the
handlers of risk in production, and risk information example of soybean rust, uncertainty of the disease
indirectly delivered to them decreases the effec- risk depends on our understanding of dispersal
tiveness and quality of risk management. Overre- potential of the disease. Before entry of this dis-
actions to soybean rust were common among the ease, dispersal information was critical to deter-
US soybean producers in the first two years. mine when the disease would arrive in the United
It is important to disseminate risk information States. Such information is critical for chemical
differentiated by simplicity or complexity to dif- companies to determine when to stockpile fungi-
ferent receivers to avoid confusion or panic and to cide for disease control. For commodity groups,
maintain the credibility of the research commu- entry time is used to embargo the importation of
nity. One example was the early release of spore soybean from occurrence countries, a temporary
dispersal assessment. In 2003 winter, a USDA strategy to raise the local market price.
trajectory analysis for the Western Hemisphere Once a disease establishes in the new country, its
was prematurely released. The results showed that damage potential to the production region de-
air parcels carrying fungal spores from lower ele- pends on the seasonal dispersal of airborne spores
vations in Brazil soybean production regions could from overwintering regions into major production
reach the United States. Unfortunately, this regions. Models for the dispersal potential are still
statement was interpreted by receivers as indicat- in their infancy, and the prediction needs to
ing the possibility of soybean rust occurrence in determine the source area and its relationship to
the coming season in the United States, which inoculum density in the receiving region, after
caused unnecessary panic in some US growers who long-distance dispersal. Such a relationship should
consequently purchased fungicides for the pre- exist in a pathosystem as demonstrated by Zeng
dicted arrival of spores during the next season. (1988) in wheat stripe rust.
Latent period after entry. There are two critical
times after the entry of an exotic plant pathogen
Future research for risk assessment into a new geographic region: time of first detec-
tion and time of first outbreak. For risk mitiga-
Our knowledge in epidemiology has enabled us to tion, when the pathogen can be detected after
assess environmental suitability, establishment entry and when the first outbreak would occur are
potential, and survival potential after the intro- key questions. Practically no introduced diseases
duction of a plant disease. By adding yield loss have produced severe, region-wide epidemics in
models, yield loss potential can be determined. The the first season of detection in the United States.
outcome of risk assessment from the above-men- Similarly, soybean rust in South America did not
tioned components represents the worst-case sce- cause economically significant losses in the 2000–
nario. For soilborne diseases, the establishment 2001 growing season after it was first found in
range of a disease is equivalent to the distribution southern Brazil and Paraguay. It caused losses
range or the range suitable to disease occurrence, >$125 million in the second year of detection
with the damage range being smaller than the (Yorinori et al., 2003). However, predicting when
pathogen distribution range. Assessment with the an introduced pathogen would become prevalent
worst-case scenario may not approach the realistic after subsequent entry and reach outbreak levels is
34

critical to disease prevention or risk management. Patil VS, Wuike RV, Thakare CS and Chirame BB (1997)
The available response period depends on detec- Viability of uredospores of Phakopsora pachyrhizi syd. at
different storage conditions. Journal of Maharashtra Agri-
tion efficiency—and the earlier the better. cultural University 22: 260–261.
Pivonia S and Yang XB (2004) Assessment of potential year-
round establishment of soybean rust throughout the world.
Plant Disease 88: 523–529.
References Rosenzweig C, Iglesias A, Yang XB, Epstein PR and Chivian E
(2001) Climate changes and US agriculture: The impacts of
warming and extreme weather events on productivity, plant
Bromfield KR, Melching JS and Kingsolver CH (1980) Viru-
diseases, and pests. Global Change and Human Health 2:
lence and aggressiveness of Phakopsora pachyrhizi isolates
90–103.
causing soybean rust. Phytopathology 70: 17–21.
Roy KW, Rupe JC, Hershman DE and Abney TS (1997) Sudden
Bromfield KR (1984) Soybean rust. Monograph No. 11. Am.
death syndrome of soybean. Plant Disease 81: 1100–1111.
Phytopathol. Soc., St. Paul, MN.
Scherm H and Yang XB (1995) Interannual variations in wheat
Hundekar AR and Hiremath PC (2001) Development, preva- rust development in China and the United States in relation
lence and seasonal severity of soybean rust caused by
to the El nino/southern oscillation. Phytopathology 85:
Phakopsora pachyrhizi Syd. Karnataka Journal of Agricul-
970–976.
tural Science 14: 71–75.
Scherm H and Yang XB (1999) Risk assessment for sudden
IPCC (2001) Agriculture in a changing climate: Impacts
death syndrome of soybean in the north-central United
and adaptation. In: Watson RT, Zinyowera MC and Moss
States. Agricultural Systems 59: 301–310.
RH (eds) Climate Change 1995. Impacts, Adaptations
Sutherst RW and Maywald GF (1985) A computerized system
and Mitigation of Climate Change: Scientific-Technical
for matching climates in ecology. Agricultural Ecosystems
Analyses (pp. 427–467) Cambridge University Press,
and Environment 13: 281–299.
Cambridge.
Sutherst RW and Maywald GF (1991) Climate modeling and
Keogh RC (1974) Studies on Phakopsora pachyrhizi Syd.: The
pest establishment. Plant Protection Q. 6: 3–7.
Causal Agent of Soybean Rust. M.S. Thesis, University of Weltzien HC (1972) Geophytopathology. Annuual Review of
Sydney, Sydney, Australia.
Phytopathology 10: 277–298.
Kochman JK (1979) The effect of temperature on development
Wilkerson GG, Jones JW, Boote KJ and Mishoe JW (1985)
of soybean rust. Australian Journal of Agricultural
SOYGRO, University of Florida, Gainesville.
Research 30: 273–277.
Yang XB, Dowler WM and Tschanz AT (1991) A simulation
Kuchler F, Duffy M, Shrum RD and Dowler WM (1984)
model for assessing soybean rust epidemics. Journal of
Potential economic consequences of the entry of an exotic
Phytopathology 133: 187–200.
fungal pest: The case of soybean rust. Phytopathology 74:
Yang XB and Feng F (2001) Ranges and diversity of soybean
916–920.
fungal diseases in North America. Phytopathology 91: 769–
Luo Y, Morgan DP and Michailides TJ (2001) Risk analysis of
775.
brown rot blossom blight of prune caused by Monilinia
Yang XB and Scherm H (1997) El Nino and infectious disease.
fructicola. Phytopathology 91: 759–768. Science 275: 739.
Madden LV and Van den Bosch F (2002) A population-dynamics
Yang XB (2003) Risk assessment: Concepts, development and
approach to assess the threat of plant pathogens as biological
future opportunities. Plant Health Progress, http://
weapons against annual crops. Bioscience 52: 65–74.
www.plantmanagementnetwork.org/sub/php/review/2003/
Madden LV and Wheelis M (2003) The threat of plant concepts/.
pathogens as weapons against US crops. Annual Review
Yorinori JT, Pavia WM, Frederick RD, Costamilan LM,
of Phytopathology 41: 155–176.
Bertagnolli PF, Hartman G, Godoy C and Nunes J Jr
Marchetti MA, Markle GM and Bromfield KR (1976) effects of
(2003) Epidemics of soybean rust (Phakopsora pachyrhizi)
temperature and dew period on germination and infection
in Brazil and Paraguay from 2001 to 2003. Phytopathology
by uredospores of Phakopsora pachyrhizi. Phytopathology
93:S103.
66: 461–463.
Zeng S (1988) Inter-regional spread of wheat yellow rust in
Meching JS, Dowler WM, Koogle DL and Royer MH (1989)
China. Phytopathologica Sinica 18: 219–223.
Effects of duration, frequency, and temperature of leaf
Zeng (2003) Discussion on the macro-phytopathology. Review
wetness periods on soybean rust. Plant Disease 73:
of China Agricultural Science and Technology 5: 3–7.
117–122.
European Journal of Plant Pathology (2006) 115:35–51
Springer 2006
DOI 10.1007/s10658-005-4050-2

Ecological genomics and epidemiology

K.A. Garrett1,*, S.H. Hulbert1, J.E. Leach2 and S.E. Travers1

1
Department of Plant Pathology, Kansas State University, Manhattan, KS 66506, USA; 2Department of
Bioagricultural Sciences and Pest Management, Colorado State University, Ft. Collins, CO, 80523, USA;
*Author for Correspondence (Fax: +1-785-532-5692; E-mail: [email protected])

Accepted 13 October 2005

Abstract

The huge amount of genomic data now becoming available offers both opportunities and challenges for
epidemiologists. In this ‘‘preview’’ of likely developments as the field of ecological genomics evolves and
merges with epidemiology, we discuss how epidemiology can use new information about genetic sequences
and gene expression to form predictions about epidemic features and outcomes and for understanding host
resistance and pathogen evolution. DNA sequencing is now complete for some hosts and several pathogens.
Microarrays make it possible to measure gene expression simultaneously for thousands of genes. These
tools will contribute to plant disease epidemiology by providing information about which resistance or
pathogenicity genes are present in individuals and populations, what genes other than those directly
involved in resistance and virulence are important in epidemics, the role of the phenotypic status of hosts
and pathogens, and the role of the status of the environmental metagenome. Conversely, models of group
dynamics supplied by population biology and ecology may be used to interpret gene expression within
individual organisms and in populations of organisms. Genomic tools have great potential for improving
understanding of resistance gene evolution and the durability of resistance. For example, DNA sequence
analysis can be used to evaluate whether an arms race model of co-evolution is supported. Finally, new
genomic tools will make it possible to consider the landscape ecology of epidemics in terms of host
resistance both as determined by genotype and as expressed in host phenotypes in response to the biotic and
abiotic environment. Host phenotype mixtures can be modeled and evaluated, with epidemiological pre-
dictions based on phenotypic characteristics such as physiological age and status in terms of induced
systemic resistance or systemic acquired resistance.

Introduction ated with genomics and other ‘‘-omics’’ (e.g. pro-

teomics, metabolomics) have allowed an
The field of plant disease epidemiology has incor- unprecedented collaboration among scientists
porated new technologies and perspectives on working at different biological scales and have
biology as they have become available, from fostered a new science, ecological genomics. In this
computer simulation modeling to automated ‘‘preview’’, we discuss how these new approaches
environmental sensing. Over the past decade, the may dovetail with plant disease epidemiology.
study of DNA within all areas of biology has gone Epidemiology has already benefited from
through a revolution, providing new types and information about the population genetics of
new quantities of genomic data for epidemiologi- pathogens, as reviewed by Milgroom and Peever
cal analyses. Given the advent of new technologies (2003). By simultaneously studying how pathogen
associated with rapid analysis and miniaturization, gene frequencies change within and among popu-
informatics, and molecular biology, it is now lations as a result of both natural selection and
possible to expand the scale of studies of both gene flow, and how pathogen populations grow
agricultural and wild species to include entire and spread, it has been possible to track disease
genomes. The high-throughput advances associ- outbreaks (e.g., Zwankhuizen et al., 1998),
36

develop predictions about sources of inoculum and sequencing allow analysis of great numbers of
pathogen life cycles (e.g., Cortesi et al., 2000; ‘‘markers’’ with added information about their
Cortesi and Milgroom, 2001), understand the likely role through reference to databases such as
evolution of virulence (Escriu et al., 2000a, b, GenBank (Black et al., 2001), thus revealing the
2003), and make predictions about the durability gene content of particular organisms.
of resistance in crop genotypes (Escriu et al., Functional genomics, or the use of genomic
2000a, b). Ultimately, modeling plant disease epi- technologies (e.g. microarrays) to find genes and
demics and pathogen evolution depends on a polymorphisms that affect traits of interest and to
complete understanding of both plant and patho- characterize the mechanisms underlying those
gen traits that influence the dynamics between a effects, has been applied effectively in agricultural
pathogen and its host. To completely understand contexts and has potential in natural systems.
any trait and its significance in a dynamic interplay Functional genomics moves beyond simple
between species requires the simultaneous use of sequence analysis to evaluate the function of par-
molecular, cellular, organismal, population and ticular DNA sequences through, for example, gene
ecological approaches. Past efforts to combine knockout mutants or gene activation mutations.
epidemiology and population genetics have come These techniques have natural applications for the
up against an upper limit on the number of eco- study of resistance and virulence, but might also be
logically important genes that could be surveyed usefully applied in the study of other
or lack of information on gene function and sig- epidemiological features. By simultaneously scan-
nificance. Yet, host plants, as well as pathogens, ning thousands of plant genes for changes in
exist in a matrix of hundreds or thousands of other expression in response to variables of interest (e.g.
taxa and their genes. Population changes in stress, infection) it has been possible to identify
pathogens, reproduction and dispersal will all candidate loci or suites of genes and molecular
depend on the interactions among these organisms mechanisms involved in the phenotypic expression
that can influence the dynamics of resistance evo- of key traits of economically important crop spe-
lution and direct effects on pathogen populations cies (Frick and Schaller, 2002; Jones et al., 2002;
(Antonovics, 2003). Mysore et al., 2003). A great deal has been learned
The developing synthesis of a functional about plant defense against disease through the
genomics approach combined with a population use of functional genomics and model plant sys-
and ecological perspective promises to lead to new tems such as Arabidopsis (Wan et al., 2002; Schenk
avenues of research and understanding of plant/ et al., 2003; Whitham et al., 2003a).
pathogen interactions. Evolutionary and ecologi- An intriguing area of epidemiology that will
cal functional genomics or EEFG (Feder and develop with the availability of new tools for
Mitchell-Olds, 2003) has as a goal to understand studying gene expression is the study of pheno-
ecological and evolutionary processes that main- typic resistance and its responses to the biotic and
tain genotypes and phenotypes. The emphasis so abiotic environment. Infection with an incompat-
far has been on wild species, but agricultural sys- ible pathogen, or a virulent pathogen that causes
tems offer both an important application and cell death, can make a plant more resistant to
relatively well-characterized systems for experi- subsequent infection by the same or different
mentation. The field of ecological genomics will pathogens, a phenomenon designated systemic
address new types of questions beyond applica- acquired resistance (SAR; Durrant and Dong,
tions based on molecular markers. Microarrays 2004). The SAR response in Arabidopsis confers
allow synoptic measurements of gene expression in resistance to several diseases (Ryals et al., 1996).
tens of thousands of genes. Real-time PCR allows Resistance to pathogens can also be influenced by
highly accurate quantitative evaluation of gene non-pathogenic organisms; systemic changes in
expression at many time steps. It will also be disease resistance in response to colonization by
possible to identify hundreds or even thousands of rhizosphere-colonizing Pseudomonas bacteria have
organisms simultaneously from individual samples been well-documented and are commonly referred
as microarrays are developed with sequences rep- to as induced systemic resistance (ISR; Iavicoli
resentative of desired sets of species, potentially et al., 2003; Cui et al., 2005). Dissection of the
including non-culturable species. Advances in SAR and ISR signaling systems in Arabidopsis
 37

indicate they are controlled by different pathways genotypic and phenotypic resistance, can now be
and signaling molecules with some common com- studied more thoroughly as it affects epidemics. In
ponents. Understanding which genes are expressed addressing these topics, we will emphasize genes
during specific defense responses can provide that influence the relationship between plant host
indications of what pathways are activated in dif- and pathogen, but the same general concepts
ferent biotic environments (Pieterse and van Loon, would apply to interactions between plant species,
1999). Tools are now available to begin studying between plants and insect herbivores, etc.
these phenomena more widely in epidemiology.
This paper will address the following topics in
ecological genomics. (1) Population genetics and How population genetics and population genomics
population genomics can inform epidemiology to can inform epidemiology
further our understanding of epidemics and to
provide insights for disease management. We will Epidemiology has traditionally used information
also consider how studies of gene expression can about host species, pathogen and vector species,
potentially add predictive power at finer spatial and environmental variables such as temperature
and temporal scales than was possible in the past. and precipitation to predict epidemic progress.
(2) Models of populations and communities may These models can be adapted to incorporate much
apply to analogous systems of gene expression more detailed information about the genomic
within organisms and in populations of organisms status of the host and pathogen communities
to inform a ‘‘population biology’’ of gene expres- considered in the context of a broadly defined
sion. (3) Genomics can contribute to understand- environment, i.e., defined to include abiotic com-
ing of resistance gene evolution and durability of ponents and potentially the complete community
resistance. (4) The landscape ecology of host metagenome of soil (Rondon et al., 2000) or other
populations and communities, in terms of both systems (Figure 1). Information about the soil

Pathogen genome(s)

Pathogen
Pathogenspecies
species

Host
Hostspecies
species Abiotic
Abiotic
environment
environment
(and
(andbiotic
bioticenvironment)
environment)

Host genome(s)
Community
Communitymetagenome
metagenome
and
and
abiotic
abiotic
environment
environment

Figure 1. The traditional disease triangle depicts prediction of epidemics based on interactions between pathogen species, host spe-
cies, and the abiotic environment. It is now much easier to obtain information about the complete genotype and current gene
expression of host and pathogen, and there is even the potential to obtain this information for complete communities such as those
in soil, the rhizosphere, and the phyllosphere, as well as endophytic communities. Models about a hierarchy of features of ‘‘geno-
mic status’’ might be experimentally evaluated in this context. For example, ‘‘The host landscape is sufficiently described to predict
epidemic features and outcomes by information about... d ... host community composition (in terms of species).’’ d ... a specific
subset of the host genotype sequence(s).’’ d ... the host genotype sequence(s).’’ d ... a subset of host gene expression.’’ d ... com-
plete profiles of host gene expression (now and/or in the past).’’
38

Table 1. The temporal and spatial scale of variation in different components of host genomic status

Component of host Temporal scale Spatial scale

genomic status

In annual monoculture
Species Cropping season Size of field in many
conventional systems
Genotype Cropping season Size of field in many
conventional systems
Gene expression Less than one hour Part of one individual
to cropping season to size of field
In unmanaged systems
Species Days to decades One individual to majority
of plant community
Genotype Days to decades (potential One individual to majority
for somatic mutation) of species (in clonal species)
Gene expression Less than one hour or until Part of one individual to majority of area
phenotype expressed (days
for defense reaction, months
for flowering, etc.)

metagenome may contribute to an understanding Characterizations of populations may include

of disease suppressive soils that develop over time the composition of both qualitative features pro-
as microbial populations respond to the buildup of duced by different genotypes and quantitative
pathogen populations. For example, soils sup- features produced by different levels of gene
pressive to the wheat take-all pathogen and potato expression in what may be the same genotype.
scab have been described, with fluorescent pseu- Evaluation of qualitative features might be per-
domonads and streptomycetes, respectively, being formed using marker or sequence studies, while
the likely causes of suppressiveness (Weller et al., evaluation of quantitative features might be
2002). Advances in genomics also make it possible performed using microarrays or real-time PCR.
to characterize the genomic status of host plants at Studies of gene expression in agriculturally
a much finer temporal and spatial scale than in the important host plants have expanded remarkably,
past (Table 1). The addition of gene expression as with microarrays now available for several major
a response variable or predictor variable in epi- crop species. These allow host resistance to be
demiological models has the potential to shift the assessed as an outcome of gene expression. In
scale of inquiry to hours and millimeters. Moni- addition, the expression of plant genes in response
toring the expression of genes in specific defense to non-pathogenic microorganisms may be highly
pathways, or individual genes that reflect the relevant to epidemiology, as it may provide an
expression of the pathways, could be used to pre- understanding of how plants select for rhizosphere
dict the outcome of pathogen infection in indi- flora that are antagonistic to pathogens, for
vidual plants or plant organs. For most diseases, example (Smith and Goodman, 1999). Microarray
progress in determining the efficacy of different analyses can be used to identify sets of coregulated
defense responses for controlling specific patho- genes and their common regulatory elements (e.g.,
gens and how the responses become distributed Maleck et al., 2000; Chen et al., 2002), which may
throughout the plant must be made before this both reveal different response pathways and allow
information is useful. Then epidemiologists will selection of smaller sets of indicator genes to
need to perform a range of exploratory field represent particular stress response pathways.
studies to identify the forms of predictors that are Microarrays developed using genes from one plant
most useful for inclusion in more detailed follow- species may also be applied, with some caveats, in
up studies. For example, if the early induction of studies of related species; for example Travers
senescence-related gene pathways were observed to et al. (in preparation) have applied maize micro-
occur, would this be related to reduced epidemic arrays to study gene expression in the related
potential at a field scale? tallgrass prairie grasses Andropogon gerardii and
 39

Sorghastrum nutans, and have identified statisti- mutants may give insight into cost of virulence
cally significant responses to simulated climate (Vera Cruz et al., 2000; Ponciano et al., 2004).
change in native field populations. Sequencing can also be used to evaluate the po-
New genetic information can be used to refine tential repertoire of resistance genes available, to
state transition models such as Susceptible-In- the extent that sequence similarity can predict
fected (SI) models (e.g., Otten et al., 2003). Rather functional similarity (Bai et al., 2002). Examples
than modeling host individuals as simply ‘‘sus- include NBS-LRR genes, the largest class of dis-
ceptible and uninfected’’ or ‘‘infected’’, more ease resistance genes. Plant genome projects have
details about the state of individuals could be in- indicated there are approximately 150 in
cluded. The first simple modifications might Arabidopsis and more than three times this number
include broad genotypic resistance and suscepti- in rice. Information about the number of resis-
bility. Further refinement could include transi- tance genes available may contribute to resistance
tional states of greater or lesser susceptibility gene deployment strategies. The identification of
based on physiological age, and probabilities of sequences associated with resistance genes may
exposure to other organisms that would induce also be applied to related plant species to answer
resistance. Matrix-based models of probabilities of long-standing questions about the number and
transitions from one state to another could be type of resistance genes in natural populations
applied to predict long-run states such as disease (Gilbert, 2002). Microbial genome projects are
severity or survival of different genotypes. Such providing similar estimates of the number and
models could potentially be applied to develop types of effector proteins in a single organism, such
both epidemiological theory and better disease as the number of gene products transferred into
management schemes. In the short-run, they could plant cells by the type III secretion system of
be used to ask questions about the incremental Pseudomonas syringae strain DC3000 (Collmer
benefits of adding information about host pheno- et al., 2002; Alfano and Collmer, 2004; Rohmer
type to epidemic models. In the long-run, these et al., 2004; Chang et al., 2005). These are not only
models could contribute to a much deeper under- important proteins that make the bacteria suc-
standing of epidemic dynamics. cessful pathogens, but also the targets of
The more complete genetic information from plant disease resistance proteins. These are just a
DNA sequencing can be used to study long- few examples of how partial and whole genome
standing questions of population structure, host sequencing projects can contribute to under-
specificity, and phylogenetics. Due to the growth standing host–pathogen interactions.
of sequence databases and the reduction in PCR Studies of gene expression in pathogens are still
amplification and sequencing costs, determining limited, but, where available, are being used to
the sequence of a specific gene in a microorganism understand expression of genes during plant colo-
is often the most efficient way to determine the nization, and under various cultural practices. As
species of the microorganism. Databases now more whole genomes are sequenced, microarrays
carry information on a huge number of organisms, using various platforms are becoming available for
and matching an unknown sequence to the several pathogens. As examples, arrays exist for the
sequences in a database like GenBank takes only a rice blast fungus and for several bacterial plant
few minutes, although one must keep in mind pathogens. Techniques other than microarrays are
that not every sequence accession in GenBank is also being applied to understand gene expression;
annotated correctly. Reduced sequencing and for example, serial analysis of gene expression
DNA amplification costs make the identification (SAGE) has been applied to study gene
of components of large microbial populations expression in response to rice blast infection
feasible. Entire fungal or bacterial communities (http://www.mgosdb.org/). Microarrays can also
can be characterized taxonomically by incorpo- be used in comparative genomics studies of closely
rating new techniques such as ‘‘shotgun sequenc- related pathogens using full genome sequences. For
ing’’ of a community’s collective genome and using example, the gene content of the human pathogen
genome database searches to identify species and Yersinia pestis has been studied as an indicator for
predict gene function (Venter et al., 2004). At a adaptation (Chain et al., 2004). Genomes have
finer scale, sequencing specific genes in pathogen been evaluated to determine what is missing in a
40

fastidious, xylem limited species like Xylella genomics’’ to inform ‘‘comparative epidemiol-
fastidiosa by comparison to other less fastidious ogy’’. For example, Kranz (2002) discusses several
bacteria (Van Sluys et al., 2002). The genomes of disease parameters influenced by host plant resis-
Xanthomonas oryzae pv. oryzae and X. oryzae pv. tance that together predict epidemic rates and
oryzicola are being compared for insights into why outcomes: disease intensity, incubation period,
the first is systemic in xylem while the second grows latent period, infection efficiency, disease effi-
in mesophyll (A. Bogdanove, pers. comm.). ciency, infection rate, lesion size, infectious period,
The greater availability of genetic information and sporulation intensity. In comparative epide-
will allow plant pathologists to move ‘‘beyond the miology, the differences in these parameters
inoculation experiment’’ in studies of the genetic between host–pathogen systems can be evaluated
features of host–pathogen interactions. In the past, both in terms of their typical values and the fre-
painstaking and expensive analyses of genetic quency distribution of these values in response to
expression in host–pathogen interactions have typical forms of resistance. The availability of gene
generally been applied to studies of pathogens expression data will also make it possible to study
introduced to hosts either in highly conducive disease parameters as a function of measures of
environments, in the case of rust fungi, for exam- gene expression, given a particular genotype
ple, or directly inoculated into host tissues, in the (Figure 2), in the same way that the expression
case of many bacterial pathogens. In contrast, it levels of key genes associated with the initiation of
would be extremely interesting and valuable to flowering have been used to predict flowering time
have a greater understanding of the genetic basis (Welch et al., 2003, 2005).
for the broad range of other epidemiological fea- There is a basic need in epidemiology for
tures that are important in determining popula- improved diagnostic systems and genomic ad-
tion-level interactions between host and pathogen. vances will greatly expand the tools available. For
For example, from the standpoint of the pathogen, example, as models of the risk of invasion by
aside from direct effects on virulence or aggres- particular plant pathogens are constructed, their
siveness, what are the genes most important for validation depends on researchers’ ability to
features such as survival in soil or on plant sur- determine precisely the abundance of pathogens in
faces, tolerance for temperature extremes, dis- a range of environmental settings. In their simplest
persal capability, or other specialized features such form, such studies require the ability to detect and
as conversion from production of urediniospores identify particular species of pathogens. Diagnosis
to production of teliospores in rust fungi? At lar- may also be taken to more sophisticated levels
ger epidemic scales, the genetic characteristics through the ability to detect particular genotypes,
most important to dispersal might be those that
affect survival of propagules under challenging
environmental conditions. These characteristics Allele A
Epidemic parameter
(e.g., latent period)

would help determine whether the long-distance

transport events so important to establishment of Allele B
epidemics in new areas occur or not. From the
standpoint of the host, what genes are most
important for predicting epidemics aside from
direct resistance genes, including features such as
the probability of escape through faster or slower Allele C
movement through developmental stages, ‘‘leak-
ing’’ of compounds in the phyllosphere or rhizo- Expression level of one or more genes
related to disease resistance
sphere, and architectural features that affect
microclimate? Such information would be useful Figure 2. Schematic of possible relationships between gene
both for applied crop plant breeding programs and expression levels and epidemic parameters such as latent peri-
for understanding resistance profiles in natural od, infection rates, lesion size, etc. Such a relationship could
be incorporated in models to refine predictions of epidemic
plant populations. outcomes such as disease severity or disease incidence. The
There is the potential to identify genes predictive different lines indicate possible differences in response for dif-
of epidemiological features using ‘‘comparative ferent alleles.
 41

in particular, those that are capable of causing ways. First, at the smallest scale, genes may be
disease. Presence of genes for these traits, such as conceptualized to interact within a cell comparably
genes related to pathogenicity, toxin production, to the way that species interact within an
and other epidemiological features, if known, ecosystem (Mauricio, 2005). For example, it may
could be used to more reliably measure genotypes be useful to apply such models to the interactions
in a population responsible for disease. A partic- between different defense response pathways.
ularly important application might be the identi- There is evidence the jasmonate (or ethylene) and
fication of disease, through evaluation of host or salicylic acid pathways affect somewhat different
pathogen, when infection is still at very low levels, pathogens and pests but also interact with each
to allow for early management that might, for other (Thomma et al., 1998; Glazebrook et al.,
example, allow an invasive pathogen to be eradi- 2003). Depending on the response examined, they
cated before it has become well-established. Fur- may sometimes be viewed as complementary (van
ther refinement for successful diagnosis of gene Wees et al., 2000) and in other cases as in com-
expression may allow assessment of features such petition (Spoel et al., 2003).
as quorum sensing status (van Bodman et al., Second, an individual plant may be conceptu-
2003). The use of microarrays also opens the alized as a population of cells or organs across
possibility of synoptic rapid-throughput diagnostic which gene expression occurs. It is now possible to
procedures for huge numbers of organisms for the measure gene expression in individual plant cells
study of the community characteristics of systems (Kerk et al., 2003; Nakazono et al., 2003) so the
such as disease-suppressive soils, the phyllosphere, spatial pattern of expression through an individual
and endophytic communities. These approaches host can be measured and modeled at whatever
could bring great advances in understanding spatial grain is motivated by the experimental
microbial biodiversity, including the potential questions. Spatial patterns of defense responses
to find new non-culturable putative pathogens between cells are relevant both to how effective
through scans for microbial genes used for taxo- defense responses are to pathogen challenge and to
nomic classification or even genes associated with how the host responds to adjacent or subsequent
pathogenicity. Epidemiologists might also make challenges by the same or other pathogens. Could
good use of a genomic tool that would allow them models of the dispersal of individuals through
to study the past presence of pathogens through ecological landscapes be usefully adapted to
on-going expression in host populations. Such a describe the dispersal of gene products within and
measure of pathogen ‘‘footprints’’ could support between cells? State transition models could be
studies of long-term epidemics and changes in host applied to individual plants in cases where it
resistance over time. But it appears that an indi- makes sense to treat them as a set of units, such as
cator of past infection is not readily available in different tissues and organs, each of which would
plants, or at least researchers have not yet dis- have its own expression status. This could be
covered how to recognize it. addressed using a variation on SI models. Predic-
tions based on these models might include the
predicted infection level as well as the predicted
How models of populations and communities may plant growth rate.
apply to systems of gene expression to inform Third, experiments in epidemiology might begin
a ‘‘population biology’’ of gene expression with models within individuals, predicting infec-
tion levels based on the expression of particular
A null model for how models of populations and genes, and then expand on these to predict infec-
communities apply to the study of gene expression tion rates in plant populations based on the gene
might be ‘‘consilience’’; that is, the null model expression rates in individuals. A null model for
might be that the same models will apply across such a study might be that the mean expression
scales, so tests could be developed to determine level of individual hosts is fully predictive of the
where population models do and don’t adequately level of infection in the population. In contrast to
explain patterns of gene expression. the null model, it would be interesting to determine
Models from population biology can be applied whether the frequency distribution of the expres-
in the study of gene expression in three general sion rates, and perhaps even their spatial pattern,
42

in the different host individuals would substan- sion, as affected by environment, over a period of
tially improve predictions of epidemic features, time. Younger tissues might only experience
just as different patterns of disease severity across ‘‘indirect’’ effects from past environmental condi-
individuals can result in different overall yields for tions, perhaps as an analog to maternal and
the same mean disease severity (Hughes, 1996). grandmother effects in individuals. Induced sys-
Simulation modeling might be used for initial tests temic resistance might be an example of short-term
of the sensitivity of epidemic outcomes to such gene expression in response to non-pathogens
patterns of expression. Addressing questions such while systemic acquired resistance might be an
as these with an understanding of mechanism will example of short-term gene expression in response
require considerably greater understanding of the to pathogens or to chemical stimulants. The timing
relationships between gene expression and gene of infection and its effects on losses in productivity
product physiological function. This third scale is can also be evaluated through variations on
addressed further in a later section. time-of-infection models for predicting yield
The study of gene expression offers a new loss (Madden et al., 2000) that include explicit
method for measuring integrated effects of envi- descriptions of host gene expression in response to
ronmental variation (Figure 3). Environmental infection. The schematic model in Figure 3 applies
variables such as temperature and precipitation most directly in agricultural systems in which a
are standard predictors of disease progress in epi- genotype is generally maintained, at least for a
demiological models (Jeger, 2004), and integrated season, through removal of competitors. A more
forms such as ‘‘growing degree days’’ are already complicated model might be developed in which
commonly used to predict growth stage as a model host genotypes can be replaced by other plant
component. Different types of host responses may genotypes. The schematic might also be adapted to
be integrated over different time intervals. take into account the possibility of thresholds such
Growth stage, or more specialized responses like that long-term changes in phenotype could be
the development of sun and shade leaves, are the produced by short-term gene expression at critical
products of the cumulative effects of gene expres- time points in development.

Host Current host

genotype phenotype

Long-term gene expression Long-term

phenotypic
characteristics

Medium-term gene Medium-term

expression phenotypic
characteristics

Short-term Short-term
gene phenotypic
expression characteristics

A. Environmental conditions over time

B. Pathogen community over time

Figure 3. The current host phenotype, at any spatial scale within a host individual, is a form of integration of the individual’s
environment, including the composition of the pathogen community, acting on the host genotype. Long-term phenotypic character-
istics would include features such as physiological age of leaves or roots, forms of specialization such as the development of sun or
shade leaves, and other characteristics that may influence disease resistance. Short-term characteristics might include features such
as upregulation of pathways contributing to induced resistance. Of course, host gene expression will also influence pathogen popu-
lations and even the abiotic microenvironment.
 43

How genomics can contribute to understanding model, while others are not (Hulbert, 1998). High
of resistance gene evolution and durability levels of sequence variation have been observed at
of resistance most R gene loci examined. This is consistent with
the arms race model’s prediction that R genes
A major goal of agricultural plant pathology is the would evolve rapidly, creating novel alleles with
development of durable resistance to plant new specificities in response to pressure imposed by
pathogens in agricultural species. ‘‘Durable resis- rapidly evolving pathogen populations. Loci like L
tance’’ has been defined as resistance that is still of flax (Ellis et al., 1999), which is structurally
effective after it has been deployed over a wide simple but has multiple resistance alleles, exhibit
area, over a long period of time, in a disease- extremely high levels of polymorphism compared
conducive environment (Johnson, 1981). Without to most genes. At some R gene loci, the patterns of
durable resistance, plant breeders are forced to nucleotide substitution between alleles or family
continually incorporate new resistance genes in members show evidence of the types of diversifying
crop varieties as pathogen populations adapt to selection that might be predicted by an arms race
infect older varieties with previously deployed model. While polymorphic nucleotides are usually
resistance genes. An understanding of the evolu- synonymous (not affecting the encoded amino
tion of host and pathogen genes affecting host– acid) at most loci, the opposite is true of certain
pathogen interactions is needed to form strategies regions of some R gene loci. This is most often true
for the durable deployment of resistance in agri- in regions of R genes thought to code for the ligand
culture. It has long been thought that under- recognition part of the protein, like the leucine-rich
standing of the relative importance of the disease repeat regions (Parniske et al., 1997; Meyers et al.,
effector proteins from bacterial and fungal 2003). Evidence of diversifying selection in other
pathogens that are detected by R genes (i.e., the regions of R genes, like the TIR domain-encoding
products of avirulence genes) should provide in- regions of the L alleles, has suggested they may also
sight into which R genes might have more durable be involved in recognition (Luck et al., 2000).
effects, but this idea has had limited impact One interpretation of an arms race evolutionary
because of the difficulty of identifying and char- progression is that there should be little variation
acterizing these effector protein genes. Compara- at a given R gene locus at one point in time and
tive genomic approaches for identifying these that most R gene alleles should be fairly recent in
genes and functional genomic approaches to ob- their evolutionary origin. This would be expected
tain ‘knocking-outs’ of their function is making if new highly effective R genes arose periodically
this increasingly feasible (Leach et al., 2001). Some and replaced the older ‘defeated’ alleles. The
resistance genes, like mlo of barley (Buschges polymorphic nature of many R gene loci indicates
et al., 1997), may confer resistance without inter- this is apparently not the case for most of them. In
acting with specific pathogen effector proteins. fact the partitioning of polymorphism between
These genes may provide non-specific resistance by functional alleles and non-functional alleles at the
changing the physiology and gene expression of Rpm1 and Rps5 loci of Arabidopsis indicated that
the plant before pathogen challenge (Wolter et al., the classes of alleles have co-existed for a long
1993). Gene expression analysis has indicated period (Bergelson et al., 2001; Tian et al., 2002),
other resistance genes with suspected non-specific probably the result of some form of balancing
effects may alter expression of defense genes selection. While actual estimates of the age of
before pathogen challenge (Bowden and Hulbert, specific resistance gene alleles are not available,
unpublished). Such analysis should be useful in this may be an indication that some R gene alleles
identifying genes controlling non-specific and thus are ancient. In contrast, no evidence that resis-
durable resistance and also provide insight into the tance alleles are ancient has been obtained by
possible physiological cost of the resistance. sequencing the same resistance allele from multiple
The isolation and sequence analysis of several germplasm accessions. If resistance alleles are
resistance genes has provided insight into the evo- indeed ancient, it should be possible to identify
lution of disease resistance in plants (Hulbert et al., versions that have accumulated extensive neutral
2001). Some of the results of these analyses are sequence polymorphisms. This has not yet been
consistent with a classical evolutionary arms race the case in the limited experiments that have been
44

conducted (Caicedo et al., 1999; Jia et al., 2003; absolute. Genomic tools will allow discovery of the
Smith et al., 2004). The low nucleotide diversity genes responsible for quantitative traits, and may
among the functional alleles of these loci is con- make it easier to determine whether resistance
sistent with the idea that they could be recently governed by quantitative traits is truly more
evolved, although other explanations are possible. durable; whether the effects of QTL are actually
The sequence evidence collected to date implies less pathotype specific, or whether an apparently
that different R gene loci are evolving in different more durable effect may be mediated by a weaker
manners. For example, some appear to be under selection on individual pathogen genes. To the
strong diversifying selection while others do not. extent that function can be inferred from sequence,
The implications of an R genes’ evolutionary his- the response of pathogens to particular minor
tory for the stability of the resistance it confers is genes may be better predicted as this information
not clear, but the ability to predict durability based becomes available. It will be particularly useful if
on genomic analysis would be quite useful for crop comparative genomics would allow predictions of
improvement strategies. Molecular analyses of the interactions between minor resistance genes
resistance proteins and their corresponding aviru- and their responses to abiotic and biotic environ-
lence proteins have indicated that some physically ments. Functional genomics may also contribute
interact directly (Scofield et al., 1996; Tang et al., to the identification of new minor resistance genes.
1996; Jia et al., 2000, Deslandes et al., 2003) while QTL analysis or the identification of quantitative
others detect modifications of other host compo- trait loci provides a powerful tool for assessing the
nents (Mackey et al., 2003; Axtell and Staskawicz, fitness consequences for genes including resistance
2003). It is possible that whether an R gene rec- genes. For example, Newcombe and Bradshaw
ognizes effector (avirulence) genes directly or (1996) used it to identify genes of large effect that
protects host targets from modification by effector changed the resistance of poplar to pathogenic
proteins affects the type of selection pressure Septoria populicola with community level effects.
driving its evolution (Ponciano et al., 2003). This The study of pathogen and host co-evolution in
association, however, is not yet clear due to the natural plant populations is also important for
small number of interactions in which this type of understanding what role pathogens may have
information is known. For R proteins that guard played in structuring plant communities. In studies
other host components, it is not clear if the nature demonstrating the importance of genetic variation
of the host protein being guarded affects the in host plant species within a larger community
durability of the R gene, but it might be expected that includes pathogens, hybridization of host
that some targets are more important to the plants (e.g. willows, sagebrush, oaks) has led to
pathogens ability to cause disease than others. The fundamental changes in the species composition of
nature of the effector gene, particularly how the entire community (reviewed in Whitham et al.,
essential it is to pathogenicity, has been proposed 1999). This ‘‘extended phenotype’’ effect would be
by many to affect R gene durability and recent reflected in the context of epidemiology by the
data indicates this is true (Vera Cruz et al., 2000). dying out of some pathogens and replacement with
One response to the problem of rapid ‘‘break- others (Whitham et al., 2003b). Agricultural sys-
down’’ of resistance in agricultural systems has tems and unmanaged systems offer an interesting
been a shift by some plant breeders toward greater contrast, because the selection pressures in un-
use of minor resistance genes that each contribute managed systems are ‘‘direct’’ while selection
a small amount of resistance and are generally pressures in agricultural systems are mediated by
thought to be more durable (Leung et al., 2003; human decision-making.
Liu et al., 2004). However, these genes, because of
their small effects, are more difficult to study in the
field and even to recognize by the phenotypes of The landscape ecological genomics of host
individual plants. The use of genetic markers has populations and communities, in terms of both
made the incorporation of minor genes easier, but genotypic and phenotypic resistance
the problem remains that, because we do not know
what genes are responsible for quantitative traits, Once meteorological measurements could be col-
the association of the markers with the traits is not lected using automated systems, epidemiologists
 45

were faced with the question of what temporal complicated. Some models have assumed that
scale of resolution was needed for understanding epidemic ‘‘waves’’ move out from an initial point
epidemic progress. Information about variation in with constant velocity to simplify the modeling of
temperature at the scale of minutes is not generally the system, while other researchers predict that
needed for predicting epidemic features. But whe- epidemic waves accelerate (Scherm, 1996; Cowger
ther predictions are improved by resolving differ- et al., 2005).
ences in temperature at the scale of days or weeks A genomics approach applied to epidemiology
will vary from one host–pathogen system to could explore multiple spatial and temporal scales
another, based on characteristics such as pathogen as well as levels of detail in genomic status, per-
generation time, and requires attention for the haps employing cellular automata models (e.g.,
construction of good models. Information about Kleczkowski et al., 1997; Figure 4). Within a host
variation in meteorological features across space is individual, the local phenotype might be at the
still not so readily collected at very fine scales, scale of a leaf or of a cell. Local gene expression
though the increasing availability of ‘‘smart dust’’ might be at the point of infection; for example,
and other tiny wireless sensor networks may change within compared to beyond a green island of host
that (e.g., http://webs.cs.berkeley.edu/). The same tissue formed around an infection by a rust fungus.
question of appropriate scale of variation to Regional gene expression within an individual
include for modeling will arise for the new spatial might be expression in tissues adjacent to infec-
maps of meteorological features. Similarly, the tion. Within a host individual and its immediate
potentially huge amount of information about gene environment, a wide range of pathogens may be
expression will require research to determine what present, specializing on different host tissues.
scale of variation is important to include in pre- Competition between particular pathogens may
diction of epidemics for particular host–pathogen play out differently depending on the time of
systems. The cost of microarray analyses limits the infection and the type of plant tissue (Adee et al.,
number of samples in time and space for now, but 1990; Al-Naimi et al., 2005). The question for
as costs become less limiting, epidemiological re- epidemiologists will be what spatial and temporal
search will focus more on determining optimal resolution is needed for predicting epidemics with
scales of variation in expression data to include in the new and upcoming abundance of data, as op-
predictive models. posed to averaging over host and pathogen indi-
Plant disease epidemiology has developed models viduals’ genomic status across space and time.
of disease foci and how these foci expand in time In host populations, ‘‘expression foci’’ in which
and space (Zadoks and Vandenbosch, 1994; host individuals share altered gene expression
Waggoner and Aylor, 2000), including studies of patterns may form around inoculum sources, with
the spatial pattern of disease used to draw infer- properties related to those of disease foci. Gene
ence about modes of dispersal and to devise opti- expression changes in hosts in response to expo-
mal sampling strategies. Landscape ecology also sure to pathogens and other microbes may range
offers methods for studying spatial features with from increased resistance through SAR or ISR to
models for describing the relationships between increased susceptibility because of weakened tissue
organisms in landscapes and for describing the integrity. The effect of exposure to pathogens that
connectivity of features (With, 2002). In agricul- do not infect has the potential to be substantial, at
tural systems, the spatial pattern of host genotypic least temporarily; Calonnec et al. (1996) estimated
resistance is sometimes manipulated through the that the infection efficiency of Puccinia striiformis
construction of intercropping systems and/or use was reduced by 44% when plants were previously
of mixed genotypes within a crop species (Garrett exposed to an ‘‘inducer race’’ of the pathogen. At
and Mundt, 1999; Zhu et al., 2000; Mundt, 2002). increasing distances from a primary inoculum
And, of course, most unmanaged systems include source, exposure to inoculum may have occurred
a mixture of plant species that, with few exceptions at more recent time points, potentially resulting in
(e.g., Phytophthora cinnamomi), do not tend to waves of different expression patterns surrounding
share the same pathogen species. Mixtures of the initial source areas. Spatial patterns of abiotic
susceptible and other genotypes make models of features, such as differences in topography that
disease spread through space somewhat more produce cooler or wetter local conditions, may
46

Host
Host individual
landscape
“Plant-wide” gene expression

Genotype “Organ-wide” gene expression

Pathogen
landscape
Local gene expression

Environmental
landscape
Local phenotype

Figure 4. Each host individual is potentially influenced by the landscape of hosts, pathogens, and other biotic and abiotic environ-
mental features. Within a host individual, these influences may play out through ‘‘plant-wide’’, ‘‘organ-wide’’, or more local gene
expression, depending on the scale of variation of each feature in the landscape and how it acts upon the host individual. ‘‘Plant-
wide’’ gene expression might include responses to factors such as drought and disease that alters water relations within the host.
‘‘Organ-wide’’ gene expression might include responses to factors such as stem or petiole lesions. Local gene expression might in-
clude responses such as localized forms of induced resistance. Models related to spatial scale and scale of genetic detail that could
be experimentally evaluated as predictors of epidemic features include the following, presented in a hierarchy of increasing com-
plexity. ‘‘The host landscape is sufficiently described to predict epidemic features and outcomes by information about... d ... the
abundance of host species’’ d ... the abundance of host genotypes’’ d ... the abundance and spatial pattern of host species’’ d ...
the abundance and spatial pattern of host genotypes’’ d ... the mean level of gene expression among host individuals’’ d ... the
spatial pattern of gene expression among host individuals’’ d ... the spatial pattern of gene expression within host individuals.’’

