Published 1985

3
Larry A. Nelson
North Carolina State
University
Raleigh, North Carolina
Agronomic and Statistical
Evaluation of Fertilizer Regis D. Voss
Iowa State University
Response Ames, Iowa

John Pesek
Iowa State University
Ames, Iowa

The existing trend for farming operations to involve larger land and capital
commitments requires unbiased estimates of the input/output relation-
ships in fertilizing various crop-soil combinations. These relationships can
best be estimated by conducting fertilizer experiments in the field over a
series of sites and years. These experiments should be planned by an inter-
disciplinary team of agronomists, statisticians, and economists because the
validity of statistical and economic interpretations depends on optimum
design of experiments and reliable data that result from carefully applied
experimental technique. The economic interpretations also depend on
proper application of statistical techniques such as those for model selec-
tion and for parameter estimation. For comprehensive discussions of how
the agronomic, statistical, and economic aspects of fertilizer response re-
search relate to one another, see Baum (1956, 1957) and Status and Meth-
ods of Research in Economic and Agronomic Aspects of Fertilizer Response
and Use published by the National Academy of Sciences-National Re-
search Council (1961). Much of the material in these three references is
relevant to the topic of this chapter, so our major goal will be to amplify
and to update the information that was presented in them.
Sites for fertilizer response experiments must be chosen in a manner
that makes them representative of the soils and environmental conditions
under study. Yield responses need to be related to both edaphic and cli-
matic variables in order to permit extrapolation of experimental results to
individual farm situations. This implies that sites be selected to provide the
necessary ranges in these properties, which will ensure that a regression
relationship between yields and the variables can be estimated. It is also

Copyright 1985 © Soil Science Society of America, 677 South Segoe Road, Madison, WI
53711, USA. Fertilizer Technology and Use (3rd Edition).

53
S4 NELSON, VOSS, & PESEK

necessary to measure these variables during the period in which the experi-
ments are being conducted.
Fertilizer recommendations are developed using response surface
functions to estimate the input-output relationships. These functions,
which usually incorporate variables representing controlled variables and
those representing random uncontrolled (but measured) variables, are of-
ten called covariance models. These functions take on a variety of forms
depending on assumed relationships between yield, fertilizer nutrients and
edaphic climatic variables. The statistical assumptions also vary with
models, but in general, the usual multiple regression assumptions apply,
i.e., the X;'s (fertilizer variables) are fixed and measured without error, the
deviations from regression are normally and independently distributed
with mean 0 and constant variance, (12. The response surface that is esti-
mated from the data is comprised of average values of the Y variables
(yield) estimated from given combinations of the Xi.
Knowledge of the response functions is very useful for determining
fertilizer rates for unlimited and limited capital situations and for linear
programming to determine how resources should be allocated among fer-
tilization and other enterprises in the overall farm program.

I. DESIGN OF FERTILIZER EXPERIMENTS

A. Importance of Planning

Planning is one of the most important and yet neglected phases of fer-
tilizer response experimentation. It assures that the population of sites
within a given soil about which inferences are drawn is the population of
interest and also that statistical analyses will not be more complicated than
necessary. In addition, it assures that probability statements made about
parameters estimated from the experimental data will be correct or nearly
so. Planning assures that the derived fertilizer response surface relating
yield to fertilizer rate is an unbiased estimate of the true surface and that an
estimate of experimental error is available from the experimental data.
Planning also helps to assure that the size of an individual experiment and a
series of experiments will be appropriate for a particular situation. Under-
sized experiments often result in wide confidence intervals for estimates of
model parameters or for the response surface itself. Oversized experi-
ments are costly.

B. Steps in Planning

1. Selection of a Series of Sites at Which Experiments will be Conducted

Baird and Fitts (1957) discussed various agronomic aspects of design-
ing and conducting a series of fertilizer trials involving corn (Zea mays L.)
in North Carolina. Some important points concerning these aspects will be
discussed here. It is necessary to establish the universe to be explored be-
EVALUA TION OF FERTILIZER RESPONSE ss
fore selecting a series of sites on which to conduct experiments. This would
usually be a soil or a soil association in a region of a state or country. Soil
samples would then be taken from potential sites within that soil area and
would be characterized for chemical properties. This characterization in-
formation would be helpful in making the final selection of sites on which to
conduct fertilizer experiments. In order that a series of experiments be
most useful in describing the response to an added nutrient, it is desirable
to select sites that represent a range of soil levels of that nutrient from the
most responsive to just where no response occurs. Selection of sites repre-
senting only low or a narrow range of soil nutrient levels will not provide an
accurate estimate of the coefficient for the interaction of soil by fertilizer
source of nutrient; thus, the predicted yield response for other soil nutrient
levels will be biased high or low depending on the narrow range selected.
The site should show sizable response to more than one nutrient to be use-
ful in describing the production function. One would use existing informa-
tion on the correlation of yields with soil test results in the selection of po-
tential sites.
Once a set of sites (which appear likely to respond to fertilizer) has
been identified, one should consider the distribution of edaphic properties
other than those used to predict fertilizer response. In this connection, it is
useful to construct a multiway table showing ranges of predetermined
edaphic properties exhibited by sites that might be selected for the sample
and visually determine what combinations of property values would best
cover the universe. Sites possessing these combinations of properties
would then be purposely selected for the series.
Certain principles are helpful in selecting the number of sites and the
number of replications per site. First, it is necessary to use enough sites to
provide stable estimates of the parameters for the relationship between
yield and the most important site variables and to provide suitable degrees
of freedom (df) for estimating the standard errors of the parameter esti-
mates. If possible, the number of site variables should be < 25% of the
number of sites. If there were four important site variables, perhaps 12
sites would be adequate. It is also necessary to provide stable estimates of
the interaction between the added nutrient and site variables, although
presumably this would automatically be satisfied if one used the rule of
thumb given above for selecting the number of sites based on the number
of important site variables. Another consideration is that the sample size is
adequate to provide stable estimates of error terms that will be used for
testing hypotheses about parameters or placing confidence limits on them
or on predicted yields. A minimum of 10 to 15 df for each of three estimates
appears necessary. The width of the confidence intervals for the contours
representing sections through the response surface is also affected by the
number of sites. If a reasonably good estimate of experimental error is
available and if the number of replications per site has been established
(see below), one is able to calculate the widths of the confidence intervals
for the yield contours for different numbers of sites to see how they are
affected by this variable.
56 NELSON, VOSS, & PESEK

It is usually advisable to have a self-contained experiment at each site

so replication at each site is necessary to estimate experimental error. For a
fixed resource level, a greater number of replications per site comes at the
expense of fewer sites. Therefore it becomes important to weigh the rela-
tive importance of having some self-containment within sites against the
possibility of having more precision on the estimates obtained over all
sites. Three replications per site would seem to be a reasonable compro-
mise, but in some situations this may be prohibitive in cost and time. In
such cases, two would be used.