also produce expression foci relevant to epidemics. epidemic rates on susceptible genotypes that would
Studies of gene expression in landscapes may be predicted by the presence of other genotypes
develop distinctions analogous to the distinction will be reduced if autoinfection rates are high;
between a dispersal gradient and a disease gradi- more inoculum will land on susceptible host indi-
ent. Disease gradients may differ markedly from viduals rather than being lost through dispersal to
dispersal gradients if the success rates per unit of non-hosts (Garrett and Mundt, 1999; Mundt,
inoculum are low, particularly if the reproductive 2002). It may prove useful to develop analogous
rates of the pathogen are density dependent concepts for gene expression, so that ‘‘autoinduc-
(Garrett and Bowden, 2002). There may be similar tion’’ of gene expression would occur when mi-
effects for gene expression, such that thresholds of crobes associated with a particular plant
exposure to pathogen inoculum, for example, must individual disperse to other organs within that
be exceeded before substantial gene expression individual to induce SAR, ISR, or other reactions.
results. At much smaller spatial scales, gene By comparison, ‘‘alloinduction’’ would occur
expression in bacterial populations may give in- when microbes are dispersed to a different plant
sights into quorum sensing and its implications for individual to induce these reactions. Higher rates
density dependent reproduction (van Bodman of alloinduction compared to autoinduction would
et al., 2003). tend to result in higher mean levels of SAR or ISR
Epidemiologists have developed the terms in populations, and the rate of alloinduction
autoinfection and alloinfection to describe infec- would be a function of host size and the dispersal
tion of a target host individual by inoculum pro- properties of the relevant microbe populations.
duced on the same target host individual vs. Feedback between host and pathogen would
infection by inoculum produced on other host occur as pathogens disperse, infect or elicit other
individuals, respectively (Robinson, 1976). The responses in plants, and then disperse further
rate of autoinfection can be an important predic- through a landscape of phenotypic resistance that
tor for epidemics of non-systemic disease in mixed has potentially been altered in response to previ-
genotype host populations. If some host genotypes ous dispersal. Good models of such a sys-
are susceptible and others are not, the reduction in tem would require the ability to predict plant
 47

phenotypic resistance levels in response to the genomic tools will also allow great advances in the
biotic and abiotic environment, pathogen pheno- study of phenotypic resistance. It will finally be
types in response to ‘‘non-host’’ environmental possible to thoroughly evaluate the many ideas put
features, plant phenotypic responses to exposure forward about age-related resistance and the effects
to pathogens, etc. (Figure 1). Of course, one of the biotic and abiotic environment on pheno-
challenge is simply to be able to describe the cur- typic resistance. Conversely, epidemiology pro-
rent level of phenotypic resistance of a host indi- vides the context for understanding the role and
vidual. Adding the spatial pattern of host significance of pathogen genes and plant genes re-
genotypes provides additional modeling chal- lated to pathogen reproduction and also provides
lenges. The level of aggregation of susceptible models for evaluating landscapes of plant pheno-
hosts will produce a particular ‘‘genotype unit types.
area’’ (Mundt and Leonard, 1986), or area occu-
pied by a single host genotype, and help to deter-
mine the extent to which microbial populations are Acknowledgements
spread between host species/genotypes or tend to
be dispersed within host species/genotypes. This Thanks to J. Bai, R. L. Bowden, A. H. C. van
pattern of host genotypes will also influence the Bruggen, members of the KSU Ecological
pattern of expression in response to microbes Genomics Community, and members of the KSU
associated with a particular host genotype. The Plant Disease Ecology Lab for helpful discussions
combination of host genotype spatial patterns and of these ideas, to EJPP reviewers for helpful com-
the spatial pattern of the biotic and abiotic envi- ments, and to S. P. Dendy for assistance prepar-
ronment will result in a host ‘‘phenotype mixture’’. ing this manuscript. This work was supported by
Just as the effects of genotype diversity vary for the U.S. National Science Foundation under
different host–pathogen systems (Lannou et al., Grant Nos. DEB-0130692 and DEB-0516046, by
1994; Mundt et al., 1995; Ngugi et al., 2001; the Ecological Genomics Initiative of Kansas
Mitchell et al., 2002; Cox et al., 2004), the com- through NSF Grant No. EPS-0236913 with
plex communities of plants and microbes involved matching funds from the Kansas Technology
in induced resistance may experience different Enterprise Corporation, by the NSF Long Term
patterns of spatial effects on induction. Models of Ecological Research Program at Konza Prairie,
epidemics in genotype mixtures will be useful in by the U.S. Department of Energy Office of Sci-
this context, but new models will also be needed. ence (Program in Ecosystem Research) Grant
No. DE-FG02-04ER63892, by the US Smithso-
nian Tropical Research Institute (Center for
Conclusion Tropical Forest Science), and by USDA Grant
No. 2002-34103-11746. This is Kansas State
Epidemiology will benefit from new genomic Experiment Station Contribution No. 06-22-J.
technologies in several ways. New diagnostic
techniques will make the development of a ‘‘com-
munity epidemiology’’ more practical, through References
providing the ability to characterize thousands of
microorganisms simultaneously as well as identi- Adee SR, Pfender WF and Hartnett DC (1990) Competition
fying particular genes and alleles. New techniques between Pyrenophora tritici-repentis and Septoria nodorum
in the wheat leaf as measured with de Wit replacement
will make it easier to extend genetic analyses of series. Phytopathology 80: 1177–1182
pathogens beyond virulence genes, by facilitating Alfano JR and Collmer A (2004) Type III secretion system
the study of the population structure and evolution effector proteins: double agents in bacterial disease and
of genes important for other important features plant defense. Annual Review of Phytopathology 42: 385–
414
such as the ability to survive in non-conducive
Al-Naimi FA, Garrett KA and Bockus WW (2005) Competi-
environments. Functional genomic analysis of tion, facilitation, and niche differentiation in two foliar
pathogen virulence genes and host resistance and pathogens. Oecologia 143: 449–457
defense response genes will enable better Antonovics J (2003) Toward community genomics? Ecology 84:
predictions of the durability of resistance. New 598–601
48

Axtell MJ and Staskawicz BJ (2003) Initiation of RPS2- Cortesi P and Milgroom MG (2001) Outcrossing and diversity
specified disease resistance in Arabidopsis is coupled to the of vegetative compatibility types in populations of Eu-
AvrRpt2-directed elimination of RIN4. Cell 112: 369–377 typa lata from grapevines. Journal of Plant Pathology 83:
Bai J, Pennill LA, Ning J, Lee SW, Ramalingam J, Webb CA, 79–86
Zhao B, Sun Q, Nelson JC, Leach JE and Hulbert SH Cowger C, Wallace LD and Mundt CC (2005) Velocity of
(2002) Diversity of nucleotide binding site-leucine-rich spread of wheat stripe rust epidemics. Phytopathology 95:
repeat genes in cereals. Genome Research 12: 1871–1884 972–982
Bergelson J, Kreitman M, Stahl EA and Tian DC (2001) Cox CM, Garrett KA, Bowden RL, Fritz AK, Dendy SP and
Evolutionary dynamics of plant R-genes. Science 292: Heer WF (2004) Cultivar mixtures for the simultaneous
2281–2285 management of multiple diseases: Tan spot and leaf rust of
Black WC IV, Baer CF, Antolin MF and DuTeau NM (2001) wheat. Phytopathology 94: 961–969
Population genomics: genome-wide sampling of insect Cui J, Bahrami AK, Pringle EG, Hernandez-Guzman G,
populations. Annual Review of Entomology 46: 441–469 Bender CL, Pierce NE and Ausubel FM (2005) Pseudomo-
Buschges R, Hollricher K, Panstruga R, Simons G, Wolter M, nas syringae manipulates systemic plant defenses against
Frijters A, van Daelen R, van der Lee T, Diergaarde P, pathogens and herbivores. Proceedings of the National
Groenendijk J, Topsch S, Vos P, Salamini F and Schulze- Academy of Sciences of the United States of America 102:
Lefert P (1997) The barley Mlo gene: a novel control 1791–1796
element of plant pathogen resistance. Cell 88: 695–705 Deslandes L, Olivier J, Peeters N, Feng DX, Khounlotham M,
Caicedo AL, Schaal BA and Kunkel BN (1999) Diversity and Boucher C, Somssich I, Genin S and Marco Y (2003)
molecular evolution of the RPS2 resistance gene in Arabid- Physical interaction between RRS1-R, a protein conferring
opsis thaliana. Proceedings of the National Academy of resistance to bacterial wilt, and PopP2, a type III effector
Sciences of the United States of America 96: 302–306 targeted to the plant nucleus. Proceedings of the National
Calonnec A, Goyeau H and de Vallavielle-Pope C (1996) Academy of Sciences of the United States of America 100:
Effects of induced resistance on infection efficiency and 8024–8029
sporulation of Puccinia striiformis on seedlings in varietal Durrant WE and Dong X (2004) Systemic acquired resistance.
mixtures and on field epidemics in pure stands. European Annual Review of Phytopathology 42: 185–209
Journal of Plant Pathology 102: 733–741 Ellis JG, Lawrence GJ, Luck JE and Dodds PN (1999)
Chain PSG, Carniel E, Larimer FW, Lamerdin J, Stoutland Identification of regions in alleles of the flax rust resistance
PO, Regala WM, Georgescu AM, Vergez LM, Land ML, gene L that determine differences in gene-for-gene specific-
Motin VL, Brubaker RR, Fowler J, Hinnebusch J, Mar- ity. The Plant Cell 11: 495–506
ceau M, Medigue C, Simonet M, Chenal-Francisque V, Escriu F, Fraile A and Garcia-Arenal F (2000a) Evolution of
Souza B, Dacheux D, Elliott JM, Derbise A, Hauser LJ and virulence in natural populations of the satellite RNA of
Garcia E (2004) Insights into the evolution of Yersinia Cucumber mosaic virus. Phytopathology 90: 480–485
pestis through whole-genome comparison with Yersinia Escriu F, Fraile A and Garcia-Arenal F (2003) The evolution of
pseudotuberculosis. Proceedings of the National Academy virulence in a plant virus. Evolution 57: 755–765
of Sciences of the United States of America 101: 13826– Escriu F, Perry KL and Garcia-Arenal F (2000b) Transmissi-
13831 bility of Cucumber mosaic virus by Aphis gossypii correlates
Chang JH, Urbach JM, Law TF, Arnold LW, Hu A, Gombar with viral accumulation and is affected by the presence of its
S, Grant SR, Ausubel FM and Dangl JL (2005) A high- satellite RNA. Phytopathology 90: 1068–1072
throughput, near-saturating screen for type III effector Feder ME and Mitchell-Olds T (2003) Evolutionary and
genes from Pseudomonas syringae. Proceedings of the ecological functional genomics. Nature Reviews Genetics
National Academy of Sciences of the United States of 4: 651–665
America 102: 2549–2554 Frick UB and Schaller A (2002) cDNA microarray analysis of
Chen W, Provart NJ, Glazebrook J, Katagiri F, Chang HS, fusicoccin-induced changes in gene expression in tomato
Eulgem T, Mauch F, Luan S, Zou G, Whitham SA, plants. Planta 216: 83–94
Budworth PR, Tao Y, Xie Z, Chen X, Lam S, Kreps JA, Garrett KA and Bowden RL (2002) An Allee effect
Harper JF, Si-Ammour A, Mauch-Mani B, Heinlein M, reduces the invasive potential of Tilletia indica. Phytopa-
Kobayashi K, Hohn T, Dangl JL, Wang X and Zhu T thology 92: 1152–1159
(2002) Expression profile matrix of Arabidopsis transcrip- Garrett KA and Mundt CC (1999) Epidemiology in mixed host
tion factor genes suggests their putative functions in populations. Phytopathology 89: 984–990
response to environmental stresses. The Plant Cell 14: Gilbert GS (2002) Evolutionary ecology of plant diseases
559–574 in natural ecosystems. Annual Review of Phytopathology
Collmer A, Lindeberg M, Petnicki-Ocwieja T, Schneider DJ 40: 13–43
and Alfano JR (2002) Genomic mining type III secretion Glazebrook J, Chen W, Estes B, Chang HS, Nawrath C,
system effectors in Pseudomonas syringae yields new picks Métraux JP, Zhu T and Katagiri F (2003) Topology of the
for all TTSS prospectors. Trends in Microbiology 10: 462– network integrating salicylate and jasmonate signal trans-
469 duction derived from global expression phenotyping. The
Cortesi P, Fischer M and Milgroom MG (2000) Identification Plant Journal 34: 217–228
and spread of Fomitiporia punctata associated with wood Hughes G (1996) Incorporating spatial pattern of harmful
decay of grapevine showing symptoms of Esca. Phytopa- organisms into crop loss models. Crop Protection 15: 407–
thology 90: 967–972 421
 49

Hulbert S (1998) Arming plants against pathogens-recombina- Madden LV, Hughes G and Irwin ME (2000) Coupling disease-
tion in resistance loci. Trends in Plant Science 3: 121–122 progress-curve and time-of-infection functions for predict-
Hulbert SH, Webb CA, Smith SM and Sun Q (2001) Resistance ing yield loss of crops. Phytopathology 90: 788–800
gene complexes: evolution and utilization. Annual Review Maleck K, Levine A, Eulgem T, Morgan A, Schmid J, Lawton
of Phytopathology 39: 285–312 KA, Dangl JL and Dietrich RA (2000) The transcriptome
Iavicoli A, Boutet E, Buchala A and Métraux JP (2003) of Arabidopsis thaliana during systemic acquired resistance.
Induced systemic resistance in Arabidopsis thaliana in Nature Genetics 26: 403–410
response to root inoculation with Pseudomonas fluores- Mauricio R (2005) Can ecology help genomics: the genome as
cens CHA0. Molecular Plant–Microbe Interactions 16: ecosystem?. Genetica 123: 205–209
851–858 Meyers BC, Kozik A, Griego A, Kuang H and Michelmore
Jeger MJ (2004) Analysis of disease progress as a basis for RW (2003) Genome-wide analysis of NBS-LRR-encoding
evaluating disease management practices. Annual Review genes in Arabidopsis. The Plant Cell 15: 809–834
of Phytopathology 42: 61–82 Milgroom MG and Peever TL (2003) Population biology of
Jia Y, Bryan GT, Farrall L and Valent B (2003) Natural plant pathogens: the synthesis of plant disease epidemiology
variation at the Pi-ta rice blast resistance locus. Phytopa- and population genetics. Plant Disease 87: 608–617
thology 93: 1452–1459 Mitchell CE, Tilman D and Groth JV (2002) Effects of
Jia Y, McAdams SA, Bryan GT, Hershey HP and Valent B grassland plant species diversity, abundance, and compo-
(2000) Direct interaction of resistance gene and avirulence sition on foliar fungal disease. Ecology 83: 1713–1726
gene products confers rice blast resistance. The European Mundt CC, Brophy LS and Schmitt MS (1995) Disease severity
Molecular Biology Organization Journal 19: 4004–4014 and yield of pure-line wheat cultivars and mixtures in the
Johnson R (1981) Durable resistance: definition of, genetic presence of eyespot, yellow rust, and their combination.
control, and attainment in plant breeding. Phytopathology Plant Pathology 44: 173–182
71: 567–568 Mundt CC and Leonard KJ (1986) Effect of host genotype unit
Jones PG, Allaway D, Gilmour DM, Harris C, Rankin D, area on development of focal epidemics of bean rust and
Retzel ER and Jones CA (2002) Gene discovery and common maize rust in mixtures of resistant and susceptible
microarray analysis of cacao (Theobroma cacao L.) varie- plants. Phytopathology 76: 895–900
ties. Planta 216: 255–264 Mundt CC (2002) Use of multiline cultivars and cultivar
Kerk NM, Ceserani T, Tausta SL, Sussex IM and Nelson TM mixtures for disease management. Annual Review of
(2003) Laser capture microdissection of cells from plant Phytopathology 40: 381–410
tissues. Plant Physiology 132: 27–35 Mysore KS, D’Ascenzo MD, He X and Martin GB (2003)
Kleczkowski A, Gilligan CA and Bailey DJ (1997) Scaling and Overexpression of the disease resistance gene Pto in tomato
spatial dynamics in plant–pathogen systems: from individ- induces gene expression changes similar to immune
uals to populations. Proceedings of the Royal Society of responses in human and fruitfly. Plant Physiology 132:
London B 264: 979–984 1901–1912
Kranz J, (2002) Comparative Epidemiology of Plant Diseases. Nakazono M, Qiu F, Borsuk LA and Schnable PS (2003)
Springer Laser-capture microdissection, a tool for the global analysis
Lannou C, de Vallavieille-Pope C, Biass C and Goyeau H of gene expression in specific plant cell types: identification
(1994) The efficacy of mixtures of susceptible and resistant of genes expressed differentially in epidermal cells or
hosts to two wheat rusts of different lesion size: controlled vascular tissues of maize. The Plant Cell 15: 583–596
condition experiments and computerized simulations. Jour- Newcombe G and Bradshaw HD (1996) Quantitative trait loci
nal of Phytopathology 140: 227–237 conferring resistance in hybrid poplar to Septoria populico-
Leach JE, Vera Cruz CM, Bai J and Leung H (2001) Pathogen la, the cause of leaf spot. Canadian Journal of Forest
fitness penalty as a predictor of durability of disease Research 26: 1943–1950
resistance genes. Annual Review of Phytopathology 39: Ngugi HK, King SB, Holt J and Julian AM (2001)
187–224 Simultaneous temporal progress of sorghum anthracnose
Leung H, Zhu Y, Revilla-Molina I, Fan JX, Chen H, Pangga I, and leaf blight in crop mixtures with disparate patterns.
Vera Cruz C and Mew TW (2003) Using genetic diversity to Phytopathology 91: 720–729
achieve sustainable rice disease management. Plant Disease Otten W, Filipe JAN, Bailey DJ and Gilligan CA (2003)
87: 1156–1169 Quantification and analysis of transmission rates for soil-
Liu B, Zhang SH, Zhu XY, Yang Q, Wu S, Mei M, Mauleon R, borne epidemics. Ecology 84: 3232–3239
Leach J, Mew T and Leung H (2004) Candidate defense Parniske M, Hammond-Kosack KE, Golstein C, Thomas CM,
genes as predictors of quantitative blast resistance in rice. Jones DA, Harrison K, Wulff BBH and Jones JDG (1997)
Molecular Plant–Microbe Interactions 17: 1146–1152 Novel disease resistance specificities result from sequence
Luck JE, Lawrence GL, Dodds PN, Shepherd KW and Ellis JG exchange between tandemly repeated genes at the Cf-4/9
(2000) Regions outside of the leucine-rich repeats of flax locus of tomato. Cell 91: 821–832
rust resistance proteins play a role in specificity determina- Pieterse CMJ and van Loon LC (1999) Salicylic acid-
tion. The Plant Cell 12: 1367–1377 independent plant defence pathways. Trends in Plant
Mackey D, Belkhadir Y, Alonso JM, Ecker JR and Dangl JL Science 4: 52–58.
(2003) Arabidopsis RIN4 is a target of the type III virulence Ponciano G, Ishihara H, Tsuyumu S and Leach JE (2003)
effector AvrRpt2 and modulates RPS2-mediated resistance. Bacterial effectors in plant disease and defense: keys to
Cell 112: 379–389 durable resistance? Plant Disease 87: 1272–1282
50

Ponciano G, Webb K, Bai J, Vera Cruz C and Leach JE (2004) Kitajima JP and Simpson AJ (2002) Comparative genomic
Molecular characterization of the avrXa7 locus from analysis of plant-associated bacteria. Annual Review of
Xanthomonas oryzae pv. oryzae field isolates. Physiological Phytopathology 40: 169–189
and Molecular Plant Pathology 64: 145–153 van Wees SCM, de Swart EAM, van Pelt JA, van Loon LC and
Robinson RA (1976) Plant Pathosystems. Springer-Verlag, Pieterse CMJ (2000) Enhancement of induced disease
New York resistance by simultaneous activation of salicylate- and
Rohmer L, Guttman DS and Dangl JL (2004) Diverse jasmonate-dependent defense pathways in Arabidopsis tha-
evolutionary mechanisms shape the type iii effector viru- liana. Proceedings of the National Academy of Science of
lence factor repertoire in the plant pathogen Pseudomonas the United States of America 97: 8711–8716
syringae. Genetics 167: 1341–1360 Venter JC, Remington K, Heidelberg JF, Halpern AL, Rusch
Rondon MR, August PR, Bettermann AD, Brady SF, Gross- D, Eisen JA, Wu D, Paulsen I, Nelson KE, Nelson W,
man TH, Liles MR, Loiacono KA, Lynch BA, MacNeil IA, Fouts DE, Levy S, Knap AH, Lomas MW, Nealson K,
Minor C, Tiong CL, Gilman M, Osburne MS, Clardy J, White O, Peterson J, Hoffman J, Parsons R, Baden-Tillson
Handelsman J and Goodman RM (2000) Cloning the soil H, Pfannkoch C, Rogers YH and Smith HO (2004)
metagenome: a strategy for accessing the genetic and Environmental genome shotgun sequencing of the Sargasso
functional diversity of uncultured microorganisms. Applied Sea. Science 304: 66–74
and Environmental Microbiology 66: 2541–2547 Vera Cruz CM, Bai J, Oña I, Leung H, Nelson RJ, Mew TW
Ryals JA, Neuenschwander UH, Willits MG, Molina A, Steiner and Leach JE (2000) Predicting durability of a disease
H-Y and Hunt MD (1996) Systemic acquired resistance. resistance gene based on an assessment of the fitness loss
The Plant Cell 8: 1809–1819 and epidemiological consequences of avirulence gene muta-
Schenk PM, Kazan K, Manners JM, Anderson JP, Simpson tion. Proceedings of the National Academy of Sciences of
RS, Wilson IW, Somerville SC and Maclean DJ (2003) the United States of America 97: 13500–13505
Systemic gene expression in Arabidopsis during an incom- vonBodman SB, Bauer WD and Coplin DL (2003) Quorum
patible interaction with Alternaria brassicicola. Plant Phys- sensing in plant-pathogenic bacteria. Annual Review of
iology 132: 999–1010 Phytopathology 41: 455–482
Scherm H (1996) On the velocity of epidemic waves in model Waggoner PE and Aylor DE (2000) Epidemiology: a science of
plant disease epidemics. Ecological Modelling 87: 217–222 patterns. Annual Review of Phytopathology 38: 71–94
Scofield SR, Tobias CM, Rathjen JP, Chang JH, Lavelle DT, Wan J, Dunning MF and Bent AF (2002) Probing plant–
Michelmore RW and Staskawicz BJ (1996) Molecular basis pathogen interactions and downstream defense signaling
of gene-for-gene specificity in bacterial speck disease of using DNA microarrays. Functional and Integrative Ge-
tomato. Science 274: 2063–2065 nomics 2: 259–273
Smith KP and Goodman RM (1999) Host variation for Welch SM, Roe JL, Das S, Dong Z, He R and Kirkham MB
interactions with beneficial plant-associated microbes. (2005) Merging genomic control networks and soil-plant-
Annual Review of Phytopathology 37: 473–491 atmosphere-continuum models. Agricultural Systems 86:
Smith SM, Pryor AJ and Hulbert SH (2004) Allelic and 243–274
haplotypic diversity at the Rp1 rust resistance locus of Welch SM, Roe JL and Dong Z (2003) A Genetic neural
maize. Genetics 167: 1939–1947 network model of flowering time control in Arabidopsis
Spoel SH, Koornneef A, Claessens SMC, Korzelius JP, Van thaliana. Agronomy Journal 95: 71–81
Pelt JA, Mueller MJ, Buchala AJ, Métraux JP, Brown R, Weller DM, Raaijmakers JM, McSpadden-Gardener BB and
Kazan K, Van Loon LC, Dong X and Pieterse CMJ (2003) Thomashow LS (2002) Microbial populations responsible
NPR1 modulates cross-talk between salicylate- and jasm- for specific soil suppressiveness to plant pathogens. Annual
onate-dependent defense pathways through a novel func- Review of Phytopathology 40: 309–348
tion in the cytosol. The Plant Cell 15: 760–770 Whitham SA, Quan S, Chang HS, Cooper B, Estes B, Zhu T,
Tang X, Frederick RD, Zhou J, Halterman DA, Jia Y and Wang X and Hou YM (2003a) Diverse RNA viruses elicit
Martin GB (1996) Initiation of plant disease resistance by the expression of common sets of genes in susceptible
physical interaction of AvrPto and Pto kinase. Science 274: Arabidopsis thaliana plants. The Plant Journal 33: 271–283
2060–2063 Whitham TG, Martinsen GD, Floate KD, Dungey HS, Potts
Thomma BPHJ, Eggermont K, Penninckx IAWA, Mauch- BM and Keim P (1999) Plant hybrid zones affect biodiver-
Mani B, Vogelsang R, Cammue BPA and Broekaert WF sity: tools for a genetic-based understanding of community
(1998) Separate jasmonate-dependent and salicylate-depen- structure. Ecology 80: 416–428
dent defense-response pathways in Arabidopsis are essential Whitham TG, Young WP, Martinsen GD, Gehring CA,
for resistance to distinct microbial pathogens. Proceedings Schweitzer JA, Shuster SM, Wimp GM, Fischer DG,
of the National Academy of Sciences of the United States of Bailey JK, Lindroth RL, Woolbright S and Kuske CR
America 95: 15107–15111 (2003b) Community genetics: a consequence of the ex-
Tian D, Araki H, Stahl E, Bergelson J and Kreitman M (2002) tended phenotype. Ecology 84: 559–573
Signature of balancing selection in Arabidopsis. Proceedings With KA (2002) The landscape ecology of invasive spread.
of the National Academy of Sciences of the United States of Conservation Biology 16: 1192–1203
America 99: 11525–11530 Wolter M, Hollricher K, Salamini F and Schulze-Lefert P
VanSluys MA, Monteiro-Vitorello CB, Camargo LEA, Menck (1993) The mlo resistance alleles to powdery mildew
CFM, da Silva ACR, Ferro JA, Oliveira MC, Setubal JC, infection in barley trigger a developmentally controlled
 51

defense mimic phenotype. Molecular and General Genetics CC (2000) Genetic diversity and disease control in rice.
239: 122–128 Nature 406: 718–722
Zadoks JC and Vandenbosch F (1994) On the spread of plant Zwankhuizen MJ, Govers F and Zadoks JC (1998) Develop-
disease – a theory on foci. Annual Review of Phytopathol- ment of potato late blight epidemics: disease foci, dis-
ogy 32: 503–521 ease gradients, and infection sources. Phytopathology 88:
Zhu Y, Chen H, Fan J, Wang Y, Li Y, Chen J, Fan J, Yang S, 754–763
Hu L, Leaung H, Mew CW, Teng PS, Wang Z and Mundt
This page intentionally blank
European Journal of Plant Pathology (2006) 115:53–59
Springer 2006
DOI 10.1007/s10658-005-2025-y

The practical considerations of scale in plant pathology

W.W. Turechek
USDA-ARS Fruit Laboratory, Building 010A – BARC-W, Baltimore Avenue, 10300 Beltsville, MD 20705-
2350, USA (E-mail: [email protected])

Accepted 28 June 2005

Key words: quantitative epidemiology, strawberry

Abstract

The concept of scale has only recently gained recognition as a central theme in ecology. The rise in
significance of scale in ecology can be attributed to the increase in hypothesis-driven experimental ecology
over the last quarter century, and the realization that experimental results do not sufficiently explain past,
or predict future observations in nature. Plant pathologists, who rely heavily on hypothesis-driven research,
have confronted these same issues for nearly a century. In this paper, I will provide a concise presentation
and discussion of the important concepts of scale and how they apply to the discipline of plant pathology.

Introduction In what follows, I will provide a concise pre-

sentation and discussion of the important concepts
The concept of scale has only recently gained of scale. Most of what I will present is drawn from
recognition as a central or unifying theme in the ecological literature, as ecologists are generally
ecology. In an extensive review of the ecological at the forefront of advancing concepts and our
literature, Schneider (2001a) showed a dramatic knowledge of scale. Three books that have helped
increase in publications of scale-related research shape my understanding of scale are ‘Quantitative
during the 1990s. The rise in importance of scale Ecology: Spatial and Temporal Scaling’ by Schneider
can be attributed to the increase in hypothesis- (1994), ‘Ecological Scale: Theory and Applica-
driven, experimental ecology over the last quarter tions’ edited by Peterson and Parker (1998), and
century and the realization that experimental ‘Scaling Relations in Experimental Ecology’ edited
results do not sufficiently explain past or predict by Gardner et al. (2001). I recommend these books
subsequent observations in nature. to anyone wishing to gain an ecological perspec-
Plant pathologists have confronted these same tive on scale.
issues for nearly a century. With few exceptions,
this important component of systems, observation, Definitions
and analysis has been conveniently – unknowingly
is perhaps a better choice of words – disregarded, ‘‘Scale has a good start on contesting niche as one
sometimes to the detriment of the hypothesis. of the vaguest yet most often used words in ecol-
Unfortunately, the common leap from laboratory/ ogy’’ (Wiens, 2001). Scale can be defined correctly
glasshouse to the field without sufficient consider- in a number of ways. The definition that likely
ation of how outcomes may differ vastly when comes to mind when used in everyday conversa-
rescaled is a theme repeated to this day. Scaling is tion is that of cartographic scale. Cartographic
generally of little concern to the organismal biol- scale is the ratio of the distance on a map to the
ogist where processes under study are often clearly distance on the ground (Schneider, 2001a). An-
defined by the organism’s size. other common usage defines scale as the
54

‘‘. . .physical dimensions of observed entities and first category is the phenomenon (process or
phenomena’’ (O’Neill and King, 1998). Merriam structure) under study. For example, the spatial
and Webster offer this definition of scale ‘‘to pattern of plant disease and the processes that
arrange in a graduated series’’ (Merriam-Webster generate it. The second is the experimental or
online dictionary). In practice, most ecologists sampling units used to acquire information or data
would argue against this latter definition because, about the phenomenon under study. The third is
unlike the previous two, the notion of quantifiable the analyzes used to summarize the data to
measurement (i.e., distance, dimension) is not describe the phenomenon. From herein, I will refer
stated explicitly. A number of ecologists have ar- to these as the system scale, observational scale,
gued recently, and I agree, that any definition of and analysis scale, respectively.
scale must consider scale as a quantity and involve, In all cases, scale is bounded by grain at the
or at least imply, measurements or measurement lowest extreme and extent at the other. When
units (O’Neill and King, 1998). applied to the systems scale grain refers to the
Hierarchical scale equates the organizational le- smallest ecological unit to which a phenomenon
vel in a hierarchy to independent or individual affects a structure or process; extent is the total
scales. Hierarchical scale forms the essence of area affected by the phenomenon, process, or
hierarchy theory (Allen and Starr, 1982) where it is structure. Applied to the observational scale grain
argued that ecosystem processes operate in a way refers to the resolution of measurements (Dungan
such that upper level processes, structure, etc. reg- et al., 2002). The frequency of sampling in space
ulate and/or constrain processes at lower scales in a and/or time determines the resolution of the data.
quantifiable manner. Some ecologists have argued Extent refers to the range over which experimental
that hierarchical levels should not be thought of as measurements are taken (Wiens, 2001). For
scales because ‘level’ is often an arbitrary, ambig- example, consider measurements of precipitation.
uous or unquantifiable term (O’Neill and King, This is an important variable in many disease
1998; Wiens, 2001). Again, the notion of quantity is forecasters and one where estimates often need to
not defined explicitly in the definition, e.g., in what be sufficiently accurate. The recording frequency
units does one quantify the ‘leaf’ level? (Turechek of the rain gauge (i.e., hourly, daily, weekly, etc.)
and Madden, 2001). sets the temporal resolution or grain. The duration
An idealistic goal in experimental design is to of time over which precipitation is measured
conduct an experiment where measurements or determines the extent. Applied to the analysis
observations are taken at the organisms or phe- scale, calculating means and grouping sampling
nomenons ‘characteristic scale’. This is defined as units coarsens grain; subsampling reduces the
the system scale at which all relevant ecological extent. These are a few examples of how analytical
and biological processes of a population or com- tools alter scale (O’Neill and King, 1998).
munity occur. This ideal, however, is likely not The scope is defined at the observational scale
achievable as biological processes occur and for experimental design, surveys, model building,
interact over a range of scales. A better way to etc. as the ratio of extent to grain, and is an
approach this concept is to envisage ecological important and useful measurement for comparing
phenomena occurring within upper and lower scales across studies. ‘‘Scope can be thought of as
limits (Schneider, 1994). Even still, this may be the number of steps, once we know the step size’’
easily definable for certain phenomena, such as (Schneider, 1994). For example, the temporal
spore dispersal, but not others, e.g., patterns of scope of an experiment is the ratio of the time from
plant disease. the beginning to the end of the experiment, to the
time-step of a single measurement. The spatial
Scaling concepts scope of a survey is the ratio of the maximum
length or distance between measurements relative
Given an acceptable definition, a logical next to the minimum length or distance of a single
question is how to apply the term scale in practice. measurement.
In other words, what exactly needs to be scaled? Disease surveys are used commonly in plant
Dungan et al. (2002) distinguished three categories pathology to estimate some unknown value or
to which scale-related terms are applicable. The quantity of a population, typically disease severity
 55

or incidence (Turechek and Mahaffee, 2004). In assuming a random distribution of infected plants.
the terminology of survey sampling, the scope of a The magnification factor can be reduced by sam-
survey is the ratio of the sampling frame to the pling more plants, but at a cost. Cluster sampling
sampling unit, where the sampling unit is the allows the sampler to observe a greater number of
smallest item sampled and the sampling frame is plants with the same number of sampling units
the total number of possible sampling units in the (Hughes et al., 1996). This, in effect, will reduce
survey area. It is always the case that the precision the MF at the cost of decreasing resolution and,
of the estimate increases, and the uncertainly consequently, will give a reduced scope. The ben-
decreases, as the number of samples or observa- efit of the trade-off needs to be determined for each
tions is increased. Thus, the challenge is to collect study and can be represented in a scope diagram.
an ample number of samples to sufficiently mini- A scope diagram is one way of displaying or
mize uncertainty under the logistical and natural comparing the scope among different phenomena,
constraints that often limit the scope of most de- events, or studies. Schneider (1994) demonstrates
signed surveys. how a simple line diagram can be used to represent
For example, a typical raised-bed field of the scope of a survey. The line can be partitioned
strawberry will contain about 43,750 plants ha)1. into two components, one representing the scope
An individual strawberry plant could be consid- of the sample and the remainder the inference
ered the sampling unit in a sampling frame of component of the sample represented by MF. This
43,750 plants. In a survey for Colletotrichum simple diagram can be used to compare survey
crown rot (caused by C. acutatum) a sample of 500 strategies and help decide the best approach.
plants gives a sampling fraction (i.e., the ratio of Continuing with the strawberry example, assume
the number of samples taken to the sampling that in a hectare of strawberries a single plant
frame) of 500/43750 or 1.14%. The inverse of the occupies a space of 0.1 m2. Figure 1 depicts the
sampling fraction represents the magnification scope diagrams for three 100 unit samples where:
factor (MF). The magnification factor magnifies (1) single plants are observed or collected in a
the result of the sample into an estimate for the random sample, (2) groups of 10 plants are col-
entire population. For example, say 20% of the lected in a cluster sample, and (3) 10 m2 grids are
sample (100 plants) was diagnosed with crown rot. observed via aerial sampling. In this example, the
Multiplying the number of diseased plants by the number of samples collected or observed remains
MF (i.e., 87.7) informs us that we should expect to the same, but the size of the sampling unit changes
find 8770 plants with crown rot in the population, resulting in a smaller magnification factor and a

Figure 1. Scope diagrams for three 100 unit samples where: (1) single plants are observed or collected in a random sample, (2)
groups of 10 plants are collected in a cluster sample, and (3) 10 m2 grids are observed via aerial sampling. The starting point of
each diagram represents the grain or resolution of the data at the observational scale, and the ending point represents the extent of
the study (the sampling frame). The sample size is denoted by N and the magnification factor by MF.
56

reduced scope. Scope diagrams can be much more define and are not necessarily constrained by
complex. Seem (2004) gives a good example of quantifiable units (Naeem, 2001). For example, the
scope diagram and its applicability in disease taxonomic groupings of race, subspecies, species,
forecasting. genus, and sometimes higher groupings can often
be ambiguous. At the population level, the
Scaling in practice boundaries between patches, communities, meta-
populations, etc. is often blurred due to fluctua-
The practice of scaling involves relating measure- tions in growth, decline, and interactions among
ments made at one scale to measurements or pre- populations.
dictions made at another. A scale-dependent
process is one in which the process (e.g., rate) or Considerations in experimental design
property (e.g., density) changes with a change of
grain or extent. An example of scale dependency ‘‘The near absence of prescriptions for incorpo-
encountered often in plant pathology is scale- rating scale in experimental design may partially
dependent patterns of plant disease, i.e., patterns explain why explicit consideration of scale is not
differing with sampling unit size and/or extent of more prevalent in the design of terrestrial field
the survey. Although these scaling dependencies experiments’’ (King et al., 2001). Readers inter-
can be represented through various modelling ested in a statistical treatment of this topic are re-
procedures (Allen, 2001), episodic dynamics, non- ferred to two chapters by Dutilleul (1998a, b). I will
Euclidean structures, and/or biotic discontinuities touch on a broader aspect of design: the desire to
that are typically associated with ecological pro- have experimental results relate directly to natural
cesses or phenomena make modelling (scaling) a observation. Every experiment is subject to a
challenge. compromise between what Manly (1992) defines as
Episodic dynamics generally make it difficult to internal vs. external validity (Naeem, 2001).
scale temporal events. For example, 30 mm of ‘‘Internal validity concerns whether the apparent
rainfall over the course of an hour implies a rain- effects or lack of effects shown by the experimental
fall rate of 0.5 mm min)1; an increase in disease results are due to the factor being studied, rather
severity from 3 lesions/leaf to 52 lesions/leaf over than some alternative factor. External validity
the period of 7 days implies a rate of disease concerns the extent to which the results of an
progress of 7 lesions day)1. The process of calcu- experiment can be generalized to some wider
lating rates masks the fact that precipitation and population of interest.’’ Naeem (2001) groups
lesion development occur as concentrated events experiments into three general classes: (1) field, (2)
(episodes) over some defined time period. Non- model-ecosystem (micro-, meso-, and macrocosm),
Euclidean structures (such as landscape surfaces) and (3) simulation. Field experiments have the
and patchy or heterogeneous environments gen- highest level of external validity and, consequently,
erally make it difficult to scale spatial properties the lowest level of internal validity. Conversely,
because they lead to fractal dimensions (i.e., simulation experiments have the highest level
greater than a plane, but less than a volume). For of internal validity and lowest level of external
example, wind patterns contributing to larger scale validity.
patterns of spore dispersal and microclimatic var- More often than not, the type of experiment and
iability are directly affected by variations in the choice of scale (determined by plot size,
topography in a largely unpredictable manner, duration, sampling extent, etc.) are a function of
much more so than if these events were to occur pragmatism. Available funding, personnel, exper-
over a strictly two-dimensional surface. imental costs, measurement technology, etc. play a
Biotic boundaries or discontinuities further more central role in the design of experiments than
contribute to modelling headaches. For example, a does the theoretical consideration of scale and
change of 3 C from 27 to 30 C straddles the validity. In any event, it is likely (and is often the
critical temperature for conidial germination of case) that to gain a full understanding of a process,
Podosphaera macularis, the causal agent of hop sets of experiments that span the range of what
powdery mildew (Mahaffee et al., 2003). Other Naeem (2001) refers to as the ‘‘scale-validity ma-
biotic boundaries are often much more difficult to trix’’ must be conducted. That is, a reasonable set
 57

of field, microcosm, or simulation experiments lated from grouped or lagged measurements or

conducted at different spatial, temporal, and pos- observations. (The term lag refers to the interval or
sibly biotic scales are often necessary to fully spacing between neighbouring units.) Imagining a
interpret a process or to explain what was grid; grouping occurs when contiguous squares of
observed naturally. the grid (sampling units) are combined to form
larger units and the quantities are combined via
Multiscale analysis juxtaposition (added); under these conditions the
resolution of the data changes. Variances are
In the (failed) quest for a ‘characteristic scale’, obtained by re-calculating the variance of the
multiscale analysis has evolved to play a central combined quantities, and comparing it to the ori-
role in scaling. Multiscale analysis is defined as an ginal or ungrouped variance and to variances
analysis with respect to multiples of a unit of calculated from successively larger groupings. The
measurement (Schneider, 1994). In general, this blocked quadrat-variance methods, such as the
type of analysis is done by first defining subsystems two-term local quadrat variance (TTLQV) method
within a system. In a survey, this might be and the paired-quadrat variance (PQV) method,
accomplished by superimposing a grid over the are examples of analyses that use grouped vari-
region of study or spatially referencing sampling ances (Ludwig and Reynolds, 1988). Lagging, on
units in the survey area. As one example, summing the other hand, results from calculating deviances
the components of the subsystem, grid, or sam- between grid components (sampling units) at
pling units (with correction factors introduced as increasingly greater separations (lags) across the
needed) can be used to scale to larger areas grid. Again, these variances are compared to the
(Schneider, 2001b). Quantities can be summed by original sample variance as well as to variances
either juxtaposing or superposing values. Sum- calculated at different lag distances. Autocorrela-
mation by juxtaposing values extends the range of tion and semivariogram analyses are examples of
scale; summation by superposing leaves the scale analyses that use lagged variances (Cressie, 1991).
unchanged. Summing the number of diseased Variances calculated according to the lag
plants in a series of contiguous plots is an example manoeuvre can be used to calculate variances that
of juxtaposing; summing the number of newly would be obtained via grouping using a Fourier
diseased plants to the number of previously dis- transformation (Schneider, 1994).
eased plants in a single plot is an example of Scaled quantities cannot be treated as unitless
superposing. numbers. The process of summing, multiplying,
The variance plays a central role in multiscale and taking derivatives of scaled quantities should
analysis. ‘‘One of the major research challenges in not be done independently of the unit. For
ecology is understanding the creation and erosion example, the sum of 52 lesions/leaf and 6 diseased
of spatial variability as a function of spatial scale. trees/orchard is non-sensical. Biologically inter-
Included in this challenge is the question of the pretable sums of scaled quantities are referred to
degree to which variance generated at one scale is as ensemble quantities (Schneider, 1994). Spatial
transformed into variance at another scale’’ and temporal averages, variances, and covariances
(Schneider, 1994). Across many disciplines, are typical ensemble quantities. This definition
including plant pathology, methods have been differs from the traditional in which an ensemble
developed for relating and/or predicting variance refers to a collection of ‘realizations’ of an event or
across scales (Hughes et al., 1997; Turechek and process; the ensemble average, for example, is the
Madden, 2003). However, the mere ability to mean of the realizations. Although this concept is
model these relations should not be mistaken as an evident in plant pathology research, the terminol-
understanding of how these relations came to be. ogy is infrequently used (one exception is
For the most part, the mechanisms or biological Ferrandino, 2004).
processes generating these differences are only
partially understood. Statistical tools
The sample variance is only one measure of
spatial variability and has limited interpretation in Over the past 20 years, the variety of statistical
multiscale analysis. Variances can also be calcu- tools available for multiscale analyses has
58

increased tremendously; many have been applied Willocquet and Savary (2004) designed a simula-
to characterize spatial patterns of plant disease. tion model to determine how varying auto-, allo-
The tests, however, can be categorized based on leaf, and allo-plant-deposition rates of infective
the general type of analysis, point-pattern vs. propagules affected epidemic development
correlation (Upton and Fingleton, 1985), or on observed at the plant and leaf level (in both
whether the data consist of mapped or unmapped examples, measurement units are implied). As
observations (Diggle, 1983). The point-pattern discussed above, simulation studies have a high
approach is based on the location of individuals degree of internal validity and allow researchers to
over an area of interest and analyses are conducted explore a range of conditions that may otherwise
either using the distances between individuals take years to observe.
(Perry, 1995; Ferrandino, 1998), or using the
counts of individuals within sampling units such as Conclusion
quadrats (Pielou, 1977; Madden and Hughes,
1995). The latter methods include the distribu- The intent of this paper was to provide an over-
tional approach that involves fitting observed view of scale-related concepts and how they might
frequencies of counts per sampling unit to well- apply to plant pathology. Although I did not
defined statistical distributions (e.g., Poisson, provide prescriptive advice on how to include scale
negative binomial, binomial, and beta-binomial). in designed studies, I hope I made it obvious where
The methods based on counts per sampling unit scaling is naturally applied in our discipline. I also
provide direct measurements of heterogeneity of hope that I impressed the importance of being
the data at the scale of the sampling units and vigilant in reporting the scale (grain, extent, and
below, but they do not explicitly define the spatial scope) of experiments and surveys to allow for
arrangement of the sampling-unit counts unless drawing valid comparisons across studies.
several sampling units are grouped in a series of To summarize briefly, before designing any
steps (Ludwig and Reynolds, 1988). experiment or survey, it would be prudent to
Spatial autocorrelation and semivariograms acquire preliminary data on the structure of the
(Cressie, 1991) use lagged variances to produce population under study so that sampling units (i.e.,
explicit information on the degree of association of grain) and the extent of the study can be appropri-
disease intensity among sampling units. Unlike the ately chosen (Legendre and Legendre, 1998).
distributional methods, these methods yield dif- Realize that this information may indicate that a
ferent results for different arrangements of counts single study may not be sufficient to gain a full
within a field, although they are not specifically understanding of the process or characteristic under
developed for counts within sampling units. Spa- study. It is also important to consider the units of
tial Analysis by Distance IndicEs (SADIE) is a measurement and how easily information can be
class of tests developed recently to quantify spatial rescaled. Pathologists should avoid describing the
pattern in the spirit of spatial autocorrelation, but scale of study as, for example, the ‘leaf scale’ unless
uses data in the form of counts (Perry, 1995; Xu a unit of measurement is clearly implied. In many
and Madden, 2004). The correlation-based meth- cases, it is likely that the scale can be defined within
ods characterize pattern at the scale of the some narrow range of values, and these should be
sampling unit and above. The results from used to identify the scale (grain) of study. Lastly, be
point-pattern and correlation-type analyses can aware that many field studies or surveys of disease
jointly be used to better interpret patterns and are the result of the interaction between two popu-
possibly describe the biological phenomenon lations: the population of the host and the popula-
responsible for generating the observed pattern tion of the pathogen. The scale at which the host
(Turechek and Madden, 1999). population exists should be an important consid-
Simulation and randomizations have also been eration of the pathogen population, because the
used to study scale-related processes. For example, host represents the possible extent of the pathogen.
Turechek and Madden (2001) used Monte–Carlo For example, knowing the scale of spore dispersal
methods and randomizations to determine how distance is not very informative unless the distance
the variability of strawberry leaf blight at a lower significantly overlaps the range of the host. Con-
scale impacted the variability at higher scales. sidering these aspects of scale in the design of an
 59

experiment will help to minimize the possibly large Mahaffee WF, Turechek WW and Ocamb CM (2003) Effect of
discrepancies in scale between what is observed and variable temperature on infection severity of Podosphaera
macularis on hops. Phytopathology 93: 1587–1592
what is being tested. Manly BFJ (1992) The Design and Analysis of Research
Studies, Cambridge University Press, Cambridge
Naeem S (2001) Experimental validity and ecological scale as
References criteria for evaluating research programs. In: Gardner RH,
Kemp WM, Kennedy VS and Peterson JE (eds) Scaling
Allen TFH (2001) The nature of the scaling issue in experi- Relations in Experimental Ecology. Columbia University
mentation. In: Gardner RH, Kemp WM, Kennedy VS and Press, New York
Peterson JE (eds) Scaling Relations in Experimental O’Neill RV and King AW (1998) Homage to St. Michael; or,
Ecology. Columbia University Press, New York why are there so many books on scale? . In: Peterson DL
Allen TFH and Starr TB (1982) Hierarchy, University of and Parker VT (eds) Ecological Scale: Theory and
Chicago Press, Chicago Applications. Columbia University Press, New York
Cressie N (1991) Statistics for Spatial Data, Wiley, New York Peterson DL and Parker VT (eds) (1998) Ecological Scale:
Diggle PJ (1983) Statistical Analysis of Spatial Point Patterns, Theory and Applications. Columbia University Press, New
Academic Press, London York
Dungan JI, Perry JN, Dale MRT, Legendre P, Citron-Pousty S, Perry JN (1995) Spatial analysis by distance indices. Journal of
Fortin M-J, Jakomulska A, Miriti M and Rosenberg MS Animal Ecology 64: 303–314
(2002) A balanced view of scale in spatial statistical Pielou EC (1977) Mathematical Ecology, Wiley, New York
analysis. Ecography 25: 626–640 Schneider DC (1994) Quantitative Ecology: Spatial and Tem-
Dutilleul P (1998a) Incorporating scale in ecological experi- poral Scaling, Academic Press, San Diego
ments: Study design. In: Peterson DL and Parker VT (eds) Schneider DC (2001a) The rise of the concept of scale in
Ecological Scale: Theory and Applications. Columbia ecology. Bioscience 7: 545–553
University Press, New York Schneider DC (2001b) Spatial allometry: Theory and applica-
Dutilleul P (1998b) Incorporating scale in ecological experi- tion to experimental and natural aquatic ecosystems. In:
ments: Data analysis. In: Peterson DL and Parker VT (eds) Gardner RH, Kemp WM, Kennedy VS and Peterson JE
Ecological Scale: Theory and Applications. Columbia (eds) Scaling Relations in Experimental Ecology. Columbia
University Press, New York University Press, New York
Ferrandino FJ (1998) Past non-randomness and aggregation to Seem RC (2004) Forecasting plant disease in a changing
spatial correlation: 2DCORR, a new approach for discrete climate: A question of scale. Canadian Journal of Plant
data. Phytopathology 88: 84–91 Pathology 26: 274–283
Ferrandino FJ (2004) Measuring spatial aggregation in binary Turechek WW and Madden LV (1999) Spatial pattern analysis
epidemics: Correlative analysis and the advantage of of strawberry leaf blight in perennial production systems.
fractal-based sampling. Phytopathology 94: 1215–1227 Phytopathology 89: 421–433
Gardner RH, Kemp WM, Kennedy VS and Peterson JE (eds) Turechek WW and Madden LV (2001) Effect of scale on plant
(2001) Scaling Relations in Experimental Ecology. Columbia disease incidence and heterogeneity in a spatial hierarchy.
University Press, New York Ecological Modeling 144: 77–95
Hughes G, Madden LV and Munkvold GP (1996) Cluster Turechek WW and Madden LV (2003) A generalized linear
sampling for disease incidence data. Phytopathology 86: modeling approach for characterizing disease incidence in a
132–137 spatial hierarchy. Phytopathology 93: 458–466
Hughes G, McRoberts N, Madden LV and Gottwald TR Turechek WW and Mahaffee WF (2004) Spatial pattern
(1997) Relationship between disease incidence at two levels analysis of hop powdery mildew in the Pacific Northwest:
in a spatial hierarchy. Phytopathology 87: 542–550 Implications for sampling. Phytopathology 94: 1116–1128
King AW, Gardner RH, Hatfield CA, Naeem S, Peterson JE Upton G and Fingleton B (1985) Spatial Data Analysis by
and Wiens JA (2001) Terrestrial perspectives on issues of Example, Vol 1. Point Pattern and Quantitative Data, John
scale in experimental ecology. In: Gardner RH, Kemp Wiley & Sons, West Sussex
WM, Kennedy VS and Peterson JE (eds) Scaling Relations Wiens JA (2001) Understanding the problem of scale in
in Experimental Ecology. Columbia University Press, New experimental ecology. In: Gardner RH, Kemp WM,
York Kennedy VS and Peterson JE (eds) Scaling Relations in
Legendre P and Legendre L (1998) Numerical Ecology, Elsevier Experimental Ecology. Columbia University Press, New
Science, Amesterdam York
Ludwig JA and Reynolds JF (1988) Statistical Ecology: A Willocquet L and Savary S (2004) An epidemiological simula-
Primer on Methods and Momputing, Wiley-Interscience, tion model with three scales of a spatial hierarchy.
New York Phytopathology 94: 883–891
Madden LV and Hughes G (1995) Plant disease incidence: Xu XM and Madden LV (2004) Use of SADIE statistics to
Distributions, heterogeneity, and temporal analysis. Annual study spatial dynamics of plant disease epidemics. Plant
Review of Phytopathology 33: 529–564 Pathology 53: 38–49
This page intentionally blank
European Journal of Plant Pathology (2006) 115:61–73
Springer 2006
DOI 10.1007/s10658-005-3682-6