2. Selection ofExperimental Design

An essential step in planning is the selection of a suitable experimental
design. It is important to use a design that is common to the entire series
and also of the simplest form that will achieve the objectives desired. The
design should be compatible with the development of a general model that
contains both controlled and random uncontrolled variables. The design
should provide for error control, which should provide a good estimate of
experimental error. For most fertilizer experiments, the randomized com-
plete block design seems to be a reasonable choice. It is simple and flexible
and the blocking provides adequate precision. Missing plots do not cause
undue difficulty in the analysis of the data.
There are several design-model relationships that should be taken into
consideration when choosing a design. First, one needs to consider what
questions are to be answered from the study. If interactions among factors
are to be studied, a factorial arrangement is suggested. The complete fac-
torial is also appropriate if the data are to be used to confirm the original
choice of model.
If prediction is most important, the variances of the estimates of the
parameters of the model are not as important as if the objective of the study
is to determine what variables belong in the general fertilizer response
model. For the predictive model case, treatment design would not neces-
sarily be chosen on the basis of minimizing variances of the individual re-
gression coefficients.
Many fertilizer experiments have been conducted using factorial ar-
rangement of treatments. With only two levels per factor, an exploratory
2n fertilizer response experiment may be conducted to obtain the absence
or presence of effects of a number of nutrients. This would help eliminate
factors that are not relevant and to provide information about the range in
rates for more intensive future experiments. With a minimum of three lev-
els per factor, it is possible to estimate a response surface. Parameter esti-
mates for the response function, however, may not have an optimal degree
of precision with such a limited number of levels. For example, the pure
quadratic terms (1311,1322, etc.), which are often of major interest, are esti-
mated with less precision than the mixed quadratic (second-order) terms
(1312,1313, etc.).
Factorial experiments have the advantage of hidden replication (i.e.,
the other factor( s) serve( s) as a source of replication for the first). Interpre-
EVALUATION OF FERTILIZER RESPONSE 57

tation of results from complete factorials that contain four or more levels of
each factor also have the advantage that it is possible to evaluate the appro-
priate model throughout the entire design space because all rows and
columns have the same number of levels of the other factor(s). It is also
possible with factorials of this size to give a better estimate of the optimal
fertilizer rates than with the 3 n system, for which it is difficult to choose and
space the levels to provide good estimates of the regression coefficients in
the response function. With four or more levels of each factor, the number
of treatments becomes large if there are several factors. The block size
problem may ultimately decrease the precision of the experiment. This is
the greatest disadvantage of factorial experiments. This problem may be
obviated to a certain extent by using incomplete block designs that do pro-
vide more error control for large experiments (say number of treatments
;;. 27) at the expense of partial or complete confounding of certain higher-
ordered interactions with blocks. Single or fractionally replicated factorial
experiments are other possibilities. These reduce the ratio of the number
of plots in the experiment to the number of effects being estimated, and yet
the precision of these estimates is not greatly reduced. A defined estimate
of error is not available, but by making certain assumptions and using cer-
tain generally accepted result-guided procedures, a reasonably good esti-
mate of experimental error can be obtained. See Cochran and Cox (1957)
for a comprehensive discussion of factorial experiments including the sin-
gle and fractional replication versions and the incomplete block designs.
Split-plot designs are often used for fertilizer response investigations
because it is desirable to incorporate into the experiment a second factor
such as crop variety or tillage with a different optimum plot size than that
for a fertilizer treatment. From the standpoint of precision, fertilizer rate
treatments should be placed into the subplots, although mechanical con-
siderations may dictate otherwise. Generally, the whole-plot treatments
are arranged in randomized complete blocks.
There is a class of incomplete factorial designs that were developed
primarily for exploring polynomial response surfaces for two or more con-
trolled variables. The designs for exploring second-order surfaces are ap-
propriate for fertilizer response evaluation. The response surface designs
have the advantage over complete factorials in that they require considera-
bly less experimental material. The coefficients of the response model are
estimated with the same or even greater degree of precision (expressed on
a per-observation basis) than those estimated from a complete factorial.
But one limitation of these designs is that the precision is redistributed due
to the fact that there are fewer points at the extremes of the permissible
ranges of the factors. They were originally developed for use in industrial
experiments, but were later adapted for use in fertilizer response studies in
the field. Basically, the response surface designs often consist of two geo-
metric figures such as a cube and an octahedron having the same center,
hence the term composite design. Many of the industrial experiments were
replicated only on the center treatment, but for field use, a randomized
complete block design usually has been used, each block of which accom-
modates a complete set of response surface treatments, there being only a
58 NELSON, VOSS, & PESEK

single center treatment per block. Experimental error is then estimated

from the complete set of treatments rather than just from the center treat-
ment. This provides a more stable estimate of experimental error, which
represents the entire design space and which is based on adequate degrees
of freedom for a field experiment.
Box and Hunter (1957) introduced the concept of a response surface
design having a "spherical variance function," meaning that the variance
of estimated response at a given point has a value that is dependent only on
the distance of that point from the center of the design and not on the direc-
tion. Such designs are called rotatable designs because they are insensitive
to rotation with respect to the original coordinate axes. There seems to be a
case for using a design that has at least some degree of rotatability, espe-
cially if there is uncertainty where the optimum combination of fertilizer
levels will be within the design space. Rotatability may be achieved by
proper choice of spacing of octahedral points from the origin. It is also pos-
sible to assure that the variance of the estimated response value is approxi-
mately uniform over the design region by proper choice of the number of
replications of the center treatment. See Cochran and Cox (1957) for axial
spacings that give rotatability and numbers of center points that give uni-
form variance for various numbers of controlled variables used in connec-
tion with certain standard designs. Hader et al. (1957) gave a description of
the experimental design and statistical methods for characterizing a re-
sponse surface for a set of data from a biological experiment. A typical re-
sponse surface design for fertilizer response investigations is shown in Fig.
3-1.
A different and basically intuitive approach to the selection of incom-
plete factorials in fertilizer work has been employed in recent years. Cady
and Laird (1973) described a number of different incomplete factorials that
would not be classed as composite response surface designs, but that do
give a good exploration of the response surface space. One, which has de-
sirable spatial characteristics, low bias error, and low variance error, and
which has been used by workers in Latin America, is the 13-treatment de-
XI X2 X3

-I -/ -I
1 -/ -I
-I / -I
1 1 -I
-I -I 1
1 -I 1
0 -I 1 1
1 1 /
0 0 0
-(JI. 0 0
Ol 0 0
0 -Ol 0
0 Ol 0
0 0 -(JI.
0 0 (JI.

Fig. 3--1. The central composite design that is formed from a cube plus an octahedron and is
frequently used in fertilizer response investigation. The a's are coordinate values for axial
points; a for rotatability = 2kl4 = 1.68 for k = three factors.
EV ALVA TION OF FERTILIZER RESPONSE 59
tl
• •
~ • •
a:
o
I- 0
u
it:
• •

1
• •
+1 o -I
FACTOR B
Fig. 3-2. Treatment design for 52 partial factorial modification by Escobar (1967).

sign developed by Escobar (1967), which consists of 13 selected treatments

from a 5 x 5 factorial. The spatial location of the treatments of this design
is shown in Fig. 3-2. Because two or three levels of each factor are studied
at each of five levels of the other factor, a rapid graphic evaluation of the
effects of the two factors can be obtained. This is especially useful in inter-
preting data from individual sites before doing a combined analysis.

3. Selection of Treatments
Another step in planning a fertilizer experiment is the selection of
treatments. One purpose of treatments is to provide evidence of the ab-
sence or presence and, if present, the magnitude of the fertilizer effect. Of-
ten, however, they are used to provide an estimate of the response surface.
In view of these two purposes, treatments should be chosen so that they
best estimate the fertilizer effect or the response surface.
Several aspects of treatment design need to be taken into consider-
ation when planning a fertilizer response investigation. Treatment design
refers to how a limited number of points (fertilizer rate combinations)
should be distributed within a given factor space. The object is to select
that particular treatment design that will best give the specific information
on plant response to fertilizer. Cady and Laird (1973) gave the following
criteria for choosing a treatment design:

1. Interpretable data without extensive analysis

2. Relatively small number of treatment combinations
3. Low variance of estimated coefficients and, if hypothesis testing is
important, independent estimates
4. Variance of the predicted values small over the central part of the
factor space
5. Variance of the estimated response function slopes small over the
central part of the factor space
6. Bias of the predicted values small over at least the central part of
the factor space
7. Measure of lack of fit available
8. Check plot included in the design
60 NELSON, VOSS, & PESEK

9. The combinations of the zero level of one factor and high level of
another factor are excluded
They thoroughly discussed minimization of variance (4) and bias (6). From
a purely statistical point of view, the treatment designs may be altered to
give either minimum variance or minimum bias error by changing the num-
bers of levels and their spacings. Bias error refers to the failure of the
model being used to predict the response curve over the entire range of the
controlled variable. The treatment designs that produce the minimum of
one type of error are not necessarily the ones that produce the minimum of
the other type of error. Designs were chosen on the basis of variance con-
siderations in the past because it is possible to obtain a good idea of the
variance before the experiment is run by examining the properties of the
inverse matrix; also, the true model has to be known to study bias, a situa-
tion not found in practice.
It is very important to be careful in choosing the fertilizer rates to be
used in an experiment. The question of the optimal number of incremental
applications and their placement when considering rates to be used must
also be addressed. The lowest and highest rates of each fertilizer nutrient
are completely at the discretion of the researcher, but they should be cho-
sen so that the responsive portion of the range is bracketed and there is at
least one rate beyond the responsive range. The center of the range should
be in the vicinity of the anticipated optimum rate for that particular fertil-
izer nutrient. This is because, statistically, the variance of predicted yield is
smallest at the center.
Centering of the range in the vicinity of the anticipated optimum may
force one to use a rate other than zero for the lowest rate. In this case, a
check treatment will need to be included in addition to the response sur-
face design points. The spacing of the fertilizer rates will be determined by
the particular treatment design being used.

II. FIELD PLOT TECHNIQUE

A. Decisions Related to Use of Randomized Complete Block Design

The randomized complete block design is a good choice for fertilizer
response investigations because the blocks control site variation, thereby
increasing precision. To minimize the extent of the area involved and thus
keep the experimental area relatively uniform, block size should be held to
15 to 20 plots if possible. The blocks should be square or nearly square in
shape and should be oriented in such a way that, although the blocks may
differ considerably from one another, the units within blocks are relatively
uniform. Usually the blocks are placed side by side so that they control var-
iation along a fertility or moisture gradient.

B. Randomization of Treatments
Randomization is employed in connection with use of the randomized
complete block design. Randomization refers to the process of assigning
EV ALVA TION OF FERTILIZER RESPONSE 61

treatments to the plots in such a way that all treatments have an equal
chance of being assigned to a particular plot. Randomization provides as-
surance that a treatment is not continually favored or handicapped in vari-
ous replications due to some extraneous source of variation, known or un-
known.
Randomization for each experiment in a series of experiments should be
performed independently. This will assure that no systematic biases will occur
throughout the series and that estimates of site, year, site x treatment,
year x treatment, and site x year x treatment effects will be valid.

C. Size and Shape of Plots

The plot size and shape that will give minimum error variance, and
hence can be called optimal, will vary depending on the crop being studied.
In general, plots should be long and narrow rather than square. The long
dimension of the plot should be parallel to the fertility gradient if one ex-
ists. Four-row plots 6 to 10 m long with only the center two rows harvested
for experimental purposes are usually adequate for row crops such as corn
and soybean [ Glycine max (L.) Merr.]. One should refer to the literature
to find appropriate field plot technique recommendations for a particular
crop. An example is the handbook on rice (Oryza sativa L.) experimenta-
tion by Gomez (1972). There are quantitative methods for determining op-
timal plot size empirically based on the original work of Smith (1938) and
later that of Hatheway and Williams (1958). Optimal plot sizes for various
crops, which were determined using their techniques, have been reported
in the literature.
Mechanical considerations will often dictate what size and shape will
be used, in spite of what is considered to be an ideal plot size and shape.
Uniform spacing of plants within the row should reflect the plant popula-
tion desired as well as results of past investigations concerning spacing x
fertilizer interaction. Often the plants will need to be sampled before ma-
turity to determine quality and plant tissue composition. This can best be
accomplished by taking a systematic sample with a random start. To imple-
ment this, one would choose a plant at one end of the plot randomly and
then take every kth plant (e.g., every 10th plant) along the row to include
in the sample. See Federer (1955) for more detailed information on size
and shape of plots as well as the statistical aspects of sampling within plots.

D. Use of Border Rows

Border rows are necessary in situations where the researcher con-

siders that fertilizer treatments imposed on one plot might affect neighbor-
ing plots. A correct judgment of the degree of this influence is important.
Providing insufficient border area can produce interplot interference that
may cause representational or cryptic error. Having too much area in bor-
der rows would result in wasted space and an increase in area in the experi-
ment, which could result in a large error variance.
62 NELSON, VOSS, & PESEK

E. Use of Uniform Experimental Technique Throughout the Series

Overall precision can be increased by using a uniform experimental

technique throughout the series of experiments. Some ways of standardiz-
ing technique are to:
1. Write out procedures for conducting various phases of the experi-
ments and a time schedule for their execution
2. Make all personnel dealing with the treatments, plots, and data
aware of the various sources of error and the need for good tech-
nique
3. Apply the treatments uniformly
4. Exercise sufficient control over external influences so that every
treatment produces its effect under controlled, comparable condi-
tions. For example, having each of three persons harvest an individ-
ual replication is better than having all three harvest all replications
as a team. Still better is to have one person harvest all three replica-
tions. If it is impossible to control environmental conditions, obser-
vations on major environmental variables should be taken and used
as covariables
5. Devise suitable unbiased measures of the effects of treatments
6. Prevent gross errors

F. Recording of Site Data for Predetermined

and Random Uncontrolled Variables
It is necessary to characterize each site used for experimentation for
principal soil and meteorological characteristics that might affect yield,
quality, and plant composition. Selection of site variables to be measured
should be based on agronomic criteria. These site data are very important
in a fertilizer response study; care should be taken in both the planning and
measurement phases so that the relevant and reliable site data will be avail-
able at the time of data analysis. Usually one will need to take readings on
many more variables than will appear as site variables in the general
model. The most critical stages in the growth of a particular plant species
are the times at which it will be important to make observations on specific
variables (such as temperature). The dates of specific physiological devel-
opment stages such as anthesis and maturity can be very important in ex-
plaining treatment effects. An incorrect judgment concerning timing could
render the variable useless in the data analysis phase. Observations on ran-
dom uncontrolled events such as pest attacks, flooding, etc. should also be
made and carefully recorded.
G. Measurement of Soil Properties of Individual Plots
If there is a potential for sizable variability in chemical and physical
properties within a site, the site should be characterized by individual plot.
It is then possible to estimate the within-site interactions involving the soil
properties and the added nutrients in addition to among-site interactions.
EVALVA TION OF FERTILIZER RESPONSE 63

III. PHYSICAL RESPONSE MEASUREMENTS

Although yield is the most common physical response measured, mea-

surement of certain other characteristics may provide useful information
on the mechanism of plant response. Variables such as chemical composi-
tion of plant tissue, both grain and vegetative, are useful in this connection.
In some cases, complete biomass yields are also measured because they
provide more complete information on utilization of nutrients. Physical
and biological quality measurements of the product are also important be-
cause they affect its nutritive and economic value.
Plant survival is another response that is measured and used for con-
ducting a covariance analysis to adjust yield to a constant stand. Causes of
stand losses also must be recorded.