Trends in theoretical plant epidemiology

H. Scherm*, H.K. Ngugi and P.S. Ojiambo

Department of Plant Pathology, University of Georgia, Athens, GA 30602, U.S.A
(Fax: +1-706-542-1262; E-mail: [email protected])

Accepted 16 September 2005

Key words: decision analysis, mathematical model, population genetics, spatial structure, statistical epi-
demiology, stochasticity

Abstract

We review trends and advances in three specific areas of theoretical plant epidemiology: models of temporal
and spatial dynamics of disease, the synergism of epidemiology and population genetics, and progress in
statistical epidemiology. Recent analytical modelling of disease dynamics has focused on SIR (susceptible–
infected–removed) models modified to include spatial structure, stochasticity, and multiple management-
related parameters. Such models are now applied routinely to derive threshold criteria for pathogen
invasion or persistence based on pathogen demographics (e.g., Allee effect or fitness of fungicide-resistant
strains) and/or host spatial structure (e.g., host density or patch size and arrangement). Traditionally
focused on the field level, the scale of analytical models has broadened to range from individual plants to
landscapes and continents; however, epidemiological models for interactions at the cellular level, e.g.,
during the process of virus infection, are still rare. There is considerable interest in the concept of scaling,
i.e., to what degree and how data and models from one scale can be transferred to another (smaller or
larger) scale. Despite assertions to the contrary, the linkages between epidemiology and population genetics
are alive and well as exemplified by recent efforts to integrate epidemiological parameters into population
genetics models (and vice versa) and by numerous integrated studies with an applied focus (e.g., to quantify
sources and types of primary and secondary inoculum). Statistical plant epidemiology continues to rely
heavily on the medical and ecological fields for inspiration and conceptual advances, as illustrated by the
recent surge in papers utilizing ROC (receiver operating characteristic), Bayesian, or survival analysis.
Among these, Bayesian analysis should prove especially fruitful given the reliance on uncertain and sub-
jective information for practical disease management. However, apart from merely adopting statistical
tools from other disciplines, plant epidemiologists should be more proactive in exploring potential appli-
cations of their concepts and procedures in rapidly expanding disciplines such as statistical genetics or
bioinformatics. Although providing the scientific basis for disease management will always be the raison
d’eˆtre for plant epidemiology, a broader perspective will help the discipline to remain relevant as more
resources are being devoted to genomic and ecosystem-level science.

Introduction observers allows us to take a bird’s-eye view of

recent developments in the area and assess their
It is perhaps somewhat atypical that this com- impact on the science and practice of plant
mentary on theoretical plant epidemiology is pathology without being influenced by predeter-
authored by a group of investigators who consider mined notions.
themselves experimentalists rather than theoreti- The Encyclopaedia Britannica defines a scien-
cians. However, our role as dispassionate tific theory as a ‘systematic ideational structure of
62

broad scope, conceived by the human imagination, examples reflect our personal views of what is
that encompasses a family of empirical (experien- interesting and important in theoretical epidemi-
tial) laws regarding regularities existing in objects ology. We limit our discussion largely to work
and events, both observed and posited’ (Anon., published since the last International Workshop
2005). In practice, this ‘systematic ideational on Plant Disease Epidemiology in Ouro Preto,
structure’ is usually formalized as a model, either Brazil, in 2001. Selected aspects of theoretical
conceptual or mathematical. Both forms of models work carried out during the 1990s have been syn-
have been influential in plant epidemiology (Jeger, thesized recently (Jeger, 2000; Gilligan, 2002).
2000; Zadoks, 2001). Although theory can be
developed without mathematical models, the two Models of temporal and spatial disease dynamics
concepts are often used synonymously in the epi-
demiological literature. The development of mathematical models to
One of the most commonly voiced criticisms describe disease dynamics has been and continues
surrounding the use of theory and mathematical to be the mainstay of theoretical epidemiology.
models in the broader ecological literature has Recent research in the area has focused on incor-
been the lack of interaction between modellers and porating spatial structure, elucidating the conse-
experimentalists during model development, test- quences of stochasticity and spatial scale,
ing, and validation (Caswell, 1988; Hall, 1988b). identifying threshold criteria for pathogen or
Theoretical epidemiology has largely escaped this strain establishment, and predicting the effects of
controversy, presumably because some of the most selected management strategies on disease
influential epidemiological modellers in plant dynamics. A detailed account of the use of ana-
pathology (or at least some members of their lytical models to address these objectives has been
laboratories) are superb experimentalists in their given by Gilligan (2002). Based on his review and
own right. The resulting synergism between mod- the subsequently published literature, a number of
els and experimental data in advancing theory is trends may be inferred.
typified in the work of JM Jeger and CA Gilligan,
among others (e.g., Jeger, 2000; Gilligan, 2002). It SIR models have entered the mainstream and be-
is important to note that testing and validation of come more versatile
models need not occur at the same time for a In its most basic form, an SIR model consists of a
model to be useful. For instance, the theory of set of linked differential equations describing the
dispersive epidemic waves (focal epidemics that dynamics of susceptible (healthy), infected, and
spread with increasing frontal velocity) was for- removed (post-infectious) host tissue; commonly,
malized by Ferrandino (1993) based on physical the infected tissue is divided into exposed (latently
principles of spore transport, with limited empiri- infected) and infectious compartments, leading to
cal support. Although additional observational an SEIR model (Madden, 2005). This type of
(Scherm, 1996) and experimental (Frantzen and analytical model, first formalized by Kermack and
van den Bosch, 2000) backing for this theory was McKendrick (1927) for human diseases, was
presented in the interim, it took more than a dec- popularized by Jeger (1982) for use in plant epi-
ade after publication of Ferrandino’s paper until demiology. Almost 20 years later, Segarra et al.
large-scale disease gradient experiments by Cow- (2001) formally derived the SEIR model for plant
ger et al. (2005) demonstrated convincingly that epidemics from the more general Kermack–
epidemics of wheat stripe rust spread consistently McKendrick model based on first principles. In
with increasing frontal velocity. addition, Segarra et al. (2001) provided a detailed
In what follows we consider current trends in comparison of the behaviour of the latter two
three specific areas of theoretical plant epidemiol- models with that of Van der Plank’s widely used
ogy: models of temporal and spatial dynamics of differential-delay equation (Van der Plank, 1963).
disease, the synergism of epidemiology and popu- Recent work has added considerable complexity
lation genetics, and advances in statistical epide- to SIR-type models (Gilligan, 2002), including
miology. The purpose here is not to provide a demographic and environmental stochasticity
comprehensive review, but rather to give selected (Park et al., 2003; Gibson et al., 2004; Otten et al.,
examples illustrating these trends. Inevitably, these 2004a), seasonal disturbance and multi-year
 63

disease dynamics (Madden and van den Bosch, 1988a). Perhaps we need to remind ourselves
2002), dynamics of host growth and susceptibility occasionally that the purpose of theory is to
(Gibson et al., 2004), virus vectoring mode and explain rather than to predict, and that theoretical
vector performance (Madden et al., 2000; Holt problems without practical applications are just as
and Colvin, 2001), interactions with biocontrol legitimate as empirical studies that do not con-
agents (Gibson et al., 2004), spatial structure and tribute to the development of new theories.
metapopulation dynamics (Park et al., 2001,
2003), and the presence of pesticide-resistant sub- Broadened scale of investigation
populations (Hall et al., 2004), among others. The With few exceptions, epidemiological models have
inclusion of stochasticity and spatial structure is traditionally focused on the field scale, a logical
especially significant as models featuring these choice considering the importance of individual
attributes can produce qualitatively very different fields as the spatial unit for tactical disease man-
predictions regarding pathogen establishment and agement by growers. In recent years, however, the
persistence than their deterministic mean-field scale of analysis has broadened to include both
counterparts. Most importantly, invasion thresh- finer and larger scales. At one end of the spectrum
olds in stochastic models are higher and the is the individual plant scale, where theoretical
pathogen or strain may be unable to persist fol- models have been developed, for instance, to
lowing successful invasion due to chance events, describe transmission of Rhizoctonia solani from
especially at low population densities (Gilligan, an infected to a healthy plant based on models of
2002). hyphal and colony growth of the fungus through
The increased complexity of contemporary SIR soil (Stacey et al., 2001; Otten et al., 2004b). At the
models adds realism and allows their application cellular level, effects of phenomena such as viral
to a wider range of problems. Indeed, models are cross-protection (Zhang and Holt, 2001) and syn-
now routinely formulated to accommodate ergism among different viruses (Zhang et al., 2000;
parameters useful in exploring specific manage- Naylor et al., 2003) have been modelled with
ment strategies (Jeger, 2000; Gilligan, 2002; Stacey respect to their effects on field-level disease
et al., 2004). For instance, a linked African cassava dynamics. However, epidemiological models that
mosaic virus–whitefly vector model (Jeger et al., explicitly describe molecular processes and inter-
2004) includes four management-related parame- actions within individual plant cells, e.g., during
ters, namely the roguing rate of infected host virus replication or virus- or transgene-induced
plants, the insecticide-induced death rate of the gene silencing, are still lacking in plant pathology,
vector, and the virus acquisition and transmission even though they are common in medical epide-
rates of the vector, both of which are determined miology (e.g., Phillips et al., 2001).
by the level of host resistance. Analysis of this At the other end of the spectrum are models for
model indicated that roguing applied once per disease development at landscape (Park
month in combination with a modest level of host et al., 2001, 2003; Otten et al., 2004a; Stacey et al.,
resistance (specifically one that reduces the prod- 2004) and continental (van den Bosch et al., 1999)
uct of acquisition rate and transmission rate below scales. With the rising interest in area-wide pest
80% of the value of the susceptible host) is suffi- management and the increasing exotic species
cient to eradicate the disease, while a combination problem (Scherm and Coakley, 2003), this scale of
of roguing and insecticide application is less investigation will become more important in the
effective. This example illustrates that analytical future. In landscape models, spread among fields
models have come a long way in their capacity to has been implemented via percolation theory (Ot-
provide specific management recommendations ten et al., 2004a), cellular automata (Gilligan,
that have traditionally been considered in the 2002), or in a metapopulation framework (Park
realm of more complex simulation models. et al., 2001, 2003) in which habitable patches are
Nonetheless, a few words of caution are made up of aggregates of susceptible fields. The
appropriate as there are some well publicized models allow for the analysis of disease spread in
examples from the broader ecological literature relation to within-patch pathogen dynamics, the
where the extension of theoretical models to strength of coupling among patches, and patch
management has met with disastrous results (Hall, size, density, and arrangement.
64

Closely related to the issue of scale is the con- suggested by a straight line in a log–log plot of
cept of scaling, i.e., to what degree and how data the measure of interest against the scale of
and models from one scale can be transferred to observation. The slope of the line is interpreted as
another (smaller or larger) scale. This has been an the fractal dimension, which summarizes the
active area of research in both theoretical and properties of the pattern across scales. In general,
statistical epidemiology (Turechek, 2005). For scale-invariance might be expected for organisms
instance, statistical procedures to extrapolate dis- occurring at a population density near their lower
ease incidence data from a lower hierarchical level critical threshold, e.g., due to human intervention
(e.g., leaves) to a higher level (e.g., shoots) and vice (Cousens et al., 2004). In a recent pest manage-
versa have been developed (Hughes et al., 1997; ment-related example, Cousens et al. (2004)
Hughes and Gottwald, 1998; Madden and counted numbers of five agricultural weeds in up
Hughes, 1999; McRoberts et al., 2003; Turechek to 202,500 contiguous 0.20.2-m quadrats in a
and Madden, 2003) and are now increasingly being single arable field. Counts from adjacent quadrats
applied to develop more efficient sampling and were pooled into progressively larger quadrats
disease assessment protocols through approaches with up to 90 m-long sides. This allowed for the
such as cluster sampling and group testing. In a calculation of incidence values for different quad-
recent example, Xu et al. (2004) used a distribu- rat sizes and an understanding of how these inci-
tion-based approach to derive relationships be- dence values vary with scale. Calculation of the
tween the incidence of spikelet infection and the fractal dimension showed that spatial patterns of
more easily determined incidence of ear infection those weed species that were most aggregated and/
for the Fusarium head blight pathosystem on or occurred at the lowest densities were scale-
wheat. These particular relationships may be use- invariant, indicating that patterns observed at
ful for making decisions in cases where manage- small scales repeated themselves at progressively
ment thresholds are based on the incidence of larger scales. Although there are theoretical rea-
infected spikelets. sons why such scale-invariance would be unlikely
On a more process-based level, Stacey et al. for plant pathogens (e.g., different mechanisms for
(2001) developed a mathematical model to scale up long- vs. short-distance dispersal along with
from the behaviour of individual hyphae (of changes in the physical environment at different
R. solani in this example) to fungal colony growth spatial scales), it would be interesting to test the
through soil and to infection of individual plants. null hypothesis of scale-variance for different types
The approach was based on a spatially explicit of pathogens, e.g., those causing aerial vs. soil-
model of hyphal expansion incorporating the borne or monocyclic vs. polycyclic diseases. Scale-
relationships between hyphal growth and fungal invariance, if it occurs in plant pathogens, would
biomass as well as between fungal biomass, prox- allow for extrapolation and prediction over a wide
imity of the mycelium to a susceptible root, and range of spatial scales with potentially useful
the probability of disease transmission. A sto- applications in areas such as precision agriculture.
chastic, cellular automaton-based model for scal- Ferrandino (2004) recently proposed a sampling
ing up from individual plants to plant populations approach for disease incidence based on a nested
infected with R. solani had been developed previ- fractal design, i.e., one in which sampling points at
ously (Kleczkowski et al., 1997), and it may be distances of, say, 1, 2, 4, 8, and 16 m are repre-
possible to combine this probabilistic model with sented equally. Using simulated spatial epidemics,
the more detailed fungal growth-based model of he showed that this design was more efficient in
Stacey et al. (2001) to arrive at estimates of both detecting aggregation than either regular, random,
the mean and variance of the spatio-temporal or spatially clustered sampling designs, in addition
dynamics of R. solani. to providing spatial information over a wider
In the broader ecological literature, fractal range of scales.
geometry has been applied for scaling among dif-
ferent spatial or temporal hierarchies if the pattern Fascination with thresholds
or process of interest is scale-invariant, i.e., repeats Van der Plank (1963) expressed his threshold
itself at progressively larger scales (Brown et al., theorem as iRc>1, which states that an epidemic
2002; Li, 2000). In practice, scale-invariance is will not occur unless the product of infectious
 65

period i and basic corrected infection rate Rc

exceeds 1. For consistency with the medical and
ecological literature, the theorem has been rewrit-

Population growth rate

ten as R0>1, where R0 is the basic reproductive
number, i.e., the number of new infected individ-
uals resulting from one introduced infected indi-
vidual (Madden, 2005). Although the interest in
thresholds for plant epidemics has been long-
0
standing (e.g., Jeger and van den Bosch, 1994a, b),
we note a recent surge of activity in this area,
mostly derived from analyses with SIR-type
models. This has included derivation of threshold
0
criteria based not only on pathogen demographics Population size
(e.g., fitness of fungicide-resistant strains; Hall
Figure 1. Graphical representation of the Allee effect showing
et al., 2004), but also on host spatial structure
an intermediate optimum in the relationship between popula-
(e.g., host density, patch size, and coupling among tion growth rate and population density, e.g., due to the diffi-
patches; Bailey et al., 2000; Gubbins et al., 2000; culty to find a compatible mating partner at very low
Park et al., 2001; Otten et al., 2004a). One of the population densities for species with an obligate sexual cycle.
reasons for the current preoccupation with
thresholds in plant epidemiology is the wider
availability of stochastic models, which allows for but also concepts and tools from population
the calculation of not only the risk of pathogen genetics. The main focus of population genetics is
invasion, but also the probability of subsequent to understand the evolutionary processes driving
persistence in the face of chance events that can and maintaining genetic variation within and
lead to extinction at low population densities. among populations (McDonald, 2004). Because
Apart from stochastic forces, the establishment host and pathogen populations consist of distinct
of a pathogen following its successful introduction genetic entities, the fundamental theory of their
may be limited by certain demographic features, population dynamics in space and time must
such as the difficulty to find a compatible mating coincide with that of their genetic composition.
partner at very low population densities for species Conceptually, there is thus a considerable overlap
with an obligate sexual cycle (Taylor and Hastings, between epidemiology and population genetics.
2005). This feature leads to an intermediate opti- Here, we take an heuristic look at the interplay
mum in the relationship between population between the two disciplines in the development
growth rate and population density (Allee effect; and application of epidemiological theory and
Figure 1). In a deterministic population model of highlight areas that may best be served by an
the heterothallic Karnal bunt fungus Tilletia interdisciplinary approach. An in-depth review of
indica, inclusion of an Allee effect resulted in a the nature of the synergy between epidemiology
teliospore threshold for establishment about two and population genetics is outside the scope of this
orders of magnitude higher than in a version of the paper but is available elsewhere (Milgroom, 2001;
model without this constraint (Garrett and Bow- Milgroom and Peever, 2003).
den, 2002). This finding has potentially important It has been suggested that, over the past
implications for risk assessments of T. indica and 20 years, a schism appears to have developed
other quarantine pathogens with an obligate between epidemiology and population genetics
sexual cycle. (Milgroom, 2001; Milgroom and Peever, 2003).
Indeed, it is tempting to conclude that such a split
Pathogen population biology was a consequence of both disciplines becoming
more specialized as they responded to new tech-
Epidemiology is a holistic discipline (Zadoks, nologies; epidemiology to the availability of
1990), and the development of epidemiological advanced modelling techniques and increased
theory thus requires an interdisciplinary approach. computing power, population genetics to advances
This includes not only mathematics and statistics, in molecular biology. In practice, however, the two
66

disciplines often have been utilized jointly to However, ascospores contributed appreciably
address applied epidemiological questions such as to the genetic composition of the pathogen
source and type of primary inoculum (Gobbin population (as indicated by the proportion of
et al., 2003; Peever et al., 2004), dispersal of sec- sexual descendants among lesions at the end of the
ondary inoculum (Cortesi et al., 2000; Cortesi and season), especially in dry conditions unfavourable
Milgroom, 2001; Loskill et al., 2004), or host for the dispersal of pycnidiospores.
specificity (Peever et al., 2000; Akimitsu et al., With regard to analytical modelling approaches,
2003; Flier et al., 2003). A notable example in a one of the key challenges has been to integrate
theoretical sense is the recent work by plant epidemiological parameters into population
pathologists on the appropriateness of the appli- genetics models (and vice versa) while at the same
cation of measures of genotypic diversity to time keeping model complexity at a manageable
microbial populations (Grünwald et al., 2003; level. Jeger (1997) illustrated this by incorporating
Kosman and Leonard, 2005). From these exam- host-pathogen gene-for-gene interactions into an
ples it should be obvious that plant epidemiology analytical SIR model. This resulted in a set of six
can benefit greatly from concepts and tools linked differential equations, one each for homo-
developed in population genetics (and vice versa), zygous and heterozygous genotypes of both host
including in studies designed to test theoretical and pathogen. Although the model was not very
ideas. tractable analytically, it allowed for the derivation
Eriksen et al. (2001) used numerical simulations of threshold criteria for persistence of specific
to address a question that had been the subject of pathogen and host genotypes. Subsequent simpli-
vigorous theoretical discussions. The problem, fication of the model allowed the effects of host
broadly put, was to determine the role of ascosp- density dependence, fitness cost for virulence in the
ores in development and microevolution of pathogen, and fitness cost for host resistance to be
septoria tritici blotch of wheat caused by My- incorporated and analyzed.
cosphaerella graminicola. Population genetics Durability of host resistance, a key concept in
studies in the United States in the 1990s had pro- population genetics, also has been examined from
vided indirect evidence for sexual reproduction by an epidemiological perspective (van den Bosch and
M. graminicola during the growing season Gilligan, 2003). This analysis considered three
(McDonald et al., 1995; Chen and McDonald, epidemiologically based measures of durability of
1996). An important question was how to deter- resistance: (1) time to invasion by a virulent
mine the relative contribution of immigration pathogen genotype; (2) time taken for the virulent
(gene flow) and sexual reproduction to the genetic genotype to dominate the pathogen population;
structure of the pathogen during the course of an and (3) time until a threshold proportion of the
epidemic, and which of these two evolutionary host population becomes diseased (‘additional
forces is of greater epidemiological importance uninfected crop growth days’). These metrics differ
within a season. This was resolved, not without conceptually from conventional population
some debate, through mark-release-recapture genetics-based measures of resistance durability in
experiments (Zhan et al., 1998) and a theoretical that they emphasize quantitative rather than
analysis (a subject of two letters to the editor in qualitative aspects, i.e., they focus on the duration
Phytopathology) of the data to estimate the rates of of resistance utility rather than the conditions
recombination and migration (Brown, 2000; Zhan under which durability is maintained. The model
et al., 2000). Nonetheless, these studies did not showed that if the virulent pathogen genotype is
answer the question of the relative contribution of not already present, and the time between intro-
ascospores vs. pycnidiospores to disease develop- duction (by mutation or immigration) and estab-
ment, nor of the extent of genetic recombination. lishment is considered as a metric of resistance
Through simulation modelling Eriksen et al. durability, both low and high proportions of
(2001) showed that the extended latent period of resistant genotypes in the crop can prolong dura-
pseudothecia compared with that of pycnidia leads bility. This observation might explain the oft-
to the release of ascospores too late in the season encountered difficulty in trying to predict the
to have a major effect on final severity of septoria durability of resistance genes (Hovmøller et al.,
tritici blotch epidemics in northern Europe. 1997; Brown, 2002; Burnett, 2003). The results
 67

also showed that the metric representing addi- aggressiveness is favoured in susceptible host
tional crop growth days without disease is genotypes. Although the use of spore production
unaffected by the proportion of the resistant host as a measure of aggressiveness may be subject to
genotype in a cultivar mixture, thus concurring debate, the study marks an important step toward
with data from experimental field studies in which documenting a virulence-aggressiveness trade-off
varying the proportion of mixture constituents (of for which previous evidence has been weak
sorghum in this case) had no effect on time of (Mundt, 2002), especially in natural systems.
disease onset (Ngugi et al., 2001).
Gudelj et al. (2004) investigated evolution of Statistical epidemiology
sibling fungal plant pathogens from an epidemio-
logical perspective using adaptive dynamics Apart from forming a crucial link between theory
methodology. They focused on the role of multiple and data, statistical concepts – in their own right –
host species involving a trade-off between the may result in new theoretical knowledge about
evolutionary benefit of being specialized and its plant pathosystems and plant epidemiology. For
cost (reduced virulence on other hosts). The results instance, distribution-based methods to charac-
showed that this infectivity trade-off accounted for terize disease aggregation in a spatial hierarchy
the evolution of only those pathogen siblings with (Hughes et al., 1997; Madden and Hughes, 1999)
non-overlapping host ranges (i.e., a high degree of have led to novel, testable hypotheses regarding
host specialization such as observed with obligate disease dynamics in time and space, e.g., for inci-
parasites), and that other mechanisms (ecological dence–severity relationships. With ever increasing
and/or epidemiological) must account for the computing power and a better understanding of
evolution of generalists with overlapping host how to utilize contemporary statistical tools, new
ranges and that of groups containing both gener- opportunities for the application of statistics in
alist and specialist siblings. plant epidemiology, both theoretical and applied,
Several important generalizations about the role continue to emerge.
of spatial structure in host-pathogen coevolution
can be drawn from the work by PH Thrall and JJ Generalized linear mixed models
Burdon, who integrated population genetics and Garrett et al. (2004) highlighted several statistical
spatio-temporal analysis of epidemics in natural methods that are used relatively little but have the
pathosystems (Burdon and Thrall, 1999, 2004). potential to improve inference from a range of
Key among these is that disease patterns in host– epidemiological studies. Foremost among these
pathogen metapopulations are spatially and tem- are mixed-effects models, i.e., models to analyze
porally asynchronous, whereby the magnitude of data with fixed and random effects. At a theoret-
pathogen fluctuations varies between host popu- ical level, the nature and properties of generalized
lations but there is clustering of disease levels linear mixed models (GLMMs) have been under-
among populations. This prediction is supported stood for decades (McCulloch and Searle, 2001),
by results of experimental studies (Burdon and but until recently, without significant input from a
Thrall, 2000; Thrall and Burdon, 2000, 2003; Bock specialist statistician, mixed-effects modelling has
et al., 2002; Thrall et al., 2002). Further, disease been very difficult in practice. Now, an increasing
persistence, and hence its impact on coevolution, is number of articles in application-oriented journals
higher at the local level. As a consequence, there is provide guidance for setting up mixed models and
a tight evolutionary link between resistance and for implementing them in off-the-shelf statistical
virulence of associated host–pathogen pairs packages (Piepho, 1999; Madden et al., 2002;
whereby pathogen virulence (ability to infect many Piepho et al., 2003; Spilke et al., 2005). One of the
host genotypes) increases with increasing mean most important advantages of these models is their
resistance of the host sub-population (Thrall and applicability to unbalanced designs, for which
Burdon, 2003). These studies also provided evi- exact statistical tests are usually not available.
dence for a trade-off between virulence and Therefore, one needs to resort to approximate
aggressiveness (defined here as spore production methods such as the restricted maximum likeli-
per pustule), whereby selection for the former is hood approach. Even for experimental designs for
favoured in resistant host genotypes while that for which traditional general linear models (GLMs)
68

are appropriate, analysis using the GLMM can olds for making decisions. An ROC curve is a plot
produce more robust results when variances are of the true positive rate (sensitivity) as a function
unequal and/or sample sizes are small (Piepho, of the false positive rate (1 – specificity) at all
1999). Madden et al. (2002) evaluated various possible decision thresholds of the risk algorithm.
GLMMs and recommended the fixed residual This curve allows one to identify trade-offs be-
variance model (which is also the simplest tween liberal and conservative thresholds in an
GLMM) for analyzing disease incidence data from attempt to identify the most suitable decision
designed experiments. threshold for a given application. ROC analysis is
best suited for responses that are inherently
dichotomous, for instance the decision whether or
Survival analysis
not to apply a fungicide. In a recent example,
Data on the occurrence and timing of events such
Dewdney et al. (2002) used ROC analysis and
as sclerotium germination, disease onset, or leaf
historical data to evaluate parameters of
abscission are routinely encountered in epidemio-
MARYBLYT (a forecaster for fire blight of apple
logical studies. With such time-to-event data, sev-
and pear) and to identify where key improvements
eral problems can arise that limit the usefulness of
were needed. MARYBLYT and Cougarblight
traditional statistical methods: (1) the times are
(another fire blight forecaster) have been com-
unlikely to be distributed normally; (2) the data set
pared using ROC analysis and found to have
will likely contain censored observations, i.e.,
equivalent action thresholds and thus perform
observations for which the event has not occurred
similarly in their ability to predict blossom blight
when the study was completed; and (3) the
(Dewdney et al., 2003).
response may be influenced by time-dependent
ROC analysis also can be applied in situations
covariates, i.e., independent variables whose
where the response is not dichotomous (Patil,
values change during the study period. Because of
1991), for instance the decision on how much
these properties, time-to-event data are now
fertilizer to apply or how many fungicide appli-
increasingly being modelled using survival analysis
cations to make. However, in plant epidemiology,
(Scherm and Ojiambo, 2004). This set of statistical
ROC analysis of responses on a non-dichotomous
methods not only allows the comparison of time-
scale has yet to be demonstrated.
to-event distributions among treatment groups,
but also the development of models for the effects
Bayesian analysis
of discrete and/or continuous covariates on event
The evaluation of plant disease forecasters based
times. Recently published examples include anal-
on ROC analysis may be improved further when
yses of the effects of landscape attributes on the
conducted in a Bayesian framework (Yuen and
time to invasion by an exotic plant pathogen (Jules
Hughes, 2002). This is accomplished by consider-
et al., 2002); of orchard characteristics, environ-
ing the prior probability of disease occurrence in
ment, and disease status of neighbouring trees on
addition to the likelihood ratios for positive and
the time of virus infection of individual orchard
negative predictions by the risk algorithm. The
trees (Dallot et al., 2004); and of disease severity
latter two are calculated directly from sensitivity
and other leaf attributes on the time of premature
and specificity of the forecaster, while the former
defoliation of diseased plants (Ojiambo and
may be based either on the historical prevalence of
Scherm, 2005).
the disease in the region of interest, or on growers’
subjective estimates of disease risk. In either case,
Decision analysis the result is a posterior probability of disease
Important advances have been made in the area of occurrence given the prediction by the forecaster.
decision analysis for disease management, espe- Yuen and Hughes (2002) illustrate this approach
cially in relation to the quantitative evaluation of by means of risk algorithms for eyespot of wheat
risk algorithms such as disease forecasters (Yuen and Sclerotinia stem rot of canola (oilseed rape).
et al., 1996; Hughes et al., 1999; Yuen and Apart from its application in the specific
Hughes, 2002; Madden, 2005). Increasingly, ROC example of ROC analysis discussed above, Bayes’s
(receiver-operating characteristic) analysis is being theorem presents a general framework for incor-
employed to optimize risk algorithms and thresh- porating uncertainty and prior information into
 69

epidemiological analyses and for updating current individuals are selected stochastically and
knowledge as new information becomes available modified (mutated or recombined) to form a new
(Mila and Carriquiry, 2004). The key feature is the population. Although the concept of evolutionary
calculation of posterior probabilities for the computing may be intuitively appealing to biolo-
parameter of interest based on empirically derived gists, the approach is computationally intensive
prior probabilities in conjunction with the condi- and effectively treats the optimization problem as
tional probability of each possible outcome. This a black box. Its theoretical and practical impact on
use of prior probabilities represents a powerful plant epidemiology remains to be seen.
mechanism for incorporating subjective informa- Apart from merely adopting statistical tools
tion such as growers’ perceptions. This is illus- from other disciplines, plant epidemiologists
trated in the work of Mila et al. (2003), who should be more proactive in exploring potential
examined the effect of soybean growers’ produc- applications of their concepts and procedures in
tion decisions on Sclerotinia stem rot incidence rapidly expanding disciplines such as statistical
using decision theory under uncertainty. Predic- genetics or bioinformatics. Conceptually, for
tions of stem rot incidence and soybean yield instance, there are many parallels between the
based on regression-type models were updated dynamics of plant pathogens in populations of
with growers’ subjective estimates of disease inci- plants and those of genetic loci or markers within a
dence via Bayes’s theorem. The resulting posterior genome (Delwiche, 2004). The key here is to
probabilities were then used to derive management remain imaginative and keep an open mind toward
criteria for profit maximization. broader applications, without being confined
Economic criteria (which often exhibit consid- to the organismal level that has historically
erable uncertainty) and growers’ perceptions are dominated statistical applications in plant
among the most important drivers affecting dis- epidemiology.
ease management decisions, yet they are routinely
ignored by plant pathologists developing decision
algorithms. The continued penetration of Bayesian Conclusions
analysis into the epidemiological mainstream
should lead to a greater appreciation of the Based on the selected examples given above there
importance of these drivers and – it is hoped – can be little doubt that significant progress has
their more widespread incorporation into disease been made in theoretical plant epidemiology since
management models. the turn of the century. New theories and models
continue to be developed, and sincere efforts are
Statistical genetics and bioinformatics being made to relate them to the broader field of
As shown in the above examples, statistical plant theoretical biology on one hand and practical
epidemiology has relied heavily on the medical and disease management on the other. As analytical
ecological fields for inspiration and conceptual models of plant disease dynamics have become
advances. This trend will likely continue in the more realistic, they also have become considerably
future as plant epidemiologists become more more complex, and solutions often can be
familiar with the theories and tools of statistical obtained only numerically. As such, the division
genetics and bioinformatics. In a recent example, between analytical and simulation models, an
Parsons and Te Beest (2004) used genetic algo- important distinction some 20 years ago (Jeger,
rithms to optimize fungicide applications on win- 1986), is narrowing. It seems that we have come
ter wheat relative to spray date as well as choice, back full circle to the medium-sized models
number, and dose of active ingredients. Genetic advocated by Botkin’s (1977).
algorithms use biologically derived concepts such Although theoretical problems need not be tied
as inheritance, mutation, natural selection, and to practical applications to be valid, the image of
recombination to ‘evolve’ a large population of theoretical epidemiology within the larger field of
possible solutions (‘individuals’) to the best (‘fit- plant pathology could benefit from a clearer doc-
test’) solution (‘survivor’). The evolution starts umentation of its impact on practical disease
from a population of completely random individ- management. In medical epidemiology, such
uals, and in each subsequent generation multiple evaluations are commonly achieved by comparing
70

model outputs with long-term morbidity data sets, using AFLP fingerprinting. Mycological Research 106:
e.g., in the case of models for the impact of 428–434
Botkin DB (1977) Bits, bytes and IBP. Bioscience 27: 385
vaccination on childhood diseases (Rohani et al., Brown JH, Gupta VK, Li BL, Milne BT, Restrepo C and West
2000). We would like to call attention to the need GB (2002) The fractal nature of nature: Power laws,
for similar analyses in plant epidemiology, espe- ecological complexity and biodiversity. Philosophical
cially with pathosystems for which long-term data Transactions of the Royal Society of London, Series B
are available (e.g., the cereal rusts). Establishment 357: 619–626
Brown JKM (2000) Estimation of rates of recombination and
of additional long-term data collection standards, migration in populations of plant pathogens. Phytopathol-
even if only for a limited number of pathosystems, ogy 90: 320–323
would provide a more solid data base from which Brown JKM (2002) Yield penalties of disease resistance in
to evaluate the impact of interventions suggested crops. Current Opinions in Plant Biology 5: 339–344
by current theoretical knowledge. Burdon JJ and Thrall PH (1999) Spatial and temporal patterns
in coevolving plant and pathogen associations. American
While plant epidemiology, by definition, is Naturalist 153(Suppl): S15–S33
concerned with the study of populations of Burdon JJ and Thrall PH (2000) Coevolution at multiple spatial
pathogens in populations of plants, there exists scales: Linum marginale–Melampsora lini – from the indi-
ample opportunity to broaden the scale of inves- vidual to the species. Evolutionary Ecology 14: 261–281
Burdon JJ and Thrall PH (2004) Genetic structure of natural
tigation and apply the concepts of theoretical
plant and pathogen populations. In: Ehler LE, Sforza R
epidemiology to both sub-organismal and ecosys- and Mateille T (eds) Genetics, Evolution, and Biological
tem scales. Examples of such non-traditional Control (pp 1–17) CABI Publishing, Wallingford, UK
applications could include models of virus cross- Burnett J (2003) Fungal Populations and Species. Oxford
protection in individual plant cells, temporal and University Press, Oxford, UK
spatial dynamics of molecular markers or of mol- Caswell H (1988) Theory and models in ecology: A different
perspective. Ecological Modelling, 43: 33–44
ecules such as mycotoxins, biotechnology risk Chen RS and McDonald BA (1996) Sexual reproduction plays a
assessment, microbial forensics, or the quantitative major role in the genetic structure of populations of the
analysis of ecosystem health. We would argue that fungus Mycosphaerella graminicola. Genetics 142: 1119–1127
plant epidemiologists, including theoreticians, are Cortesi P and Milgroom MG (2001) Outcrossing and
not yet taking advantage of these new opportuni- diversity of vegetative compatibility types in populations
of Eutypa lata from grapevines. Journal of Plant Pathol-
ties sufficiently. Although providing the scientific ogy 83: 79–86
basis for disease management will always be the Cortesi P, Fischer M and Milgroom MG (2000) Identification
raison d’eˆtre for plant epidemiology, a broader and spread of Fomitiporia punctata associated with wood
perspective will help the discipline to remain rele- decay of grapevine showing symptoms of esca. Phytopa-
thology 90: 967–972
vant as more efforts and resources continue to be
Cousens R, Wallinga J and Shaw M (2004) Are the spatial
devoted to genomic and ecosystem-level science. patterns of weeds scale-invariant? Oikos 107: 251–264
Plant epidemiology, both theoretical and applied, Cowger C, Wallace LD and Mundt CC (2005) Velocity of
will remain as integrating a discipline as it has ever spread of wheat stripe rust epidemics. Phytopathology 95:
been, but the individual components that require 972–982
integration are changing. Dallot S, Gottwald T, Labonne G and Quiot JB (2004) Factors
affecting the spread of Plum pox virus strain M in peach
orchards subjected to roguing in France. Phytopathology
94: 1390–1398
References Delwiche CF (2004) The genome palimpsest: Genomics in
evolution and ecology. Bioscience 54: 991–1001
Akimitsu K, Peever TL and Timmer LW (2003) Molecular, Dewdney MM, Biggs AR, Lightner G and Turechek WW
ecological and evolutionary approaches to understanding (2002) MARYBLYT: Can there be improvements? (Abstr)
Alternaria diseases of citrus. Molecular Plant Pathology 4: Phytopathology 92: S19
435–446 Dewdney MM, Biggs AR and Turechek WW (2003) A
Anon. (2005) Encyclopaedia Britannica Online. http://www. statistical comparison of MARYBLYT and Cougarblight
britannica.com (accessed 28 August 2005) using Receiver Operator Characteristic (ROC) analysis.
Bailey DJ, Otten W and Gilligan CA (2000) Saprotrophic (Abstr) Phytopathology 93: S20
invasion by the soil-borne fungal pathogen Rhizoctonia solani Eriksen L, Shaw MW and Østergård H (2001) A model of the
and percolation thresholds. New Phytologist 146: 535–544 effect of pseudothecia on genetic recombination and
Bock CH, Thrall PH, Brubaker CL and Burdon JJ (2002) epidemic development in populations of Mycosphaerella
Detection of genetic variation in Alternaria brassicicola graminicola. Phytopathology 91: 240–248
 71

Ferrandino FJ (1993) Dispersive epidemic waves. I. Focus Hughes G and Gottwald TR (1998) Survey methods for
expansion within a linear planting. Phytopathology 83: assessment of Citrus tristeza virus incidence. Phytopathol-
795–802 ogy 88: 715–723
Ferrandino FJ (2004) Measuring spatial aggregation in binary Hughes G, McRoberts N and Burnett FJ (1999) Decision-
epidemics: Correlative analysis and the advantage of making and diagnosis in disease management. Plant
fractal-based sampling. Phytopathology 94: 1215–1227 Pathology 48: 147–153
Flier WG, Grünwald NJ, Kroon LPNM, Sturbaum AK, Hughes G, McRoberts N, Madden LV and Gottwald TR
vanden Bosch TBM, Garay-Serrano E, Lozoya-Saldana (1997) Relationships between disease incidence at two levels
H, Fry WE and Turkensteen LJ (2003) The population in a spatial hierarchy. Phytopathology 87: 542–550
structure of Phytophthora infestans from the Toluca Valley Jeger MJ (1982) The relation between total, infectious, and
of central Mexico suggests genetic differentiation between postinfectious diseased plant tissue. Phytopathology 72:
populations from cultivated potato and wild Solanum spp. 1185–1189
Phytopathology 93: 382–390 Jeger MJ (1986) The potential of analytic compared with
Frantzen J and van den Bosch F (2000) Spread of organisms: simulation approaches to modeling plant disease epidemi-
Can travelling and dispersive waves be distinguished? Basic ology. In: Leonard KJ and Fry WE (eds) Plant Disease
and Applied Ecology 1: 83–91 Epidemiology: Population Dynamics and Management (pp
Garrett KA and Bowden RL (2002) An Allee effect reduces the 255–281)1 Macmillan, New York, USA
invasive potential of Tilletia indica. Phytopathology 92: Jeger MJ (1997) An epidemiological approach to modelling the
1152–1159 dynamics of gene-for-gene interactions. In: Crute IR,
Garrett KA, Madden LV, Hughes G and Pfender WF (2004) Holub EB and Burdon JJ (eds) The Gene-for-Gene
New applications of statistical tools in plant pathology. Relationship in Plant–Parasite Interactions (pp 191–209)
Phytopathology 94: 999–1003 CABI Publishing, Wallingford, UK
Gibson GJ, Kleczkowski A and Gilligan CA (2004) Bayesian Jeger MJ (2000) Theory and plant epidemiology. Plant Pathol-
analysis of botanical epidemics using stochastic compart- ogy 49: 651–658
mental models. Proceedings of the National Academy of Jeger MJ and van den Bosch F (1994a) Threshold criteria for
Sciences USA 101: 12120–12124 model plant disease epidemics: I. Asymptotic results.
Gilligan CA (2002) An epidemiological framework for disease Phytopathology 84: 24–27
management. Advances in Botanical Research 38: 1–64 Jeger MJ and van den Bosch F (1994b) Threshold criteria for
Gobbin D, Pertot I and Gessler C (2003) Genetic structure of a model plant disease epidemics: II. Persistence and endemic-
Plasmopara viticola population in an isolated Italian moun- ity. Phytopathology 84: 28–30
tain vineyard. Journal of Phytopathology 151: 636–646 Jeger MJ, Holt J, van den Bosch F and Madden LV (2004)
Grünwald NJ, Goodwin SB, Milgroom MG and Fry WE Epidemiology of insect-transmitted plant viruses: Model-
(2003) Analysis of genotypic diversity data for populations ling disease dynamics and control interventions. Physiolog-
of microorganisms. Phytopathology 93: 738–746 ical Entomology 29: 291–304
Gubbins S, Gilligan CA and Kleczkowski A (2000) Population Jules ES, Kauffman MJ, Ritts WD and Carroll AL (2002)
dynamics of plant-parasite interactions: Thresholds for Spread of an invasive pathogen over a variable landscape:
invasion. Theoretical Population Biology 57: 219–233 A non-native root rot on Port Orford cedar. Ecology 83:
Gudelj I, Fitt BDL and van den Bosch F (2004) Evolution of 3167–3181
sibling fungal plant pathogens in relation to host special- Kermack WO and McKendrick AG (1927) A contribution to
ization. Phytopathology 94: 789–795 the mathematical theory of epidemics. Proceedings of the
Hall CAS (1988a) An assessment of several of the historically Royal Society of London, Series A 115: 700–721
most influential theoretical models used in ecology and of Kleczkowski A, Gilligan CA and Bailey DJ (1997) Scaling and
the data provided in their support. Ecological Modelling 43: spatial dynamics in plant-pathogen systems: From individ-
5–31 uals to populations. Proceedings of the Royal Society of
Hall CAS (1988b) What constitutes a good model and by whose London, Series B 264: 979–984
criteria? Ecological Modelling 43: 125–127 Kosman E and Leonard KJ (2005) Similarity coefficients for
Hall RJ, Gubbins S and Gilligan CA (2004) Invasion of drug molecular markers in studies of genetic relationships
and pesticide resistance is determined by a trade-off between individuals for haploid, diploid, and polyploid
between treatment efficacy and relative fitness. Bulletin of species. Molecular Ecology 14: 415–424
Mathematical Biology 66: 825–840 Li BL (2000) Fractal geometry applications in description and
Holt J and Colvin J (2001) Observation and theory of whitefly- analysis of patch patterns and patch dynamics. Ecological
borne virus disease epidemics. In: Jeger MJ and Spence NJ Modelling 132: 33–50
(eds) Biotic Interactions in Plant–Pathogen Associations Loskill B, Gobbin D and Berkelmann-Löhnertz B, (2004)
(pp 331–343) CABI Publishing, Wallingford, UK Vertikale Verteilung bodenbürtiger Infektionen durch
Hovmøller MS, Østergård H and Munk L (1997) Modelling Plasmopara viticola. (Abstr) Mitteilungen aus der
virulence dynamics of airborne plant pathogens in relation Biologischen Bundesanstalt für Land- und Forstwirtschaft,
to selection by host resistance in agricultural crops. In: Berlin-Dahlem 396: 113
Crute IR, Holub EB and Burdon JJ (eds) The Gene-for- Madden LV (2005) Botanical epidemiology: Some key advances
Gene Relationship in Plant-Parasite Interactions (pp 173– and its continuing role in disease management. European
190) CABI Publishing, Wallingford, UK Journal of Plant Pathology (this volume)
72