IV. STATISTICAL ANALYSIS OF FERTILIZER RESPONSE DATA

A. Adjustment of Yield Data to a Specified Moisture Content

Grain yields are often adjusted to a standard moisture content to make

them comparable from one treatment or study to another and relate them
to the market. This is accomplished by use of the following formula:

Yield at standard moisture content = W [(100- A )/(100- M)]

where W = weight of harvested grain, A = actual moisture content of

grain (% ), and M = standard moisture content of grain (% ).
As an example, suppose one plot yielded 25 kg corn that had a mois-
ture content of20.0%. This yield in kilograms corn per plot at 15.5% mois-
ture is given by:

Yield at 15.5% moisture = 25 [100%-20.0% )/(100%-15.5%)]

= 23.7 kg/plot.
The moisture adjustment must be done separately for each plot in an ex-
periment.

B. Analysis of Variance for Individual Site-Year Data

One type of data analysis that should be routine is an analysis of vari-

ance of data for each individual site-year. The analysis normally would in-
clude single degree of freedom components representing terms of a
second-order response surface either in the original X scale or on a trans-
formed scale as subdivisions of the treatment sum of squares. A typical
analysis of variance key-out for aN, P, and K fertilizer experiment using a
randomized complete block design and fitting a second-order response sur-
64 NELSON, VOSS, & PESEK

Table3-1. Analysisofvariancekey-outfora4 x 4 x 4N,P,andKfertilizer

experiment according to a randomized complete block design with five blocks,
assuming the fitting of a second-order response surface.
Source df
Blocks 4
Treatments 63
N 3
NJin 1
Nfu'd 1
Ot er 1
P 3
PJin 1
P quad 1
Other 1
NP 9
NJin X PJin 1
Other 8
K 3
KJin 1
Kfu'd 1
Ot er 1
NK 9
Nlin X KJin 1
Other 8
PK 9
PJin x KJin 1
Other 8
NPK 27
Error 252

face is shown in Table 3-1. F tests of significance are conducted for treat-
ments, and all single degree of freedom components. The sources labeled
Other and NPK are usually pooled to give a "Lack of Fit" source of varia-
tion and this mean square is tested with the error mean square to check for
inappropriateness of the model. For the example given in Table 3-1, the
Lack of fit has 54 df. The results give an indication as to which nutrients the
crop responded and how well the model fits.

C. Combined Analysis of Variance Over Locations and/or Years

A combined analysis of variance for data from all sites and/or years is
conducted after performing the individual site-year analyses of variances.
This analysis is necessary to test for treatment x environment interactions.
It also allows a test of the treatment main effects averaged over all sites
and/or years. In addition, it provides estimates of the error terms that will
be used in deriving errors for testing various terms in the response func-
tion. For example, the error term for testing site variable terms is usually
different from that used to test fertilizer main effect terms.
The format for an analysis of variance of data combined over sites
within a year is shown in Table 3-2. An analysis of variance of data com-
bined over sites and years has the form shown in Table 3-3. Treatments for
both of the analyses are subdivided in the same manner as shown in the
analysis of variance key-out for the individual site. The interaction terms
EVALVA nON OF FERTILIZER RESPONSE 6S

Table 3-2. Format for an analysis of variance key-out for data combined
over sites within a year. t
Source df
Sites(S) 8-1
Replications in site 8(r-l)
Treatment (T) t-l
SXT (8 -1)(t -1)
Error 8 (r - l)(t - 1)

t Capital letter abbreviations refer to the name of the effect. Lower case letter abbrevia-
tions refer to the number of levels.

Table 3-3. Format for an analysis of variance key-out for data combined
over sites and years. t
Source df
Sites (S) 8-1
Years(Y) y-l
SX Y (8 -1)(y - 1)
Replications in (S X Y) 8y(r-l)
Treatment (T) t-l
TXS (t - 1)(8 - 1)
TX Y (t - l)(y - 1)
TX YXS (t - 1) (y - 1) (8 - 1)
Error sy (t - l)(r - 1)
t Capital letter abbreviations refer to the name of the effect. Lower case letter abbrevia-
tions refer to the number of levels.

may also be subdivided to show how individual components of treatments

interact with sites and years.
Error terms from the individual site-years are pooled to obtain the
sum of squares for Error in both of the analyses. Prior to pooling, there
should be a reasonable indication that the variances from the individual
site-year analyses are homogeneous. This can be checked with a test for
homogeneity of variance such as Bartlett's (1937) or Hartley's (1950) tests.
There will often be a few site-years within a series that should be consid-
ered as outliers and deleted from the combined data set. The results of one
of these tests of homogeneity of variance as well as examination of the data
scatter will help to identify these sites.
The appropriate mean square to use as the denominator in the F ratio
having a given mean square as numerator may be inferred by writing out
expectations of the mean squares for the various sources of variation.
Rules for writing these are given by Schultz (1955) and Steel and Torrie
(1980). One should allow room for some flexibility in methodology in test-
ing procedures based on the unique set of circumstances connected with a
particular series of experiments.

D. Analysis of Covariance in Analyzing Fertilizer Response Data

There are two situations in which analysis of covariance may be used

to advantage in the analysis of fertilizer response data. This technique is
sometimes used to provide estimates of missing plot values that result in
66 NELSON, VOSS, & PESEK

unbiased estimates of treatment sums of squares. It is also used to adjust

yield or other response variables for differences in the level of some
observed uncontrolled variables such as plant population per plot. In con-
nection with the second type of application, the precision of the analysis is
often improved considerably. Heady and Pesek (1957) found the relation-
ship between plant population and corn yield to be highly significant in the
analysis of data from an experiment on a Haynie soil (coarse-silty, mixed
[calcareous], mesic Mollie Udiftuvent) in Iowa; therefore, they included
this as a term in the response function derived from that particular set of
data.

E. Response Surface Fitting to Data From Individual Site-Years

The treatment sum of squares should be subdivided in the analysis of

variance into components for terms of the response surface model (see sec-
tion B above). Our discussion here deals with the choice of an appropriate
model that will be used for fitting data from the individual site-years and
ultimately the general response function. Mead and Pike (1975) stated that
the choice of a particular response function in the analysis of data is often
determined by several conflicting objectives. These usually include a
purely descriptive reduction in the data, a short -term prediction for partic-
ular combinations of factor levels, a long-term prediction for the general
pattern of response, and a simple interest in the "true" pattern of re-
sponse. Models have been classified into two broad categories, empirical
and biological. The parameters of an empirical model do not necessarily
have biological meaning, whereas those of a biological model are consid-
ered to be approximations of some aspect of a biological process. At this
point the distinction between empirical and biological models seems more
tenuous than it was previously considered to be. Perhaps the problem rests
more in the difficulty with defining the complex underlying biological proc-
esses and their interactions than in dealing with their mathematical quanti-
fication.
The most frequently used response function in fertilizer response
studies is the polynomial of degree, p. Many workers consider this an em-
pirical model, although Niklas and Miller (1927) contended that there was
a biological basis for use of a second-order polynomial model. Polynomial
models are easily fitted by least squares methods ~nd are easily generalized
to a multinutrient relationship represented by Y regressed on several X
variables. Mead and Pike (1975) state that apart from a straight line, the
quadratic (p = 2) is the most commonly used polynomial. It has the fol-
lowing form for a single nutrient:
[1]
In particular, for fitting a response surface for N, P, and K the form is as
follows:
Y = bo + b1N + b2P + b 3K + b ll N2 + b22 P2 + b33 K2
+ b12NP + b13NK + b23 PK. [2]
EV ALVA TlON OF FERTILIZER RESPONSE 67