Madden LV and Hughes G (1999) An effective sample size for Rhizoctonia solani through structured soil. Soil Biology
predicting plant disease incidence in a spatial hierarchy. and Biochemistry 36: 203–210
Phytopathology 89: 770–781 Park AW, Gubbins S and Gilligan CA (2001) Invasion and
Madden LV and van den Bosch F (2002) A population- persistence of plant parasites in a spatially structured host
dynamics approach to assess the threat of plant pathogens population. Oikos 94: 162–174
as biological weapons against annual crops. Bioscience 52: Park AW, Gubbins S and Gilligan CA (2003) Extinction times
65–74 for closed epidemics: The effects of host spatial structure.
Madden LV, Jeger MJ and van den Bosch F (2000) A Ecology Letters 5: 747–755
theoretical assessment of the effects of vector–virus trans- Parsons DJ and Te Beest D (2004) Optimising fungicide
mission mechanism on plant virus disease epidemics. applications on winter wheat using genetic algorithms.
Phytopathology 90: 576–594 Biosystems Engineering 88: 401–410
Madden LV, Turechek WW and Nita M (2002) Evaluation of Patil GP (1991) Encountered data, statistical ecology, environ-
generalized linear mixed models for analyzing disease mental statistics, and weighted distribution methods. Envi-
incidence data obtained in designed experiments. Plant ronmetrics 2: 377–423
Disease 86: 316–325 Peever TL, Olsen L, Ibañez A and Timmer LW (2000) Genetic
McCulloch CE and Searle SR (2001) Generalized, Linear, and differentiation and host specificity among populations of
Mixed Models. Wiley, New York, USA Alternaria spp. causing brown spot of grapefruit and
McDonald BA (2004) Population genetics of plant pathogens. tangerine  grapefruit hybrids in Florida. Phytopathology
The Plant Health Instructor, doi:10.1094/PHI-A-2004- 90: 407–414
0524-01 Peever TL, Salimath SS, Su G, Kaiser WJ and Muehlbauer FJ
McDonald BA, Pettway RE, Chen RS, Boeger JM and (2004) Historical and contemporary multilocus population
Martinez JP (1995) The population genetics of Septoria structure of Ascochyta rabiei (teleomorph: Didymella rabiei)
tritici (teleomorph Mycosphaerella graminicola). Canadian in the Pacific Northwest of the United States. Molecular
Journal of Botany 73(Suppl 1): S292–S301 Ecology 13: 291–309
McRoberts N, Hughes G and Madden LV (2003) The Phillips AN, Youle M, Johnson M and Loveday C (2001) Use
theoretical basis and practical application of relationships of a stochastic model to develop understanding of the
between different disease intensity measurements in plants. impact of different patterns of antiretroviral drug use on
Annals of Applied Biology 142: 191–211 resistance development. AIDS 15: 2211–2220
Mila AL and Carriquiry AL (2004) Bayesian analysis in plant Piepho HP (1999) Analysing disease incidence data from
pathology. Phytopathology 94: 1027–1030 designed experiments by generalized linear mixed models.
Mila AL, Carriquiry AL, Zhao J and Yang XB (2003) Impact Plant Pathology 48: 668–674
of management practices on prevalence of soybean Scle- Piepho HP, Büchse A and Emrich K (2003) A hitchhiker’s
rotinia stem rot in the north-central United States and on guide to mixed models for randomized experiments. Journal
farmers’ decisions under uncertainty. Plant Disease 87: of Agronomy and Crop Science 189: 310–322
1048–1058 Rohani P, Earn DJD and Grenfell BT (2000) Impact of
Milgroom MG (2001) The synthesis of genetics and epidemi- immunisation on pertussis transmission in England and
ology: Contributions of population biology in plant pathol- Wales. The Lancet 355: 285–286
ogy. Journal of Plant Pathology 83: 57–62 Scherm H (1996) On the velocity of epidemic waves in model
Milgroom MG and Peever TL (2003) Population biology of plant disease epidemics. Ecological Modelling 87: 217–222
plant pathogens: The synthesis of plant disease epidemiol- Scherm H and Coakley SM (2003) Plant pathogens in a
ogy and population genetics. Plant Disease 87: 608–617 changing world. Australasian Plant Pathology 32: 157–165
Mundt CC (2002) Use of multiline cultivars and cultivar Scherm H and Ojiambo PS (2004) Applications of survival
mixtures for disease management. Annual Review of analysis in botanical epidemiology. Phytopathology 94:
Phytopathology 40: 381–410 1022–1026
Naylor M, Godfray HCJ, Pallett DW, Tristem M, Reeves JP Segarra J, Jeger MJ and van den Bosch F (2001) Epidemic
and Cooper JI (2003) Mutualistic interactions amongst dynamics and patterns of plant diseases. Phytopathology
viruses? In: Hails RS, Beringer JE and Godfray HCJ (eds) 91: 1001–1010
Genes in the Environment (pp 205–225) Blackwell, Oxford, Spilke J, Piepho HP and Hu X (2005) Analysis of unbalanced
UK data by mixed linear models using the MIXED procedure
Ngugi HK, King SB, Holt J and Julian AM (2001) Simulta- of the SAS system. Journal of Agronomy and Crop Science
neous temporal progress of sorghum anthracnose and leaf 191: 47–54
blight in crop mixtures with disparate patterns. Phytopa- Stacey AJ, Truscott JE and Gilligan CA (2001) Soilborne
thology 91: 720–729 fungal pathogens: Scaling-up from hyphal to colony
Ojiambo PS and Scherm H (2005) Survival analysis of time to behaviour and the probability of disease transmission.
abscission of blueberry leaves affected by Septoria leaf spot. New Phytologist 150: 169–177
Phytopathology 95: 108–113 Stacey AJ, Truscott JE, Asher MJC and Gilligan CA (2004) A
Otten W, Bailey DJ and Gilligan CA (2004a) Empirical model for the invasion and spread of rhizomania in the
evidence of spatial thresholds to control invasion of fungal United Kingdom: Implications for disease control strate-
parasites and saprotrophs. New Phytologist 163: 125–132 gies. Phytopathology 94: 209–215
Otten W, Harris K, Young IM, Ritz K and Gilligan CA Taylor CM and Hastings A (2005) Allee effects in biological
(2004b) Preferential spread of the pathogenic fungus invasions. Ecology Letters 8: 895–908
 73

Thrall PH and Burdon JJ (2000) Effect of resistance variation in blight in wheat. European Journal of Plant Pathology 110:
a natural plant host-pathogen metapopulation on disease 959–971
dynamics. Plant Pathology 49: 767–773 Yuen JE and Hughes G (2002) Bayesian analysis of plant
Thrall PH and Burdon JJ (2003) Evolution of virulence in a plant disease prediction. Plant Pathology 51: 407–412
host–pathogen metapopulation. Science 299: 1735–1737 Yuen JE, Twengström E and Sigvald R (1996) Calibration and
Thrall PH, Burdon JJ and Bever JD (2002) Local adaptation in verification of risk algorithms using logistic regression.
the Linum marginale–Melampsora lini host–pathogen inter- European Journal of Plant Pathology 102: 847–854
action. Evolution 56: 1340–1351 Zadoks JC (1990) Whithering plant disease epidemiology?.
Turechek WW (2005) Spatial and temporal scales in plant Plant Disease 74: 82
disease epidemiology. European Journal of Plant Pathology Zadoks JC (2001) Plant disease epidemiology in the twentieth
(this volume) century. A picture by means of selected controversies. Plant
Turechek WW and Madden LV (2003) A generalized linear Disease 85: 808–816
modeling approach for characterizing disease incidence in a Zhan J, Mundt CC and McDonald BA (1998) Measuring
spatial hierarchy. Phytopathology 93: 458–466 immigration and sexual reproduction in field populations of
vanden Bosch F and Gilligan CA (2003) Measures of durability Mycosphaerella graminicola. Phytopathology 88: 1330–1337
of resistance. Phytopathology 93: 616–625 Zhan J, Mundt CC and McDonald BA (2000) Estimation of
vanden Bosch F, Metz JAJ and Zadoks JC (1999) Pandemics of rates of recombination and migration in populations of
focal plant disease, a model. Phytopathology 89: 495–505 plant pathogens – a reply. Phytopathology 90: 324–326
Vander Plank JE (1963) Plant Diseases: Epidemics and Control, Zhang XS and Holt J (2001) Mathematical models of cross
Academic Press, New York, USA protection in the epidemiology of plant virus diseases.
Xu XM, Parry DW, Edwards SG, Cooke BM, Doohan FM, Phytopathology 91: 924–934
vanMaanen A, Brennan JM, Monaghan S, Moretti A, Zhang XS, Holt J and Colvin J (2000) Mathematical models of
Tocco G, Mule G, Hornok L, Giczey G, Tatnell J, host plant infection by helper-dependent virus complexes:
Nicholson P and Ritieni A (2004) Relationship between Why are helper viruses always avirulent? Phytopathology
the incidences of ear and spikelet infection of Fusarium ear 90: 85–93
This page intentionally blank
European Journal of Plant Pathology (2006) 115:75–93
Springer 2006
DOI 10.1007/s10658-005-4039-x

Establishing priorities for plant science research and developing world food securityq

Robert W. Herdt
Adjunct International Professor of Applied Economics and Management, Cornell University, Ithaca, NY,
14853, USA (Phone: +1-607-255-8802; E-mail: [email protected])

Accepted 30 September 2005

Key words: congruence, crop loss, development assistance, hunger, research priorities, technological
change

Abstract

This paper begins with a broad review of food security in the developing world. I argue that technological
change has made a key contribution to improving food security wherever it has been achieved and that
plant sciences can contribute in the future. Potential contributions by plant scientists will have to be funded
through development assistance. A perspective on development assistance and the role of assistance to
agricultural research in particular provides a useful background to the consideration of how to set priorities
for research using information on what is needed and what can be done. Optimizing the contributions of
research entails five steps: (1) determine the specific objective, (2) identify alternatives to address the
objective, (3) choose a method by which to set priorities, (4) apply the selected method to quantify pri-
orities, (5) allocate available funds among the priority alternatives. Finally, it is important to take a long-
term view and continue supporting the research long enough to make a difference. The paper discusses
these steps, illustrates how such an approach might be applied and demonstrates the importance of
applying economic criteria to research resource allocation.

Introduction bour, irrigation and fertilizer have contributed to

the progress that has been made; in addition,
There are nearly 800 million hungry people intangible factors like efficient marketing sys-
including 185 million seriously malnourished pre- tems, dynamic production technology and higher
school children in the developing world. All lack education have played an equally important role
adequate food, water and protection from food- in generating long-run growth in agricultural
related disease, but without the great strides that production (Eicher and Staatz, 1998; Hayami
have been made in reducing hunger in Asia and and Ruttan, 1985; Mellor, 1966). These intangi-
Latin America over the past 50 years, there would ble factors are the major differences between the
be millions more. Unfortunately, progress has not low-productivity, traditional production systems
been achieved everywhere; in many African that still prevail in much of Africa and the dy-
countries food output per person has fallen over namic, high-input, high-output systems that
the last decade and in India and Bangladesh large increasingly prevail in Asia. Development assis-
numbers of hungry people remain despite the tance has contributed to Asia’s ability to keep
substantial gains in per capita food production. pace with its demand for food by helping Asian
Analysis of food production growth of the scientists develop suitable new agricultural tech-
past 50 years shows that increases in land, la- nology; appropriate development assistance
could help Africa begin the same process and
q
Prepared for the International Epidemiology Workshop, agricultural research could play a part (Sachs,
Rennes, France, April 13, 2005 2005).
76

This paper briefly reviews the process by which tific revolutions were applied to food production;
most of the world has achieved food security, farm incomes grew and per capita food supplies
identifies the focus of remaining needs, sketches increased (Johnson, 2000).
the contributions of development assistance and Today the developing world has twice the
considers how agricultural research priorities population it had in 1960, 150 million fewer
should be established. It discusses five steps in- hungry people, real prices of food grain one-
volved in priority setting: determining the objec- third as high, and 20% more food available per
tive, identifying alternative activities, choosing a person. These great advances in food security
method for setting priorities, implementing the resulted from a combination of technology,
method, and allocating resources among the policies and institutions that encouraged pro-
alternatives according to the priorities established. duction growth in agriculture. As explained by
The importance of using economic considerations T.W. Schultz, developing world farmers while
along with crop loss estimates for establishing re- poor, use the resources and technology available
search priorities is demonstrated. to them efficiently, but without the innovations
in policy, institutions and technology needed to
generate the ability to accelerate food production
Food security and development assistance and the incentives to use those innovations they
are unlikely to increase production much faster
The long view than needed to meet their own needs (Schultz,
1964).
Before the middle of the nineteenth century Technology embodied in fertilizer and machin-
‘hunger and premature death’ was the norm for ery drove the increases in food security in the
most of humanity. Eighty percent or more of all industrialized countries from about 1850 onwards.
people in Europe, North America and elsewhere But when high rates of fertilizer were applied to
in the world were farmers, as in today’s poor rice grown in the tropics at mid-20th century, they
countries. A gradual process of agricultural caused the plants to grow rank and fall over rather
development and demographic transformation than produce more grain (Herdt and Mellor,
from 1800 to 1950 was required to largely 1964). It took the green revolution of the 1970s to
achieve food security in industrialized Europe provide new varieties sufficiently productive under
and North America nearly doubling life expec- tropical conditions to generate a growth spurt in
tancy from 35 to 68 years (Fogel, 2004). But in Asian agriculture.
1950 ‘hunger and premature death’ was the Complementary policies to assure greater secu-
norm in Egypt, India, China and most of the rity of tenure and more stable prices helped. As
rest of the developing world, with life expectancy farming becomes more technologically advanced it
hovering around 40 years. Hunger, stunting, requires capital investments like wells and build-
nutritional deficiencies and diseases were wide- ings that are attached to the land. To encourage
spread. Then around the middle of the 20th farmers to make such investments they must have
century this began to change for many poorer assured rights to the land; alternatively, govern-
countries; average food consumption increased ments may invest in irrigation systems and other-
by 20%, real prices of food fell despite a dou- wise subsidize agricultural investments. The
bling of population and life expectancy increased institutions that assure land rights, incentive prices
from 40 to 64 years (FAO, 2002). This remark- and a steady stream of new technology are critical
able achievement took one-third the time re- for agricultural development. All these require-
quired by the industrialized countries of the ments can only be achieved in a stable, non-
north. Chronic food shortages, as manifested in oppressive political, social, and economic context.
protein-energy malnutrition, fell in much of Asia Hence well-functioning governments that under-
and Latin America, in large part because grain stand the importance of agriculture, make the
yields and farm incomes increased through a necessary investments in agricultural infrastruc-
very similar process used in Europe, North ture and human capital, and encourage a balance
America and Japan during the previous between markets and the state, are critical (Hay-
150 years. Products of the industrial and scien- ami, 2001).
 77

International contributions In sub-Saharan Africa there were limited con-

tributions from the green revolution. New varieties
Crop yields in the developing world have increased of most crops did not exceed 30% of planted area
substantially over the past 50 years. Wheat, maize and fertilizer application rates remain at five to
and rice yields have more than doubled in most 10% of the levels used in Asia. Much of the output
regions with greater increases in Asia and Latin increase that did occur was achieved by extending
America than in Africa; yields of other crops like the area under cultivation and mining the soil of
sorghum and potato have also increased signifi- plant nutrients through shorter fallow periods.
cantly. New modern crop varieties together with Food production did not keep pace with popula-
fertilizer and irrigation drove these gains; the tion growth and a decade-long drop in per capita
greater the adoption of these technologies, the food production continues. Today, Africa faces a
greater the yield increases. Fertilizer consumption food crisis and an environmental crisis, both
grew at over 10% per year in the 1990s and resulting from low input, low yield agriculture.
reached 225 kg per hectare of arable land in East While technological change was central to agri-
and South East Asia and 110 kg in South Asia; it cultural development, aid for technological change
was stagnant in Sub-Sahara Africa at less than has been a small fraction of agricultural aid and
10 kg in 2000 (FAO, 2005a). From 1960 to 2000, agricultural aid has been a small fraction of total
public breeding programmes in over 100 develop- aid. Between 1973 and 2005, total development
ing countries released over 8,000 new varieties of assistance varied between about $40 billion and
the major food crops (Evenson and Gollin, 2003). $60 billion annually in 2002 dollars, according to
More than 35% of these varieties were based on data compiled by the Organization for Economic
crosses made at the international agricultural re- Cooperation and Development (OECD Develop-
search centres funded by development assistance ment Assistance Committee). Development assis-
and many of the others were made by plant tance to agriculture from all wealthy countries
breeders trained at those centres or stimulated to grew from $4.7 billion in 1973 to over $12 billion
emulate or exceed the achievements of the centres. per year in 1983–87 but since then has fallen back
Since the 1990s private, local and international to about the 1973–1977 amounts (Table 1). In the
seed companies have begun creating varieties for most recent period around one-quarter of all aid to
developing countries based on ‘platform’ varieties agriculture went to what OECD calls ‘agricultural
generated by these public sector breeding pro- sector policy, planning and programmes; aid to
grammes. agricultural ministries; institution capacity

Table 1. Official development assistance to agriculture sub-sectors, annual average constant $ 2002, million

1973–1977 1978–1982 1983–87 1988–1992 1993–1997 1998–2002

Agricultural policy & administration 421 359 857 1468 562 1614
Agricultural water resources 1097 2207 2114 1699 1061 660
Agricultural development & general 735 1251 2307 1188 1081 647
Forestry, not research 149 369 613 880 468 354
Crop production 331 1173 1028 724 388 258
Fisheries, not research 192 471 400 408 285 235
Research 63 275 456 375 184 201
Agricultural inputs 313 684 552 317 309 186
Agricultural land resources 204 253 795 417 271 178
Agricultural finance and coops 425 1127 1549 895 209 132
Extension 104 235 514 230 77 99
Livestock production + vet services 274 379 331 312 124 94
Agricultural services 426 544 1035 840 167 71
Agrarian reform 0 38 31 440 143 63
Total agriculture 4735 9371 12596 10201 5353 4813
Food Aid (not included above) 2681 2858 3000 1502 524 1383

Source: Extracted from OECD (2005), deflated by the total DAC deflator.
78

building and advice; unspecified agriculture.’ The investments intended to generate additional in-
second largest amount went to ‘agricultural water come and low economic rates of return were of
resources,’ and the third largest to ‘integrated much less concern than several decades earlier.
projects’ and ‘farm development.’ Agricultural Integrated agricultural or rural development
research has received modest support over the projects made up the second largest area of agri-
years, seldom exceeding 10% of agricultural aid. It cultural development assistance in the 1960s and
is, however, impossible to identify the develop- 1970s. USAID experience was positive but after
ment assistance support for the plant sciences, emphasizing such projects in many countries for
which presumably is a fraction of research. about a decade, they fell out of favour (Kumar,
1987). Such projects achieved roughly the same
The effect of agricultural development assistance rate of ‘success’ in World Bank evaluations as
irrigation projects. In 1993 the World Bank’s data
The effect of this development assistance varies indicated an overall success rate of 49% for such
across sub-sectors of agriculture. Irrigation and area development projects (World Bank Opera-
drainage projects were the largest sub-sector for tions Evaluation Department, 1993). On average
thirty years through the mid-1980s and evaluation they generated a 10.4% economic rate of return,
indicates aid for irrigation was usually effective. with just over half giving an economic rate of re-
USAID evaluations of irrigation projects showed turn over 10% (the other half characterized as
that while many problems had to be overcome, the ‘failures’ because they produced below 10%).
results encouraged continuing investment. For Failures in area development projects were most
example, a report that summarized the agency’s frequent in Eastern and Southern Africa. Area
30 year experience, including evaluations of AID’s development projects went out of favour in the
projects in Sudan, Senegal, Egypt, Morocco, 1980s but recently have reemerged in the form of
Turkey, Pakistan, Korea, The Philippines, and participatory rural development and poverty alle-
Indonesia avoided any quantitative generalizations viation work.
about the rates of return to the aid investments but Projects to provide subsidized credit and build
indicated that while there was evidence that irri- agricultural cooperatives comprised the third
gation’s contribution to rice yields accounts for largest proportion of development assistance to
about 30% of the factors involved in the Philip- agriculture – over 10% of the development assis-
pines, it is dangerous to generalize about the re- tance portfolio in the 1970s and 1980s. A summary
turns for other areas or other crops (Steinberg view of experienced analysts based on many eval-
et al., 1983). uations of such projects found that despite the
A 1995 review of irrigation project evaluation by optimistic expectations of their sponsors, the re-
the World Bank focused on 208 Bank-funded sults of such programmes were disappointing.
irrigation projects. Evaluations rated 67% satis- Loan-default problems were serious, poor farmers
factory, comparable to the average of 65% for all remained unable to obtain loans, and those who
Bank-supported agriculture projects but worse did get credit were often unnecessarily and ineq-
than the average of 76% for all Bank projects uitably subsidized. Many agricultural banks and
(World Bank Operations Evaluation Department, other specialized formal lenders serving rural areas
1995). A later review of the Bank’s strategy for were floundering as a result of the requirements of
water management summarized results for 336 the programmes and as a result often limited the
World Bank water projects completed from 1988 range of services they provide (Adams et al., 1984;
through to 1999 and indicated that their perfor- Meyer and Nagarajan, 1996). Credit projects lost
mance was below the Bank average, based on the favour in the late 1980s and 1990s and currently
assessment of project results along three related make up less than 3% of the agricultural assistance
dimensions – outcome, institutional development portfolio.
impact, and sustainability of project benefits (Pit- Assistance to agricultural research absorbed
man, 2002). Just over 40% had satisfactory ratings around 4% of agricultural development assistance
in 1988; that increased to 53% by 1996. By the over the past 25 years. Many analytical estimates
1990s the World Bank considered water projects as of the economic rates of return to agricultural
part of the social support system rather than as research have been made and, contrary to the
 79

conclusion reached for other kinds of agricultural research on their ‘most important’ unsolved
assistance, over 95% of the studies show sub- problems; but if researchers have no way to
stantial positive economic return on investments address the unsolved problems, there can be no
(Alston et al., 2000). Careful examination of nearly effective research. On the other hand if
300 studies reporting over 1800 individual rates of researchers discover something for which users
return indicate no support for the idea that returns have no need, it is of no value (although it adds
have fallen over time, but rather that returns vary to knowledge and may be valuable ‘basic’ re-
in other ways that make intuitive sense. In par- search).
ticular, research on commodities with longer pro- Having users decide seems eminently reason-
duction cycles like livestock and more diffuse able, but in the case of publicly funded research,
effects like natural resource management have becomes circular as the bureaucracy involved in
lower rates of return. Overall, the median rate of directly funding research seeks the optimal allo-
return to agricultural research investments is cation by appealing to both the users and doers of
nearly 50% and the median rate of return to re- research. Dalrymple (2005) provides a useful dis-
search and extension combined is nearly 40%. cussion distinguishing between researchers who
Studies examining the relationship of agricultural provide the supply of scientific goods and the users
growth to research, education, roads, and other who represent the demand for such goods. The
important factors in India and China reinforce the best approach would take both positions into ac-
importance of research for growth (Fan et al., count, perhaps through a process something like
1999, 2002; Rosegrant and Evenson, 1992). Much that reflected in Figure 1.
attention has been focused on variety develop- Figure 1 identifies a ‘political-bureaucratic str-
ment but it is clear that pathology, entomology, ucture’ that interprets the latent demand for
epidemiology and other plant sciences play an innovations generated by farmers, consumers,
important role in the development of new crop processors and other actors in the ‘ socio-economic
varieties. structure.’ This political-bureaucratic structure
might also be characterized as a decision-maker
who generates the actual demand for innovations.
Establishing priorities for plant science research
This structure distributes funds to the ‘innovation-
producing institutions’ that pay researchers to
Plant scientists interested in contributing to food
conduct research and thereby generate the supply
security in the developing countries face a simple
of innovations. As those innovations are used by
question: ‘What should we do?’ International
the socio-economic structure they generate actual
assistance can most effectively address research
payoff. The supply and demand analogy has some
questions while control of epidemics is largely the
appeal, but even the elaborated view depicted in
responsibility of national authorities and except
Figure 1 breaks down because there is no equili-
for certain critical pests like desert locust. In con-
brating price mechanism for publicly funded re-
trast, research has been one of the most effective
search so supply and demand are not the right
areas of development assistance. The question of
terms. Nonetheless, it seems clear that the two
how assistance might be allocated to various re-
aspects – what new knowledge is needed by users
search options in the plant sciences is the subject of
and what can be done – should be considered in
the balance of this paper.
setting research priorities. In the procedure out-
The question of ‘how’ to allocate research re-
lined here, both are. Five steps are required to
sources is difficult to separate from ‘who’ should
produce an answer to the question of how re-
allocate them and there are two views on who
sources should be allocated:
should set priorities. One holds that priorities
should be set by those who do the research while 1. Determine the objective,
the other holds that priorities should be set by 2. Identify alternatives, assemble data for each,
those who benefit from it or by those who pay for 3. Choose a method for setting priorities,
it. But ‘who’ largely implies ‘how.’ If researchers 4. Establish priorities among the alternatives,
decide, they will favour what they believe they can 5. Allocate available resources among alterna-
most effectively do. If users decide they will favour tives.
80

Perceived Expected
Socioeconomic Structure payoff
latent supply of
Government policies, factor and (ex ante)
innovations
product prices and markets,
access to institutions (credit,
information, education), state of Actual
technology, etc. payoff
Actual (ex post)
supply of
innovations

Innovation-Producing
Institutions
Public research institutes; public
interaction with private sector;
basic and applied research.

Politico-Bureaucratic Structure
Effective (actual) Social pressure system; Latent
demand for electoral-bureaucratic reward demand for
innovations system; appropriation-legislative innovations
system.

Figure 1. Generalized model of the supply and demand for technological innovation in the public sector.
Source: Adapted by Dalrymple (2005) from: Alain de Janvry, ‘‘Social Structure and Biased Technical Change in Argentine Agri-
culture,’’ in Hans Binswanger and Vernon Ruttan (eds.), Induced Innovation: Technology, Institutions and Development. Johns Hop-
kins University Press, Baltimore, 1978, pp. 301–303. Original referred to both technological and institutional innovation.

Determine the objective people should be the primary objective. One might
accommodate this concern by weighting the in-
The objective of the whole exercise is presumed to come of the poor more heavily in setting priorities,
be to put available plant science research resources or consider only the farm income of the poorest
to their ‘best use.’ That is, to optimize the use of farmers, ignoring the income of others. If so, it is
those resources in generating ‘something’ – but, important to define who the poor are. Some ana-
exactly what is the something? One option might lysts focus on the one-fifth of the population with
be to maximize added output and loss prevented for the lowest incomes – the low income quintile. But
all crops. But crops are different, so it is not is this the low income quintile in each country or in
appropriate to simply add together the prevented the developing world as a whole? An alternative is
losses of grains, vegetables, coffee, and cotton. It is to consider the contribution to equity – that is the
easy enough to aggregate different crops by valu- income of the poor relative to the wealthy. This is
ing each agricultural product at its price and sometimes done by considering the ratio of in-
adding to get total value of agricultural output. comes of the lowest income quintile to that of the
However, production of different crops requires highest income quintile. The Gini coefficient is a
different inputs. For example, it takes much more measure of equity that reflects the relative income
capital to produce a ton of wine grapes than a ton of all units in the population, not just the highest
of rice; more to produce a ton of coffee than a ton and lowest quintiles. However, it is seldom prac-
of lentils. Such differences in costs suggests the tical to use the Gini coefficient because of the
objective to be maximized might be the value difficulty in obtaining the data to calculate it.
remaining after subtracting the cost of inputs, in One might prefer to focus directly on the con-
other words the net value of farm output or net tribution of research to nutritional adequacy of the
farm income. poor. Like the income of the poorest quintile, this
Some advocate that publicly-supported research avoids data problems associated with measuring
should have a special focus on the poor, arguing equity but introduces complications associated
that maximizing net farm income of the poorest with defining nutritional adequacy. How can
 81

improvements in calories be aggregated with gains considered together. An omitted option cannot
in vitamins or increased protein intake? An index simply be added later because all interact and,
of contribution to nutrition might be devised but depending on relationships, an allocation to a new
in reality the contribution of any particular food to option does not necessarily reduce all others in the
any individual’s nutrition depends on that indi- same proportion. The scope of alternatives will
vidual’s current nutritional status, and, using such depend on how broadly one defines the problem.
a set of weights for aggregating different nutrients For example, if plant science is taken to include
is no less arbitrary than applying a set of prices to the economics of plant protection, such matters
aggregate across commodities and involves many must be included as alternatives. If sociology re-
more computational steps. search on movements to ban chemicals in favour
These complexities and others lead to using of green agriculture is an alternative that might be
monetary terms to value the productivity of re- funded by the decision maker, then research on
search, which are in any case needed to account for such topics must be on the list. If plant breeding
input costs. It introduces the challenge of defining for genetic resistance is an option, it must be in-
the price for each commodity and input. Com- cluded. Whether such topics should be included in
modity prices are different in every territory and allocating research resources for plant sciences is a
most fluctuate on a day-to-day basis and over prior decision. Here we make the assumption that
longer periods. Surprisingly, there is no readily the universe of alternatives can be defined along
applicable set of international prices by which to the dimensions of: pest or causal organism, crops/
value agricultural commodities. The World Bank host plants, locations, tasks and approaches.
tabulates monthly prices for rice, maize, wheat,
soybeans, rubber, sugar, but not for all agricul- Manageable interest
tural commodities. In abstract terms, there is almost no limit to the
Another issue is which price along the market- alternatives that might be considered. In practical
ing chain should be used for aggregation. Price to terms, however, one should restrict the alternatives
producers differs from price to consumers by the considered to the set over which the decision-ma-
amount of marketing costs. Marketing costs are ker has a manageable interest. A manageable
likely to be relatively similar among the grains but interest is the set of issues over which a decision-
marketing costs for perishable fruits differ signifi- maker can make and implement a decision. In
cantly from those for grains. Low income con- other words, a decision-maker with responsibility
sumers may have different relative values for for one province has a manageable interest in
grains and fruits than high income consumers. A alternatives for that province and should restrict
commodity’s value in one country may differ from considerations to alternatives within the province,
its value in another. Finally, the poorest consum- while a decision maker with responsibility for a
ers use a high proportion of their incomes simply nation must deal with all alternatives relevant for
to obtain food so the purchase price of food is an that nation. Likewise, a decision-maker responsi-
important factor in the real incomes of the poor. ble for cereals should deal only with alternatives
For the purpose of this discussion it is assumed relevant for cereals, while one with responsibility
that the issues identified above are resolved and for all crops has a much larger set of alternatives.
there is agreement among the stakeholders and In recent times agricultural research decision-
decision maker that the objective to be maximized makers have become more attuned to the views
is the contribution of each alternative to the real net of a broad range of people and groups who
income of the lowest income quintile in the least express interests in food-related matters because
developed countries. For convenience, call this the of their interests as consumers or simply as
real income of the poor. members of civil society. These groups, together
with farmers, food processors, researchers, tax-
Identify alternatives payers, research organizations and others are
considered as ‘stakeholders’ in the decisions
It is essential to begin with a comprehensive list of made about the allocation of public resources
research alternatives. The allocation process re- and decision-makers often seek ways to incor-
quires that similar information on all options be porate stakeholder views into both the definition
82

of alternatives and in setting priorities among are to be used in that way. For our discussion the
the alternatives. number of territories is called: G.

Plant hosts Possible research activities

One dimension defining the universe is the set of Contemporary efforts to understand the challenges
crops or plant hosts that will be included. It is plant diseases pose to the global food supply
presumed that the interest is with plants of eco- roughly follow the above approach of identifying
nomic importance, but this covers a wide range. the gains (and losses prevented) from controlling
Even limiting consideration to agricultural crops is specified sources of loss on specified plants in
challenging because there are many ‘minor’ crops specified countries. For example, the objective of
that are of importance in some particular situa- one ambitious study on the subject reports the scale
tions. A recent global effort to define plants of of losses caused by plant pathogens, animal pests
international agricultural importance resulted in a and weeds on eight crops in seven global regions
list of 64 species (FAO, 2005b). For this exercise (Oerke et al., 1994). It seems appropriate to follow
the number of host plants is called: H. this lead and define research activities through the
target intersections of causal organism, plant host
Organisms included and location. For convenience we call the inter-
A second dimension is the set of pest and disease sections ‘research tasks,’ and their number is: R.
causal organisms to be included. That is, are all Hence:
plant diseases to be included – bacterial, viral, R¼NHG ð1Þ
fungal and idiopathic? What about nematodes?
Will vectors of all diseases be included or only Diseases are controlled through host resistance,
vectors of major diseases? If vectors carry human or pesticides and cultural practices, but all three are
animal diseases as well as plant diseases, will re- probably involved in most successful control sys-
search on those animal diseases be included as part tems. Each of the technological control ap-
of the allocation problem? Will priorities be defined proaches may entail distinctly different activities.
strictly for plant diseases or will non-vector insects For example, host resistance may be pursued
and weeds be included? In reality it is difficult to through conventional plant breeding or through
separate out these causal organisms, especially be- genetic engineering and may be polygenic or
cause when a new epidemic breaks out the causal monogenic (Sorho et al., 2005). Biological control
organisms are largely unknown and in many cases a may be pursued using native or exotic organisms.
single event has multiple causes. For convenience of The technology for each approach requires quite
discussion call the number of causal organisms: N. different resources and, most effective control en-
tails several approaches. The number of such
technologies is called: T.
Geography
The total possible number of research activities,
The third dimension is geographic: over what set
A, is then:
of agroecologies, countries or territories are the
allocations to be made? Assuming an interest in
developing countries, are all developing countries A¼TR or: A¼TNHG
to be covered? The World Bank defines least ð2Þ
developed, low-income and middle-income devel-
oping countries. Should only countries with a de- A useful notation is to allow each of the elements
fined minimum amount of crop land be included? T, N, H and G assume the form of a subscript that
Should the former Eastern Bloc countries in- runs from 1 to t; 1 to n; 1 to h and 1 to g. Then any
cluded? Given the importance of climate in plant individual research activity can be designated as A
diseases, one might argue that it makes most sense with the appropriate subscripts, or in general as:
to use agroecological regions. Logical though it is, Atnhg
the problem introduced by this is that most data
are available for political regions and must be The allocation problem is: to determine the pri-
transformed into agroecological categories if they orities for research among all possible research
 83

activities, that is, among all possible intersections make to several dimensions and then aggregating
of causal organisms, host plants, geographies, and those contributions. For example, individual
technologies. To get some idea of the magnitude of scores could be assigned to dimensions like output,
the allocation task, suppose that for the whole equity, geographic distribution, women’s income,
developing world, there are 50 major causal food security or others, and those scores aggre-
organisms, 25 plant hosts, 10 territories and 5 gated. The aggregation may be through simple
technologies, then there will be 62,500 research addition or alternatively through weighted aggre-
activities among which to allocate resources. This gation. Table 2 shows how such a system might
appears to be an overwhelming task, but of course, work.
some combinations will be ‘empty’ and others will The first section shows the scores for two re-
most efficiently be combined into one activity search activities; one is high on women’s income
thereby reducing the number of alternatives. Still, and low on output and food security; the second
the number will be large, requiring a systematic alternative scores high on output and food security
procedure for organizing all the applicable infor- and the same as the first alternative on the other
mation. three characteristics. The aggregate is the simple
average or the aggregated value using equal
Priority setting methods weights. The second section illustrates the effect of
differential weights where equity and women’s in-
Three broadly different methods have been used to come have higher weights and other characteristics
set priorities among research alternatives: scoring, have lower weights. With the unequal weights the
congruence and benefit:cost approaches (Norton aggregate score of the first research alternative is
and Pardey, 1987). Each has a number of varia- much closer to the second, reflecting the greater
tions. weights given for two of the characteristics. In this
system both weights and the scores each contribute
Scoring approaches to the aggregate score.
The simplest possible approach is to group alter- Any number of variations of scoring approaches
natives in priority categories such as high, medium can be devised. For example one might use data on
and low or rate each alternative on a one to five, production in geographic regions to score the
one to ten, or some other numerical scale that di- geographic characteristic and value of output to
rectly indicates priority by the score of each score the output variable. A number of possible
alternative. More challenging is ranking alterna- weighting schemes may be devised; and a large
tives numerically from the ‘most important’ to the number of different characteristics may be used as
‘least important,’ giving each alternative a unique weights. The weights can be determined in a sep-
numerical rank indicating its priority. arate exercise from the scores so stakeholders can
Often people are not comfortable with a single be involved where they have special knowledge or
number because they believe either there is ‘no interests (e.g. in the scores) without completely
basis’ for making such a judgment or there are determining the outcome.
several different dimensions to alternatives that The same versatility that permits the intro-
would rate differently. They prefer scoring or rat- duction of many characteristics and variations
ing the contribution each alternative is expected to on weighting is one of the limitations of scoring

Table 2. Illustration of alternative scoring approaches

Activity Output Equity Geo-graphic Women’s income Food security Aggregate score

Simple average
Score 1 2 2.5 3 4 2 2.7
Score 2 4 2.5 3 4 4 3.5
Weighted
Weight 0.1 0.3 0.1 0.4 0.1
Score 1 2 2.5 3 4 2 3.05
Score 2 4 2.5 3 4 4 3.45
84

approaches. One must be careful not to design the of the A intersections defined in (1) above. Intui-
weights to skew the results in a particular direction tively, crop losses are the amount of crop lost to
and recognize that the greater the number of various pests or because production factors are
weighting characteristics, the more difficult it is to used at less than maximum output levels. Here the
trace the links between characteristics, weights and focus is on losses from pathogens and pests. As
aggregate score (Alston et al., 1995). with most seemingly simple concepts, complexities
lie below the surface, as the literature on crop
Congruence-based allocation losses makes clear. Figure 2 illustrates this point.
Another approach is based on the view that re- For any crop a physiologically defined theoret-
search resources ought to be allocated in propor- ical yield potential can be associated with any
tion to the ‘importance’ of each activity as genotype and climate regime, unimpeded by limi-
reflected in the value of crop production or, in the tations of water, nutrients and pests. In any
case of plant protection, the value of crop losses practical situation there is some unavoidable crop
attributed to various problems. This approach is loss, given the impossibility of controlling all fac-
known as the ‘congruence’ method of allocating tors that lead to losses. This defines an attainable
research resources. yield. In general, attaining that yield requires
Using Y to represent production loss prevented expenditures on inputs or control measures in ex-
or yield increase obtained, for every Atnhg there is a cess of the profitable levels and so one can define
corresponding Ytnh. A critical relationship is the economically non-recoverable loss and hence an
contribution each research activity (Atnhg) is ex- economic yield. That is the level one would expect
pected to make to increasing Ytnh and in turn, the to observe if all farmers applied all crop loss
contribution that increased output makes to the control measures at the economically optimal le-
incomes of the poorest quintile. Implementing this vel, but generally the actual yield is somewhat
approach requires information like: applying $X below that level. The actual yield reflects the yield
to Atnhg over a period of Y years will prevent losses response to crop protection actually used that is,
or increase production by Ytnh and raise real in- the prevented loss. Still lower, assuming some
come of the poor by $Ztnh per year over sub- effectiveness of current practices is the yield with-
sequent years. out crop protection. The distance between these
An obvious starting point is to know the Ytnh – differently defined yields reflect the various loss
the yield loss or potential yield increase – for each concepts.

Theoretical yield potential

unavoidable crop loss

Attainable yield

economically non-recoverable loss

Economic yield

potential actual loss

loss
Actual Yield

Prevented loss
(response to
Yield without protection protection)

Zero yield

Figure 2. Conceptual model for crop loss assessments (adapted from Oerke et al., 1994 modification of Zadoks and Schein, 1979).
 85

Of course, except for actual yield, all the yield diversity of clientele and the complexity of the
levels and losses identified above are concepts that systems necessary to integrate the number and
cannot be observed under production agriculture. diversity of client views make such approaches
But the concepts are so intuitively helpful that inherently difficult to structure. These difficulties
some broadly accepted conventions have been could, of course, be overcome and estimates of the
developed that permit estimates to be generated. relative importance of conducting research in each
In studies of crop losses, ‘attainable yield’ is de- of the Atnhg could be generated using participatory
fined or computed using crop growth models methods. In practice, more efforts seem to have
taking into account the climate, water availability, been devoted toward participatory research than
yield potential of varieties grown, rates of fertilizer toward participatory priority setting.
application and other cultural techniques like
seedbed preparation and crop density (Oerke Estimated crop losses
et al., 1994). The difference between the estimated An important stimulus to crop loss measurement
attainable yield and actual yield provides an esti- was given by several major symposia on the sub-
mate of ‘actual loss’ attributed to pests and ject in the last century. The first was organized in
pathogens. It is also possible to estimate the the 1960s by the Food and Agriculture Organiza-
‘prevented loss’ from knowledge about the plant tion (FAO, 1967) and a second took place in the
protection measures used and their effectiveness. 1970s in honour of E.C. Stakman (Teng and
In many cases the results from plant protection Krupa, 1980). A more recent study of crop losses
experiments are used in making such estimates. (Oerke et al., 1994) provides access to a large
An alternative, more participatory approach to amount of systematically organized information.
using crop modelling and experiments is to draw This work was designed to stimulate research on
on the knowledge of farmers who are producing the causes of losses, improve methods to protect
crops in the areas of interest. Clearly, those who crops, enhance the effectiveness of control meth-
make their livelihoods through farming have an ods, integrate plant protection with other man-
interest in anything that reduces yields and they agement practices to optimize methods of crop
would seem an important resource for identifying production, and help generate support for research
yield losses and potentials for increases. As intui- on effective crop protection. Table 3 provides a
tively appealing as this is, a review of the literature summary.
reporting such activities identifies at least six lim- The immense amount of work and detailed,
itations (Dalrymple, 2005). First, farmers are likely country-by-country, crop by crop information that
to be highly influenced by their immediate and lies behind the table cannot be overstated. Based
highly visible problems and are likely to have a on that work, global crop losses are estimated to
short-term outlook and be less concerned with or be about 75% as large as actual production, with
aware of the opportunities offered by longer-term the losses almost equally attributed to pathogens,
research. Second, farmers are more likely to favour pests and weeds. The lowest estimated losses,
research that generates benefits they receive rather about 30%, are in Europe and North America,
than broadly-adapted research that generates while Africa and Asia each are estimated to lose
product price reductions and benefits to consum- nearly 50% of their attainable production. Nearly
ers. Third, in most developing nations elites 60% of global losses occur in Asia, far more than
dominate and will naturally direct attention to in any other region. This is because Asia produces
research that favours them over less powerful nearly half of global agricultural production and
groups. Fourth, those who favour participatory has a higher rate of loss than other regions.
approaches in setting priorities generally ignore Using a simple congruence approach to set pri-
consumers and consider only farmer participation, orities based on these data would suggest that 60%
despite the evidence showing that consumers are of research resources should be allocated to pre-
the main beneficiaries of much agricultural re- venting losses in Asia and within that allocation,
search. Fifth, as the geographic scope of the allo- the resources allocated to pathogens, animal pests
cation exercise is enlarged it becomes increasingly and weeds should be in the ratio of approximately
difficult to get a comprehensive and unbiased view 14:18:14, the proportion of loss to the three main
from farmers and consumers. Finally, the wide causal agents. The balance of available resources
86

Table 3. Estimates of crop losses, in financial terms (US$), occurring during the production of the eight principal food and cash
crops in the years 1988–1990, by continent

Continent Actual Loss (%) of production Loss, overall

production due to

$ bn. % Pathogen Animal pests Weeds % $ bn. % of global total

Africa 13.3 4.0 15.6 16.7 16.6 48.9 12.8 5.3

N. America 50.5 15.1 9.6 10.2 11.4 31.2 23.0 9.4
L. America 30.7 9.2 13.5 14.4 13.4 41.3 21.8 8.9
Asia 162.9 48.6 14.2 18.7 14.2 47.1 145.3 59.6
Europe 42.6 12.7 9.8 10.2 8.3 28.2 16.8 6.9
U.S.S.R. 31.9 9.5 15.1 12.9 12.9 40.9 22.1 9.1
Oceania 3.3 1.0 15.2 10.7 10.3 36.2 1.9 0.8
Total 335.2 100.0 13.3 15.6 13.1 42.1 243.7 100.0

Source: Oerke et al. (1994, p. 749); final column added by the author of this paper. Prices used in valuing production (by Oerke et al.)
were: wheat: US$ 136.2/t; rice: US$ 209.1/t; barley: US$ 79.5/t; maize: US$ 98.1/t; potatoes: US$ 128.7/t; soybeans: US$ 236.1/t;
cotton: US$ 490.6/t; coffee: US$ 1934.4/t.

would be allocated similarly, following the pro- findings or ‘solutions’ that prevent loss, is defined
portion of losses in each region to each source. as the research input–output function. The input–
One can argue that congruence with crop losses output function for each Atnhg should reflect the
is a good approach if dollars spent on every Atnhg difficulty and time required to find a solution
have the same effectiveness in increasing the in- through that activity; input–output functions for
comes of the poor (given that objective). This is different activities will reflect differences in the
likely to hold if no Atnhg are ‘harder’ or ‘easier’ difficulty or time needed for various Atnhg. The
than others so that a given amount of money spent input–output function provides an estimate of
on each would make the same contribution to the what research may actually contribute towards the
value of each crop and that each commodity objective while crop loss estimates represent the
makes the same contribution to real incomes of the opportunity for research to contribute – these are
poor. But, the relationship between research input the two key factors: what can be done and what is
and loss prevented is likely to be complex – some needed. In Dalrymple’s terms, the input–output
research challenges are harder than others. The function reflects the supply of research findings
research continuum from basic through applied to while increased real incomes of the poor from the
adaptive implies that a higher degree of uncer- crop loss thereby prevented reflect the demand for
tainty is associated with ‘more basic’ research research findings.
activities that are likely to take longer. ‘More dif- To illustrate: a set of pesticides can be screened
ficult’ research problems are likely to require more for their effectiveness against a particular pest in a
resources and time to generate usable results but relatively few growing cycles, say in a matter of 2–
are also likely to have the potential to generate 5 years. If the pesticide has been approved, a
higher returns. The contribution of losses pre- control practice can be recommended to farmers
vented for a commodity important in the con- shortly thereafter. An alternative approach, the
sumption of the poor or in generating income for development of cultivars with genetic resistance to
the poor is more important for the objective than the pest, may take 6–10 years from the beginning
for a commodity not important to the poor. of research to release to farmers. Even if the two
Technologies not well-suited for adoption by the activities give the same yield effect and remain
poor would contribute less than those especially effective for the same period, they have different
well-suited. input–output functions. Some kinds of research
may have a greater inherent requirement for inputs
Input–output function of Atnhg such as laboratory equipment, experimental fields
The relationship between research input, expressed and labour; costs of land, labour and capital vary
in researcher time and funds, and the expected across locations and other factors affect the cost of
 87

any particular research activity, but these differ- different approach expected to be less successful in
ences can be incorporated in an input–output preventing the yield losses; B2 is expected to take
function. less time than B1 but the latter will ultimately
A number of things can be inferred about the prevent more of the losses, although not as much
relationship between research input and expected as the approach used in A1 and A2. Intuitively,
prevented yield loss or expected output (Ytnhg). such ideas seem consistent with the ways scientists
First, at the beginning of the process and with zero think about research alternatives and incorporate
input, expected prevented loss is zero. Second, no more information in the allocation process than
matter how great the resources or time taken, there the congruence approach.
is some maximum value of expected Ytnhg
depending on actual losses or yield potential for Benefit:cost
each Atnhg,. Third, at the beginning of the research In addition to incorporating the effects of different
with small inputs the probability of finding a input–output relationships, benefit:cost ap-
successful ‘solution’ is small and therefore the ex- proaches can incorporate variations in resource
pected prevented Ytnhg is small. The expected value use, time lags and uncertainty into priority setting.
likely increases slowly until some critical minimum To illustrate, the following example may be help-
amount of resources are applied and at that point ful. Suppose two different research activities could
increases rapidly over some range of research in- be targeted at preventing a $900,000 annual crop
puts. Beyond some level of resources the expected loss. Suppose the first, Atnh1, costs $50,000 a year,
Ytnhg is likely to begin to increase at a declining will be completed in 5 years and is expected to
rate. prevent half the potential loss while Atnh2 costs
Figure 3 illustrates a few of the many possible $25,000 a year, will take 10 years and is expected
input:output relationships consistent with the to prevent 80% of the potential loss after 10 years.
inferences stated above. Each curve portrays the In both cases the technologies are assumed to re-
relationship for a different research activity de- main effective for 10 years after introduction.
signed to prevent losses experienced in a particular They are illustrated in Table 4.
intersection defined by equation (1). Research in- The first line in the Table shows the research
put consists of money, people and time reflected cost per year (all numbers in ‘000). For simplicity
on the horizontal axis as cost per year. Curves A1 cost is assumed to be constant for a defined
and A2 use the same research approach but with number of years, but that assumption is easy to
more resources applied each year in the case of A1, relax. The research is aimed at preventing the po-
so the solution is expected to be found sooner. tential loss depicted in the second line. The percent
Both are expected to generate knowledge to pre- expected prevented loss shown in the third line is
vent the entire loss. Curves B1 and B2 represent a the concept introduced in Figure 3, in percentage

% expected
prevented
loss

actual loss A1 A2

B1
A1: Process 1, high input/yr
A2: Process 1, low input/yr
B2 B1: Process 2, high input/yr
B2: Process 2, low input/yr
a1…b2: optimal input levels

b2 a1 b1 a2
$/yr research input

Figure 3. Hypothetical research input/output relationships.