These authors listed the following reasons for the general popularity of the
quadratic model: (i) it involves merely the addition of an extra term to the
straight line relationship which, for most people, makes it the simplest cur-
vilinear relationship; (ii) it has a simply defined maximum at X = - b]/2b]"
and (iii) the method of least squares produces estimates of parameters
without complex calculations. Another advantage is that the scope of yield
response patterns that may be fitted within the polynomial family using
least squares procedures is broad due to the possibility of making various
transformations primarily of the X variables.
One disadvantage of the quadratic model is that it does not allow for
asymmetry around the optimum in the yield response pattern. This asym-
metry often occurs in actual practice. Also extrapolation outside of the ex-
perimental range of X values is impossible.
Some have used models containing polynomials in transformed X var-
iables. The most commonly used transformation model is the square root
model (X is replaced by X 1/2 in the quadratic model). The equation for this
model in the single nutrient case is

[3]

The curve reaches a maximum at X = (-b,/2b ll )2 and then declines. It is

asymmetrical around the maximum. When fit to the same data, predicted
values from the square root equation rise more rapidly than those for the
quadratic because the slope of the square root curve tends to infinity as X
tends to zero. The curve for the predicted values of the square root model
often tapers off to a broad area in the region of the maximum that has less
slope than the quadratic curve predicted from the same data. The physical
maximum may be at either a higher or lower value than that for the qua-
dratic, depending on the yield pattern for the high rate points.
Fitting a response surface for N, P, and K by the mathematical form of
the square root response surface gives:

[4]

The square root model, mentioned by Heady et a1. (1961) for fertilizer
recommendation purposes, was also favored by Abraham and Rao (1965).
More recently it has been used by the Australian National Soil Fertility
Project for the analysis of field data (Colwell, 1979), and it was the princi-
pal model discussed in Colwell's (1978) book on design and analysis of field
trial data.
Heady et a1. (1961) used a polynomial model having a linear term, but
also a 1.5 or 3/2 power op the second term. Anderson and Nelson (1971)
used models of the type Y = b1(X + d) + b2(X + d)h, where Xis added
nutrient, d is soil nutrient (estimated from the data), and h is an exponent
that was given each of the following values: 0.50 (square root model), 0.75,
68 NELSON, VOSS, & PESEK

0.95,1.25,1.50,1.75, and 2.00 (quadratic model). In another approach to

the estimation woblem, h, an optimal value of h, was determined simul-
taneously with d for each site-replication using iteration techniques. A re-
striction that 0.25 ~ II ~ 2.00 was imposed. The criterion for optimality
was minimum residual sum of squares. The value of II varied considerably
within a site, implying that the same polynomial model would not be ap-
propriate for all replications, even at the same site. There was almost as
much variation in II within sites as among sites.
Neider (1966) developed a group of empirical models called inverse
polynomials, which he claimed are more flexible and realistic than ordi-
nary polynomials. These have the general mathematical forms:
/'...
y-I = aX-I + b [5]
and .A.
y-I = aX- I + b + eX [6]

for a saturation effect curve and a toxicity effect curve, respectively. They
have been used primarily by British scientists. The second curve is analo-
gous to an ordinary quadratic polynomial but is not constrained to be sym-
metrical. These are somewhat more complicated mathematically than the
ordinary polynomials, both from the standpoint of assumptions about the
distribution of errors and the fitting procedures.
The power function (Cobb-Douglas) is an empirical model that has
been employed in a number of fertilizer response investigations. It has the
general form for a single nutrient, such as nitrogen (N)

[7]

where Yis the predicted yield, a and b are constants to be estimated from
the data, and N is the rate of added nutrient. The equation may be trans-
formed to linear form
/'....
(log Y) = log a + b log N. [8]

The Cobb-Douglas function when generalized to the multinutrient case

takes on the following form:

[9]

where a, b, and N are as defined above, P is phosphorus, and K is potas-

sium. The constants for P and K are e and d, respectively. The most serious
limitation of this function is that it does not provide for a maximum yield
and hence cannot adequately describe the complete range of responses ob-
served. It is not even asymptotic to the maximum.
Boyd (1972), Anderson and Nelson (1975), and Boyd et al. (1976) re-
ported the use of intersecting straight-line models, which was prompted by
their observation that the quadratic and square root models gave biased
estimators of the optimal fertilizer rates when the yield had a plateau re-
EV ALUA nON OF FERTILIZER RESPONSE 69

sponse pattern. The advantage is flexibility of the model, i.e., a model or

submodel that fits the data well is chosen from a family of models and this is
more apt to provide unbiased estimates of optimal fertilizer rates in pla-
teau response situations than would traditional models. Also, the models
are readily fit using unsophisticated computing equipment. The reported
disadvantages are that these models are mathematically unsatisfying and
that the discontinuities that they have at the intersection points are unreal-
istic. In addition, these models do not lend themselves to situations where
interactions between nutrients exist or where a general model that includes
site year x added nutrient terms is developed for a combined data set from
a series of site years.
Perhaps the model that most workers have traditionally identified
with the "biological model' concept is the Mitscherlich equation (Mits-
cherlich, 1909), which is also called the negative exponential model. This
equation was based on Mitscherlich's observation that the response to an
increment of fertilizer was proportional to the decrement from the maxi-
mum yield. He formulated the following differential to express this rela-
tionsh~p

8 YI8X = c(A - Y), [10]

where 8 Y 18 X is the differential of yield (Y) with respect to added nutrient

(X), c is a response coefficient, and A is the asymptotic maximum yield.
When originally developed to apply to plant response in sand culture, the
mathematical equation resulting from integration of Eq. [10] was
Y = A[l-exp( -cX)), [11]

where Y = predicted yield and the other terms are as previously defined.
This equation was later modified by Baule (1918) for use with a series of
soils to include the initial amount of nutrient in the soil, d, which is ex-
pressed in fertilizer equivalent units
Y = A {l-exp[ -c(X + d)]}. [12]

The two variable generalization of this model is

Certain attributes of the Mitscherlich equation originally seemed at-

tractive from a biological point of view. The primary one was the concept
of a maximum yield, A, which is asymptotically approached as fertilizer is
added. There were statistical limitations, however. This equation is more
difficult to fit (even with modern digital computers) than models that may
be fit by least squares techniques. There may be problems achieving con-
vergence if the data are imprecise or if there are only a few levels of treat-
ments. The Mitscherlich equation is not as flexible as some of the other
models such as the quadratic in terms of developing a general model that
includes variables to explain variation among site-years.
70 NELSON, VOSS, & PESEK

Some workers contend that the MitscherIich equation is not really a

biological model but rather an empirical model. The fact that there have
been two revisions of the mathematical form of the original model based on
modified views of the biological processes and how they should be quan-
tified lends support to this view. The first revision resulted from the empiri-
cal observation that in many cases the yield decreased beyond a certain X.
This so called second approximation, which was formulated by Mits-
cherlich (1930), took the form

Y= A[I-exp( -eX)] [exp (-kX")], [14]

where k is the damage factor and other terms are as before.

von Boguslawski and Schneider (1962) proposed a third approxima-
tion to the Law of Yield, which was a follow-up to the first two approxima-
tions. It has the form

Y = MlO-Z{log[(X + i)/(m + i))}n,

where M is the maximum yield obtainable; m is the corresponding input of

X; i can be interpreted with some approximation as the fertilizer equiva-
lent of the soil; Z and n are parameters controlling the shape of the curve;
X is the amount of applied nutrient; and Y is the predicted yield. This
model is extremely flexible but involves a large number of parameters that
must be estimated.
Balmukand (1928) developed another so-called biological model that
is based on a completely different theory of biological response. Like the
Mitscherlich, it has an asymptotic maximum to which the predicted Y con-
verges as the nutrient rate approaches infinity. It has a hyperbolic form,
which is related to NeIder's (1966) inverse polynomial models. It assumes a
biological analogy to Maskell's electrical resistance formula. The underly-
ing supposition is that each activity of the plant (yield, etc.) is determined
by a potential set of resistances, each of which represents one of the exter-
nal factors.
Balmukand's equation for the relationship of yield to N, P, and K is
/'..