88

Table 4. Illustration of the calculation of present value of net benefits of two research alternatives at a discount rate of 10%

Year 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

Alternative 1
1. Cost 50 50 50 50 50 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
2. Potential loss 900 900 900 900 900 900 900 900 900 900 900 900 900 900 900 900 900 900 900 900
3. % exp. Prev loss 0 0 0 0 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0 0 0 0 0
4. Loss prevented 0 0 0 0 450 450 450 450 450 450 450 450 450 450 450 0 0 0 0 0
5. Adoption % 0 0 0 0 0 0.3 0.6 0.9 0.9 0.9 0.9 0.9 0.9 0.9 0.9
6. Expected LP 0 0 0 0 0 135 270 405 405 405 405 405 405 405 405 0 0 0 0 0
7. Exp. net benefit )50 )50 )50 )50 )50 135 270 405 405 405 405 405 405 405 405 0 0 0 0 0
8. PV of net benefit )45 )41 )38 )34 )31 76 139 189 172 156 142 129 117 107 97 0 0 0 0 0
NPV 1134
Alternative 2
1. Cost 25 25 25 25 25 25 25 25 25 25 0 0 0 0 0 0 0 0 0 0
2. Potential loss 900 900 900 900 900 900 900 900 900 900 900 900 900 900 900 900 900 900 900 900
3. % exp. Prev loss 0 0 0 0 0 0 0 0 0 0.8 0.8 0.8 0.8 0.8 0.8 0.8 0.8 0.8 0.8 0.8
4. Loss prevented 0 0 0 0 0 0 0 0 0 720 720 720 720 720 720 720 720 720 720 720
5. Adoption % 0 0 0 0 0 0 0 0 0 0 0.3 0.6 0.9 0.9 0.9 0.9 0.9 0.9 0.9 0.9
6. Expected LP 0 0 0 0 0 0 0 0 0 0 216 432 648 648 648 648 648 648 648 648
7. Exp. net benefit )25 )25 )25 )25 )25 )25 )25 )25 )25 )25 216 432 648 648 648 648 648 648 648 648
8. PV of net benefit )23 )21 )19 )17 )16 )14 )13 )12 )11 )10 76 138 188 171 155 141 128 117 106 96
NPV 1161

terms for ease of computation. For a greater re- ogy is assumed to be obsolete and no longer
search cost the expected prevented loss will be prevents yield losses (one could, of course, have a
achieved sooner and possibly to a greater extent so technology with slowly declining benefits). Line 8
that for every different research input:output shows the ‘present value’ of expected net benefits,
relationship the values of the first and the third a concept designed to reflect the idea that pre-
lines of the table will be different. The fourth line is venting losses sooner rather than later is better and
the quantity of loss prevented (LP), the product of that money spent today in preventing such losses is
lines 2 and 3. The time path of expected farmer worth more than money spent for the same pur-
adoption is incorporated in line 5 of the analysis pose in the future. The ‘discount rate’ reflects the
and reflects the percent of line 4 that is realized relative value of solutions today compared to
each year. If adoption of the results of various solutions in the future. If it makes no difference
Atnhg takes different time paths, incorporating the when the solution is obtained the discount rate is
adoption lag will differentially affect the benefits zero, but in most cases, people would prefer
obtained in any year. The sixth line is the product solutions sooner rather than later and the stronger
of lines 4 and 5 and reflects the expected amount of that desire, the higher the discount rate. Of course,
loss prevented each year. Multiplying this line by all the relationships and parameters in the table
the output price would give value but for sim- are illustrative and the result of simplifying
plicity here we assume a price of 1 per unit of assumptions.
output. The seventh line is line 6 minus line 1, the The top part of the table illustrates the early
value of expected net benefit in the year in which return case: $50,000 is expended each year for
benefits are expected to be obtained. Net benefits 5 years, at which point the expected results are
are, of course, negative in years 1 through 5 until obtained. In years 6 through 15 the loss prevented
results are obtained; zero in year six when the re- could be $450,000 but because adoption takes time
search is completed but the results have not yet the expected loss prevented is as shown in line 6.
been adopted; and are assumed to reach 30%, The expected net benefit, the difference between
60% and finally 90% over the next three years as cost and expected LP, is equal to cost in the first
adoption occurs. Thereafter net benefits are con- 5 years; afterwards costs go to zero and net benefit
stant for the next 6 years after which the technol- is the loss prevented. The values in line 7 are
 89

‘discounted’ at 10% to get the present value (PV) global long term downward trend in grain prices
in line 8 and those values are added together to get has been ascribed to the success of research in
their sum for the entire period, called the net increasing the productivity of grain production in
present value or NPV. In the bottom panel alter- many locations throughout the world. In contrast,
native 2 is shown. In terms of Figure 3, alternative the prices of food legumes show no such long term
1 might be represented by a curve like B2 while decline, in part because there have been relatively
alternative 2 is more like curve B1. modest productivity gains.
In the numerical example the NPV of the first By incorporating appropriate assumptions
alternative is 1134 and the NPV of the second is about the way consumers respond to additional
1161. Thus, even though the first alternative has a supplies (demand elasticities) it is possible to esti-
shorter time to solution (5 vs. 10 years), the second mate the impact of a productivity gain on prices
has a slightly greater present economic value, in and incorporate that into the estimates of benefits
part because it is lower cost (25 vs. 50) and in part and costs. In addition it is possible to calculate
because it prevents a greater proportion of the how much of the productivity gain remains in the
potential loss (80% vs. 50%). hands of producers and how much goes to con-
Each of the numbers that goes into the compu- sumers. Economists call these the producers’ sur-
tation has an impact on the NPV. For example, if plus and consumers’ surplus and commonly use
the discount rate is 5% rather than 10%, the NPV such concepts in estimating the benefits from
of the first alternative is 1936 and of the second is technological change (Alston et al., 2000); they can
2506 – with a lower discount rate the NPV of the also be used in research resource allocation.
first alternative increased by 70% and that of the An additional elaboration has been developed
second by 115%, illustrating the differential effect to accommodate the idea that it is difficult to give a
discount rate has on more distant income. In the point estimate of the likelihood that any particular
extreme case, if future costs and benefits are not research activity will be successful. This incorpo-
discounted then the NPV is simply the sum of the rates a ‘triangular distribution’ into the input–
stream of expected net benefits; in this case 3395 output function using estimates of the maximum
and 5582. likelihood of success, the minimum and the most
With the discount rate at 10%, if the time to likely probability of success (Mills, 1998). These
success and adoption are shortened by one year numbers are then aggregated into a single one used
and obsolescence still occurs after 10 years, then as the ‘probability of success’ in the table.
the NPV of the first alternative is 1266 rather than
1134. If the solution using the second alternative is Allocate resources among alternatives
found after 8 years rather than 10 and adoption
and obsolesce patterns are unchanged, its NPV is As illustrated, priorities can be established in sev-
1543 rather than 1161. If, in the second alternative, eral ways, from categories to rankings to benefit:
costs increase to 30 in year two, 35 in year 3, 40 in cost (with or without considering economic sur-
year 4, 45 in year 5 and 50 in year 6 and beyond, plus) to subjective scores. Regardless of the
NPV becomes 1068 rather than 1161. Hence, dif- method the result will be a set of numbers repre-
ferent time paths to success or adoption or costs senting the priority of each Atnhg. However, no
generate different patterns of returns and higher or matter which approach is used, that set of numbers
lower NPV. While the approach seems complex does not imply any specific allocation of resources.
and requires the specification of numerical values Any of the sets of numbers could be used to allo-
to concepts that normally are little more than cate resources proportionately, but each would be
‘hunches’ of scientists, it has successfully been arbitrary, given what is recognized about the re-
applied to help guide resource allocation in a search input–output function. Alternatively, the
$110 million programme (Herdt, 1991). numbers generated can be interpreted as a ranking
of importance if a technique is consistently ap-
Elaborations to benefit:cost plied, however such a ranking still does not
Increases in output that are large relative to cur- translate into a particular resource allocation.
rent supply may have the effect of reducing the Various options might be used to translate
price of the commodity in question. In fact, the the priority into an allocation. One extreme would
90

be to allocate all available resources to the top $50/year and $517 at $10/year. Although we
ranked activity – a position not likely to get much computed the NPV for alternative 1 only for one
sympathy. In a strict capital budgeting problem investment level in Table 4, other levels would
where the NPV is independent of the size of the generate a series of NPVs to trace out the curve
investment, funds would be allocated to the alter- shown. A similar curve of NPV vs. annual research
native with the highest NPV, then the second investment is shown for one additional alternative
highest, etc. until all funds are used up. Another and could be plotted for every possible research
option is to argue that all alternatives should get alternative.
equal allocations. This may be the most politically Such curves provide the key to solving the re-
appealing and would make much of steps three source allocation problem. Notice that each curve
through to five unnecessary! However, using some is increasing but at some point the rate of increase
concepts from economic theory, one can easily falls and the curve eventually flattens out as
show that a non-equal allocation can generate researchers run out of good ideas to investigate
higher benefits. and the work becomes less productive. In other
Economic theory shows that the optimal pattern words, the increase in NPV for higher and higher
of investment would be to invest in each alterna- increments of investment eventually declines (and
tive just the amount that provides an equal incre- may even become zero or negative). Economic
mental (in economic jargon, marginal) return to theory says that the greatest total expected gain
each alternative and at the same time uses up all will be obtained when the additional NPV from
the available resources. Applying this concept re- each alternative is equated and all resources are
quires data on the marginal return to each alter- used. This can be illustrated as follows.
native, which can be derived as follows. A larger Suppose the research manager has $150 to invest
annual investment in a particular alternative is each year among the three alternatives shown in
likely to shorten the time until success (although Figure 4. If it is invested equally, $50 in each
there is a limit to how short the time can get). On alternative, Alternative 1 generates an expected
the other hand, a smaller annual investment is NPV of 1134, alternative 2 an expected NPV of
likely to lengthen the time to success (although if 1605 and alternative 3 an expected NPV of 1175.
the annual investment gets too small the proba- The total of the three is 3914. On the other hand, if
bility of success may become zero). The following instead of equal allocation, alternative 3 gets $75/
illustrates what larger or smaller annual invest- year its NPV goes up quite a lot (the curve is steep)
ments would do for one research alternative. and if at the same time alternative 1 gets $25/year
Suppose that with a larger annual investment its NPV goes down by a lesser amount (its curve is
($50,000) the research phase can be shortened from less steep).
10 to 8 years and adoption speeded up so that 10% The changes in NPV are given in Table 5; each
of farmers adopt in the 8th year and 40% in the 9th, column shows the change in NPV from the lower
etc. As a consequence the NPV would increase from
1161 to 1605 (details not shown). On the other
hand, if the annual investment is smaller ($10,000) NPV
and the research phase is consequently stretched 2500
out to 15 years with a similar relative pattern of
2000 Alternative 2
adoption as originally, then the NPV falls to 517.
Following this procedure one can estimate the 1500 Alternative 3

NPVs that correspond to a series of different annual Alternative 1

research investments for a given research alterna- 1000
tive. These values can be plotted as in Figure 4,
500
alternative 2. In a similar way, for each research
alternative there exist a series of NPVs corre- 0
sponding to different levels of research investments. 0 10 25 50 75 100 125
$/year
Figure 4 shows NPV curves for three research
alternatives. Alternative 2 is the case we have been Figure 4. Hypothetical net present value (NPV) of three re-
following with NPV of 1161 at $25/year, 1605 at search alternatives.
 91

Table 5. Change in NPV from changing annual research in- Middle East, things were not much changed from
vestment 150 years earlier. Poverty, short lives, high rates of
hunger and low-yield, low technology agriculture
Investment 0 10 25 50 75 100 125
prevailed except for a few enclaves. A remarkable
Alternative 1 625 309 200 50 20 0 change has occurred since 1950. The world’s
Alternative 2 517 544 444 195 100 50 population has doubled but there are 150 million
Alternative 3 400 410 365 225 100 75
fewer hungry people; per capita food availability
in the developing world has increased by 20% and
investment level to the specified one. For example, the real price of food worldwide has fallen by half.
the first $10 of investment generates additional World food production has more than kept pace
expected NPV of $625 in alternative 1, $517 in with growing food demand. Still, there are far too
alternative 2 and $400 in alternative 3. Consider many poor, hungry and ill-clothed people in
$50 invested in each alternative: switching $25 developing countries, with by far the greatest
from alternative 1 to alternative 3 reduces expected proportion in sub-Saharan Africa and South Asia.
NPV from alternative 1 by $200 and increases Technology, policies and institutions designed to
expected NPV from alternative 3 by $225 – a net encourage economic growth of agriculture and en-
increase of $25. Of course, if the research decision sure the poor are included in growth are the
maker had more funds available, say $225, it important necessary conditions to overcome hun-
would be better to invest $50, $75 and $100 in ger and poverty. Far from being tradition-bound
alternatives 1, 2 and 3 respectively. The differences and resistant to change, millions of farmers in poor
are modest because the three curves are quite countries have accepted new technologies in the
similar; the more different the input–output curves form of seed varieties, fertilizers and irrigation and
are, the greater will be the difference in expected driven the rate of food production ahead of the
output from applying the economic decision rule demand for food. Experience shows that such
compared to equal allocation. technology must be carefully designed to fit the
In this example, the three alternatives (Atnh1, situations where it is to be used, but once systems
Atnh2, Atnh3) all apply to preventing losses from a for doing such research became operational, a green
single Ytnh of 900 per year. Incorporating all pos- revolution spread through Asia and Latin America.
sible Ytnh and all possible research alternatives to But the necessary combination of policies, tech-
prevent those potential losses would provide a full nology and government institutions have proven
solution to the allocation problem that follows the elusive in sub-Saharan Africa. That part of the
rule suggested by economic theory and would word remains the challenge for the 21st Century.
maximize the expected value of prevented losses. Development assistance from wealthy countries
Of course, implementing such an allocation pro- has contributed in significant ways to help improve
cedure requires input–output functions for all conditions in poor countries with agricultural re-
possible research alternatives and a computational search among the most successful of aid efforts.
algorithm to solve the entire system. But with The technology and cultivation practices devel-
modern computers this is possible. oped by the international agricultural research
centres of the CGIAR spread widely through Asia,
the Middle East and Latin America and provided
Summary the basis for a green revolution in many countries.
While poverty still is the lot of too many, food
In 1800 most people in most countries were availability and incomes are much improved.
chronically hungry, life expectancy was 35– Plant science research has contributed to the
40 years and misery was the accepted lot of most improved management and control of many plant
people. By 1950 a few countries in Europe and pests and diseases but crop losses continue to
North America had achieved a remarkable claim over 40% of potential production having an
improvement in living standards; food production estimated value of nearly $250 billion. Appropri-
and consumption reached adequate levels for most ately directed research could develop systems and
and people lived to their mid-60s on average. But products to save much of that potential. In order
in Asia, Africa, most of Latin America and the to best allocate available research resources to
92

address those challenges, decision makers must: Evaluation and Priority Setting. Cornell University Press,
carefully define objectives; specify possible Ithaca, NY
Dalrymple DG (2005) Setting the agenda for science and
research alternatives in quantitative terms; choose technology in the public sector: The case of international
among several different approaches for setting agricultural research. Science and Public Policy, forth-
priorities; apply the method to establish priorities; coming
and allocate the resources among alternatives. The Eicher C and Staatz J (1998) International Agricultural Devel-
optimal allocation of research resources can only opment. The Johns Hopkins University Press, Baltimore and
London
be established by applying economic principles to Evenson RE and Gollin D (eds.) (2003) Crop Variety Improve-
estimates of the research input–output functions ment and its Effect on Productivity. CABI Publishing,
that quantify how alternatives are expected to Wallingford, United Kingdom
prevent crop losses. Fan S, Hazell P and Thorat S (1999) Linkages Between
Most allocations simply take the previous year’s Government Spending, Growth, and Poverty in India.
International Food Policy Research Institute, 110, Wash-
resources and make small adjustments; some ington, DC
allocations use scoring approaches; some allocate Fan S, Zhang L and Zhang X (2002) Growth, Inequality, and
resources in proportion to value of production, Poverty in Rural China. International Food Policy Re-
contribution to incomes of the poor or in pro- search Institute, 125, Washington, DC
FAO (1967) Background Papers for the Symposium on Crop
portion to the value of crop losses. These intui-
Losses. FAO Symposium on Crop Losses. Rome. Food and
tively appealing procedures all have the drawback Agriculture Organization of the United Nations
of failing to take into account either the likely FAO (2002) World Agriculture: Toward 2015/2030. Food and
degree of success research may have in addressing Agriculture Organization of the UN, Rome, Italy
each alternative or the importance of each possible FAO (2005a) Compendium of Agriculture – Environmental
solution to poor farmers and consumers. Indicators (1989–1991 to 2000). Food and Agriculture
Organizations of the United Nations
Economic principles offer tools that can incor- FAO (2005b) International Treaty on Plant Genetic Resources.
porate many considerations important to stake- Food and Agriculture Organization of the United Nations
holders. Allocations that use marginal productivity Fogel RW (2004) The Escape from Hunger and Premature
variations on benefit:cost approaches require large Death 1700–2100: Europe. America and the Third World.
amounts of data and require researchers to make Cambridge University Press
Hayami Y (2001) Development Economics: From the Poverty
their assumptions explicit. These are difficult to to the Wealth of Nations. Oxford University Press,
apply and are seldom used. However, they would London
keep decision makers from overlooking poten- Hayami Y and Ruttan V (1985) Agricultural Development: An
tially large contributions to the ultimate goal of International Perspective. The Johns Hopkins University
Press, Baltimore and London
improving the lives of the poor through agricul-
Herdt RW (1991). Research priorities for rice biotechnology.
tural research. The paper demonstrates that two In: Khush G and Toenniessen G (eds.) Rice Biotechnology
aspects related to future research – what can be done CAB International, Wallingford, UK
and what is needed by users of new knowledge – Herdt RW and Mellor JW (1964) The contrasting response of
should be considered by any decision-maker in rice to nitrogen: India and the United States. Journal of
setting research priorities and that using economic Farm Economics 46: 150–160
Johnson DG (2000) Population, food and knowledge. Amer-
principles together with such information generates ican Economic Review 90: 6–21
a higher expected return on research investments Kumar K (1987) AID’s experience with integrated rural
than alternative methods. development projects. USAID Program Evaluation Report
No. 19, July 1987, Washington, DC
Mellor JW (1966) The Economics of Agricultural Development.
References Cornell University Press, Ithaca, NY
Meyer RL and Nagarajan G (1996) Evaluating credit guarantee
Adams DW, Graham DH and Pischke JDV (eds.) (1984) programs in developing countries. Department of Agricul-
Undermining Rural Development with Cheap Credit. tural Economics and Rural Sociology Working Paper 2322 .
Westview Press, Boulder, CO Ohio State University, Columbus, Ohio
Alston JM, Chang-Kang C, Marra MC, Pardey PG and Wyatt Mills B (ed.) (1998) Agricultural Research Priorities: Informa-
T (2000) A Meta-Analysis of Rates of Return to Agricul- tion Investments for the Improved use of Research
tural R&D: Ex-Pede Herculem? International Food Policy Resources. International Service for National Agricultural
Research Institute, 113, Washington, DC Research, The Hague
Alston JM, Norton GW and Pardey PG (1995) Science Under Norton GW and Pardey PG (1987) Priority-setting Mecha-
Scarcity: Principles and Practices for Agricultural Research nisms for National Agricultural Research Systems: Present
 93

Experience and Future Needs. International Service for Sorho F, Pinel A, Traore O, Bersoult A, Ghesqiere A, Hebrard E,
National Agricultural Research, 7, The Hague Konate G, Sere T and Fargette D (2005) Durability of natural
OECD Development Assistance Committee (2005) Interna- and transgenic resistances in rice to Rice yellow mottle virus.
tional Development Statistics on-line; CRS online database European Journal of Plant Pathology 112: 349–359
in aid activities. Organization for Economic Cooperation Steinberg DI, Clapp-Wincek C and Turner AG (1983) Irriga-
and Development tion and AID’s Experience: A consideration based on
Oerke E-C, Dehne H-W, Schonbeck F and Weber A (1994) evaluations. USAID, Washington, DC
Crop Production and Crop Protection: Estimated Losses in Teng PS and Krupa SV (1980). Assessment of crop losses which
Major Food and Cash Crops. Elsevier, Amsterdam constrain production and crop improvement in agriculture
Pitman GK (2002) Bridging Troubled Waters: Assessing the and forestry. In: Teng PS and Krupa SV (eds.) E. C.
World Bank Water Resources Strategy. The World Bank, Stakman Commemorative Symposium. University of Min-
OED, Washington, DC nesota, Minneapolis, Minnesota (pp. 327) Department of
Rosegrant MW and Evenson RE (1992) Agricultural produc- Plant Pathology, University of Minnesota,
tivity and sources of growth in South Asia. American World Bank Operations Evaluation Department (1993) Annual
Journal of Agricultural Economics 74: 757–761 Review of Evaluation Results 1992. The World Bank,
Sachs JD (2005) The End of Poverty: Economic Possibilities for Washington, DC
Our Time. The Penguin Press, New York World Bank Operations Evaluation Department (1995) The
Schultz TW (1964) Transforming Traditional Agriculture . Yale World Bank and Irrigation. The World Bank, 14908,
University Press, New Haven Washington, DC
This page intentionally blank
European Journal of Plant Pathology (2006) 115: 95–103
Springer 2006
DOI 10.1007/s10658-005-1230-z

Disease assessment concepts and the advancements made in improving the accuracy
and precision of plant disease data

Forrest W. Nutter Jr.1 Paul D. Esker1 and Rosalee A. Coelho Netto2

1
Department of Plant Pathology, Iowa State University, 351 Bessey Hall, Ames, IA, 50011 USA
(Fax: +1-515-294-9420; E-mail: [email protected]); 2National Research Institute of the Amazon, Caixa
Postal 478, CEP 69011-670, Manaus, AM, Brazil

Accepted 13 July 2005

Key words: phytopathometry

Abstract

New concepts in phytopathometry continue to emerge, such as the evolution of the concept of pathogen
intensity versus the well-established concept of disease intensity. The concept of pathogen severity, defined
as the quantitative measurement of the amount of pathogen per sampling unit has also emerged in response
to the now commonplace development of quantitative molecular detection tools. Although the concept of
disease severity, i.e., the amount of disease per sampling unit, is a well-established concept, the accuracy
and precision of visual estimates of disease severity is often questioned. This article will review disease
assessment concepts, as well as the methods and assessment aides currently available to improve the
accuracy and precision of visually-based disease severity data. The accuracy and precision of visual disease
severity assessments can be improved by quantitatively measuring and comparing the accuracy and pre-
cision of rates and/or assessment methods using linear regression, by using computer-based disease
assessment training programmes, and by developing and using diagrammatic keys (standard area
diagrams).

Introduction theory and practice (Shokes et al., 1987; Nutter

et al., 1991; Nutter and Schultz, 1995). An integral
component of studying the interactions of host
‘‘How can plant pathologists apply advanced statistical and pathogen populations in time and space is the
procedures or develop quantitative predictive models ability to accurately discriminate between levels of
based upon disease assessment data of unknown accu- injury (disease intensity) caused by plant
racy and precision?’’ David R. Mackenzie, 1979 pathogens.
The efficient application of any integrated disease Disease intensity is a generic term for the
management programme requires accurate and amount of disease in a host population. Disease
precise information concerning the quantitative intensity can be either the independent variable or
measurement of the disease and/or pathogen the dependent variable in stimulus–response
population, yet the accuracy and precision of models; however, in both cases, disease intensity
quantitative disease/pathogen assessments in plant needs to be quantified with a high degree of
pathology is often taken for granted (Main, 1977; accuracy and precision if meaningful predictive
Zadoks and Schein, 1979; Gaunt, 1987; Kranz, models are to be developed (Nutter, 1990; O’Brien
1988; Nutter et al., 1991; Nutter and Schultz, and van Bruggen, 1992; Nutter and Gaunt, 1996;
1995; Nutter and Gaunt, 1996). Accurate and Guan and Nutter, 2003). For example, quantifying
precise disease (or pathogen) assessments provide disease intensity–crop yield (loss) relationships
a quantitative link between disease management demands a high degree of accuracy and precision
96

with regards to disease assessments because disease disease epidemics result from the interaction of
intensity is used as the independent variable (X) in host and pathogen populations in time and space,
single-point or area under the disease progress as affected by the environment, quantification of
curve (AUDPC) – crop yield (loss) regression the disease population usually involves an assess-
models (Guan and Nutter, 2001, 2004). The stim- ment of visible injury (disease symptoms). This is
ulus (disease intensity) must be measured accu- true because disease injury is often directly pro-
rately in order to develop yield response or yield portional to the size of the pathogen population
loss models that have adequate predictive (Nutter et al., 1991; Nutter and Guan, 2001). On
capabilities. the other hand, pathogen assessments can be
The symbol Y is often used to represent a obtained by directly measuring the pathogen
measure of disease (or pathogen) intensity because population (e.g. the number of spores, sclerotia,
disease intensity assessments are often graphed on nematodes, etc.) per unit area or volume, or the
the y-axis with respect to time (X-axis). The use of a detection method to determine the pres-
graphical representation of disease intensity versus ence or absence of a pathogen for each sampling
time is referred to as a disease progress curve, unit (e.g., ELISA or PCR to detect the presence of
whereas the graphical representation of pathogen a pathogen in or on a sampling unit). Thus,
intensity versus time is referred to as a pathogen researchers can perform disease assessments or
progress curve (Nutter, 2001). A disease (or pathogen assessments (or both); however, these
pathogen) progress curve is the signature of an terms should not be used interchangeably because
epidemic and represents the integration of all host, they represent different populations being assessed
pathogen, and environmental effects (including (Nutter, 1997b, 1999, 2001).
pathogen vectors and disease management tactics)
that occur during the period of host–pathogen Disease versus pathogen intensity
interaction (Campbell and Madden, 1990; Nutter,
1997b). Disease intensity (Y) is a general (generic) term
Quantification of disease intensity also requires used for quantifying the amount of disease in a
a high degree of accuracy and precision when population (Campbell and Madden, 1990; Nutter
disease intensity is the dependent variable (Y) to et al., 1991; Nutter and Gaunt, 1996). In plant
quantify the rate of disease progress with respect pathology, the three most common measures of
to time (X), or the change in disease intensity with disease (and pathogen) intensity are: (i) preva-
respect to distance (X). In both cases, X (time or lence, (ii) incidence, and (iii) disease severity. Dis-
distance) can be measured with great accuracy, ease prevalence is a term that is often used
and therefore, the accuracy and precision of dis- interchangeably (and mistakenly) with disease
ease intensity assessments (Y) directly affect how incidence. Prevalence is defined as the number of
much of the variation in Y can be explained by X geographical sampling units (fields, farms, coun-
in such models. ties, states, regions, etc.) where a disease or path-
ogen has been detected, divided by the total
Disease assessment defined number of geographical sampling units assessed
(Zadoks and Schein, 1979; Campbell and Madden,
The branch of plant pathology that deals with the 1990; Nutter et al., 1991; Nutter and Gaunt, 1996).
theory and practice of quantitative disease (and/or It is important to distinguish disease prevalence
pathogen) assessment is known as phytopathom- from pathogen prevalence. Disease prevalence
etry (Main, 1977; Zadoks and Schein, 1979; measures the proportion (or percentage) of geo-
Campbell and Madden, 1990). Disease assessment graphical sampling units (fields, counties, coun-
is defined as the act (or process) of quantitatively tries, etc.) where a disease (expressing symptoms)
measuring disease intensity (Campbell and Mad- has been found to occur, divided by the total
den, 1990; Nutter et al., 1991; Nutter and Gaunt, number of geographical sampling units inspected
1996). In plant pathology, there are two basic and or surveyed, whereas pathogen prevalence is a
distinct populations that can be quantitatively as- measure of the number of geographical sampling
sessed: the pathogen population and the disease units where the pathogen has been detected (e.g.,
population (Nutter, 1997b, 1999). Because plant by direct plating, inspections for the presence of
 97

pathogen signs, ELISA, PCR, etc.), divided by the sampling unit), or the area of non-green tissue of a
total number of geographical sampling units that sampling unit divided by the total area of the
were inspected, tested, or indexed (Nutter, 2001). sampling unit  100 (Nutter, 2001). Disease
Prevalence data are often multiplied by 100 to severity could be also defined as the volume of a
express as a percentage. A single diseased or in- disease-induced gall (using the equation and
fected plant (or plant part) is all that is required to dimensions for a cylinder or a sphere, etc.), as is
change the status of a geographic sampling unit done in human epidemiology for cancerous tumors
from negative ()) to positive (+), provided the (Nutter, 1999).
sensitivity of the method is sufficient to detect the The concept of pathogen severity is becoming
presence of a pathogen in a bulked (diluted) more widespread as new methods are developed to
sample. Bulking samples is particularly useful quantify the amount of a pathogen present in a
when pathogen incidence is low because the sampling unit. Examples include the use of quan-
number of bulked samples tested or indexed is titative PCR methods to estimate the amount of
often less than the number of individuals sampled virus (or pathogen) present per gram of leaf tissue,
and processed, thus reducing the cost of detection or the number of nematodes per gram of root
per sampling unit (Nutter and Gaunt, 1996; tissue. In spite of advances concerning the concept
Nutter, 1997b). of pathogen severity, the majority of severity
Disease incidence is defined as the number of assessments employed today usually involve visual
sampling units (e.g., leaflets, leaves, stems, tillers, estimates of disease severity (l2/L2100), and yet, a
whole plants, seeds, etc.) that are diseased number of critical questions still remain. At the
(expressing symptoms), divided by the total num- top of the list of such questions is: how can we do
ber of sampling units sampled and assessed (Nut- better to improve the accuracy and precision of
ter et al., 1991; Nutter, 1997b, 1999). As with visual disease assessments? This article will address
prevalence, it is important to make a clear dis- specific ways to improve the accuracy and preci-
tinction as to whether incidence is based on sion of diseases assessments. These are: (i) use of
detection of the pathogen or on the basis of disease regression to quantify the precision of disease
(visual symptoms) in a host population (Nutter, raters and/or assessment methods, (ii) the use of
2001). Progress curves based on pathogen detec- computer-based assessment training programmes,
tion (indexing) methods, such as ELISA, may and (iii) the use of standard area diagrams
closely mirror progress curves based on disease (diagrammatic keys) in colour.
symptoms (Padgett et al., 1990; Nutter 2001);
however, in many instances, the use of different
disease assessment or pathogen detection methods Use of linear regression to assess and compare the
may result in progress curves with quite different precision of assessment methods
shapes and rates (Nutter, 1997b, 2001).
Disease severity is a measure of the amount of The expenditure of time and money to develop,
disease per sampling unit and it is this particular evaluate, and compare disease assessment methods
type of measurement that this article will focus can prevent serious flaws (e.g., rater bias) in data
upon (Nutter et al., 1991; Nutter, 1997b). acquisition. Disease assessment methods should
Researchers should clearly define disease severity provide accurate and precise information that sat-
by providing not only a descriptive definition, but isfies the goals and needs of the research (Nutter
also an operational definition that includes the and Gaunt, 1996). Campbell and Madden (1990)
dimensions that were used to assess disease. In have defined precision as the lack of variation in
plant pathology, disease severity is most often disease estimates when the same sampling units are
operationally defined as the diseased leaf area (l2), evaluated by other raters. However, this definition
divided by the total leaf area of a leaf or sampling of precision excludes another potential source of
unit (L2)100, i.e., (l2/L2100) to obtain percent- error, i.e., the repeatability of individual raters
age disease severity (James, 1971; Nutter et al., (Nutter et al., 1993). Shokes et al. (1987) proposed
1991). Other common measures of disease severity using a test-retest procedure using correlation
include the number of lesions/leaf (or sampling analysis to quantify rates repeatability; however,
unit), the number of lesions/cm2 of leaf (or other this method provides a measure of precision
98

(agreement) among raters and does not quantify Disease assessment training with computer
the degree of bias among raters. programmes
Simple linear regression provides a powerful
method to quantify the degree of error (bias) due The accuracy and precision of disease severity
to raters or assessment methods (Nutter et al., assessments have come into question due to the
1993). Regression analysis has been used to measurable bias that different raters have shown
determine the relative precision of a visual assess- when evaluating the same set of diseased sampling
ment method (disease severity) and a remote units (Sherwood et al., 1983; Forbes and Jeger,
sensing assessment method (reflectance at 600 nm) 1987; Kranz, 1988; Nutter et al., 1993). Accuracy
in which a hand-held, multispectral radiometer can be defined as the measure of the closeness of a
was employed (Nutter et al., 1993). The disease disease assessment to the true value (Nutter et al.,
assessed was dollar spot of bentgrass (caused by 1991; Zadoks and Schein, 1979). When assessing
Sclerotinia homoeocarpa). The precision of differ- disease severity, the stimulus (X) is the actual dis-
ent disease assessment methods and raters can be ease severity of a sampling unit and the rater’s
evaluated and compared by operationally defining estimate of disease severity (Y) is the response. For
intra-rater repeatability and inter-rater reliability each 1% increase in actual severity, we would ex-
(Nutter and Schultz, 1995). Intra-rater repeat- pect a rater to also to estimate a 1% increase, i.e.,
ability for different assessment methods can be the slope should be equal to 1.0 (no systematic
determined by regressing one set of measurements bias) and the intercept should not be significantly
(Y) (obtained by each rater) with a repeated set of different from zero (no constant bias present).
measurements (X) performed on the same set of Accuracy cannot be properly evaluated unless the
sampling units. The parameters and statistics used researcher is confident that the actual (true) dis-
to compare the intra-rater repeatability of different ease severity can be measured absolutely. This is
assessment methods and/or raters are: slope, easily achieved using computer-generated images
intercept, coefficient of determination (R2), coeffi- of diseased leaves because the computer can be
cient of variation (CV), and the standard error of programmed to calculate the number of non-green
the estimate for Y (SEEy) (Nutter et al., 1993; (diseased) pixels in an image, divided by the total
Nutter and Schultz, 1995). A slope significantly number of pixels in the image  100 to obtain a
less than or greater than 1.0 would indicate the true measure of percentage disease severity (Nutter
presence of systematic bias and the greater the and Litwiller, 1998; Nutter et al., 2000). The use of
deviation from 1.0, the greater the systematic bias computer programmes to enhance learning has
for a specific rater and/or method. This is because become widely accepted for several reasons
for each 1% increase in estimated disease severity (Nutter, 1997a). One advantage of computer-aided
the first time a set of sampling units is assessed, disease assessment training is that a full range of
there should be a corresponding 1% increase in disease severity levels can be presented as stimuli
estimated disease severity when the same set of to which operators of the programme respond.
sampling units are assessed a second time by the Nutter and Worawitlikit (1990) built upon the
same rater or method (Nutter et al., 1993). An computer-based disease assessment training con-
intercept significantly different from zero indicates cept by developing a computer programme to as-
the presence of another form of bias that is con- sess diseases of peanut called Disease.Pro.
stant for all disease levels evaluated. The use of R2, Recognizing the tremendous potential to improve
CV, and SEEy values to quantify and compare the the accuracy and precision of disease assessments
precision of disease assessment methods or raters through computer-based training programmes,
has been previously described (Nutter et al., 1993; Nutter and Litwiller (1998) later developed a more
Nutter and Schultz, 1995; Nutter, 2001). Likewise, generic disease assessment training programme
linear regression can be used to quantify precision (Severity.Pro) that allowed the user to select from
among raters or methods (inter-rater reliability) by a menu of leaf shapes (alfalfa, apple, barley,
having two or more raters (and/or methods) assess cucumber, grape, tomato, etc.) and lesion types
the same set of sampling units, and then evaluating (anthracnose, blotch, downy mildew, target spot,
the slopes, intercepts, R2, CV, and SEEy values powdery mildew, etc.) to mimic almost any foliar
(Nutter et al., 1993). pathosystem. Severity.Pro was recently rewritten
 99

in Java to be more compatible with present-day Computer-based disease assessment training

operating systems. programmes provide a useful platform for teach-
The most current version of Severity.Pro all- ing disease assessment theory and hands-on prac-
ows raters to: (i) choose whether or not they tice. For example, the results of computer training
want the actual severity to be immediately dis- for six raters using Severity.Pro are shown in
played (feedback), (ii) choose the number of Table 1. The six raters evaluated computer-gen-
leaves to be assessed and the size of the lesions erated images of grapevine leaves infected by
that will appear on diseased leaf images (small, downy mildew by assessing 30 images before and
medium, large, or random), (iii) view graphs of after training. In pre-tests, one rater (Rater 2)
the absolute error (Y) versus the actual severity generally overestimated downy mildew severity
(X) (absolute error is defined as the estimated throughout the range of the severities tested, with
severity minus the corresponding actual severity), rater error being as high as 21% (solid circles,
and (iv) regress the rater’s severity estimates (Y), Figure 1a). Following training, Rater 2’s estimates
against the actual disease severities (X). These were within 5–10% of the actual severity levels
changes allowed for a more powerful training (open circles, Figure 1a). Figure 1b shows that this
tool because: (i) raters can take a pre-test (with- rater also had a constant bias of 6.8% prior to
out feedback before training) to provide a base- training (Y-intercept) and that this bias was
line of how different disease severity levels are reduced to near zero ()0.14%) after computer-
perceived, (ii) the data can be viewed in graphical based training using Severity.Pro. Based on R2
form and analyzed by regression, (iii) raters can values, the precision of Rater 2 was also signifi-
execute a drill and practice session and receive cantly improved following training (R2 was 95%
feedback as to the actual level of disease severity following training compared with 85% prior to
immediately after the estimated severity is keyed training). As a group, five of the six raters dem-
in, (iv) rater improvement and, more importantly, onstrated improvement in precision following
the degree of rater improvement can be computer training, as measured by improvement
documented by having raters take a post-test in the coefficients of determination (R2) (Table 1).
(without feedback after training) and then com- Three of the raters had R2 values that were 6–11%
pare pre- and post-test regression parameters and higher after training. The other three raters (rater
statistics, and (v) the results of pre- and post-tests 1, rater 4, and rater 6) were already highly precise
can also be used to evaluate and compare rater and their R2 values increased 2, 1, and 0%,
performance. respectively.