[lIY] = e + [ani (n + N)] + [a p / (p + P))/ + [ak/ (k + K)),

"'-
where [lIY] is the predicted reciprocal of the yield for a given combination
of added N, P, and K; e is a general constant; n, p, and k are soil levels of
the nutrients that are parameters to be estimated from the data; and N, P,
and K are added levels of the nutrients. The am ap , and ak are constants that
express the importance of the three nutrients to the crop. They are esti-
mated from the data.
Although the literature on fertilizer response modeling is extensive,
there are few studies dedicated to comparing different models. The reason
is that statistically valid comparisons among mathematical models of dif-
fering functional form are difficult to make. It is also difficult, if not impos-
EVALUA TlON OF FERTILIZER RESPONSE 71

sible, to compare models on the basis of correctness of fertilizer recom-

mendations when in fact the true recommendation is never known. It does
appear, however, that the choice of model has a considerable effect on the
estimates of optimal fertilizer rates.
Cady and Laird (1973) studied the effect of treatment design and pos-
tulated model on the bias error. By bias error they meant the integrated
area between the true response curve and the curve for the model being
used for fitting the data. Of the factors studied, they found that choice of
the model had the most important effect on bias error.
Abraham and Rao (1965) compared the following functions as to suit-
ability for fertilizer recommendation purposes: Mitscherlich, Cobb-
Douglas, hyperbola, quadratic, and square root transformation of the qua-
dratic. All models fit reasonably well to rice yields. The estimated optimal
rates from the quadratic and Mitscherlich were close to each other and
their estimated standard errors smaller than those for the other models.
On the basis of the ease of fit and results of tests of hypotheses about model
parameters, the quadratic model was favored.
Jonsson (1974) studied the effect of varying N ranges on the choice of
an optimum model among the quadratic, square root, and an extended hy-
perbola model. It was concluded that at a narrow range in N, the quadratic
and square root equations are similar and superior in describing the data,
whereas the hyperbola is inferior. At a wider range of N rates, the super-
iority of the square root and extended hyperbola over the quadratic was
noted. The parameter estimates of the quadratic equation were found to be
very dependent on the number of N rates used in the fit. This would suggest
that the fertilizer response curves used in practice were not symmetric and
this would militate against the quadratic model, which assumes symmetry.
Anderson and Nelson (1975) compared several models with respect to
the optimum amount of nutrient to be applied. They generalized by stating
that if one is interested in obtaining the optimal fertilizer recommendation
(Xo), an estimating procedure that gives a good Xo estimate, rather than a
minimal residual sum of squares, should be developed.
Mombiela and Nelson (1980) compared the quadratic, square root,
and Mitscherlich functions on the basis of the functions' recommendation
bias, variability of recommendation, and expected economic loss for data
generated from an underlying response function for P whose parameters
were prespecified. They concluded that response surfaces should be mech-
anistic rather than empirical if good performance over a wide variety of
situations is desired. The quadratic surface was found to be too rigid to ac-
commodate non symmetrical shapes and its performance was reported to
be greatly affected by price ratio fluctuations. The performance of the
square root surface was satisfactory at medium and high price ratios if the
soils involved had medium to high soil test values and, thus, small yield
responses.
Heady et al. (1961) used corn response data from an Ida silt loam soil
(fine-silty mixed [calcareous], mesic Typic U dorthent) to compare the fol-
lowing models with respect to shapes of isoquants and isoclines as well as
72 NELSON, VOSS, & PESEK

estimates of economic optima: square root, quadratic, 1.5 power model,

and Cobb-Douglas model. They concluded that while each might be ap-
propriate for certain purposes, the square root function appeared to have a
slight superiority in statistical efficiency in predicting the production sur-
face for the particular environmental conditions under which the experi-
ment was conducted.
In summary, concerning model selection, one cannot recommend a
single model for all situations. The nature of the actual crop response and
the relevant area of interest should be considered in choosing a model. It
would seem that the square root transformation of the quadratic model
would be a reasonable choice in many situations unless there is a definite
"plateau" response pattern. In this case, some of the straight-line models
would seem applicable. If the quadratic model is being considered, great
care should be taken to see that the model will not introduce upward biases
into the estimate of the optimal fertilizer rate. Use of the Mitscheriich
equation would seem to be impractical from the standpoint of a general
model that includes several fertilizer nutrients, several site-years, and in-
teractions between the nutrients and the site-year variables.
Using least squares, one is able to estimate the regression coefficients
of the polynomial models and the variances of these estimates. The pre-
dicted yield for a given combination of the Xi and its variance are also avail-
able. This latter variance is a function of the particular combination of the
Xi for which the response is estimated, but it is also dependent on the de-
sign used for the experiment.
If we assume the general model is Y = b o + b,X + b 11 X2, then the es-
timate of optimal fertilizer rate is (R-b 1 )/2b J1 , with R being the cost of a
unit of fertilizer/price of a unit of product. An exact estimate of the vari-
ance of this optimal rate is calculable, but because the denominator, which
is very small and unstable, has an important effect on this variance, this
variance is often quite large. The regression coefficients and their standard
errors are more involved to calculate for the inverse polynomial, the Mits-
cheriich, and Balmukand Resistance model. The latter two models require
iterative fitting techniques.

F. Two-Dimensional Contour Plots of Yield for Individual Site Year Data

Heady and Dillon (1961) have described the properties of the re-
sponse surfaces generated by the most frequently used response models
with two variates. These are the Cobb-Douglas equation, the Mitscherlich-
Spillman equation, the resistance equation proposed by Balmukand, and
the quadratic and square root transformation ofthe quadratic equations. It
is possible to plot contours in the space of X" X 2 , ••• X k along which the
yields are constant (called isoquants). If the design is rotatable and the con-
tours are circles, spheres, or hyperspheres around the center of the design
region the variance of predicted Yvalues should be equal along a contour
meaning that the predictions are equally reliable. In Fig. 3-3 is shown a
two-dimensional plot of yield contours for a response surface estimated
from a N x P experiment. The visual representation gives an idea as to how
EVALUATION OF FERTILIZER RESPONSE 73

' \\
\
"' \
\, \.
\
"
\
\
\
\
\\

/"---- ............
/ "
I
/ " ",
I \
I \
I \
I .... - , \
I
I
: I
,I '"
" \ \
\

\
I
I
I \ \\ PRATE
I I \ \
I I \ \
I \ \ \
I \ \ \
\ \
\ \ \ \
\ \ \ \
\ \ I \
\130 \140 \150 kg/ho yield I \
\
\
\ \ I:
\ \ " / I
\ \, " ,_...... / I
,, \
\ , I
'- \
, I
\
... \
\
\
\
, I
I
I
, \

., ,
\

\
\\

'\
\
\
\
\
"
" ' ............
J
__ ....... /
/
/

N RATE
Fig. 3-3. Two-dimensional plot of ellipsoidal yield contours for a response surface estimated
from a N x P fertilizer experiment.

responsive the yield measure is to these two fertilizer nutrients and how
these nutrients interact with each other. For cases involving three or more
nutrients, a two-dimensional plot may be done at each level of the third
factor.
There are four fundamental and limiting surfaces generated by a
second-degree equation in two dimensions and they have distinctly differ-
ent appearances (Fig. 3-4). Figure 3-4a is a system of closed ellipsoids,
which has a unique maximum in the region investigated. This is a desirable
surface to obtain from the standpoint of estimating the optimal fertilizer
rate. Figure 3-4b depicts a saddle point (sometimes called a color mini-
max). Such a surface would be difficult to interpret agronomically. Data
that give rise to such a surface could result from an experiment that had a
poor choice of treatment levels and/or that lacked precision. Substantia-
tion of this type of pattern would be called for before extending the appli-
74 NELSON, VOSS, & PESEK

(a) Maximum, (b) Saddle point i

Eliptical contours Hyperbolic contours

'~
~ ...
(c) Stationary ridOei (d) RIslno ridoei
'.

No unique maximum Parabolic contours

Fig. 3-4. Four fundamentaiiimiting surfaces generated by a second-degree equation in two

dimensions.

cation very far. When conducting the substantiation experiment, more

treatment combinations in the vicinity of the assumed optimum would
seem in order.
Figure 3-4c and 4d represent essentially limiting cases. Figure 3-4c
shows a stationary ridge pattern, which might be regarded as a surface such
as 4a and 4b, which is infinitely attenuated along the X 2 axis. Figure 3-4d
shows a rising ridge in which the contours of the ridge are parabolas . The
center is at infinity. For a more complete discussion of the forms of re-
sponse surfaces see Box (1954).
With modern computing equipment and software packages it is possi-
ble to plot the contours in two-dimensions or even to plot a three-
dimensional surface. These plots provide information that should supple-
ment the results of the statistical applications such as tests of significance of
the regression coefficients of the model.

G. Response Surface Fitting to Combined Data From all Sites and Years

A response surface may be fitted to data from all sites and years using a
model that includes the same applied nutrient terms as in the individual site
model along with terms to explain among site variation, among year varia-
tion, site x nutrient, year x nutrient, and site x year x nutrient variation.
This usually is the model that is used for making general recommendations
for a crop on a specific soil. The inclusion of site and year terms often
causes sizeable adjustments in the applied nutrient regression coefficients
so that estimates of optimal fertilizer rates will be very sensitive to the par-
ticular site and year variables and their interactions with added nutrient,
which are included in the model.
EVALUATION OF FERTILIZER RESPONSE 75

In spite of this model specification difficulty, it would seem that mak-

ing recommendations from such a general response surface model would
be more efficient than obtaining estimates of optima from individual sites
and then averaging them. It would also permit tailor-making predictions to
individual conditions by substituting values for individual site-year condi-
tions into the model. Obtaining optima estimates from main effect treat-
ment means averaged over sites and years would seem to gloss over the
very real possibility of interaction between added nutrients and site and
year variables.
Given that a model has been chosen to explain the variation attribut-
able to added nutrients for an individual site-year, there is some uncer-
tainty as to which site-year variables and site-year variable x added nutri-
ent terms should be included and in what form they should be in the model
being developed for the combined data.
One question that usually arises is whether to combine soil nutrient
with added nutrient or to carry the former in the model as a covariable. For
a single nutrient the two forms would be

[15]
or

where X = added nutrient, d = soil nutrient, and the bi and b; are regres-
sion coefficients estimated from the data. Equation [15] would be much
more efficient from a statistical and agronomic point of view in terms of low
variances of regression coefficients and predicted yields, compactness of
the model, and agronomic realism. The exact soil nutrient level is never
known, however, so it must either be estimated from the data iteratively in
fertilizer equivalent units or else the soil nutrient levels must be estimated
from soil tests and then combined with the added nutrient levels. Soil test
values are only indexes of nutrient availability and are not equivalent units
of the fertilizer source of the nutrient. Hildreth (1957) dealt with the prob-
lem of the estimates of soil nutrient obtained by tests having a different
fertilization efficiency than the added nutrients. Mombiela et al. (1981)
more recently suggested an approach to determine from the data the math-
ematical form of the relationship between the d estimated from the data
and soil test results. Once a reasonably good estimate of the form of this
function is known, one can replace d by the function in the model that com-
bines soil and added nutrient.
The form given in Eq. [16] has been used by Voss and Pesek (1967),
Voss et al. (1970), Ryan and Perrin (1973), and Colwell (1967,1968,1970,
1979). It has the advantage of being convenient and easy to fit by least
squares methods. At this point, there seems to be no overwhelming advan-
tage of one form of model over the other.
One may use an agronomic approach to the selection of a general
model that would initially involve the controlled variable terms plus other
variables that agronomic theory would suggest are important and subse-
quently those whose relationships with yield are not supported by the data
76 NELSON, VOSS, & PESEK

would be removed. The agronomic terms would include the soil nutrient
variables as described above. Certain more permanent soil properties such
as cation exchange capacity and mineral index (14An A peak height ratios
on x-ray diffraction patterns) were also found by Nelson and McCracken
(1962) to be useful for explaining among site variation.
Another class of variables that is necessary in most general models is
those that explain weather effects. Sopher et al. (1973) discussed three ap-
proaches to reducing daily weather data to a reasonable number of mean-
ingful variables. These included (i) use of seasonal, monthly, bimonthly,
or weekly climatic data in the model; and (ii) use of a polynomial regres-
sion method devised by Fisher (1924) in which the growing season is di-
vided into n short intervals, a kth degree polynomial is fit to the n intervals,
and the k orthogonal polynomial regression coefficients are then entered
into the general model as independent variables. Such an approach was
used by Hendricks and Scholl (1943), Runge and Odell (1958), and Runge
(1968). The third method is employment of the plant-available soil mois-
ture holding capacity, potential evapotranspiration, and rainfall data to
calculate a water budget for a specific soil. From the water budget, the
number of moisture deficit days in each of several crop growth periods is
noted. These moisture deficit days in each growth period are then used as
independent variables in a general model. Mason and Cooper (1958), as
quoted in Sopher et al. (1973), incorporated the numbers of drought days
(van Bavel, 1953) for four selected physiological growth periods in the
growing season of corn into a single-variable drought index. This was
achieved by regressing yield of corn on the number of drought days in each
of the four physiological periods, and then obtaining a weighted average of
the number of drought days using the regression coefficients for the four
periods as relative weights. The weighted average is the drought index.
The index obtained by the above procedure was then used as an indepen-
dent variable in the general model that was fit to data combined across site-
years. Sopher and McCracken (1973) used this drought index in develop-
ing yield response models for corn production in the Coastal Plain of North
Carolina. Others have found that the stress day concept describes the reac-
tion of plants to climate quite well (Denmead & Shaw, 1962). In general
terms, a stress day is a day when conditions of atmospheric demand for
moisture, water availability in the soil, and stage of crop development are
such that adequate moisture for meeting evapotranspiration demands can-
not be extracted from the soil by the crop and the plants begin to wilt.
Sopher et al. (1973) were careful to point out that the degrees of free-
dom associated with the estimation of the drought index should be equal to
the number of individual drought components composited in the index plus
one for the index itself. They also made the observation that excess mois-
ture and cool temperatures should be included along with drought mea-
surements early in the growing season.
Once it has been established which variables agronomic theory sug-
gests should be included in the general agronomic model, some objective
methodes) should be employed to determine which variables are contribut-
ing to the prediction of yield and therefore should be retained in the regres-
EVALUATION OF FERTILIZER RESPONSE 77