Table 1. Y-interceptsa, slopesb, coefficients of determination (R2)c, and standard errors of the Y-estimate (SEEY)d for six raters be-
fore training (pretest) and after training (posttest) using a computer programme that simulates downy mildew of grapevines (adap-
ted from Nutter, 2001)

Rater Pretest Posttest

Intercept Slope R2 SEEY Intercept Slope R2 SEEY

1 )7.19 1.11 0.94 2.74 )1.52 1.01 0.96 1.95

2e 6.83 1.02 0.85 3.80 )0.14 0.94 0.95 1.97
3 )6.30 0.91 0.91 2.83 7.21 0.83 0.97 1.41
4 1.89 1.06 0.91 2.97 0.61 0.82 0.92 2.25
5 )1.13 1.27 0.84 4.69 )8.46 1.05 0.94 2.61
6 1.77 1.01 0.97 1.62 )3.10 1.03 0.97 1.76
Improved 3/6 2/6 5/6 5/6
a
Y-intercepts that deviate from zero indicate the presence of a constant source of rater bias with regards to accuracy.
b
Slopes that deviate from 1.0 indicate the presence of a systematic source of rater bias with regards to accuracy.
c
The higher the coefficient of determination (R2), the higher the precision of rater estimates.
d
The lower the standard error of the Y-estimate, the higher the precision of rater estimates.
e
Data for Rater 2 are shown in graphical form in Figure 1 as this data would appear in the disease assessment computerized training
programme Severity.Pro (Nutter and Litwiller, 1998).
100

(a) 30 the accuracy and the precision of disease assess-

ment data.
Rater 2 estimated - actual severity (%)

20
Use of standard area diagrams to improve the
10 accuracy and precision of disease severity
assessments
0

Disease assessment keys, also known as diagram-

-10 matic keys or standard area diagrams are pictorial
diagrams that depict the true amount of injury
-20 Pretraining results
Posttraining results
(usually disease severity) on individual sampling
units (quadrats, whole plants, leaves, fruit, tubers,
-30
0 10 20 30 40 50 60 70 80 90 etc.). Disease severity of each individual diagram is
expressed as a percentage of the total surface area
(b) 100
Pretraining of each sampling unit (disease area/total area of
Y = 6.8 + 1.02 (X), R2 = 0.85
the image  100) (Nutter and Esker, 2001). Stan-
Posttraining
80 dard area diagrams (SADs) provide raters with a
Rater 2 estimated severity (%)

Y = -0.14 + 0.94 (X), R2 = 0.95

series of reference images that are accepted to be
60 the truth in terms of the actual amount of injury
(severity) depicted on each disease diagram. Clive
James developed and marketed the first series of
40
black and white standard area diagrams (James,
1971). More recently, Nutter and Litwiller (1998)
20 developed and tested a computer programme
(Severity.Pro) that generates standard area dia-
0 grams in colour. Thus, Severity.Pro provides a
0 10 20 30 40 50 60 70 80 90
Actual severity (%)
powerful tool to generate, capture and print dis-
eased leaf images with known severity levels in
Figure 1. Improvement in the (a) absolute error (estimated colour (Nutter et al., 1998). This enables
minus actual disease severity) and (b) accuracy (slope, inter- researchers to create a series of pictorial colour
cept) and precision (R2, SEEy) of Rater 2 (from Table 1),
diagrams that can be used as an assessment aid to
before and after disease assessment using the computer
program Severity.Pro (Nutter and Litwiller, 1998). improve the accuracy and precision of disease
assessment data. Although it has long been as-
sumed that the use of standard area-diagrammatic
As stated earlier, the standard error of the keys will help to improve the accuracy and preci-
Y-estimate is another important measure of rater sion of visual disease severity assessments per-
precision. This statistic provides information con- formed by raters (James, 1971; Horsfall and
cerning the degree of error associated with a pre- Cowling, 1978; Kranz, 1988), only recently have
dicted value of Y. Therefore, the lower the SEEY, definitive studies been conducted to demonstrate
the higher the precision (Nutter and Schultz, that the accuracy and precision of disease assess-
1995). For five of the six raters, SEEY values ments are actually improved when standard area-
decreased following disease assessment training by diagrammatic keys are used (Godoy et al., 1997;
an average of 40% (3.41 to 2.04). Rater 6 showed Nutter et al., 1998; Leite and Amorim, 2002;
no significant change in pre-test versus post-test Gomes et al., 2004). As part of a class exercise for
SEEY values because this rater was already very students enrolled in a course in plant disease epi-
precise (i.e., rater 6 had very low SEEY values in demiology at Iowa State University, 10 raters were
both pre- and post-training tests). Thus, comput- asked to assess 30 diseased leaf images (repre-
erized disease assessment training programmes senting a range of disease severities) of downy
provide an important educational tool that can be mildew of grape, both with and without the use of
used to teach disease assessment theory and colour-standard area diagrams (Nutter and
concepts, as well as to substantially improve both Litwiller, 1998). When individual rater estimates
 101

were regressed against the true disease severity and SEEy values were lower when standard area
levels as calculated by the computer programme, it diagrams were used, indicating there was a signif-
was found that rater estimates of disease severity icant increase in the precision of the assessment
were much closer to the actual (true) severity levels data when using the standard area diagrams.
when raters used standard area diagrams as an As a class, statistical analyses for accuracy
assessment aid to assess disease severity (Nutter showed that eight of the ten raters achieved
and Esker, 2001). intercepts closer to zero (less constant bias) and
As an example, Figure 2 shows a typical situa- that seven of the ten raters achieved slopes closer
tion regarding accuracy and precision of visual to 1.0 (less systematic bias) when standard area
assessments performed by one rater with, and diagrams were used (Nutter and Schultz, 1995).
without, the use of standard area diagrams for Statistical analyses for precision showed that seven
grapevine downy mildew. When using the stan- of the ten raters achieved higher coefficients of
dard area diagrams, this rater had greater accuracy determination (R2), and eight of the ten raters had
(less systematic and constant bias) as indicated by lower standard errors of the estimate for Y (SEEy)
a slope closer to 1.0 (1.05) and a Y-intercept closer when standard area-diagrammatic keys were used
to zero ()0.28%) compared to the slope (0.86) and as an assessment aid. Thus diagrammatic standard
intercept (5.34%) when not using the standard area-assessment keys can substantially improve
area diagrams. Moreover, R2 values were higher both the accuracy and the precision of visual
disease assessments.
(a) 100
Without using standard area diagrams
Fitted line: Y = 5.34 + 0.86(X), R2 = 89.0%
80 Summary and conclusions
Estimated severity (%)

60
The potential for rater bias (under- or over-esti-
mation of the actual level of disease severity) is an
40
ever-present concern that should receive serious
consideration by researchers when raters are
making visual disease assessments and will use that
20
information as the basis to develop stimulus–re-
sponse models, or to evaluate and compare disease
0
0 20 40 60 80 100
management tactics, strategies, or integrated dis-
Actual severity (%) ease management systems (Zadoks and Schein,
(b) 100
1979; Gaunt, 1995; Nutter, 1997b, 1999, 2001).
Using standard area diagrams Rater bias, however, can be effectively reduced.
Fitted line: Y = -0.28 + 1.05(X), R2 = 94.6% Disease assessment training programmes using
80
computer-generated images of disease leaves have
Estimated severity (%)

been shown to improve both accuracy and preci-

60
sion (Nutter and Schultz, 1995; Nutter and Parker,
1997; Nutter and Litwiller, 1998). Moreover,
40
studies by Godoy et al. (1997), Gomes et al. (2004),
Nutter (2001), and Nutter and Esker (2001) have
20
documented that the use of standard area diagrams
as an assessment aid for visually assessing disease
0 severity can also significantly improve the accuracy
0 20 40 60 80 100
Actual severity (%) and precision of disease severity assessment data.
The use of both computer-based disease assessment
Figure 2. (a) Estimated severity of grapevine downy mildew training programmes and standard area diagrams
compared with actual (true) severity when assessing computer
to improve the accuracy and precision of disease
images without the use of standard area diagrams and (b)
Estimated versus actual severity when using standard area assessment data are not mutually exclusive, as both
diagrams. Improvements were apparent as both systematic methods should be used to obtain the best disease
and constant bias were reduced. assessment data possible. Finally, the use of
102

regression to evaluate and compare the accuracy Kranz J (1988) Measuring plant disease. In: Kranz J and
and precision of: (i) colour disease assessment Rotem J (eds) Experimental Techniques in Plant Disease
Epidemiology (pp. 35–50) Springer-Verlag, New York.
protocols (e.g., use of a linear scale versus a loga- Leite RMVBC and Amorim L (2002) Development and
rithmic disease assessment scale) (Nutter and Es- validation of a diagrammatic scale for Alternaria leaf spot
ker, 2005), (ii) disease assessment instruments (e.g., of sunflower. Summa Phytopathologica 28: 14–19.
image analysis or remote sensing sensors and Main CE (1977) Crop destruction – The raison d’ être of plant
instruments) (Nutter, 1990; Guan and Nutter, pathology. In: Horsfall JG and Cowling EB (eds) Plant
Disease: An Advanced Treatise (pp. 55–78) Academic
2004), and/or (iii) disease raters, can provide Press, New York.
researchers with statistical methods to determine Nutter FW, Jr. (1990) Remote sensing and image analysis for
the accuracy and precision of disease assessment crop loss assessment. In: Crop Loss Assessment in Rice (pp.
data (Nutter et al., 1993; Nutter and Littrell, 1996; 93–105) International Rice Research Institute, Manila,
Nutter, 1997a; Guan and Nutter, 2003). Thus, Philippines.
Nutter FW, Jr. (1997a) Disease severity assessment training. In:
researchers can and should place greater focus Francl L and Neher D (eds) Exercises in Plant Disease
upon evaluating, comparing, and selecting the best Epidemiology (pp. 1–7) APS Press, St. Paul, MN.
disease assessment protocols, instruments, and/or Nutter FW, Jr. (1997b) Quantifying the temporal dynamics of
raters that best meet the goals of the research. plant virus epidemics: a review. Crop Protection 16: 603–
618.
Nutter FW, Jr. (1999) Understanding the interrelationship
References between botanical, human, and veterinary epidemiology:
The Ys and Rs of it all. Ecosystem Health 5: 131–140.
Campbell CL and Madden LV (1990) Introduction to Plant Nutter FW, Jr. (2001) Disease assessment terms and concepts.
Disease Epidemiology, John Wiley, New York. In: Maloy OC and Murray TD (eds) Encyclopedia of Plant
Forbes GA and Jeger MJ (1987) Factors affecting the estimation Pathology (pp. 312–323) John Wiley and Sons, Inc, NY.
of disease intensity in simulated plant structures. Zeitschrift Nutter FW, Jr. and Esker PD (2006) Disease assessment keys.
fur Pflanzen Krankheiten und Pflanzenschutz 94: 113–120. In: Maloy OC and Murray TD (eds) Encyclopedia of Plant
Gaunt RE (1995) New philosophies and technologies in disease Pathology (pp. 323–326) John Wiley and Sons, Inc, NY.
measurement and appraisal. Canadian Journal of Plant Nutter FW, Jr. and Esker PD (2005) The role of psychophysics
Pathology 17: 185–189. in phytopathology: the Weber-Fechner law revisited. Euro-
Godoy CV, Carneiro MTPG, Iamauti MT, Dalla Pria M, pean Journal of Plant Pathology 114: 199–213.
Amorim L, Berger RD and Bergamim Filho A (1997) Nutter FW, Jr. and Gaunt RE (1996) Recent developments in
Diagrammatic scales for bean diseases: development and methods for assessing disease losses in forage/pasture crop.
validation. Journal of Plant Diseases and Protection 104: In: Chakraborty S, Leath DT, Skipp RA, Pederson GA,
336–345. Bray RA, Latch GCM and Nutter FW Jr. (eds) Pasture
Gomes AMA, Michereff SJ and Mariano RLR (2004) Develop- and Forage Crop Pathology (pp. 93–118) American Society
ment and validation of a diagrammatic key for Cercospora of Agronomy, Crop Science Society of America, and Soil
leaf spot of lettuce. Summa Phytopathologica 30: 38–42. Science Society of America, Madison, WI.
Guan J and Nutter FW, Jr. (2001) Factors affecting the quality Nutter FW, Jr., Gleason ML, Jenco JH and Christians NC
and quantity of sunlight reflected from alfalfa canopies. (1993) Assessing the accuracy, intra-rater repeatability and
Plant Disease 85: 865–874. inter-rater reliability of disease assessment systems. Phyto-
Guan J and Nutter FW, Jr. (2003) Quantifying the intra-rater pathology 83: 806–812.
repeatability and inter-rater reliability of visual disease and Nutter FW, Jr. and Guan J (2001) Disease losses. In: Maloy
remote sensing assessment methods in the alfalfa foliar OC and Murray TD (eds) Encyclopedia of Plant Pathology
disease pathosystem. Canadian Journal of Plant Pathology (pp. 340–351) John Wiley and Sons, Inc, NY.
25: 143–149. Nutter FW, Jr. and Littrell RH (1996) Relationships among
Guan J and Nutter FW, Jr. (2004) Comparison of single-point percent defoliation, canopy reflectance, and pod yield in
alfalfa yield models based on visual disease intensity and the peanut-late leafspot system. Crop Protection 15:
remote sensing assessments. Canadian Journal of Plant 135–142.
Pathology 26: 314–324. Nutter FW, Jr. and Litwiller D (1998) A computer program to
Gaunt RE (1987) Measurement of disease and pathogens. In: generate standard area diagrams to aid raters in assessing
Teng PS (ed.) Crop Loss Assessment and Pest Management disease severity. Phytopathology 88: S117.
(pp. 6–18) The American Phytopathological Society, St. Nutter FW, Jr., Litwiller D and Wegulo SN (1998). Do
Paul, MN. standard area diagrams improve the accuracy and precision
Horsfall JG and Cowling EB (1978) Pathometry: the measure- of disease assessments? Proceedings of the 7th International
ment of plant disease. In: Horsfall JG and Cowling EB Congress of Plant Pathology. p. 2.1.16.
(eds) Plant Disease: An Advanced Treatise. Vol. 2 (pp. 120– Nutter FW, Jr. and Parker SK (1997) Fitting disease progress
136) Academic Press, New York. curves using EPIMODEL. In: Francl L and Neher D (eds)
James C (1971) A Manual of Assessment Keys for Plant Exercises in Plant Disease Epidemiology (pp. 24–28) APS
Disease. Canadian Department of Agriculture Publishing. Press, St. Paul, MN.
 103

Nutter FW, Jr. and Schultz PM (1995) Improving the accuracy Proceedings of the American Peanut Research and Ed.
and precision of disease assessment: selection of methods Society 22: 38.
and use of computer-aided training programs. Canadian O’Brien RD and van Bruggen AHC (1992) Accuracy, precision,
Journal of Plant Pathology 17: 174–184. and correlation to yield loss of disease severity scales for
Nutter FW, Jr., Teng PS and Shokes RM (1991) Disease corky root of lettuce. Phytopathology 82: 91–96.
assessment terms and concepts. Plant Disease 75: 1187– Padgett GB, Nutter FW, Jr. and Kuhn CW (1990) Quantifi-
1188. cation of disease resistance that reduces the rate of tobacco
Nutter FW, Jr., Wegulo SN, Litwiller D, Emmitt RW and etch virus epidemics in bell pepper. Phytopathology 80:
Magarey PA (2000) A computer program to improve the 451–455.
accuracy and precision of severity assessments in grape- Sherwood RT, Berg CC, Hoover MR and Zeiders KE (1983)
vines. In: Magarey, PA, Thiele, SA, Tschirpig, KL, Emmet, Illusions in visual assessment of Stagonospora leaf spot of
RW, Clark, K and Magarey RD (eds) Proceedings of the orchardgrass. Phytopathology 73: 173–177.
Third International Workshop on Grapevine Downy and Shokes FM, Berger RD, Smith DH and Rasp JM (1987)
Powdery Mildew (pp. 64–69). South Australian Research Reliability of disease assessment procedures: A case study
and Development Institute Series Report No. 50. with late leafspot of peanut. Oleagineux 42: 245–251.
Nutter FW, Jr. and Worawitlikit O (1990) Disease. Pro: A Zadoks JC and Schein RD (1979) Epidemiology and Plant
computerized disease assessment training program. Disease Management, Oxford University Press, New York.
This page intentionally blank
European Journal of Plant Pathology (2006) 115:105–122
Springer 2006
DOI 10.1007/s10658-005-1250-8

Relation between soil health, wave-like fluctuations in microbial populations,

and soil-borne plant disease management

Ariena H.C. van Bruggen1, Alexander M. Semenov2, Anne D. van Diepeningen1, Oscar J. de Vos1 and
Wim J. Blok1
1
Biological Farming Systems, Wageningen University, Marijkeweg 22, 6709 PG Wageningen the Netherlands;
(Fax: +31-317-468213; E-mail: [email protected]); 2Department of Microbiology, Biological
Faculty, Moscow State University, 119899 Moscow, Russia;

Key words: biological control, disease management, harmonic fluctuations, resilience, soil health, soil-borne
pathogens

Abstract

A healthy soil is often defined as a stable soil system with high levels of biological diversity and
activity, internal nutrient cycling, and resilience to disturbance. This implies that microbial fluctuations
after a disturbance would dampen more quickly in a healthy than in a chronically damaged and
biologically impoverished soil. Soil could be disturbed by various processes, for example addition of a
nutrient source, tillage, or drying-rewetting. As a result of any disturbance, the numbers of hetero-
trophic bacteria and of individual species start to oscillate, both in time and space. The oscillations
appear as moving waves along the path of a moving nutrient source such as a root tip. The phase
and period for different trophic groups and species of bacteria may be shifted indicating that suc-
cession occurs. DGGE, Biolog and FAME analysis of subsequent populations in oscillation have
confirmed that there is a cyclic succession in microbial communities. Microbial diversity oscillates in
opposite direction from oscillations in microbial populations. In a healthy soil, the amplitudes of these
oscillations will be small, but the background levels of microbial diversity and activity are high, so
that soil-borne diseases will face more competitors and antagonists. However, soil-borne pathogens
and antagonists alike will fluctuate in time and space as a result of growing plant roots and other
disturbances, and the periods and phases of the oscillations may vary. As a consequence, biological
control by members of a single trophic group or species may never be complete, as pathogens will
encounter varying populations of the biocontrol agent on the root surface. A mixture of different
trophic groups may provide more complete biological control because peaks of different trophic
groups occur at subsequent locations along a root. Alternatively, regular addition of soil organic
matter may increase background levels of microbial activity, increase nutrient cycling, lower the
concentrations of easily available nutrient sources, increase microbial diversity, and enhance natural
disease suppression.

Abbreviations: BCA – biocontrol agent; CFUs – colony forming units; DGGE – denaturing gradient gel
electrophoresis; FAME – fatty acid methyl esters; GFP – green fluorescent protein

Introduction recognized that the state of health of terrestrial,

edaphic and aquatic ecosystems is important.
Health is a necessary condition for the survival of However, it is not so easy to define, let alone
individual living organisms, communities, ecosystems, measure, ecosystem health. This is also true for
and for nature in general. Ecologists have long soil, which is considered as a living system, where
106

many physical and chemical properties are medi- conclusions will be presented regarding soil health,
ated by biota, which are primarily responsible for microbial oscillations and soil-borne diseases.
its health (Brussaard et al., 2004). The living
components of a soil possess both stable and
dynamic characteristics. Recently, we proposed to Soil health and disease suppression
use the resistance and resilience of microbial
communities in response to a disturbance as Rapport (1995) defined a healthy ecosystem as an
quantitative indicators for soil health (van ecosystem with the following characteristics: (1)
Bruggen and Semenov, 1999, 2000). integrity of nutrient cycles and energy flows, (2)
Soil as a productive system was developed and biological diversity, (3) interconnectedness
continues to develop jointly with plants that between functional units, (4) stability and resil-
inhabit the soil. The condition of the soil, includ- ience when faced with a disturbance or stress, and
ing its microbial community, reflects the condition (5) limited plant and animal disease outbreaks. A
of the (past) vegetation. In agricultural ecosys- soil ecosystem is considered healthy if it has a good
tems, especially in conventional farming systems, balance of mineral and organic substances and
most natural plants are considered weeds. living components. Such a balance is reached when
Decreasing the vegetation diversity leads to an ecosystem comes to a climax condition, char-
pauperization of soil inhabitants, decreasing of acterized by high biodiversity and low concentra-
interconnectedness and functional interchange- tions of easily available organic and inorganic
ability. An extremely simplified vegetation, such as nutrients (van Bruggen and Semenov, 1999, 2000).
a monoculture, selects a specific microbial com- To maintain soil health, it is necessary to promote
munity, including plant pathogenic microorgan- high primary productivity, high microbial bio-
isms and sometimes also their parasites or mass, activity and diversity, high nutrient turnover
antagonists. However, such a simplified ecosystem rates, and low residual nutrient pools; in other
may be very sensitive to the slightest disturbance words, oligotrophic conditions. In particular, los-
and cannot be considered healthy. ses of mineral nitrogen and dissolved organic
High quality soil is the main production carbon from soil and soil biological complexity
resource for many societies. However, this re- have been used to assess the functioning of soil
source is disappearing at an alarming rate due to ecosystems (Liiri et al., 2002).
loss of organic matter as a result of erosion, oxi- Soils of natural ecosystems are generally
dation, compaction, and biological impoverish- thought of as being healthier than those of agro-
ment. In particular, agricultural systems with ecosystems. Indeed, cultivated soils generally have
minimal biological diversity and large inputs of lower microbial diversities and more severe disease
synthetic fertilizers and pesticides have problems problems than they had as a natural habitat (Ko,
with poor soil health and associated plant diseases. 1982; Buckley and Schmidt, 2001). Organically
Action must be taken urgently to restore the bal- managed soils, where synthetic fertilizers and
ance of the soil ecosystem and its health status. pesticides are not used, are closer to natural soils
In this review we present a dynamic view of than conventionally managed soils even though
microbial populations, soil health, and disease soil fertility is maintained by regular additions of
suppression. In the following paragraph an intro- organic materials (van Bruggen, 1995; van
duction is given on soil health and its connection Bruggen and Termorshuizen, 2003). Especially
to disease suppression. The next two paragraphs chlorinated pesticides have had negative impacts
deal with temporal and spatial oscillatory responses on microbial diversity (Mas et al., 1996). Although
of bacterial communities to various disturbances. some authors found no differences in soil micro-
Then, we demonstrate that soil-borne pathogens bial diversity between organically and conven-
respond with similar oscillations to a disturbance tionally managed soils (Lawlor et al., 2000;
from a growing root. Next, agricultural manage- Franke-Snyder et al., 2001), most researchers
ment practices to control soil-borne diseases, like reported a higher biological diversity for organi-
the use of organic amendments, tillage and bio- cally than for conventionally managed soils with
control agents, will be discussed from the point of respect to various taxa, namely bacteria (Sivapalan
view of dynamic microbial oscillations. Finally, et al., 1993; Drinkwater et al., 1995; Mäder et al.,
 107

2002; van Diepeningen et al., 2005), arbuscular are less pronounced (Semenov et al., 1999). Sta-
mycorrhizal fungi (Ryan et al., 1994; Oehl et al., bility and resilience of microbial communities after
2003), nematodes (Mulder et al., 2003; van Die- exposure to a disturbance could possibly also be
peningen et al., 2005), earthworms (Mäder et al., related to disease suppression (van Bruggen and
2002), and arthropods (Drinkwater et al., 1995; Semenov, 1999, 2000). Indeed, soils with a higher
Mäder et al., 2002). Also, a higher microbial biological diversity and activity, such as natural or
activity (Workneh et al., 1993; Mäder et al., 2002) organically managed agricultural soils are fre-
and microbial biomass (Workneh and van Brug- quently more suppressive to soil-borne diseases
gen, 1994; Mäder et al., 2002; Mulder et al., 2003) than conventionally managed agricultural soils
were found in organically managed soils. (van Bruggen, 1995; van Bruggen and Termorshuizen,
High microbial biomass, activity, and diversity 2003).
in natural or agricultural soils have been associ-
ated with suppression of soil-borne plant diseases
(Nitta, 1991; Workneh and van Bruggen, 1994; Temporal wave-like fluctuations of microbial
Mäder et al., 2002). This kind of suppression populations
may be due to general competition or antago-
nism, which may be non-specific and active Fluctuations in soil microbial populations have
against a wide range of soil-borne pathogens been observed many times, both in laboratory
(Gerlagh, 1968; Whipps, 1997). However, in a experiments and in the field with native bacterial
few cases, no relationships were found between communities (Aristovskaya, 1980; Zvyagintsev
microbial biomass, activity or diversity and dis- and Golimbet, 1983; Semenov, 2001). Under nat-
ease suppression. Boehm et al. (1993, 1997) ural conditions, microbial fluctuations in soil
found that the level of Pythium root rot sup- appear irregular, and generally do not correlate
pression in peat mixes was not related to with variations in external environmental charac-
microbial biomass, activity or diversity but to the teristics, such as temperature and moisture content
composition of the rhizosphere bacterial popu- of the soil.
lation. The seemingly unpredictability of disease Irregular fluctuations can turn into regular
suppression in relation to microbial community oscillations with distinct waves after a disturbance
parameters may be due to a greater specificity of such as addition of fresh organic matter to soil
the relationship between pathogen and antago- (Doebeli and Ruxton, 1997, 1998; Clarholm, 1981).
nist than sometimes thought, due to influences of Soil is generally low in easily available nutrients,
varying soil physical and chemical characteristics especially fallow arable soil. Any disturbance pro-
(Hoper and Alabouvette, 1996), or due to vari- viding a nutrient impulse under these conditions,
ation in soil microbial communities in time and such as incorporation of fresh organic matter or
space (van Bruggen and Semenov, 2000). rewetting after drying, is likely to initiate a wave-
Soil microbial populations generally fluctuate, like response of the microbial community (van
and start to oscillate regularly in response to a Bruggen and Semenov, 1999; Caldéron et al.,
disturbance, such as addition of organic material to 2000). Hints of wave-like fluctuations were ob-
soil (van Bruggen and Semenov, 2000). The tained in a field experiment after incorporation of
amplitude of the waves in microbial populations (a cover crop debris into soil (van Bruggen and
measure of stability of the soil ecosystem), their Semenov, 2000), but the observations were too
frequency, and the time needed to return to initial sparse for time series or harmonical analysis
conditions before organic amendment (a measure (Shumway, 1988) to prove that regular oscillations
of the resilience of the system) may be used as occurred.
indicators for soil health (van Bruggen and Seme- Only recently, we demonstrated the occurrence
nov, 1999, 2000). The strongest wave-like response of regular oscillations over time using appropriate
of microbial communities occurs in soils low in statistical techniques (Zelenev et al., 2004). Tem-
organic matter (Semenov et al., 1999). In high-or- poral oscillations of microbial populations (CFUs
ganic matter soils with higher microbial biomass and microscopic cell counts) were observed for one
and activity, wave-like responses are also notice- month in soil amended with fresh plant material
able but the amplitudes and periods of these waves (grass–clover) incubated at constant temperature
108

and moisture. Bacterial populations fluctuated sandy soil. The response of copiotrophic bacterial
with different periods and amplitudes, depending CFUs (de Vos and van Bruggen, 2001; Zelenev et
on the specific conditions of each experiment, but al., 2005) to the disturbance was determined daily
immediately after the disturbance they revealed over a period of nine days, both for the grass–
remarkable oscillations with large amplitudes. The clover (GC) and the non-amended control series
patterns of the oscillations were quite predictable, (CO). Copiotrophic bacteria are fast-growing
always with a small and large peak within one bacteria, with a relatively low substrate affinity
week after incorporation of a grass–clover mixture and high half saturation constant. Copiotrophic
in soil (Zelenev et al., 2004). CFUs oscillated over time in a wave-like fashion
Various mechanisms underlying oscillations in after amendment of the soil, whereas in the non-
microbial populations could be envisaged. Ecolo- amended soil the CFUs fluctuated only very
gists would first of all think of predator–prey slightly (Figure 1). Microbial communities were
interactions. Total bacterial-feeding nematodes did characterized daily by determining DGGE pro-
not oscillate, but increased monotonously in the files using eubacterial primers, FAME composi-
second week after grass–clover incorporation into tion, and physiological profiles (Biolog, Hayward,
the soil (Zelenev et al., 2004). Daily changes in active CA, USA) on mixtures of copiotrophic colonies
numbers of bacterial-feeding nematodes did oscil- removed from agar plates (de Vos and van
late with a frequency similar to that of bacterial Bruggen, 2001). The patterns of DGGE bands
oscillations due to intermittent activation of the (Figure 2), fatty acid composition and Biolog
dormant juveniles (Dauerlarvae). However, the re- profiles indicated a succession in taxonomic and
sponse of bacterial-feeding nematodes was still too functional groups over time. Discriminant analy-
slow to explain the decline after the first peak in sis of the DGGE band intensities, percentages of
bacterial populations within two days after the individual fatty acids, and intensities of physio-
incorporation of grass–clover material (Zelenev logical reactions on Biolog plates (Figure 3)
et al., 2004). Similarly, protozoa were likely to be showed that there were repetitive cycles in the
too slow to be responsible for the first decline in succession of bacteria over time: communities at
bacterial populations, suggesting that bacteria ini- times when CFUs increased were more similar to
tiate the oscillations, and that their predators follow each other than to those when CFUs decreased
suit (Zelenev et al., 2004). During the experiments and vice-versa (de Vos and van Bruggen, 2001).
with grass–clover amended soil, various chemical In an attempt to relate amplitudes and periods of
and physical parameters were measured, such as the oscillations (representing stability and resilience
ammonium and nitrate concentrations, pH, and of the soil ecosystem) to soil health, grass–clover
redox potential. None of these parameters oscillated mixtures were added to c-irradiated and non-irra-
over time (Zelenev et al., 2004). In a simulation diated soils, a filtered (0.8 lm) soil suspension was
model, bacterial populations started to oscillate due added to the irradiated soil, and microbial popu-
to a temporary shortage of easily available substrate lations were enumerated daily. In the c-irradiated
(Zelenev, 2004). Indeed, substrate availability is a soil the amplitudes and periods of the wave-like
plausible explanation for initiation of the oscilla- fluctuations in microbial communities in response
tions. Yet, local oxygen deprivation after intensive to the disturbance by grass–clover were larger than
bacterial growth has not been excluded but remains those in the non-irradiated soil, supporting the
as a potential mechanism underlying the initiation notion that non-irradiated soil is healthier (Zelenev
of bacterial oscillations. et al., 2004). The amplitudes of microbial popula-
Another aspect of the mechanisms underlying tions were also generally higher in conventionally
bacterial oscillations is whether all taxa oscillate than in organically managed soils (unpublished
simultaneously, or if each peak represents a dif- results) and higher in a fallow soil than in a cover-
ferent microbial community corresponding with cropped soil after addition of the same amount of
different organic components that are decom- cover crop plant material (van Bruggen and
posed subsequently, or if there are repetitive Semenov, 2000). Thus, the amplitude and period of
successions within each peak. This question was microbial oscillations after a disturbance may in-
addressed in another time-series experiment with deed be good indicators for soil health (van Brug-
and without grass–clover incorporated into a gen and Semenov, 2000; Orwin and Wardle, 2004).
 109

100
CFU, million g-1 dry soil
80

60

40

20

0
0 2 4 6 8 10
Days after Grass-Clover incorporation
GC Incorporation Control

Figure 1. Wavelike fluctuations in numbers of copiotrophic bacteria isolated from sandy soil zero to nine days after incorporation
of a grass–clover (GC) mixture and the relative stable numbers in the non-amended control soil.

Soil health is also frequently associated with vetch/oats cover crop debris in fallowed versus
limited disease outbreaks (Rapport, 1995; van cover-cropped soil resulted in temporal fluctuations
Bruggen and Semenov, 2000), and indeed, root in copiotrophic bacterial CFUs over the next five
disease suppression is generally greater in non- weeks (van Bruggen and Semenov, 1999, 2000), and
irradiated than in c-irradiated soil (Workneh and in similar fluctuations in damping-off incidence of
van Bruggen, 1994), in natural than in agricultural tomato seedlings (caused by naturally occurring
soil (Ko, 1982), and in organic than in conventional Pythium ultimum and Pythium aphanidermatum) in
agricultural soil (van Bruggen, 1995; van Bruggen soil samples taken daily from the same experiment
and Termorshuizen, 2003). This leads to the fol- (Figure 4). The oscillations showed similar periods
lowing questions: do plant pathogens also fluctuate but were shifted in time: disease incidence increased
in soil after a disturbance, and are the amplitudes when copiotrophic – and possibly antagonistic -
greater in less healthy soils? Incorporation of populations decreased. It would be interesting to

Figure 2. DGGE patterns of PCR products derived from DNA from copiotrophic bacterial colonies. The numbers represent the
number of days after the incorporation of a grass/clover mixture in soil (GC) and in a non-amended control soil (CO). M repre-
sents a set of eubacterial marker strains. The urea/formamide denaturing gradient was between 40% and 48%. Note that the bac-
terial composition of CO does not change in time, while the composition of GC changes over time, the composition being similar
after 0, 5 and 9 days and different on the other days.
110

4 extensively (Schippers and van Vuurde, 1978; van

Vuurde and Schippers, 1980; Scott et al., 1995;
Kim et al., 1997; Semenov et al., 1999). High
2
microbial densities have generally been observed
close to the root tip and in middle and upper
Can2

0 sections of the roots, and patterns in microbial

-15 -10 -5 0 5 10 15 density along roots have been thought to be a di-
rect reflection of patterns of exudation and
-2
sloughing off of cortex cells (Rovira, 1973; Schip-
pers and van Vuurde 1978; van Vuurde and
-4 Schippers, 1980; McCully and Canny, 1985). This
Can1 is a rather static viewpoint in which growth and
Increasing CFUs Decreasing CFUs Control
death of microbial populations is not explicitly
Figure 3. Discriminant analysis of the Biolog data of the considered. After a series of experiments on the
copiotrophic bacterial population zero to nine days after distribution of microbial populations along roots,
grass–clover incorporation into a sandy soil and in the con- we arrived at a very different and dynamic concept
trol. All samplings were done in duplicate; daily data were
compared to the data on the next sampling day to determine
of microbial community development in the rhi-
whether CFUs were increasing or decreasing. The clustering zosphere, namely that bacterial communities re-
of increasing and decreasing data shows that there are differ- spond to the influx of nutrients from a root tip
ent, repetitive, stages in the succession of bacteria in time. with growth and death cycles at any location
where the root tip passes, resulting in wave-like
investigate if the decline in bacterial CFUs and the patterns along each root (Semenov et al., 1999;
increase in Pythium infections were associated with van Bruggen et al., 2000; Zelenev et al., 2000).
a decrease in oxygen availability. In the above-mentioned experiments, wheat
plants (Triticum aestivum L.) were grown in 60 or
Spatial wave-like fluctuations of microbial 90 cm long root observation boxes with soil high
populations in the rhizosphere or low in fresh organic matter. After two, three,
and five weeks, 2 cm root sections were cut at 4 cm
Distribution patterns of microbial populations intervals. Copiotrophic and oligotrophic bacteria
within root systems have been investigated were isolated from the rhizosphere and corre-

500 40
copiotrophic bacteria
400 Pythium damping off
CFUs g-1 soil (millions)

Percent damping-off

30

300
20
200

10
100

0 0
-10 0 10 20 30 40
Days after CC incorporation

Figure 4. The CFUs of copiotrophic bacteria and the percentage damping-off of tomato seedlings caused by Pythium ultimum and
Pythium aphanidermatum naturally occurring in soils collected one day before, one day after and one, two, three and five weeks
after incorporation of a vetch/oats cover crop (after van Bruggen and Semenov, 2000).
 111

sponding bulk soil on carbon-rich and carbon- any point where the nutrient source passes result-
poor media, respectively (Semenov et al., 1999). ing in waves in space (Figure 5).
For the first time, wave-like distributions of bac- Not only total bacterial communities, but also
terial populations were demonstrated along plant individual bacterial strains exhibit wave-like fluc-
roots using harmonics analysis (Semenov et al., tuations along roots. The biocontrol agent Pseu-
1999). Peaks in oligotrophic populations were domonas fluorescens 32-gfp, marked with the green
slightly shifted upwards on the root compared to fluorescent protein gene, was added to soil samples
those of copiotrophic populations, indicative of from neighbouring conventional and organic farms
the possibility that oligotrophs would follow at Heelsum, the Netherlands, and re-isolated from
copiotrophs in a succession starting from the tip. the rhizosphere along the total length of wheat
There were no (cross) correlations of either bac- roots after three weeks of growth. Both CFUs on
terial group with number of mature lateral roots selective media and fluorescent microscopic counts
per section, or with concentrations of soluble total oscillated significantly and similarly along the
organic carbon (TOC) in the rhizosphere length of the roots (Semenov et al., 2004). The
(Semenov et al., 1999). The oscillations shifted oscillations had a much greater amplitude and
from week to week, and were justifiably called period in the conventionally than in the organically
‘moving waves’. managed soil (Figure 6). In the last soil, P. fluo-
To ascertain that the spatial pattern in microbial rescens 32-gfp populations were zero towards the
populations was not related to lateral root for- root tip. The reason was the lower survival of P.
mation, we did an experiment with an artificial fluorescens 32-gfp in the organically than in the
nutrient source moving through soil. A tube, conventionally managed soil, presumably due to
through which a solution with sugars and amino more intense competition in the organic soil.
acids was pumped, was pulled at a speed of 1 or Apparently, when root tips reached a depth of 10–
4 cm per day through a dialysis sleeve buried in 35 cm below soil level, the majority of the intro-
soil. This experimental setup gave the expected
wave-like patterns in bacterial populations along
the path of the moving nutrient source similar to 107 CFUs g-1 soil
the patterns in real rhizospheres along wheat roots 3
(van Bruggen et al., 2000). Oscillations in space
2
were transformed to oscillations in time, taking the
moving rate of the tube into account. This resulted 1
in oscillations with similar periods, regardless of 0
the moving rate of the tube, indicating that the 0 20 40 60 80 100
periods are dictated by growth and death rates of 3
the bacteria, not by the growth rate of a root. 2
These experiments led to the so-called moving-
1
wave hypothesis for bacterial populations in the
rhizosphere: ‘Waves originate from bacterial 0
0 20 40 60 80 100
growth on nutrients from the root tip, followed by
death when nutrients become exhausted and re- 3
growth from recycled carbon sources plus sub- 2
strate from soil organic matter’. This hypothesis 1
was visualized by means of the results of a simu-
0
lation model (Zelenev et al., 2000). We envisage 0 20 40 60 80 100
the following scenario. As the root tip moves into Distance from the root base (cm)
bulk soil, releasing nutrients, dormant bacteria
(and probably fungi) are activated, grow, and then Figure 5. Moving waves: Oscillations in numbers of copio-
trophic bacteria along the root in distance from the root base
die as nutrients become exhausted; dead bacteria
after two, three and four weeks of growth of wheat. With the
lyse and a new generation grows on recycled growing root tip (to the right of each graph) the population
nutrients (plus additional substrate from soil and moves to a maximum, followed by a harmonic iteration of
roots). Thus, there are growth and death cycles at minima and maxima.
112

1200
CFUs g-1dry soil (millions)
1000

800

600

400

200

0
0 10 20 30 40
Distance from root tip (cm)
conventional, exp. data conventional, harmonics
organic, exp. data organic, harmonics

Figure 6. Experimental data and harmonic fluctuations of a GFP-labelled strain of biocontrol agent P. fluorescens (CFUs/g)1dry -
soil) along the root of wheat in a conventionally and an organically managed sandy soil plotted against distance (cm) from the
root tip. The conventional soil had a higher available N and K content and a higher pH than the organic soil. Organic C contents
were similar. The contribution of the harmonics to the total variance was 75.6% for the conventional soil and 82.5% for the
organic soil. The amplitudes of P. fluorescens in the organic soil were lower than in the conventional soil and P. fluorescens could
not be detected any more around the organic root tip, probably due to a reduced survival in the organic soil compared to the con-
ventional soil.

duced cells had died, so that no cells or CFUs were opposite phase with CFU oscillations (26% of
detected in this region at the time of sampling. the total number of bands, representing only
Furthermore, we investigated if the bacterial 25.0% of the total band intensity), and remaining
communities fluctuated as a whole along the bands whose intensity showed no relationship
wheat root or whether there is a succession in with CFU oscillations or that were restricted to
bacterial composition from peak to peak or certain root zones (54.8% of the total number of
within peaks (van Diepeningen, pers. comm.). bands, representing only 37.5% of the total band
Therefore, rhizosphere microbial communities intensity). Discriminant analysis of the bacterial
along roots of wheat were studied in detail (20–25 populations in root sections with increasing and
rhizosphere and bulk soil samples along the total decreasing phases in the oscillations showed that
root length) by colony enumeration and DGGE the community compositions of waxing popula-
analysis of extracted DNA in the same organic tions are more similar to each other than to those
and conventional soils used for the experiments of waning populations, especially in convention-
with P. fluorescens 32-gfp. Similar to our previous ally managed soil (Figure 7). Again the succes-
findings, the numbers of copiotrophic and oligo- sion appeared to be cyclic, in space as well as
trophic bacteria oscillated with significant har- over time.
monics along each root, independent of soil Two measures of bacterial biodiversity in soil,
moisture or lateral roots. The oscillations and species richness S and the Shannon index H, were
rhizosphere effects were more pronounced in the calculated based on the DGGE data. Both
conventionally managed soil. For amplified eu- biodiversity measures oscillated with significant
bacterial 16S rDNA fragments from DGGE harmonics along the root in opposite phase to
analysis three different groups could be distin- total bacterial CFUs. The bacterial diversity along
guished: those fluctuating in intensity in phase the root was negatively correlated with the num-
with CFU oscillations (19.2% of the total num- bers of oligotrophic and copiotrophic bacterial
ber of bands, representing 37.4% of the total CFUs in the conventional soil and with oligo-
band intensity); those fluctuating in intensity in trophic bacterial CFUs in the organic soil
 113

Conv decreasing
Can 2

Conv increasing
-10 0 10 Org decreasing
Org increasing

-5
Can 1
Figure 7. Discriminant analysis of populations in the increasing and decreasing parts of the wave-like oscillations in the rhizo-
sphere of wheat, in an organic and conventional sandy soil. Soil samples originated from neighbouring organic and conventional
farms. The conventional soil had a higher available N and K content and a higher pH than the organic soil. Organic C contents
were similar.

(Table 1). This indicates that a limited number of spatial–temporal distributions were described as
fast growing taxa were growing and dying over ‘running waves’ of bacteria literally moving
time and in space. towards the root surface (Kozhevin, 1989). In
Besides fluctuations in the vertical direction seawater, chemotactic bacteria were shown to oc-
along the root, running waves of bacterial popu- cur in concentric spheres with alternating higher
lations have also been observed in the horizontal and lower bacterial densities around point sources
direction away from the root surface. Kozhevin of diffusing nutrients, forming wave-like patterns
(1989) observed fluctuations in cells of introduced both in space and over time at scales of a few lm
Bradyrhizobium japonicum 1021 perpendicular to and seconds. These patterns were attributed to the
the length of soybean roots (Glycine max) at a combined effects of molecular diffusion of the
microscopic scale (up to 1 mm from the root sur- attractant, congregation and subsequent dispersal
face), using immunofluorescence. The pattern of of the motile bacteria (Blackburn et al., 1998), and
the oscillations shifted over time, and these were simulated by nonlinear diffusion-reaction
models. The running waves observed by Kozhevin
(1989) may also be the result of diffusion-reaction
Table 1. Cross-correlation coefficients (CCF) between oligo- mechanisms.
trophic and copiotrophic bacterial CFUs at various depths in
As mentioned by Kozhevin (1989), there must be
the rhizosphere of wheat grown in conventional and organic
soil, and band intensity in DGGE gels of amplified 16S a connection between bacterial oscillations in space
rDNA fragments from DNA isolated from corresponding rhi- and time. Indeed, we demonstrated that spatial
zophere samples wave-like fluctuations of microbial populations
along the path of a moving nutrient source could be
CCF laga
transformed to spatial moving waves by taking the
Conventional soil Oligotrophic CFUs )0.552b 2 rate of root growth into account (van Bruggen
Copiotrophic CFUs )0.635 2 et al., 2000). The connection between spatial and
Organic soil Oligotrophic CFUs )0.466 )1 temporal oscillations was used to create a simula-
Copiotrophic CFUs n.s.c
tion model to describe and predict microbial
Soil samples originated from neighbouring organic and con- dynamics in the rhizosphere (Zelenev et al., 2000).
ventional farms. The conventional soil had a higher available N This model could also be used to predict the dis-
and K content and a higher pH than the organic soil. Organic C tribution of infections by pathogens in a root sys-
contents were similar.
a tem, since infection could possibly take place more
One lag corresponds to 1.5 cm.
b
Significant at P=0.05. easily when microbial abundance and activity
c
Not significant. decline, at the waning phases of microbial waves.
114

Wave-like fluctuations of plant pathogens week before isolation of the pathogens from root
in the rhizosphere sections. For comparison of peaks in infection
with those in bacterial populations, the bacterial
The occurrence of microbial growth and death curves were shifted 14 cm to the right since the
cycles at any point along a root could have root tip moved down 2 cm day)1 during the week
important consequences for infection by plant since the bacterial populations had been assessed.
pathogens. Infections by plant pathogenic fungi Both pathogens oscillated in a different phase
are rarely uniformly distributed in the root sys- relative to the bacterial oscillations. There were
tem. Some fungi preferentially infect in the negative correlations between densities of copio-
vicinity of the root tip, while others infect pri- trophic bacteria and R. solani infections at 0 cm
marily older sections of the roots. Given that lag, while there were positive correlations between
there are wave-like patterns of saprotrophic copiotrophic bacteria and Pythium infections at a
microbial populations in space and time in the lag of 6 cm (Figure 8). Infection by R. solani was
rhizosphere along roots, and that pathogens often probably inhibited when large bacterial popula-
have a saprotrophic phase before infecting a host, tions were encountered on the root surface at the
it is likely that there are also wave-like patterns in time of inoculation. It is not known when P. ul-
root infections. Indeed, when sclerotia of Rhi- timum infection took place, but possibly a few
zoctonia solani were placed uniformly along the days after passing of the root tip, when the first
total length of wheat roots growing in root wave of copiotrophic bacteria was already
observation boxes, the proportions of root sec- declining. This shift in Pythium infection relative
tions (of eight roots) from which R. solani were to the first peak in copiotrophic bacteria after
isolated showed wave-like fluctuations when passing of the root tip is similar to the shift in the
detrended data were plotted versus distance from Pythium damping-off peak relative to the first
the root tip (van Bruggen et al., 2002). Similarly, peak in bacterial CFUs after grass–clover incor-
the proportions of root sections from which poration in soil, as discussed in a previous section
naturally occurring Pythium ultimum was isolated of this paper.
were distributed in a wave-like fashion along the In addition to the oscillations in root infections
root (Figure 8). The first peak in Pythium infec- along the length of a root, there is the probability
tions was closer to the root tip than that of R. of root infection by plant pathogens located at
solani (van Bruggen et al., 2002). In the same increasing distances from the root surface that
experiments, copiotrophic bacteria were enumer- can fluctuate in space. The probability of infec-
ated at the time of inoculation with R. solani, one tion generally declines with perpendicular

0.3 5.0E+07

0.2 4.0E+07
Proportion infected

CFU g-1 dry weight

0.1 3.0E+07

0 2.0E+07
0 20 40 60 80 100 120
-0.1 1.0E+07

-0.2 0.0E+00

-0.3 -1.0E+07
Distance from the root tip (cm)
Pythium Copiotrophic bacteria

Figure 8. Harmonic fluctuations of the proportion of natural Pythium infection and the numbers of copiotrophic bacteria along
the wheat root.
 115

distance of propagules from the root, but the Exudation from root tips leads to creation of
decline is generally not monotonous (Mol and waves both in the vertical direction (macro-waves
van Riessen, 1995; Gilligan and Bailey, 1997). along the root) and in the horizontal direction
Gilligan and Bailey (1997) placed propagules of (micro-waves perpendicular to the root).
R. solani at 1–4 mm intervals for a maximum The reasons for fluctuating probabilities of
distance of 15 mm horizontally from the host infection with distance of propagules from the root
(radish seed) and calculated probability of infec- are not clear. If propagules of a pathogen are
tion with distance from the host. Close to the randomly or regularly distributed in space and a
host, there was frequently a small peak in the root passes releasing exudates, then hyphae grow-
probability of infection, followed by a decline in ing towards the roots arrive at different distances
this probability with distance from the host. After from the tip depending on their original distance
detrending and estimating some missing data by from the host. Upon arrival at the root surface they
linear interpolation, we re-analyzed their data encounter more or less bacteria, depending on the
with harmonics analysis. Most of the data pre- phase of bacterial waves, so that they have a lower
sented by Gilligan and Bailey (1997) actually or higher chance of infecting the host. This would
declined in a wave-like fashion with distance from result in fluctuating probabilities of infection with
the host (Figure 9). Harmonics analysis showed distance from the root, and waves of infection
that there were significant waves in the horizontal along the length of the root.
direction, similar to the waves found for B. ja- Alternatively, if there are waves of nutrients
ponicum by Kozhevin (1989). Similarly, the data moving into the rhizosphere as a result of waves in
of Mol and van Riessen (1995) on the probability substrate utilization and release at the root surface
of infection by Verticillium dahliae seemed to or due to a day–night rhythm in exudation, the
decrease in a wave-like fashion with increasing probability of infection may also fluctuate. If
distance from the root surface of various crop zoospores, for example, are pulled three steps for-
species and potato cultivars; unfortunately, the ward and one backward as the waves pass, syn-
published data were not suitable for harmonics chronization of zoospores would take place, so that
analysis. they arrive in waves at the root surface just behind
The realization that wave-like patterns of sap- the root tip. These horizontal dynamics of patho-
rotrophic and pathogenic microorganisms occur gens would result in waves of infection along the
both in the horizontal and vertical direction in the root as the root tip moves on (provided that the
rhizosphere, led to a new view of root infection in root tip does not die from Pythium infection).
relation to microbial population dynamics. Al- Thus, the occurrence of microbial growth and
though not demonstrated, it is likely that wave-like death cycles at any point along a root does indeed
distribution patterns of the microbial community seem to have important consequences for infection
are generated along any roots, including lateral by plant pathogens along the root.
roots, initiated from the growing root tips.

Management of soil-borne disease taking microbial

0.2 oscillations into account
Experiment Harmonics
Probability of infection

0.1 Based on the premise that microbial communities

in healthy soils have strong resistance and resilience
0
against disturbances and suppress disease out-
0 5 10 15 breaks (Rapport, 1995), we would need to manage
-0.1
microbial communities so that the amplitudes of
the oscillations (resistance) and the time to return
-0.2
to quasi-stationary conditions (resilience) are
Distance inoculum-host (mm) minimized. To accomplish this, we would need to
Figure 9. The probability of infection by Rhizoctonia solani in enhance the biological buffering capacity of a soil
relation to distance from the host. Detrended values are by enhancing the background level of microbial
shown (derived form Gilligan and Bailey, 1997). activity and food web complexity, for example by
116

regular additions of relatively stable organic matter 1994; Workneh and van Bruggen 1994; Mäder
to soil. On the other hand, we may want to stim- et al., 2002; Schjønning et al., 2002; van Diepen-
ulate a specific part of the microbial community to ingen et al., 2005). It would also enhance suppres-
antagonize certain plant pathogens by applying sion of many soil-borne pathogens (van Bruggen,
appropriate external disturbances to stimulate 1995; van Bruggen and Termorshuizen, 2003). For
oscillations and succession, for example by tillage. example, densities of Phytophthora and Pythium
In this paragraph, we will only discuss those propagules in soil were lower and those of the
management strategies that influence short-term antagonist Trichoderma higher in soils amended
oscillations of microbial populations, and sapro- with various organic materials (composted cotton-
trophic growth and infection by plant pathogens. gin trash, composted yard waste, or cattle manure)
Management practices that affect long-term pop- than with synthetic fertilizer (Bulluck et al., 2002).
ulation dynamics of microbial communities, such Wave-like responses of these pathogens to intro-
as crop rotation, will not be discussed here. duction of the organic materials were not investi-
gated.
Considering the reaction to a disturbance by
Organic matter management incorporation of a winter cover crop or weeds into
soil, sowing of a subsequent crop needs to be timed
Soil organic matter has generally not been man- so that the inoculum of a facultative saprotrophic
aged explicitly by conventional farmers. Addition pathogen is not at its peak at that time. The
of synthetic fertilizers over many years has resulted quality of the organic matter in terms of easily
in loss of organic matter by stimulating decompo- available substrate and the C:N ratio, and the
sition of native soil organic matter and enhancing activity of the microbial community will determine
microbial respiration. Combined with the negative if facultative saprotrophic pathogens can multiply
effects of synthetic pesticides on components of the in this material. Pathogens such as Pythium and
microbial communities and associated food webs, Rhizoctonia species multiply easily in fresh sub-
conventionally managed soils are generally bio- strate, and may cause serious damping-off prob-
logically impoverished. Any addition of substrate lems when a crop is sown within three to four
for microbial growth such as crop residues, dead weeks after incorporation of fresh plant material
organisms after tillage, or root exudates, will result (van Bruggen and Semenov, 2000). For example,
in large oscillations in microbial populations, when a mixture of vetch and oats was incorpo-
including fast-growing plant pathogens with sap- rated in soil that had been fallow or had been
rotrophic abilities such as Pythium spp. (van cover-cropped, damping-off of tomatoes by
Bruggen and Semenov, 2000). Pythium species was most severe seven days after
Large pools of mineral nitrogen in soil may even incorporation of the plant material, and five days
exacerbate the fluctuations. Soil and plant nitro- after the first peak in copiotrophic bacteria
gen concentrations can have a profound effect on (Figure 4). The peaks in bacterial CFUs and
both ecosystem health and disease severity: high damping-off incidence were higher in the previ-
levels of nitrogen in the soil, particularly in the ously fallowed soil than in the cover-cropped soil,
form of nitrate, may enhance several fungal dis- indicating that the cover-cropped soil was more
eases (Workneh et al., 1993; van Bruggen, 1995; stable (van Bruggen and Semenov 1999, 2000). In
Tamis and van den Brink, 1998, 1999; Clark et al., another study where a vetch–oats cover crop was
1999). incorporated in organically and conventionally
Several things can be done to restore the bio- managed soils, in vitro growth of P. aphanidermatum
logical buffering capacity and enhance internal peaked after 7–10 days while that of R. solani
nutrient cycling in soil. Regular addition of fairly peaked after 21–35 days (Grünwald et al., 1997,
stable organic matter, including solid animal 2000). Microbial measurements were generally
manure, composts of plant and animal origin, and lower and in vitro growth of the pathogens higher
lignified roots of deep-rooted plants such as alfalfa, in the conventionally compared to the organically
rye or grass–clover would enhance microbial bio- managed soils, but these differences were tempo-
mass, activity, and diversity, and food web com- rarily nullified after cover crop incorporation
plexity in soil (Sivapalan et al., 1993; Ryan et al., (Grünwald et al., 2000).
 117

Different from the effects of cover crop of soil samples collected from a grassland and a
incorporation, high-nitrogen-containing organic vegetable field. Such a disturbance may also give
amendments such as cattle and poultry manure or facultative saprotrophic pathogens a chance to
soy meal had an immediate suppressive effect on grow, but wave-like fluctuations may be dampened
several root pathogens and nematodes, as a result sooner than in the case of a disturbance by fresh
of ammonia release immediately after initiation of plant materials. A report on a lower incidence of
microbial decomposition (Lazarovits et al., 2001). Pythium damping-off of sugar beet in a farm with
Bulluck et al. (2002) also documented immediate reduced tillage than in a conventional farm with
suppressive effects of various types of compost and regular tillage (El Titi and Richter, 1987) is in
cattle manure applied at moderate to high rates on agreement with the notion that reduced tillage
southern blight of tomatoes. However, these decreases the chance that Pythium would grow
materials may not be suppressive to Pythium, explosively in fresh substrate after tillage. On the
which thrives well under high ammonium concen- other hand, pathogens that survive in stubble could
trations (van Bruggen and Semenov, 1999). become problematic in no-till fields. Roget (1995)
After repeated applications of organic materials, demonstrated that after conversion from regularly
higher organic matter and microbial activity in tilled to no-till wheat production Rhizoctonia root
bulk soil would result in a ‘masking of the rhizo- rot increased in the first few years. However, this
sphere effect’ (Gilbert et al., 1994), reduce the increase was followed by a decline in Rhizoctonia
microbial oscillations along roots, and limit sub- root rot after about five years of no-till.
strate concentrations seeping into soil, thereby
reducing the attraction of root pathogens to the (Partial) soil sterilization
root surface and decreasing the chance of infection
by many pathogens. Reducing the amount of It is well known that soil-borne plant pathogens can
easily available mineral nutrients and soluble car- wreak havoc when introduced into steamed green-
bon compounds by reducing fertilizer applications house soil or fumigated field soil due to the exis-
and the addition of stable carbon sources, would tence of a biological vacuum (Bollen, 1974;
lead to oligotrophication. This promotes mycor- Kreutzer, 1965). Any disturbance of a recently
rhizal infections, which can also suppress various sterilized (or c-irradiated) and re-colonized soil
root diseases (Sharma et al., 1992; Ryan et al., leads to wild fluctuations in microbial populations
1994). Whether mycorrhizal infections also occur (Zelenev et al., 2004), and may lead to similar
in wave-like patterns along the roots is not known. fluctuations in facultative saprotrophic plant
pathogens. A good alternative to soil sterilization
Tillage may be biological soil disinfestation (Blok et al.,
2000), which does not result in enhanced disease
No-till or reduced tillage has been promoted in pressure when pathogens are re-introduced, and
recent years primarily to reduce soil erosion. provides long-lasting disease control (Goud et al.,
However, tillage practices also have pronounced 2004).
effects on survival of fungi and micro- and macro-
fauna in soil. Deep tillage can enhance the bacteria Mixed cropping
to fungi ratio and eliminate predatory nematodes,
affecting especially the k-strategists (Berkelmans Mixed cropping – a system where two or more
et al., 2003). No-till or reduced tillage is often crops are grown in the same field – can enhance
associated with higher microbial biomass and food web diversity and decrease severity of
activity and a more complex food web in the upper foliar plant diseases (Finckh and Wolfe, 1997).
soil layers compared to regular tillage, i.e. plowing Although positive correlations between above-
(van Diepeningen et al., 2005). ground and below-ground biodiversity have
Tillage is a form of disturbance resulting in clear seldom been demonstrated (De Deyn et al., 2004),
fluctuations in microbial activity and biomass. suppression of root disease (Burdon and Chilvers,
Caldéron et al. (2000) showed clear fluctuations in 1976; Villich, 1993) and enhanced soil microbial
microbial biomass during the first eight days after diversity in mixed cropping systems have some-
simulated tillage in the laboratory, namely mixing times been found (G.A. Hiddink, pers. comm.).
118

Microbial composition in the rhizosphere is successful under controlled environmental condi-

strongly dependent on plant species (Smith et al., tions using simplified potting mixes presumably
1999). Thus in a mixture of roots of different low in microbial diversity (Fravel, 1999). More-
species it may be more difficult for a pathogen to over, inoculation of soil with a single strain of a
find its host and the saprotrophic phase of biocontrol agent rarely leads to a high level of
pathogens may be limited by a greater variety of protection and often the positive effect is incon-
antagonists. sistent (Weller, 1988; Koch, 1999). Better results
have been obtained with combinations of strains
Cultivar selection or species (e.g. Pierson and Weller, 1994; Guetsky
et al., 2001, 2002; Szczech and Shoda, 2004).
Choice of crops and cultivars will influence the These results with biocontrol agents can now be
microbial communities that are selectively enhanced interpreted in view of the general occurrence of
or suppressed in the rhizosphere by the quality of microbial oscillations in time and space in the
root exudates (Grayston et al., 1998; Kowalchuk rhizosphere. Introduced biocontrol agents are
et al., 2002; Garbeva et al., 2004). Differential likely to oscillate similarly to the native soil
interactions between plant genotypes and benefi- microbial communities. Densities of Pseudomonas
cial microorganisms have been demonstrated for fluorescens introduced on wheat seed seemed to
species of mycorrhizal fungi, rhizobia, and general form wave-like patterns along the length of the
plant growth-promoting rhizobacteria (PGPR) root, the amplitudes tapering off towards the root
(Smith et al., 1999). PGPR have been used as tip (Scott et al., 1995). In an experiment in our
biofertilizers and biological control agents (Ger- laboratory with GFP labelled, phloroglucinol-
mida, 1996). They can be directly antagonistic producing P. fluorescens mixed into soil we con-
towards plant pathogens or can stimulate systemic clusively proved a wave-like distribution of this
induced resistance in the plant (Kloepper et al., bacterium along growing wheat roots (Figure 6),
1997). similar to the oscillations of native bacterial pop-
Besides exudate quality, exudation rates can ulations (Semenov et al., 1999). We also showed
also vary per cultivar, and these rates determine to that there is a succession in microbial communities
a large extent the amplitude of the ensuing within each wave, repeating from wave to wave.
microbial oscillations. However, plant breeders Thus, microorganisms that may be good antago-
have generally not taken exudation rates and nists in vitro, may take a different position in the
exudate quality into account. There would be a succession compared to the target pathogen, and
great opportunity to select cultivars for their may therefore not be effective as biocontrol agents.
ability to stimulate specific microbial communities This might mean that biocontrol can only be
that can contribute to disease suppression. accomplished if waves in populations of the bio-
control agent coincide more or less with potential
Biological control waves in pathogen populations (unless there is
systemic induced resistance). Potential biocontrol
Many microorganisms have been found with bio- agents may need to be selected so that their pop-
logical control potential against various plant ulations are maximal in the region along the root
pathogens. Biological control agents may use a where the target pathogen invades the root. Dif-
variety of inhibitory and suppressive mechanisms: ferences in succession and position along the root
(1) competition for resources and space, (2) anti- may also explain the greater success of biocontrol
biotic production, (3) removal of pathogenicity mixtures than of single biocontrol agents (Guetsky
factors produced by the pathogen, (4) production et al., 2001, 2002; Szczech and Shoda, 2004).
of degrading enzymes that target the pathogen and In organically managed soil with high micro-
(5) the induction of resistance in the host plant bial diversity and activity (and therefore low
(Whipps, 2001). However, many biological control concentrations of easily available nutrients)
agents perform poorly under field conditions introduced biocontrol agents may survive less
(Fravel, 1999) and only few biocontrol species compared to biologically impoverished conven-
have been registered for field use (Copping, 2001). tional soil (van Bruggen et al., 2000). The differ-
Biocontrol of soil-borne pathogens has been more ence between the effectiveness of biocontrol
 119

agents in biologically impoverished soil versus Single species of saprotrophic bacteria, biocon-
biologically diverse and active soils, may be ex- trol agents and phytopathogenic fungi also show
plained by reduced availability of easily decom- wave-like fluctuations in bulk soil and along plant
posable substrate in these latter soils. The roots. Different trophic groups and species may
rhizosphere effect would be masked and microbial fluctuate with different periods and phases. A
oscillations subdued in such soils (van Diepenin- cyclic succession occurs in response to nutrient
gen, pers. comm.). In biologically impoverished input; in the waxing phases of successive oscilla-
soils, high substrate concentrations can be ex- tions, microbial communities are taxonomically
pected where wildly oscillating microbial com- and physiologically more similar to one another
munities are at a minimum; biocontrol agents than to the communities in the waning phases.
with a slightly different niche compared to the This has consequences for the selection of bio-
majority of these oscillating communities may control agents and cultivars. Introduction of a
have a chance to survive and even grow. How- single biocontrol agent to a soil may not lead to
ever, they may not grow in a rhizosphere with the expected results due to wave-like fluctuations
very diverse microbial communities. An intro- in the rhizosphere of the biocontrol agent and the
duced phloroglucinol-producing, gfp marked P. target pathogen, if they are out of phase. A mix-
fluorescens strain declined faster in three organi- ture of biocontrol agents of different trophic
cally managed soils than in three neighbouring groups may be more successful.
conventionally managed soils (van Bruggen et al., The main strategies to control soil-borne dis-
2004). The same strain showed only mild wave- eases can be classified into three categories: (1)
like oscillations along a wheat root in an organ- enhancement of general microbial biomass and
ically managed soil compared to a conventionally diversity resulting in a masking of the rhizosphere
managed soil (Figure 6), and had less effect on effect, a reduction of the amplitude of wave-like
take-all disease of wheat in soils from three or- oscillations and an increase in natural disease
ganic farms than in soils from three neighbouring suppression, (2) removal of dormant propagules or
conventional farms with a lower microbial pathogens in their saprotrophic phase from their
diversity (van Bruggen et al., 2004). Thus, it is food base by stimulating wave-like fluctuations in
questionable if inundative biological control can populations of potentially competitive microor-
be effective in soils with a high microbial bio- ganisms, for example by soil tillage, and (3) aug-
mass, activity and diversity, and low levels of mentation of microbial communities by biocontrol
easily available substrate. agents, which must be able to survive and grow in
the rhizosphere. The first strategy is the main
strategy used by organic farmers, while the second
Conclusions and third strategies are typical for conventional
farms. In all cases, dynamic oscillations of micro-
In this review we showed that populations of dif- bial communities and individual species must be
ferent trophic groups of bacteria develop in a wave- taken into account. This constitutes a new view of
like fashion with repetitive growth and death cycles, plant disease control.
both in time and space after an impulse of readily
utilizable substrate. Oscillatory development of
bacterial populations may be a universal phenom-
enon after a disturbance, which could possibly be Acknowledgements
used to compare soils in terms of stability and
resilience, and consequently soil health. Indeed, the We would like to thank Vladimir V. Zelenev for
amplitudes of the oscillations are smaller and de- carrying out harmonic analyses and Gerbert
cline more quickly in soils with a high microbial Hiddink for sharing his unpublished data. This
biomass, activity and diversity, and low levels of work was partially supported by NWO Russia
easily available substrate. These are characteristics collaborative grant (Dossier number 047.014.001)
of soil health. It is argued that healthy soils are and NATO-Russia Collaborative Linkage Grant
more suppressive to soil-borne plant pathogens (RCLG) ‘Risk analysis of pathogen spread in the
than biologically impoverished soils. vegetable production and processing industry’.
120

References De Deyn GB, Raaijmakers CE and Van der Putten WH (2004)

Plant community development is affected by nutrients and
Aristovskaya TV (1980) Microbiological Processes of Soil soil biota. Journal of Ecology 92: 824–834.
Formation. Nauka, Leningrad (in Russian). De Vos OJ and van Bruggen AHC (2001) Soil microbial
Berkelmans R, Ferris H, Tenuta M and van Bruggen AHC composition as affected by grass–clover mixture incorporation
(2003) Long-term effects of crop management on trophic in the soil. ISME-9, Amsterdam, the Netherlands, P17.077.
levels of nematodes other than plant parasitic nematodes Doebeli M and Ruxton GD (1997) Evolution of dispersal rates
disappear after one year of uniform treatment. Applied Soil in metapopulation models: Branching and cyclic dynamics
Ecology 23: 223–235. in phenotype space. Evolution 51: 1730–1741.
Blackburn N, Fenchel T and Mitchell J (1998) Microscale Doebeli M and Ruxton GD (1998) Stabilization through spatial
nutrient patches in planctonic habitats shown by chemo- pattern formation in metapopulations with long-range
tactic bacteria. Science 282: 2254–2256. dispersal. Proceedings of the Royal Society of London
Blok WJ, Lamers JG, Termorshuizen AJ and Bollen GJ (2000) Series B – Biological Sciences 265: 1325–1332.
Control of soilborne plant pathogens by incorporating fresh Drinkwater LE, Letourneau DK, Workneh F, van Bruggen
organic amendments followed by tarping. Phytopathology AHC and Shennan C (1995) Fundamental differences
90: 253–259. between conventional and organic tomato agroecosystems
Boehm MJ, Madden LV and Hoitink HAJ (1993) Effect of in California. Ecological Applications 5: 1098–1112.
organic matter decomposition level on bacterial species El Titi A and Richter J (1987) Integrierter Pflanzenschutz im
diversity and composition in relationship to Pythium Ackerbau: Das Lautenbach Projekt. III. Schädlinge und
damping-off severity. Applied and Environmental Micro- Krankheiten 1979–1983. Zeitschrift für Pflanzenkrankhei-
biology 59: 4171–4179. ten und Pflanzenschutz 94: 1–13.
Boehm MJ, Wu T, Stone AG, Kraakman B, Iannotti DA, Finckh MR and Wolfe MS (1997). The use of biodiversity to
Wilson GE, Madden LV and Hoitink HAJ (1997) Cross- restrict plant diseases and some consequences for farmers
polarized magic-angle spinning 13C nuclear magnetic and society. In: Jackson LE ) Ecology in Agriculture (pp
resonance spectroscopic characterization of soil organic 203–237) Academic Press, San Diego, CA, USA.
matter relative to culturable bacterial species composition Franke-Snyder M, Douds DD, Galvez L, Phillips JG,
and sustained biological control of Pythium root rot. Wagoner P, Drinkwater L and Morton JB (2001) Diver-
Applied and Environmental Microbiology 63: 162–168. sity of communities of arbuscular mycorrhizal (AM) fungi
Bollen GJ (1974) Fungal recolonization of heat-treated glass- present in conventional versus low-input agricultural sites
house soils. Agro-Ecosystems 1: 139–155. in eastern Pennsylvania, USA. Applied Soil Ecology 16:
Brussaard L, Kuyper TW, Didden WAM, de Goede RGM 35–48.
and Bloem J (2004) Biological soil quality from biomass Fravel D (1999) Hurdles and bottlenecks on the road to
to biodiversity – importance and resilience to manage- biocontrol of plant pathogens. Australasian Plant Pathol-
ment stress and disturbance. In: Schjønning P, Elmholt S ogy 28: 53–56.
and Christensen BT (eds) Managing Soil Quality: Chal- Garbeva P, van Veen JA and van Elsas JD (2004) Microbial
lenges in Modern Agriculture (pp. 139–161). CAB Inter- diversity in soil: Selection of microbial populations by plant
national. and soil type and implications for disease suppressiveness.
Buckley DH and Schmidt TM (2001) The structure of microbial Annual Review of Phytopathology 42: 243–270.
communities in soil and the lasting impact of cultivation. Gerlagh M (1968) Introduction of Ophiobolus graminis into new
Microbial Ecology 42: 11–21. polders and its decline. Agricultural Research reports 713.
Bulluck LR, Brosius N, Evanylo GK and Ristaino JB (2002) Centre for Agricultural Publishing and Documentation,
Organic and synthetic fertility amendments influence soil Wageningen.
microbial, physical and chemical properties on organic and Germida JJ (1996) Use of rhizobacteria as biofertilizers for
conventional farms. Applied Soil Ecology 19: 147–160. enhancing growth and yield of crops. In: Scoles G and
Burdon JJ and Chilvers GA (1976) Epidemiology of Pythium- Rossnagel B (eds) Proceedings V Int. Oat Conf. & VII Int.
induced damping-off in mixed species seedling stands. Barley Genetics Symp. Univ. Extension Press.
Annuals of Applied Biology 82: 233–240. Gilbert GS, Handelsman J and Parke JL (1994) Root camou-
Calderon FJ, Jackson LE, Scow KM and Rolston DE (2000) flage and disease control. Phytopathology 84: 222–225.
Microbial responses to simulated tillage in cultivated and Gilligan CA and Bailey DJ (1997) Components of pathozone
uncultivated soils. Soil Biology and Biochemistry 32: 1547– behaviour. New Phytologist 136: 343–358.
1559. Goud JC, Termorshuizen AJ, Blok WJ and van Bruggen AHC
Clarholm M (1981) Protozoan grazing of bacteria in soil (2004) Long-term effect of biological soil disinfestation on
– impact and importance. Microbial Ecology 7: 343– Verticillium wilt. Plant Disease 88: 688–694.
350. Grayston SJ, Wang S, Campbell CD and Edwards AC (1998)
Clark MS, Horwath WR, Shennan C, Scow KM, Lantni WT Selective influence of plant species on microbial diversity in
and Ferris H (1999) Nitrogen, weeds and water as yield- the rhizosphere. Soil Biology and Biochemistry 30: 369–378.
limiting factors in conventional, low-input, and organic Grünwald NJ, Workneh F, Hu S and van Bruggen AHC
tomato systems. Agriculture, Ecosystems & Environment (1997) Comparison of an in vitro and a damping-off
73: 257–270. assay to test soils for suppressiveness to Pythium
Copping LG (2001) The Biopesticide Manual, British Crop aphanidermatum. European Journal of Plant Pathology
Protection Council, Surrey, UK. 103: 55–63.
 121

Grünwald NJ, Hu S and van Bruggen AHC (2000) Short-term Mol L and van Riessen HW (1995) Effect of plant roots on the
cover crop decomposition in organic and conventional germination of microsclerotia of Verticillium dahliae. I. Use
soils: Characterization of soil C, N, microbial and plant of root observation boxes to assess differences among crops.
pathogen dynamics. European Journal of Plant Pathology European Journal of Plant Pathology 101: 673–678.
106: 37–50. Mulder C, de Zwart D, van Wijnen HJ, Schouten AJ and
Guetsky R, Shtienberg D, Elad Y and Dinoor A (2001) Breure AM (2003) Observational and simulated evidence of
Combining biocontrol agents to reduce the variability of ecological shifts within the soil nematode community of
biological control. Phytopathology 91: 621–627. agroecosystems under conventional and organic farming.
Guetsky R, Shtienberg D, Elad Y, Fisher E and Dinoor A Functional Ecology 17: 516–525.
(2002) Improving biological control by combining biocon- Nitta T (1991) Diversity of root fungal floras: Its implications
trol agents each with several mechanisms of disease for soil-borne diseases and crop growth. Japan Agricultural
suppression. Phytopathology 92: 976–985. Research Quarterly 25: 6–11.
Hoper H and Alabouvette C (1996) Importance of physical and Oehl F, Sieverding E, Ineichen K, Mäder P, Boller T and
chemical soil properties in the suppressiveness of soils to Wiemken A (2003) Impact of land use intensity on the
plant diseases. European Journal of Soil Biology 32: 41–58. species diversity of arbuscular mycorrhizal fungi in agro-
Kim DS, Weller DM and Cook RJ (1997) Population dynamics ecosystems of central Europe. Applied and Environmental
of Bacillus sp. L324-92R12 and Pseudomonas fluorescens 2- Microbiology 69: 2816–2824.
79RN10 in the rhizosphere of wheat. Phytopathology 87: Orwin KH and Wardle DA (2004) New indices for quantifying
559–564. the resistance and resilience of soil biota to exogenous
Kloepper JW, Tuzun S, Zehnder GW, Wei G, Hoffland E, disturbances. Soil Biology and Biochemistry 36: 1907–1912.
Bakker PAHM and vanLoon LC (1997) Multiple disease Pierson EA and Weller DM (1994) Use of mixtures of
protection by rhizobacteria that induce systemic resistance- fluorescent pseudomonads to suppress take-all and improve
historical precedence. Phytopathology 87: 136–138. the growth of wheat. Phytopathology 84: 940–947.
Ko WH (1982) Biological control of Phytophthora root rot of Rapport (1995) Ecosystem health – more than a metaphor.
papaya with virgin soil. Plant Disease 66: 446–448. Environmental Values 4: 287–309.
Koch E (1999) Evaluation of commercial products for micro- Roget DK (1995) Decline in root rot (Rhizoctonia solani AG-8)
bial control of soil-borne plant diseases. Crop Protection in wheat in a tillage and rotation experiment at Avon,
18: 119–125. South Australia. Australian Journal of Experimental Agri-
Kowalchuk GA, Buma DS, de Boer W, Klinkhamer PGL and culture 35: 1009–1013.
vanVeen JA (2002) Effects of above-ground plant species Rovira AD (1973) Zones of exudation along plant roots and
composition and diversity on the diversity of soil-borne spatial distribution of microorganisms in the rhizosphere.
microorganisms. Antonie van Leeuwenhoek 81: 509–520. Pesticide Science 4: 361–366.
Kozhevin PA (1989) Microbial Populations in Nature. Moscow Ryan MH, Chilvers GA and Dumaresq DC (1994) Coloniza-
State University Press 174 p. (in Russian). tion of wheat by VA-mycorrhizal fungi was found to be
Kreutzer WA (1965). The reinfestation of treated soil. In: Baker higher on a farm managed in an organic manner than on a
KF and Snyder WC (eds) Ecology of Soil-borne Plant conventional neighbour. Plant and Soil 160: 33–40.
Pathogens (pp 495–508) University of California Press, Schippers B and van Vuurde JWL (1978). Studies of microbial
Berkeley. colonization of wheat roots and the manipulation of the
Lawlor K, Knight BP, Barbosa-Jefferson VL, Lane PW, Lilley rhizosphere microflora. In: Loutit MW and Miles JAR
AK, Paton GI, McGrath SP, O’Flaherty SM and Hirsch (eds) Microbial Ecology (pp 295–298) Springer-Verlag,
PR (2000) Comparison of methods to investigate microbial Berlin.
populations in soils under different agricultural manage- Schjønning P, Elmholdt S, Munkholm LJ and Debosz K (2002)
ment. FEMS Microbiology & Ecology 33: 129–137. Soil quality aspects of humid sandy loams as influenced by
Lazarovits G, Tenuta M and Conn KL (2001) Organic organic and conventional long-term management. Agricul-
amendments as a disease control strategy for soil-borne ture, Ecosystems & Environment 88: 195–214.
diseases of high-value agricultural crops. Australasian Plant Scott EM, Rattray EAS, Prosser JI, Killham K, Glover LA,
Pathology 30: 111–117. Lynch JM and Bazin MJ (1995) A mathematical model
Liiri M, Setala H, Haimi J, Pennanen T and Fritze H (2002) for dispersal of bacterial inoculants colonizing the wheat
Soil processes are not influenced by the functional com- rhizosphere. Soil Biology and Biochemistry 27: 1307–
plexity of soil decomposer food webs under disturbance. 1318.
Soil Biology and Biochemistry 34: 1009–1020. Semenov AM (2001) Oscillations of microbial communities in
Mäder P, Fliebbach A, Dubois D, Gunst L, Fried P and Niggli soils. Works of All-Russia Conference ‘To the 100-Anni-
U (2002) Soil fertility and biodiversity in organic farming. versary from Birthday of Academician E.N. Mishustin’,
Science 296: 1694–1697. 22.02.2001, MSU, Moscow (pp. 57–72), M. MAKC Press
Mas A, Houwen F and Verstraete W (1996) Agricultural (in Russian).
factors affecting methane oxidation in arable soil. Biology Semenov AM, van Bruggen AHC, van Diepeningen AD, Sayler
and Fertility of Soils 21: 95–102. RJ, Zelenev VV and de Vos O (2004) Wave-like develop-
McCully ME and Canny MJ (1985) Localization of translo- ment of bacterial communities and a biocontrol agent
cated 14C in roots and root exudates of field grown maize. introduced in the wheat rhizosphere. Phytopathology 94:
Physiologia Plantarum 65: 380–392. S94.
122

Semenov AM, van Bruggen AHC and Zelenev VV (1999) Moving van Diepeningen AD, de Vos OJ, Korthals GW and van
waves of bacterial populations and total organic carbon Bruggen AHC (2005) Effects of organic versus conventional
along roots of wheat. Microbial Ecology 37: 116–128. management on chemical and biological parameters in
Sharma AK, Johri BN and Gianinazzi S (1992) Vesicular- agricultural soils. Applied Soil Ecology, in press, available
arbuscular mycorrhizae in relation to plant-disease. World on line.
Journal of Microbiology and Biotechnology 8: 559–563. vanVuurde JWL and Schippers B (1980) Bacterial colonization
Shumway RH (1988) Applied Statistical Time Series Analysis, of seminal wheat roots. Soil Biology and Biochemistry 12:
Prentice-Hall, Englewood Cliffs, NJ, 379. 559–565.
Sivapalan A, Morgan WC and Franz PR (1993) Monitoring Vilich V (1993) Crop rotation with pure stands and mixtures of
populations of soil microorganisms during a conversion barley and wheat to control stem and root rot diseases.
from a conventional to an organic system of vegetable Crop Protection 12: 373–379.
growing. Biological Agriculture and Horticulture 10: 9–27. Weller DM (1988) Biological control of soilborne pathogens in
Smith KP, Handelsman J and Goodman RM (1999) Genetic the rhizosphere with bacteria. Annual Review of Phytopa-
basis in plants for interactions with disease-suppressive thology 26: 379–407.
bacteria. Proceedings of the National Academy of Sciences Whipps JM (1997) Developments in the biological control of
of the United States of America 96: 4786–4790. soilborne plant pathogens. Advances in Botanial Research
Szczech M and Shoda M (2004) Biocontrol of Rhizoctonia 26: 1–134.
damping-off of tomato by Bacillus subtilis combined with Whipps JM (2001) Microbial interactions and biocontrol in
Burkholderia cepacia. Journal of Phytopathology 152: 549–556. the rhizosphere. Journal of Experimental Botany 52: 487–
Tamis WLM and van den Brink WJ (1998) Inventarisatie van 511.
ziekten en plagen in wintertarwe in gangbare, geı̈ntegreerde Workneh F, vanBruggen AHC, Drinkwater LE and Shennan C
en ecologische teeltsystemen in Nederland in de periode (1993) Variables associated with corky root and phytoph-
1993–1997. IPO-DLO Rapport nr. 98-01, Wageningen thora root rot of tomatoes in organic and conventional
Tamis WLM and van den Brink WJ (1999) Conventional, farms. Phytopathology 83: 581–589.
integrated and organic winter wheat production in the Workneh F and van Bruggen AHC (1994) Microbial density,
Netherlands in the period 1993–1997. Agriculture, Ecosys- composition, and diversity in organically and convention-
tems & Environment 76: 47–59. ally managed rhizosphere soil in relation to suppression
vanBruggen AHC (1995) Plant-disease severity in high-input of corky root of tomatoes. Applied Soil Ecology 1:
compared to reduced-input and organic farming systems. 219–230.
Plant Disease 79: 976–984. Zelenev VV (2004) Spatial and temporal fluctuations in
vanBruggen AHC and Semenov AM (1999) A new approach to bacteria, microfauna and mineral nitrogen in response to
the search for indicators of root disease suppression. a nutrient impulse in soil. Ph.D. thesis. Wageningen
Australasian Journal of Plant Pathology 28: 4–10. University, Wageningen, the Netherlands.
vanBruggen AHC and Semenov AM (2000) In search of Zelenev VV, Berkelmans R, van Bruggen AHC, Bongers T and
biological indicators for soil health and disease suppression. Semenov AM (2004) Daily changes in bacterial-feeding
Applied Soil Ecology 15: 13–24. nematode populations oscillate with similar periods as
vanBruggen AHC, Semenov AM and Zelenev VV (2000) Wave- bacterial populations after a nutrient impulse in soil.
like distributions of microbial populations along an artifi- Applied Soil Ecology 26: 93–106.
cial root moving through soil. Microbial Ecology 40: 250– Zelenev VV, van Bruggen AHC and Semenov AM
259. (2000) ‘‘BACWAVE’’, a spatial-temporal model for
vanBruggen AHC, Semenov AM and Zelenev VV (2002) traveling waves of bacterial populations in response
Wavelike distributions of infections by an introduced and to a moving carbon source in soil. Microbial Ecology
naturally occurring root pathogen along wheat roots. 40: 260–272.
Microbial Ecology 44: 30–38. Zelenev VV, van Bruggen AHC and Semenov AM (2005)
vanBruggen AHC and Termorshuizen AJ (2003) Integrated Short-term wave-like dynamics of bacterial populations in
approaches to root disease management in organic farming response to nutrient input from fresh plant residues.
systems. Australasian Plant Pathology 32: 141–156. Microbial Ecology 49: 83–93.
vanBruggen AHC, Hiddink GA, Semenov AV and Semenov AM Zvjagintsev DG and Golimbet VE (1983) Dynamics of micro-
(2004) Suppression of take-all disease in soils from organic bial number, biomass and productivity of microbial com-
versus conventional farms in relation to native and intro- munities in soils. In: Nauka M (ed.) Successes of
duced Pseudomonas fluorescens. Phytopathology 94: S105. Microbiology, no. 18 (pp. 215–231) (In Russian).
European Journal of Plant Pathology (2006) 115:123–138
Springer 2006
DOI 10.1007/s10658-005-0651-z

Patterns and management of crop multiple pathosystems

Serge Savary1,2, B. Mille3, B. Rolland4 and P. Lucas3

1
IRD, Centre INRA, BiO3P, France (E-mail: [email protected]) 2UMR Sante´ des Plantes, Centre
INRA de Bordeaux, 33883, B.P. 81, Villenave d’Ornon, France ; 3INRA, BiO3P, F-35653, Le Rheu cedex,
France ; 4INRA, APBV, F-35653, Le Rheu cedex, France

Accepted 1 July 2005

Key words: agro-ecosystem, epidemiological guilds, guilds of harmful agents, IPM, multiple epidemics,
syndrome of disease, syndrome of production

Abstract

The study of multiple pathosystems has played a central role in the development of botanical epidemiology,
leading to a number of approaches and concepts. Multiple pathosystems are facts, which are experienced by
many non-cultivated, or cultivated, plant communities. The shapes and composition of multiple patho-
systems vary in space and time because of their inherent structure of relationships, and also in response to
management. Examples of variation in multiple pathosystems are given, of groundnut in Côte d’Ivoire, of
wheat in Brittany, and of upland rice in northern Laos. Variation in the yield-reducing effects of multiple
pathosystems is discussed, including interactions among disease elements, relationships with attainable
performances, and linkages with production situations. Progress has been achieved in understanding the
links between injury profiles, production situations, and attainable performances. Questions about
the functioning and consequences of multiple pathosystems are central to defining the scientific bases for,
the design of, and the implementing of IPM. The complexity of multiple pathosystems, however, remains a
deterrent, not a challenge, to many plant pathologists. Progress achieved in designing production systems
for hardy wheat in France, however, is very promising, because of the multidisciplinary science it involves,
and because of the promise to deliver it carries. The concepts of epidemiological guilds and of guilds of
harmful agents are offered as perspectives to address and manage syndromes of production and syndromes
of disease.

Relevance of multiple pathosystems in botanical Studies on multiple pathosystems (i) led to

epidemiology attempts to understand and manage them (e.g.,
Jörg et al., 1987; Daamen et al., 1989; Bastiaans
Multiple pathosystems as a research theme and Daamen, 1994), (ii) resulted in analyses of
case-studies, and efforts dealing with specific cases
The study of multiple pathosystems has played a and contexts (e.g., Hamelink et al., 1988; Avelino,
central role in the development of botanical epi- 1999), and (iii) often were perceived as practical
demiology. As a subject, it is the equivalent in endeavours only. Studies of multiple pathosys-
botanical epidemiology of community ecology in tems, dealing with a complex subject, inviting
general ecology. Research in the field has led to the complex analyses, and leading to complex inter-
development of a body of approaches, often sta- pretations, have arguably led to results that were
tistical and multivariate, as the objective often has difficult to share. Arguably, this type of research
been mostly descriptive, rather than explanatory. often addresses open-ended questions, not specific
124

hypotheses. Its value however is multifold: first, it Over time, e.g., during the cycle of a field crop, a
provides a framework for other, hypothesis-spe- number of diseases may appear, spread, decline,
cific, epidemiological research; second, it allows and interact among themselves and the growing
the description so furthering hypotheses for future crop. Figure 1 shows a series of principal compo-
research; and third, in some successful cases, it has nent analyses on three very different multiple
allowed escape from idiosyncrasies and has gen- pathosystems. Principal component analysis is
erated some useful generalisations. used here as one convenient means to provide a
‘‘No one can be a good observer unless he is a preliminary overview of very complex structures.
good theorizer’’ (Charles Darwin, quoted from For instance, the multiple pathosystem of
Zadoks, 1972). There are very few plant popula- groundnut in Côte d’Ivoire involves a series of
tions that are exposed to one disease only. Multi- fungal pathogens (Savary, 1987a) affecting the
ple pathosystems are where initial observations are foliage (Cercospora arachidicola, Cercosporidium
made, where incipient hypotheses are borne, where personatum, Puccinia arachidis), shoots, and stems
these are tested, and sometimes result in success in (Corticium rolfsii, Aspergillus niger), and pods
managing diseases. This paper is not intended as a (Botryodiplodia sp.). Another principal component
review of so vast a topic. Its aim is only to high- analysis highlights a series of wheat diseases in
light some important points, research issues, and Brittany: eyespot, brown rust, septoria blotch and
research avenues. Our purpose is to touch upon a yellow rust. A third example illustrates an analysis
limited number, but very different, aspects of the on upland rice injuries in northern Laos (IRRI,
subject. References therefore are used only to 1998), which involves an array of injuries by insects
illustrate research themes and approaches, and are (stem borers causing dead hearts and white heads,
given with no intention of offering a comprehen- root injuries by white grubs), foliage injuries
sive overview. (caused by several species), disease injuries caused
by fungi (neck and leaf blast caused by Magna-
Multiple pathosystems as facts porthe grisea, brown spot caused by Cochliobolus
miyabeanus, sheath blight caused by Rhizoctonia
Multiple pathosystems consist of a series of disease solani, sheath rot caused by Sarocladium oryzae),
elements that are present in the same host stand. and weed infestation by a number of species. Not

Groundnut diseases (Côte d'Ivoire) Wheat diseases (Britanny, France) Upland Rice Pests (Laos)

1 1 1

Cr
EyeSpot WEED
N P
0.5
A 0.5
Brow nRust BS
WG
axis 2 (14.7%)
axis 2 (21.9%)

axis 2 (29.1%)

B SHR
SHB
0 R 0 0
DH
-1 0 1 -1 0 1 -1 0 1
Septoria
NB

-0.5 -0.5 DEF

Yellow Rust LB

-1 -1 -1
axis 1 (41.3%) axis 1 (35.1%) axis 1 (25.6%)

Figure 1. Three multiple pathosystems portrayed by principal component analyses: groundnut diseases in savanna and forest envi-
ronments in Côte d’Ivoire, wheat diseases in Brittany, France, and upland rice in northern Laos. Left: Principal components analy-
sis on 209 farmers’ fields in several provinces (forest and savanna) of Côte d’Ivoire, 1982–1985; N: Aspergillus niger rot; Cr:
Corticium rolfsii rot; A: Cercospora arachidicola leaf-spot; P: Cercosporidium personatum (Phaeoisariopsis personata) leaf-spot; B:
Botryodiplodia pod rot; R: rust; Centre: wheat diseases in Britanny in a series of variety trials at varying levels of inputs, 2000,
2001, and 2003. Vectors indicate intensities of eyespot (Tapesia yallundae), of brown rust (Puccinia recondita), septoria
(Mycosphaerella graminicola), and yellow rust (Puccinia striiformis). Data are from 180 individual plots (2.6  15 m) combining
four crop management practices with five wheat cultivars in the replications over the 3 years. Right: upland rice pests in northern
Laos, 1996 and 1997; LB: leaf blast; DEF: defoliating insects; NB: neck blast; DH: dead heart caused by stemborers; SHB: sheath
blight, SHR: sheath rot; BS: brown spot; WEED: weed infestation; WG: white grub injury. Proportion of variances accounted for
are indicated along each axis.
 125

all the disease elements that actually were present by Magnaporthe grisea) in the upland rice patho-
in each of the three examples are represented. The system. Collinearity or non-collinearity of (dis-
three data sets used here correspond to different ease) vectors may lead to forwarding hypotheses,
contexts for data acquisition (see legend of which in turn would require additional analyses.
Figure 1). The groundnut data were collected
during a multiyear survey in several provinces of
Côte d’Ivoire, on 309 different farmers’ fields (Sa- Shapes of botanical pathosystems
vary, 1987a). The wheat data correspond to a series
of varietal trials at different levels of inputs con- Elements of multiple pathosystems
ducted over three different climatic years in Brit-
tany (Rolland et al., 2003). The upland rice data Multiple pathosystems have shapes, where indi-
were collected in a series of on-farm field experi- vidual diseases display a particular role. Several
ments, where different fertiliser regimes were tested studies have shown that multiple pathosystems
during two successive years (IRRI, 1998; Roder vary in shapes. Only two of the many reasons for
and Savary, unpublished data). change are illustrated here.
Some of the analyses involve more disease ele- Change in age, i.e., development of the host
ments than others, and one of them involves more stand, is one strong reason for change in the shape
than just plant pathogens. Deriving conclusions on of multiple pathosystems. Figure 2 shows three
the overall importance of diseases in each patho- separate principal component analyses on the
system from the mere number of elements would groundnut–leaf-spot–rust pathosystem at three
of course be incorrect. These analyses only provide different ages of the groundnut stands. Although
a view of possible associations, suggesting rela- the analyses pertain to the same farmers’ fields,
tionships, or absence of relationships, among dis- comparison of analysis of Figure 2a (young
ease elements. The relationships that seem to stands), 2b (middle-age stands), and 2c (stands
emerge from these summary analyses develop approaching or at harvest stage) indicate strong
against the background of a large number of fac- variation in relationships among variables. In
tors, including crop development stage, or crop young groundnut stands, a strong relationship
management. For instance, a linkage appears in between rust (R) and early leaf-spot (A) is appar-
the groundnut multiple pathosystem between N ent, both diseases being opposed to A. niger (N)
(A. niger) and Cr (C. rolfsii); by contrast, there wilt. In middle-age fields, a very strong association
seems to be independence between Cr and A between rust (R) and late leaf-spot (P) is indicated,
(Cercospora arachidicola) in the groundnut both diseases being opposed to early leaf-spot (A).
pathosystem, and independence between NB and In older fields, the relationship between rust and
LB (neck and leaf blast, respectively, both caused late leaf-spot has become loose, although both
1 1 1
(a) Cr P (b) Cr (c)

0.5 N N B
R
N R
axis 2 (27.3%)

axis 2 (34.8%)

axis 2 (23.6%)

A A
P
0 0 0
-1 0
R1 -1 0 1 -1 0 1

Cr
-0.5

P
A
-1 -1 -1
axis 1 (57.7%) axis 1 (50.4%) axis 1 (46.6%)

Figure 2. Patterns of change in multiple pathosystems over time and host development: principal component analyses of disease
levels in farmers’ fields in Côte d’Ivoire. (a) Groundnut field at early development stages (first trifoliate leaf – flowering); (b)
groundnut field at medium development stages (flowering – pod filling); (c) groundnut field at final development stages (pod filling
– harvest stage). Symbols for disease vectors are the same as in Figure 1.
126

remain opposed to early leaf-spot; and a linkage niger wilt and early leaf-spot in older stands). One
between A. niger and early leaf-spot is detected, important factor that drives relationships among
which did not exist at earlier ages. These shifts in foliar diseases is disease-induced-defoliation. It
relationships are reflections of the respective will be addressed later in this discussion.
dynamics of foliar diseases (rust and early leaf- Crop management is another major reason for
spot usually establish earlier in a crop stand than changes in shapes of multiple pathosystems.
late leaf-spot, while early leaf-spot generally de- Figure 3 is an illustration of the effects of four crop
clines as crop maturity approaches), and of envi- management patterns in wheat experiments.
ronmental factors that favour certain diseases at Strong shifts in disease vectors are detected in
certain stages of the crop (humid environment and Figure 3a–d, the transition from a to d corre-
contaminated seeds favour A. niger wilt and C. sponding to intensified wheat production. While in
rolfsii in the early crop stages; dense, green cano- Figure 3a all diseases, except yellow rust, appear
pies favour rust and late leaf-spot, whereas water closely associated, an opposition between eyespot
stresses and poor soil fertility favour late leaf-spot and both septoria leaf blotch and brown rust
in established stands; and more humid environ- develops in Figure 3b, which persists in Figure 3c,
ments favour Botryodiplodia pod rot, late leaf- but disappears in Figure 3d. Such sharp changes
spot, and rust while drier environments favour A. must be attributed to changes in fertiliser inputs,
1 1
(a) (c)
YellowRust
BrownRust
axis 2 (25.0%)

axis 2 (27.4%)

BrownRust Septoria
0 0
-1 0 EyeSpot Septoria 1 -1 0 1
EyeSpot

YellowRust

-1 -1

axis 1 (36.7%) axis 1 (47.5%)

1 1
(b) (d)
YellowRust
BrownRust
Septoria
axis 2 (25.0%)

axis 2 (29.0%)

EyeSpot
BrownRust
0 0
-1 EyeSpot 0 1 -1 0 1
Septoria
YellowRust

-1 -1

axis 1 (37.9%) axis 1 (43.6%)

Figure 3. Patterns of change in multiple pathosystems over crop management practices: principal component analyses of disease
levels in field trials in Britanny, France. (a) pattern of crop management A: fertiliser input with a target yield of 10 t ha)1; seeding
rate: 250 seeds m)2; use of a crop growth regulator; three fungicide applications; (b) pattern of crop management B: fertiliser input
with a target yield of 9 t ha)1; seeding rate: 250 seeds m)2; use of a crop growth regulator; two fungicide applications; (c) pattern
of crop management C: fertiliser input with a target yield of 8 t ha)1; seeding rate: 150 seeds m)2; no growth regulator; one fungi-
cide applications; (d): pattern of crop management D: fertiliser input with a target yield of 7 t ha)1; seeding rate: 150 seeds m)2; no
crop growth regulator; no fungicide applications.
 127

seeding rates, plant hormone use, and of course arachidis), early leaf-spot (Cercospora arachidico-
fungicide use. Crop management, involving pesti- la), late leaf-spot (Cercosporidium personatum), and
cide use or not, has been found to be a major factor web blight (Rhizoctonia solani) of groundnut. Sev-
for changes in shapes of multiple pathosystems, in eral techniques, including geostatistical and multi-
many diverse examples, including, e.g., wheat in variate, were used to show, as the maps strongly
Australia (Stynes, 1980) and the Netherlands suggest, that (i) rust (Figure 2a) and web blight
(Daamen et al., 1989), lowland rice in Asia (Savary (Figure 2d) are spatially strongly associated, (ii)
et al., 2000a), pea in Idaho (Wiese, 1982), early leaf-spot (Figure 2b) is more severe where rust
groundnut in Côte d’Ivoire (Savary, 1987a), or is less severe, and (iii) late leaf-spot (Figure 2c) does
coffee in Honduras (Avelino, 1999). not intensify strongly where rust or early leaf-spot
severities are extreme. The maps, which were drawn
at the end of a cropping season, also show that two
Shapes of multiple pathosystems in space very different types of epidemics developed in the
same stand, a typically focal epidemic (Figure 2d,
The spatial distributions of four different pathogens web blight), and three general epidemics
in the same crop stand are shown in Figure 4 (Figures 2a, 2b and 2c, rust, early leaf-spot, and late
(Lannou and Savary, 1991): rust (Puccinia leaf-spot, respectively) which did intensify locally.

(a) (c)

(b) (d)

Figure 4. Patterns of change in multiple pathosystems over space: spatial distribution of four diseases a groundnut plot, Côte
d’Ivoire (Lannou and Savary, 1991, modified). (a) Groundnut rust, Puccinia arachidis; (b) Early leaf-spot, Cercospora arachidicola;
(c) Late leaf-spot, Cercosporidium personatum (Phaeoisariopsis personata), (d) Web blight, Rhizoctonia solani. Disease assessments
were made at 90 days after sowing. Rust, early leaf-spot, and late leaf-spot: severity (% diseased leaf area) scales; web blight: inci-
dence (% diseased plants) scale. From Lannou and Savary, 1991, modified.
128

Analysis of spatial patterns in multiple patho- equation and systems of equations of the Lotka-
systems may lead to a number of hypotheses, as in Volterra type (Pianker, 1983, cited in Madden
the case of multiple infection of hop stands by et al., 1987). Use of this approach has been
different viruses (Pethybridge and Turechek, extensive in botanical epidemiology, and has been
2003), which can lead to experimentally testing based on both the development of simple equa-
spatial co-occurrence and variation of infections in tions that adequately describe disease progress
the host–vector–viruses system. In the case of over time (see, e.g., Madden, 1980; Campbell and
groundnut diseases, much of the spatial co-varia- Madden, 1990), and on tools to numerically inte-
tion of disease intensities may be attributed to grate sets of differential equations that constitute
competition towards vacant sites and defoliation simulation models (Zadoks, 1971; Zadoks and
of diseased tissues by some of the pathogens. Rabbinge, 1985).
Nelson and Campbell (1993) studied a far more The groundnut–leaf-spot–rust pathosystem
complex multiple pathosystem, which involves represents a good example to illustrate the ap-
eight fungal pathogens of white clover (Rhizoctonia proach. This system is fairly simple, having only
solani, Pseudomonas andropogonis, Stagonospora two disease components, rust and leaf-spot, but
meliloti, Cercospora zebrina, Curvularia trifolii, leads to considering several processes and inter-
Colletotrichum trifolii, Polythrincium trifolii, and actions: (i) a biotrophic pathogen which multiplies
Uromyces sp.), in presence or absence of three virus only on healthy, green tissues, (ii) a necrotrophic
diseases (alfalfa mosaic virus, yellow vein virus, pathogen (only Cercosporidium personatum is
and peanut stunt virus). They detected disease considered here, but both C. personatum and
aggregation of the fungal disease complex at sev- Cercospora arachidola could be considered collec-
eral scales (leaf, plant, population), which fluctu- tively) which causes extensive defoliation, which
ated over time as foliation and defoliation (iii) compounds physiological (senescence) defoli-
occurred, and which varied spatially as well. ation, (iv) competition between the two pathogens
Changes over time in the spatial aggregation of the in their access to growing crop tissues, and (v) the
foliar disease complex were associated with chan- ability of one of the two diseases (leaf-spot) to
ges in disease severity itself and defoliation. The multiply from defoliated, infectious tissues (Savary
background of varying, multiple virus infection did and Servat, 1991). Such characteristics are very
not seem to affect either the dynamics of the foliar similar to another legume-based multiple patho-
disease complex, or its spatial pattern. system, the bean–angular leaf-spot–anthracnose–
rust system (Gomes Carneiro et al., 2000;
Shapes of multiple pathosystems over time: Bassanezi et al., 2001; de Jesus et al., 2001), which
dynamics of multiple pathosystems has been extensively studied. A series of linked
differential equations representing the groundnut–
One approach to addressing the dynamics of leaf-spot–rust pathosystem are given in Table 1,
multiple pathosystems is by means of linked dif- corresponding to the overall model structure of
ferential equations. The equations themselves can Figure 5. Parameters for the model (relative rates
assume a number of shapes, but are in many re- of crop growth, of increase of both diseases, of
spects fairly similar. One important difference physiological defoliation, and of disease-induced
however among seemingly analogous systems of defoliation) were estimated (Savary and Servat,
equations is the nature of the modelled diseases 1991) using a numerical integration procedure
variables, which often are proportions (severities coupled with a sequence of two optimisation pro-
or incidences, i.e., disease densities) or less fre- cedures (Rosenbrock, 1960; Nelder and Mead,
quently (as in the model used below), amounts of 1964), applied to a set of 15 epidemics where levels
diseased tissues. This difference has important of the both diseases were artificially manipulated
consequences on the meaning of parameters used. (Savary and Zadoks, 1992a).
Use of linked differential equations derives in Simulated outputs using optimised parameters
botanical epidemiology from the approach Van are shown in Figure 6a. A regular, logistic-shaped
der Plank (1963) introduced to the field, which increase of leaf-spot is combined with a faster in-
itself is related to earlier ecological models, crease of rust, which declines in the later stage of
including especially the Verhulst-Pearl logistic the epidemic; these are coupled with a regular
 129

Table 1. Equations used in a rust–leaf-spot–groundnut multiple pathosystem simulation model

Equation Hypotheses

(1) Rate of increase of leaf area: The rate of increase of leaf area is proportional to a relative
dL=dt ¼ RRL  Lf1
½ðL þ Y þ Z þ totDÞ=Lmax g rate and the amount of (healthy) leaf, corrected for the fraction
of leaf growth, relative to a maximum. Leaf growth includes
defoliated tissues
(2) Rate of rust increase: The rate of rust increase is proportional to a relative rate
DY=dt ¼ RRY  Yf1
½ðY þ ZÞ=ðL þ Y þ ZÞg and the amount of rust-diseased tissues, corrected for the
fraction of relative growth, relative to the current total leaf
tissues
(3) Rate of leaf-spot increase: The rate of leaf-spot increase is proportional to both (1) a
DZ=dt ¼ ðRRZ  Z þ RRZDZ  DZÞ relative rate for standing diseased tissues and the amount of
 f1
½ðZ þ YÞ=ðL þ Y þ ZÞg leaf-spot-diseased tissues and (2) a relative rate for defoliated
diseased tissues and the amount of defoliated infected tissues,
corrected for fraction of relative growth, relative to the current
total leaf tissues
(4) Rate of defoliation (healthy tissues) The rate of defoliation of healthy tissues is the sum of
dD=dt ¼ RRDS  L þ RRDZ  ½Z=ðL þ Y þ ZÞ  L senescence-induced (relative rate and healthy tissues) and
indirectly leaf-spot-induced (relative rate, proportion
leaf-spot-diseased, and healthy tissues)
(5) Rate of defoliation (leaf-spot-diseased tissues) The rate of defoliation of leaf-spot-diseased tissues is
dDZ=dt ¼ (RRDS+RRDZ)  Z proportional to a relative rate (senescence and leaf-spot
accumulated) and to the amount of leaf-spot-diseased tissues
(6) Rate of defoliation (rust-diseased issues) The rate of defoliation of rust-diseased tissues is proportional to
dDY=dt ¼ (RRDS+RRDZ)  Y a relative rate (senescence and leaf-spot accumulated) and to the
amount of rust-diseased tissues

Variables (dimensions in brackets): L: healthy leaf tissues [L2]; Y: rust-diseased tissues [L2] ; Z: leaf-spot-diseased tissues [L2]; D:
healthy, defoliated tissues [L2]; DY: rust-diseased, defoliated tissues [L2]; DZ: leaf-spot-diseased, defoliated tissues [L2]; RL: rate of leaf
growth [L2 T)1]; RY: rate of rust increase [L2 T)1]; RZ: rate of leaf-spot increase [L2 T)1]; RD: rate of defoliation of healthy tissues
[L2 T)1]; RDY: rate of defoliation of rust-diseased tissues [L2 T)1]; RDZ: rate of defoliation of leaf-spot diseased tissues [L2 T)1];
RRL: relative (intrinsic) rate of leaf growth [T)1]; Lmax: maximum leaf growth [L2]; RRY: relative rate of rust increase [T)1]; RRZZ:
relative rate of leaf-spot increase from non-defoliated tissues [T)1]; RRZDZ: relative rate of leaf-spot increase from infected defoliated
tissues [T)1]; RRDS: relative rate of defoliation (senescence) of healthy tissues [T)1]; RRDZ: relative rate of defoliation of (leaf-spot)
diseased tissues [T)1]; totD: accumulated defoliation [L2].

accumulation of defoliated tissues, and a bell- remains essentially unaffected, whereas the rust
shaped green leaf area curve. When the relative epidemic is slightly enhanced. Increase of this
rate of rust increase is reduced (Figure 6b) or in- parameter (Figure 6g), on the other hand, leads to
creased (Figure 6c), a strong rust reduction, or only a slight increase in total defoliation, an
increase, respectively, is simulated, coinciding with unaffected leaf-spot epidemic, but a reduction by
opposite behaviour of the leaf-spot epidemic. about 8% of the rust epidemic.
When the relative rates of leaf-spot increase from The overall picture this model gives is that the
both non-defoliated and defoliated tissues are re- system is quite sensitive to comparatively small
duced (Figure 6d), the leaf-spot epidemic is variation (10% or less) in parameter values. Leaf-
strongly reduced, while the rust epidemic is spot is behaving as a very strong competitor for
strongly increased. When both these parameters rust, especially through leaf-spot induced-defolia-
are increased (Figure 6e), the leaf-spot epidemic is tion. Rust on the other hand can rapidly take
strongly increased, the rust epidemic is nearly advantage of seemingly small increases in the
halved, and the amount of defoliation is increased. amount of available healthy tissues. These inter-
A last set of tests, where the relative rate of defo- actions are strongly influenced by crop growth
liation induced by leaf-spot is varied is also shown. (and the effects of diseases on crop growth), and
When the leaf-spot induced-defoliation rate is de- defoliation (especially through physiological
creased (Figure 6f), the amount of total defolia- senescence). One interesting response of the system
tion is barely altered, the leaf-spot epidemic is that a reduced retention (greater disease-induced
130

totD Lmax
RRL
RRZZ RL

RRZDZ
RRY Y
Z

RZ RY
L

RDZ RD RDY

RRDS

RRDZ

DZ D DY

totD

Figure 5. Overall structure of a mechanistic simulation model incorporating two foliar diseases, rust and leaf-spot on groundnut.
State variables: L: healthy leaf tissues [L2]; Y: rust-diseased tissues [L2]; Z: leaf-spot-diseased tissues [L2]; D: healthy, defoliated tis-
sues [L2]; DY: rust-diseased, defoliated tissues [L2]; DZ: leaf-spot-diseased, defoliated tissues [L2]; Rates: RL RL: rate of leaf
growth [L2 T)1]; RY: rate of rust increase [L2 T)1]; RZ: rate of leaf-spot increase [L2 T)1]; RD: rate of defoliation of healthy tis-
sues [L2 T)1]; RDY: rate of defoliation of rust-diseased tissues [L2 T)1]; RDZ: rate of defoliation of leaf-spot diseased tissues
[L2 T)1]; Parameters: RRL: relative (intrinsic) rate of leaf growth [T)1]; Lmax: maximum leaf growth [L2]; RRY: relative rate of
rust increase [T)1]; RRZZ: relative rate of leaf-spot increase from non-defoliated tissues [T)1]; RRZDZ: relative rate of leaf-spot
increase from infected defoliated tissues [T)1]; RRDS: relative rate of defoliation (senescence) of healthy tissues [T)1]; RRDZ: rela-
tive rate of defoliation of (leaf-spot) diseased tissues [T)1]; totD: accumulated defoliation [L2].

defoliation) of leaf-spot infected tissues (Figure 6a would require a set of at least 10 equations. If this
vs. 6g) leads to a reduced rust epidemic. Selecting approach were to be taken, a complex model
varieties that shed their leaves at low leaf-spot structure would have to be designed, a large
severity might then be an efficient way of reducing number of parameters would have to be estimated,
rust epidemics. and numerical solutions would become difficult to
The white clover–foliar fungal diseases–viruses interpret. The approach chosen by Nelson and
system (Nelson and Campbell, 1993) renders the Campbell (1993) in their field work, however, was
groundnut–rust–leaf-spot system a comparatively not to consider each disease separately, but to
simple system to address. Nelson and Campbell quantify the leaf disease complex as a whole. This
question the relevance of the approach illustrated leads to the interesting avenue of perhaps consid-
above, which in the case of the clover-based system ering groups of pathogens that share similar
 131

(a) 3
AUPC(LAI) = 45.4 RRL = 0.08
AUPC(def) = 38.6 RRDS = 0.001 Healthy
RRY = 0.11 Defoliated
Leaf area (m )
2

2 RRZZ = 0.07
RRZDZ = 0.03
RRDZ = 0.15 Rust
Leafspot
1

0.4
AUPC(rust) = 5.2
AUPC(leafspot) = 5.8
Diseased Area (m )
2

0.3

0.2

0.1

0
0 10 20 30 40 50

Time (days)

(b) 3 (c) 3
AUPC(LAI) = 47.5 RRL AUPC(LAI) = 42.3 RRL
AUPC(def) = 38.9 RRDS AUPC(def) = 38.7 RRDS
RRY = 0.10 RRY = 0.12
Leaf area (m )

Leaf area (m )
2

2 RRZZ 2 RRZZ
RRZDZ RRZDZ
RRDZ RRDZ
1 1

0 0

0.4 0.4
AUPC(rust) = 3.5 AUPC(rust) = 6.8
AUPC(leafspot) = 7.1 AUPC(leafspot) = 4.5
Diseased Area (m )

Diseased Area (m )
2

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 10 20 30 40 50 0 10 20 30 40 50
Time (days) Time (days)

Figure 6. Two foliar diseases dynamically interacting, simulated rust and leaf-spot epidemics on groundnut. Each simulation is rep-
resented by two graphs of outputs, healthy and (total) defoliated area (upper half) and rust and leaf-spot severities (lower half).
Numerical values of parameters are indicated in graphs b–g only when changes from reference (optimised) values (graph a) have
been used. Calculated areas under progress curves (AUPC) of the rust epidemic, of the leaf-spot epidemic, of the leaf area index,
and of the defoliated leaf area index are indicated. (a) simulated outputs with optimised parameter values for RRL (relative rate of
leaf growth; [T)1]), RRDS (relative rate of senescence defoliation of healthy tissues [T)1]), RRY (relative rate of rust increase
[T)1]), RRZZ (relative rate of leaf-spot increase from non-defoliated tissues [T)1]), RRZDZ (relative rate of leaf-spot increase from
infected defoliated tissues [T)1]), and RRDZ (relative rate of defoliation of leaf-spot diseased tissues [T)1]). (b) simulated outputs
for a reduced relative rate of rust increase (RRY). (c) simulated outputs for an increased relative rate of rust increase (RRY). (d)
simulated outputs for reduced relative rates of leaf-spot increase (RRZZ and RRZDZ). (e) simulated outputs for increased relative
rates of leaf-spot increase (RRZZ and RRZDZ). (f) simulated outputs for a reduced relative rate of defoliation induced by leaf-
spot (RRDZ). (g) simulated outputs for an increased relative rate of defoliation induced by leaf-spot (RRDZ).

functional attributes in a community, and model One of the many criticisms of the linked
the dynamics of guilds of pathogens, rather than differential equation approach (see, e.g., Nelson
of individual diseases. and Campbell, 1993) is its inability to account for
132

3 3
(d) AUPC(LAI) = 56.4 RRL (e) AUPC(LAI) = 33.3 RRL
AUPC(def) = 33.2 RRDS AUPC(def) = 43.4 RRDS
RRY RRY
Leaf area (m )

Leaf area (m )
2

2
2 RRZZ = 0.06 2 RRZZ = 0.08
RRZDZ = 0.02 RRZDZ = 0.04
RRDZ RRDZ
1 1

0 0

0.4 0.4
AUPC(rust) = 8.3 AUPC(rust) = 2.9
AUPC(leafspot) = 2.8 AUPC(leafspot) = 7.7
Diseased Area (m )

Diseased Area (m )
2

2
0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 10 20 30 40 50 0 10 20 30 40 50

Time (days) Time (days)

(f) 3
AUPC(LAI) = 45.3 RRL
(g) 3
RRL
AUPC(LAI) = 45.5
AUPC(def) = 37.8 RRDS AUPC(def) = 39.4 RRDS
RRY RRY
Leaf area (m )
2

Leaf area (m )

2
2

RRZZ 2 RRZZ
RRZDZ RRZDZ
RRDZ = 0.14 RRDZ = 0.16
1
1

0
0

0.4 0.4
AUPC(rust) = 5.9 AUPC(rust) = 4.5
AUPC(leafspot) = 5.8 AUPC(leafspot) = 5.8
Diseased Area (m )

Diseased Area (m )
2

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 10 20 30 40 50 0 10 20 30 40 50
Time (days) Time (days)

Figure 6. Continued

spatial patterns, and the essential effects of spatial equations of Table 1 and the flowchart of Figure 5,
patterns on disease dynamics. Figure 4 is sufficient however, amounts to implicitly considering a
evidence of the fact that the groundnut–leaf-spot– growing canopy with two layers: (i) a (healthy)
rust system is no exception. The model summarised layer where defoliation is caused by senescence and
here does however include implicit assumptions distance effect of leaf-spot disease, and (ii) a (dis-
pertaining to the distribution of disease on host eased) layer, where both rust and leaf-spot lesions
tissues. As in many pathosystems involving foliar occur, and where defoliation is caused by both leaf-
diseases, a very strong vertical aggregation occurs spot and senescence effects.
in both the leaf-spot (Boote et al., 1980) and the The purpose of this type of model is to explore
rust (Savary, 1987b) pathosystems. Such aggrega- interactions within a framework of thinking de-
tion of diseases (and defoliation) along the vertical fined, and limited by, a set of hypotheses.
dimension of a crop canopy must have very strong Expanding the model to address additional, albeit
consequences on the behaviour of the multiple important, features of the considered system might
pathosystem. Considering the model outlined in prevent the interpretation of simulation results,
 133

evaluation of the model, and its use. Further dis- Quantification, analysis, and modelling damage
cussion on strategic models of this kind is given in is both the scientific-technical cornerstone for
McRoberts et al. (2003). disease management (James, 1974; Chiarappa,
1980; Madden, 1983; Teng, 1983; Zadoks, 1985;
Teng, 1987; Gaunt, 1995) and one of the very
Multiple pathosystems and damage to crops important entry points for disease management
(see, e.g., Teng and Savary, 1992). This has been a
Multiple injuries and the resulting damage very broad and active field of research for many
decades. The subject has particular relevance when
Consideration of multiple pathosystem, i.e., considering multiple pathosystems (Madden and
pathogens dynamically interacting among them- Nutter, 1995), however, and a few points are dis-
selves and a growing canopy, inevitably leads to cussed here.
questions regarding harmfulness of the multiple
pathosystem. The extent of damage (sensu Zadoks, Five directions
1985, i.e., yield reduction) caused by a multiple
pathosystem has scientific relevance of its own; it As opposed to an additive model for combined
also has very practical implications with respect to effects of injuries on damage, the model developed
the availability, efficiency, and deployment of by Padwick was a useful starting point, from
management options. Plant pathologists did not which several directions of thoughts were ex-
study multiple pathosystems because of the inter- plored.
esting interactions among competing pathogens (i) A first direction concerns the nature of dis-
that may take place, and chose them because they ease interactions in their yield-reducing effect: very
were good examples for community ecology often, a less than additive effect is observed, but
studies; they did not either address the issue with some injuries may synergistically increase yield
the prime objective of selecting plants that would losses. Less than additivity has been demonstrated
be resistant against several diseases, or crops in one of the best documented studies on damage
whose yields would be stable in the presence of caused by a multiple pathosystem, the early
several diseases. One simple, practical reason was blight–verticillium wilt–potato leafhopper of po-
the need to assess the damage caused by several tato (Johnson, 1986; Johnson et al., 1986, 1987;
injuries. Padwick (1956) made an early attempt to Johnson and Teng, 1990). Injuries caused by dis-
such quantification, with the formula: eases may, however, synergistically reduce yield.
Synergies in yield-reducing effects are found in
Percent yield loss
potato early dying caused by Verticillium dahliae
¼ f100 and Pratylenchus penetrans (Francl et al., 1990), as

½ð100
P1Þð100
P2Þ well as in combinations of infections of wheat by
 (100
PnÞ=100n
1g Septoria nodorum and Puccinia recondita (Van der
Wal et al., 1970). A first hypothesis therefore
where Pi is the percent yield loss caused by an refers to the direction, positive or negative, of
individual injury i. Padwick’s model assumes that combined injuries on damage.
the only interaction between diseases on yield is (ii) A second direction concerns the nature of
competition for the crop’s resources. This amounts competition, which may only be for ‘resource’, as
to forwarding the hypothesis that one disease Padwick’s model refers, or may involve other
cannot affect what other diseases have already in- mechanisms, resulting then in damage lower than
jured (Johnson et al., 1986). Padwick’s view expected from Padwick’s model. Such is the case in
strongly contrasted with several later analyses, the potato–early blight–verticillium wilt–leafhop-
whereby individual losses were merely accumulated per studied by Johnson (1986) and Johnson et al.
in a ‘loss profile’ (see, e.g., Pinstrup-Andersen (1986).
et al., 1976). Teng (1994) pointed to the fact that (iii) A third direction concerns the way damage
this latter reasoning may lead to the impossible measurements are expressed. Damage is com-
result that diseases, or crop harmful agents in monly reported as percentage. The reported
general, may cause losses exceeding 100%. figures (percent losses) are therefore dependent on
134

the level of the uninjured yield reference, and their lium wilt (Johnson, 1986; Johnson et al., 1986,
meaning will strongly depend on whether this yield 1987; Johnson and Teng, 1990).
reference is low or high. One of many alternatives New developments have taken place, where
to reporting damage as a proportion is expressing these five points are considered in the case of
it as a biomass, and the choice will depend on the lowland rice in Asia (Pinnschmidt et al., 1995;
end-use of the information. Willocquet et al., 2000, 2002, 2004). Simulation
(iv) A fourth direction concerns the relation- models have been developed that make use of the
ships that link damage to the various injuries (the concept of guilds of injuries (Pinnschmidt et al.,
damage function; Zadoks, 1985). The damage 1995; Willocquet et al., 2000, 2002) which have
function may involve the yield pertaining to a been used to analyse and understand the yield-
given production situation as an explanatory var- reducing effects of several pathogens, insects, and
iable (the attainable yield; Rabbinge and De Wit, weeds in the same crop. A modelling structure has
1989). As is the case with the damage caused by been designed so that it can simultaneously handle
single harmful agents (see, e.g., Rossing, 1991, for production situations (as drivers of attainable crop
the grain aphid on wheat, and Rabbinge et al., performances) and injury profiles (as drivers of
1985, for powdery mildew in wheat), the amount multiple injuries) in the very combinations where
of damage caused by a multiple pathosystem may field characterisation had shown these (production
depend on the level of attainable yield. This has situation)  (injury profile) associations occur
been exemplified in the case of the multiple (Willocquet et al., 2000, 2002)). Production situa-
pathosystem of groundnut in Côte d’Ivoire tions and their associated injury profiles were then
(Savary and Zadoks, 1992b), and in the more used as the modelling context where disease and
complex and diverse multiple pathosystem of pest management tools could be most efficiently
lowland rice in Asia (Savary et al., 2000b). deployed, and where progress should be expected,
(v) A fifth direction of thought follows the real- and so expressed in yield gains, instead of yield
isation that diseases, and any harmful agent in losses (Willocquet et al., 2004).
general, belong to one or a few categories, based on
the type of injury mechanisms they trigger. Rabb-
inge and Rijsdijk (1981) and Boote et al. (1983)
defined the limited number of ways for a harmful Multiple pathosystems and integrated pest
organism to hamper the physiological perfor- management
mances of a growing canopy. This has particular
relevance when considering multiple pathosystems A negative view
for two reasons. First, it provides a basis for
designing experiments, developing field survey In his article on the functioning and performances
procedures, and defining field measurements that of tropical ecosystems, Janzen (1973) was express-
refer not to specific diseases (or harmful agents) but ing his frustration at science not achieving its goals
to specific injuries. Diseases then are not measured in vital fields of application: ‘‘Nearly all research in
with respect to how fast they intensify, but rather to tropical agriculture is highly reductionist, parochial,
how much they may affect the performances of a and discipline-oriented ’’. At the time when Janzen
crop. Further, they need not necessarily be mea- wrote, much of the synthesis tools that now are
sured individually, but collectively, as Nelson and available to plant pathologists did not yet exist.
Campbell (1993) did. Second, this categorization A negative view, similar to Janzen’s, could be
provides a framework for modelling mechanisti- expressed considering the very slow pace of pro-
cally the physiological interactions between a crop gress that has been achieved in understanding,
stand and a multiple pathosystem. This direction of analysing, and managing multiple pathosystems.
thought has been underpinning research involving In spite of the availability of tools to address it, the
simulation models as tools for understanding complexity of these systems remains a deterrent,
damage caused by diseases and means to reduce not a challenge, to many plant pathologists. But
them (Rouse, 1988; Rabbinge et al., 1989; Gaunt, the primary reason why progress has been so slow
1995), including work conducted on multiple is the weakness of communication among disci-
pathosystems such as potato–early blight–verticil- plinary fields (McRoberts et al., 2003).
 135

Questions about multiple pathosystems are for maximum yield and maximum grain protein
central to defining the scientific bases for, the content under intensive production conditions,
designing of, and the implementing of IPM. i.e., high (nitrogen) fertiliser inputs and a pesticide
Reductionism, in its many forms, and discipline- umbrella leads to ignoring a large fraction of ge-
oriented science, are the very same reasons that netic resources available in a germplasm, and to
hamper progress in IPM (Jeger, 2000), and it restricting progress within unrealistically favour-
would seem that Joni Mitchell’s song (‘‘Hey farmer able production situations. Hardy wheat varieties,
farmer. Put that DDT now. Give me spots on my which yield-wise may perform somewhat below, or
apples. But leave me the birds and the bees. nearly as well as, conventional high yielding cul-
Please.’’; (McRoberts et al., 2003)) does not seem tivars, but exhibit a number of incomplete resis-
to be fading away anytime soon. tances and have fair performances with respect to
protein contents of the grain, are being tested in a
number of sites in France. Incomplete resistances
A positive view concern eyespot (Tapesia yallundae) and leaf
blotch (Mycosphaerella graminicola). Experiments
There has nevertheless been a change in the way do not involve varieties only, but patterns of crop
disease management has been addressed, scientifi- management; they also lead to assessment of eco-
cally and technically, over the past 50 years – the nomic performances. So far (over 3 years, 1999–
time-span covering the cycle of the International 2002), reported results are very encouraging
Epidemiology Workshops. One good example (Rolland et al., 2003): under a low input regime,
appears to be the multiple pathosystem of wheat in hardy wheat cultivars yields are reasonably stable
wheat-based systems of western Europe. (within 10 to 8 t ha)1), have acceptable grain
Initial steps were taken in the Netherlands with protein contents (10.5–11.5%), and are produced
the EPIPRE project which saw scientists sharing at costs reduced from 400 to 150 euros ha)1
experience with farmers, adapting theories to compared to an intensive production system. In 46
practice, and farmers empowered in their disease (i.e., 73%) of the 63 tested combinations of
management decisions from epidemiological and (year  crop management pattern  cultivar), low-
systems science (Zadoks, 1989). input crop management with hardy varieties gen-
This early farmers-driven project had a setting erated the highest net returns. Thus in about 73%
different from that of today. Much work has been of the cases, host plant resistance to the wheat
accomplished since, and already was on its way multiple pathosystem is mobilised as a tool for
then, to show through long-term experiments system stabilisation – instead of pesticides for
(Jordan et al., 1985; Webster, 1985; McRoberts system perturbation, and yield-driven control.
et al., 2000) that integrated production systems for Interestingly, the effort appears to be led by
field crops tend to perform better financially than breeders; that economists, agronomists, and
high-input systems when commodity prices are low pathologists contribute to designing production
(McRoberts et al., 2003). The notion that disease systems that suit a seed-based technology is mak-
management depends on production situations – ing an experience of this kind very promising, not
not only because multiple pathosystems are so only because of the science it involves, but because
sensitive in their composition to crop management, of the promise to deliver it entails.
but simply because disease management is only
part of crop management, and so, necessarily de-
pends on the socio-economic dimensions of what a Perspectives
production situation is – was no longer a theory but
a concept put into practice and a way to conduct Epidemiological guilds
research.
The old, simple, idea that stable yield, and stable The example of approach used above to model the
yield characteristics, including multiple, incom- dynamics of simultaneous disease epidemics epi-
plete, host plant resistance are criteria for selecting demics invites the question of how to best address
varieties was revisited by G. Doussinault (Dous- the temporal and spatial structure of multiple
sinault, 1998; Doussinault et al., 2001). Selecting pathosystems whose disease components may be
136

numerous, or variable. Nelson and Campbell current things. Andow and Hikada (1989) used the
(1993) addressed a very diverse multiple patho- term to both describe different patterns of man-
system. Aside from rendering their study doable, agement of rice in Japan, and the plant health
their approach to quantify and analyse several consequences this difference in production situa-
foliar diseases collectively is an important choice, tions has on rice diseases. Physicians do not differ
which carries the simplifying, and new, assumption in their use of the word (e.g., Peto, 2001; Zimmet
that several diseases may be considered as an et al., 2001). The detection of linkages between
aggregate, from an epidemiological standpoint. production situations and injury profiles is analo-
One approach to analysing complex, multiple gous to considering corresponding syndromes of
pathosystems may thus be grounded on the con- production and syndromes of diseases. This rep-
sideration of epidemiological guilds, rather than of resents an avenue towards improving plant health,
individual pathogens. Definition of these guilds i.e., better management of multiple pathosystems,
and clustering individual disease components in via improvements of management. Considering
them could make use of results achieved from epidemiological guilds and guilds of harmful
comparative epidemiological work (Kranz, 1974, agents might be an interesting direction to take
1980). Simplification of this kind, if successful, towards that aim.
would have the merit of accommodating quanti-
tative epidemiological knowledge on individual
disease components, while also generating a com- References
mon framework for understanding and manage-
ment of multiple pathosystems that differ in their Andow DA and Hikada K (1989) Experimental natural history
biological components, but share common epide- of sustainable agriculture: Syndrome of production. Agri-
culture, Ecosystems and Environment 27: 447–462.
miologically functional traits.
Avelino J (1999) Contribution à la gestion du pathosystème
Hemileia vastatrix et Coffea arabica au Honduras. Influence
Guilds of harmful agents des principaux facteurs physiques, phytotechniques et bio-
logiques. PhD Dissertation, Paris XI Orsay Univ, 127 pp.
This concept has proven useful. The notion that Babcock Grove P (ed.) (1961) Webster’s Third New Interna-
tional Dictionary. Merriam-Webster Inc. Springfield, Mas-
harmful organisms share common injury mecha-
sachusetts.
nisms (Rabbinge and Rijsdijk, 1981; Boote et al., Bastiaans L and Daamen RA (1994) The role of crop
1983) enables us to bring together organisms that husbandry practices in wheat production for disease con-
otherwise profoundly differ. This has found sev- trol. Brighton Crop Protection Conference, Pests and
eral applications, in designing field quantification Diseases – 1994. Vol. 3 (pp. 1169–1174).
Bassanezi RB, Amorim L, Bergamin Filho A, Hau B and
methods, field experiments to measure damage,
Berger RD (2001) Accounting for photosynthetic efficiency
and of course, simulation models. These in turn of bean leaves with rust, angular leaf-spot and anthracnose
have found strategic applications for research to assess crop damage. Plant Pathology 50: 443–452.
prioritisation towards multiple pathosystems, for Boote KS, Jones JW, Smerage GH, Barfield CS and Berger RD
assessing the impact of new crop management and (1980) Photosynthesis of peanut canopies as affected by
leaf-spot and artificial defoliation. Agronomy Journal 72:
changes in production situations, and of disease
247–252.
management tools. It also has allowed a shift of Boote KJ, Jones JW, Mishoe JW and Berger D (1983) Coupling
thinking, from assessment of damage, to projec- pests to crop growth simulators to predict yield reductions.
tion of yield gains (Willocquet et al., 2004), that is, 73: 1581–1587.
a shift from what has been lost because of current Campbell LC and Madden LV (1990) Introduction to Plant
Disease Epidemiology, John Wiley & Sons, New York, 532.
practices, to what could be gained from future
Chiarappa L (ed.) (1980) Crop Loss Assessment Methods.
options. Supplement No. 3 FAO/Comm. Agric., Bureaux, Farnham,
England, 123 pp.
Syndromes of production, syndromes of disease Daamen RA, Wijnands FG and Van der Vliet G (1989)
Epidemics of diseases and pests of winter wheat at different
levels of agrochemical inputs. A study on the possibilities
The term syndrome has two definitions (Babcock
for designing an integrated crop system. Journal of Phyto-
Grove, 1961): (i) a group of symptoms or signs pathology 125: 305–319.
typical of a disease, disturbance, condition, or le- deJesus Junior WC, Vale FXR, Coelho RR, Zambolim L,
sion in animals or plants, and (ii) a set of con- Costa LC and Bergamin Filho A (2001) Effects of angular
 137

leaf-spot and rust on yield loss of Phaseolus vulgaris. Kranz J (1974) Comparison of epidemics. Annual Review of
Phytopathology 91: 1045–1053. Phytopathology 12: 355–374.
Doussinault G (1998) La sélection pour les variétés adaptées à Kranz J (1980) Comparative epidemiology: An evaluation of
une agriculture durable. Dossiers de l’Environnement de scope, concepts, and methods. In: Palti J and Kranz J (eds)
l’INRA 16: 39–44. Comparative Epidemiology. A Tool for Better Disease
Doussinault G, Judeau B and Jahier J (2001) Evolution de la Management. Pudoc Wageningen. (pp. 18–28).
variabilité génétique chez le blé. Dossiers de l’Environne- Lannou C and Savary S (1991) The spatial structure of
ment de l’INRA 21: 91–104. spontaneous epidemics of different diseases in a groundnut
Francl LJ, Madden LV, Rowe RC and Riedel RM (1990) plot. Netherlands Journal of Plant Pathology 97: 355–367.
Correlation of growing season environmental variables and Madden LV (1980) Quantification of disease progression.
the effect of early dying on potato yield. Phytopathology 80: Protection Ecology 2: 159–176.
425–432. Madden LV (1983) Measuring and modeling crop loss at the
Gaunt RE (1995) The relationship between plant disease field level. Phytopathology 73: 1591–1596.
severity and yield. Annual Review of Phytopathology 33: Madden LV, Pirone TP and Raccah B (1987) Temporal
119–144. analysis of two viruses increasing in the same tobacco
Gomes Carneiro SMTP, Amorim L, Bergamin Filho A, Hau B fields. Phytopathology 77: 974–980.
and Bianchini A (2000) Dinâmica de area foliar, desfolha e Madden LV and Nutter FW Jr (1995) Modeling crop losses at
variaveis de area foliar sadia em feijoeiros com infecçoes the field scale. Canadian Journal of Plant Pathology 17:
isoladas e conjuntas de Phaeoisariopsis griseola e Colletotri- 124–137.
chum lindemuthianum. Summa Phytopathologica 26: McRoberts N, Hughes G and Savary S (2003) Integrated
406–412. approaches to understanding and control of diseases and
Hamelink J, Hau B and Kranz J (1988) Untersuchungen zur pests in field crops. Australasian Journal of Phytopathology
Entwicklung von Befalls/Verlust-Relationen bei Schader- 32: 167–180.
reger-Komplexen an Einzelpflanzen des Maises. Zeitsch- Mc Roberts N, Foster GN, Sutherland A, Willock J and
rift Fur Pflanzenkrankheiten Und Pflanzenschutz 95: 258– Edwards-Jones G (2000) An evaluation of returns from
269. crop management schemes based on farmer practices.
IRRI (1998) Program Report for 1998. A Characterization of Aspects of Applied Biology 62: 277–284.
Upland Rice Biotic Constraints in Laos. International Rice Nelder JA and Mead R (1964) A simple method for function
Research Institute, Los Banos, Philippines. minimization. Computer Journal 7: 308–313.
James CW (1974) Assessment of plant diseases and losses. Nelson SC and Campbell LC (1993) Disease progress, defoli-
Annual Review of Phytopathology 12: 27–48. ation, and spatial pattern in a multiple-pathogen disease
Janzen DH (1973) Tropical agroecosystems. Science 182: complex on white clover. Phytopathology 83: 419–429.
1212–1219. Padwick GW (1956) Losses caused by Plant Diseases in the
Jeger MJ (2000) Bottlenecks in IPM. Crop Protection 19: Tropics. Commonwealth Mycological Inst. Kew, Surrey,
787–792. Phytopath. Papers No.1. 60 pp.
Johnson KB (1986) Systems analysis and modeling of yield loss Peto J (2001) Cancer epidemiology in the last century and the
in potato caused by interacting populations of early blight, next decade. Nature 411: 390–395.
Verticillium wilt, and the potato leafhopper. PhD Disser- Pethybridge SJ and Turechek WW (2003) Analysis of the
tation, Univ. Minnesota 201 pp. association among three virus infecting hop in Australia.
Johnson KB, Radcliffe EB and Teng PS (1986) Effects of Plant Pathology 52: 158–167.
interacting populations of Alternaria solani, Verticillium Pinnschmidt HO, Batchelor WD and Teng PS (1995) Simula-
dahliae, and the potato leafhopper (Empoasca fabae) on tion of multiple species pest damage in rice using CERES-
potato yield. Phytopathology 76: 1046–1052. rice. Agricultural Systems 48: 193–222.
Johnson KB, Teng PS and Radcliffe EB (1987) Analysis of Pinstrup-Andersen P., de Londono N. and Infante M. (1976) A
potato foliage losses caused by interacting infestations of suggested procedure for estimating yield and production
early blight, Verticillium wilt, and potato leafhopper; and losses in crops. PANS 22: 359–365.
the relationship to yield. Zeitschrift für Pflanzenkrankheiten Rabbinge R and Rijsdijk FH (1981) Disease and crop physi-
und Pflanzenschutz 94: 22–33. ology: A modeler’s point of view (pp. 201–220). In: Ayres
Johnson KB and Teng PS (1990) Coupling a disease progress PG (ed.) Effects of Disease on the Physiology of the
model for early blight to a model of potato growth. Growing Plant. Cambridge University Press, Cambridge
Phytopathology 80: 416–425. 258 pp.
Jörg E, Weissenfels D and Kranz J (1987) Krankheits- und Rabbinge R, Jorristma ITM and Schans J (1985) Damage
Schadlingsbefall an Haupt- und Nebenhalmen des Weizens. components of powdery mildew in winter wheat. Nether-
Zeitschrift Fur Pflanzenkrankheiten Und Pflanzenschutz. lands Journal of Plant Pathology 91: 235–247.
94: 509–519. Rabbinge R and De Wit CT (1989) Systems, models and
Jordan VWL, Hurcheon JA, Donaldson GV, Farmer DP simulation. In: Rabbinge R, Ward SA and Van Laar HH
(1985) Research into and development of integrated farm- (eds) Simulation and Systems Management in Crop Pro-
ing systems for less intensive arable crop production. In: tection (pp. 3–15). Pudoc, Wageningen, 420 pp
Atkinson D and McKinlay RG (eds) Integrated Crop Rabbinge R, Ward SA and Van Laar HH (eds) (1989)
Protection: Towards Sustainability? Vol. 63 (pp. 287–296) Simulation and Systems Management in Crop Protection.
BCPC Symposium. Pudoc, Wageningen 420 pp.
138

Rolland B, Bouchard C, Loyce C, Meynard JM, Guyomard H, Teng PS (1994) Integrated pest management in rice. Experi-
Lonnet P and Doussinault G (2003) Des itinéraires mental Agriculture 30: 115–137.
techniques à bas intrants pour des variétés rustiques de Teng PS and Savary S (1992) Implementing the systems
blé tender: une alternative pour concilier économie et approach in pest management. Agricultural Systems 40:
environnement. Courrier de l’Environnement de l’INRA 237–264.
49: 47–62. Vander Wal AF, Shearer BL and Zadoks JC (1970) Interaction
Rosenbrock HH (1960) An automatic method for finding the between Puccinia recondita and Septoria nodorum on wheat,
greatest or the least value of a function. Computer Journal and its effect on yield. Netherlands Journal of Phytopa-
3: 175. thology 76: 261–263.
Rossing WAH (1991) Simulation of damage in winter wheat Van der Plank JE (1963) Plant Diseases. Epidemics and
caused by the grain aphid Sitobion avenae. 3. Calculation of Control. Academic Press, New York 349 pp.
damage at various attainable yield levels. Netherlands Webster JPG (1985) Analysing the economic consequences of
Journal of Plant Pathology 97: 87–103. sustainability in farming systems. In: Atkinson D and
Rouse DI (1988) Use of crop-growth models to predict the McKinlay RG (eds) Integrated Crop Protection: Towards
effects of diseases. Annual Review of Phytopathology 26: Sustainability? Vol. 63 (pp. 287–296) BCPC Symposium.
183–201. Wiese MV (1982) Crop management by comprehensive
Savary S (1987a) Enquête sur les maladies fongiques de appraisal of yield determining variables. Annual Review
l’arachide (Arachis hypogaea) en Côte d’Ivoire (I). Méth- of Phytopathology 20: 419–432.
odes d’enquête et étude descriptive: Les conditions cultu- Willocquet L, Savary S, Fernandez L, Elazegui FA and Teng
rales et les principales maladies. Netherlands Journal of PS (2000) Development and evaluation of a multiple-pest,
Plant Pathology 93: 167–188. production situation specific model to simulate yield losses
Savary S (1987b) The effect of age of the groundnut crop on the of rice in tropical Asia. Ecological Modeling 131: 133–159.
development of primary gradients of Puccinia arachidis foci. Willocquet L, Savary S, Fernandez L, Elazegui FA, Castilla N,
Netherlands Journal of Plant Pathology 93: 15–24. Zhu D, Tang Q, Huang S, Lin X, Singh HM and Srivastava
Savary S and Servat E (1991) Un modèle simplifié de simulation RK (2002) Structure and validation of RICEPEST, a
des épidémies de rouille et de cercosporioses de l’arachide. production situation-driven, crop growth model simulating
In: Savary S (ed.) Approches de la Pathologie des Cultures rice yield response to multiple pest injuries for tropical Asia.
Tropicales. Exemple de l’Arachide en Afrique de l’Ouest Ecological Modelling 153: 247–268.
(pp. 119–145). Karthala/ORSTOM, Paris 288 pp. Willocquet L, Elazegui FA, Castilla N, Fernandez L, Fischer
Savary S and Zadoks JC (1992a) Analysis of crop loss in the KS, Peng S, Teng PS, Srivastava RK, Singh HM, Zhu D
multiple pathosystem groundnut-rust-leaf-spot. I. Six and Savary S (2004) Research priorities for rice disease and
experiments. Crop Protection 11: 99–109. pest management in tropical Asia: A simulation analysis of
Savary S and Zadoks JC (1992b) Analysis of crop loss in the yield losses and management efficiencies. Phytopathology
multiple pathosystem groundnut-rust-leaf-spot. II. Study of 94: 672–682.
the interaction between diseases and crop intensification in Zadoks JC (1971) Systems analysis and the dynamics of
factorial designs. Crop Protection 11: 110–120. epidemics. Phytopathology 61: 600–610.
Savary S, Willocquet L, Elazegui FA, Teng PS, Du PV, Zhu D, Zadoks JC (1972) Methodology in epidemiological research.
Tang Q, Lin X, Singh HM and Srivastava RK (2000a) Rice Annual Review of Phytopathology 10: 253–276.
pest constraints in tropical Asia: Characterization of injury Zadoks JC (1985) On the conceptual basis of crop loss
profiles in relation to production situations. Plant Disease assessment: The threshold theory. Annual Review of
84: 341–356. Phytopathology 23: 455–473.
Savary S, Willocquet W, Elazegui FA, Castilla NP and Teng PS Zadok JC and Rabbinge R (1985) Modelling to a purpose. In:
(2000b) Rice pest constraints in tropical Asia: Quantifica- Gilligan CA (ed.) Advances in Plant Pathology. Mathe-
tion of yield losses due to rice pests in a range of production matical Modelling of Crop Diseases. Vol. 3 (pp. 231–244)
situations. Plant Disease 84: 357–369. Academic Press, London.
Stynes BA (1980) Synoptic methodologies for crop loss Zadoks JC (1989) EPIPRE, a computer-based decision support
assessment. In: Teng PS (ed.) Crop Loss Assessment (pp. system for pest and disease control in wheat: Its develop-
166–175). Misc. Publ. Univ., Minn. Agric. Exp. Stn. 7 ment and implementation in Europe. In: Leonard KJ and
Teng PS (1983) Estimating and interpreting disease intensity Fry WE (eds) Plant Disease Epidemiology. Vol. 2 (pp. 3–
and loss in commercial fields. Phytopathology 73: 29) McMillan, New York.
1587–1590. Zimmet P, Alberti KGMM and Shaw J (2001) Global and
Teng PS (ed.) (1987) Crop Loss Assessment and Pest Manage- societal implications of the diabetes epidemic. Nature 411:
ment. APS Press, St. Paul, MN 270 pp. 782–787.