sion equation. In general, those variables whose regression coefficients

(hi) are greater than their standard errors (Sbi) are contributing more than
"noise" to the prediction of yield. This "sorting out" of variables for which
the ratio = hi/ Sbi = tb ;:: 1 is a less conservative approach than making ac-
tual tests of significance to determine which variables should be retained in
the model. It is for this reason that we recommend this procedure, espe-
cially in the early stages of the variable selection process.
The agronomic approach to model selection, although reported by
Laird and Cady (1969) to be superior to an empirical approach, has the
general limitation that the functional relationship between yield and the
combined set of controlled and uncontrolled variables is not known due to
the complexity of biological systems. Another choice would be to use a
more empirical approach based on fitting alternative models not necessar-
ily having an agronomic basis to the data and evaluating the fit based on
criteria such as significance of regression coefficients, minimum mean
square error, Mallows Cp statistic, etc. Should several models perform sim-
ilarly under the previous criteria, the final choice among them is made ac-
cording to simplicity or convenience. For a discussion of the relative merits
of the agronomic and empirical models, see Laird and Cady (1969).
These two authors noted that error structure should be considered in
testing the regression coefficients in either type of model. A combined
analysis of variance similar to the ones shown in section C above should be
conducted first to obtain estimates of the errors (replications in site or rep-
lications in site-years and pooled error). The replications in site or replica-
tions in site-year should be used to test sites or site-years and their interac-
tions. Added nutrient and added nutrient x site and/or years should be
tested by pooled error.
In fitting the empirical models, one would place controlled fertilizer
variables into the model and then add additional site-year variables and
interaction terms based on their performance using empirical selection
procedures. One difficulty in choosing a model and in interpreting the im-
portance of terms contained therein is posed by the correlation (termed
multicollinearity) of the site-year variables which is often pronounced.
There are several fairly standard procedures for building a regression
model empirically and these are described in detail in Draper and Smith
(1981). Some of the more common of these are the Forward Selection Pro-
cedure, which builds a model from a few to a larger number of variables;
the Stepwise Regression Procedure, which also builds a model starting
with few variables but which reevaluates all previously entered variables as
each new variable is entered; the Backward Elimination Procedure in
which variables showing little contribution to the prediction of yield are
eliminated one-by-one from a "full model," and the Maximum and Mini-
mum R2 Improvement Procedures in which a new variable is chosen on the
basis of producing either maximum or minimum gains in R 2 as it is entered
into the model. The danger in using any of these empirical techniques is
that because they are more mechanically than agronomically based, a good
fit will be obtained to the particular set of data being analyzed, but the
equation may not predict well for sites outside of the particular set. There
78 NELSON, VOSS, & PESEK

appears to be no easy answer to the question of how to choose a model be-

cause the problem of choosing the best function is not soluble using a sim-
ple set of rules. The lack of a common definition of the term best is one of
the problems itself.
Laird and Cady (1969) and Cady and Allen (1972) used a prediction
sum of squares approach to the choice of a model rather than a residual
sum of squares approach. The principle is that when making an individual
prediction, the regression equation should be derived from data that ex-
clude the observed point corresponding to that prediction. The prediction
sum of squares is then obtained by squaring the deviations of the observed
points from the predicted points, the latter of which have been obtained
from all of the observed data except the point being predicted. This should
lead to a model that is not confined in applicability to only the available set
of data. The prediction sum of squares technique was used by Wood and
Cady (1981) in the Benchmark Soils Project of the Universities of Hawaii
and Puerto Rico to develop a model for transferring fertilizer response in-
formation from one site to another belonging to the same soil family.
Regardless of the selection procedure used, one should not be too op-
timistic about arriving at a model that is far superior in fit to all other
models using empirical techniques. In addition, one can only hope that the
variables contained in the "best model" have some agronomic rationale
and produce estimates of optimal fertilizer rates that are reasonably free of
bias.
As with the individual location models, least squares procedures may
be used to estimate the regression coefficients of the general model devel-
oped for combined data from all sites and years and to place confidence
limits on them. One must remember to use the appropriate error term for
calculating the variance of a regression coefficient depending on whether it
is a site-year, added nutrient, or site-year x added nutrient variable. One
would prefer to keep the number of site-year and site-year x nutrient terms
to a minimum to assure that the model is "compact." Extraneous terms in
the regression whose regression coefficients are not larger than their stand-
ard errors may add "unwanted noise" to the predictions.
As in the case of the individual site-year analyses, the estimate of the
optimal fertilizer rate involves a ratio of two regression coefficients and, so
again, the variance of this estimate may be quite large.

V. ECONOMIC ANALYSIS OF FERTILIZER USE DATA

A. General Techniques

The economic techniques used to determine optimal rates of fertilizer

depend heavily on the model used to fit the data. For example, Anderson
and Nelson (1975) described rather simple procedures for economic analy-
sis in connection with the family of linear-plateau models. If the slope of
one of the sloping straight lines of the model exceeds the cost of fertilizer/
EV ALVA TION OF FERTILIZER RESPONSE 79

price of product ratio, one adds fertilizer to the point where that sloping
line intersects with another line such as the plateau. In two or three sloping
line models there will be two or three slopes to compare with the cost/price
ratio. Economics of nondivisible treatments (e.g., plowing, disking) fol-
lows a similar "yes-no" decision procedure.
Most of our discussion here, however, will be based on use of a poly-
nomial or polynomial transformation model that provides the basis for us-
ing the economic principle of diminishing returns and the added cost-
added returns concept. For a much more comprehensive treatment of
these concepts and the techniques used in economic analysis see Heady et
al. (1955) and Heady and Pesek (1957).
The diminishing returns concept implies that as higher levels of a fer-
tilizer nutrient are applied, the increments of output decrease. A point is
reached on the response curve where the cost of an increment of fertilizer
applied is equal to the value of its yield return. It is at this point where the
maximum profit per unit land area will be achieved. The optimal rate of
fertilizer is the quantity offertilizer that achieves this objective. This then
is the rational upper limit of fertilizer use with unlimited capital and with-
out alternative investments. The purpose of the economic analyses then is
to find the fertilizer rate where the added returns from an increment of fer-
tilizer just equals the added cost. This is achieved algebraically in the single
nutrient case by equating the change in yield per unit change in added nu-
trient to the cost of fertilizer/price of product ratio, i.e.,

8 Y/8X = Px/ Py [17]

where Y = yield of crop at fertilizer rate X, Px = cost of unit of fertilizer,

and P y = value of an increment of yield of the crop; 8 Y/8X is defined as
the marginal physical product and it is also the slope of the response func-
tion for any particular input.
It can be readily seen that

8 Y . P y = P x ' 8X [18]
obtained from the above relationship is equivalent to the added cost equals
added returns identity.
When two or more fertilizer variables are being considered, the princi-
ple of substitution must be taken into consideration when determining op-
timal rates offertilizer. This occurs with a response surface having iso-yield
contours (isoquants). Implied is the fact that there are a number of combi-
nations of input levels of the various fertilizer nutrients that will produce a
given yield of the crop being considered. Application of this principle in-
volves comparing costs of different combinations of inputs that could be
used in obtaining a specified quantity of production and finding the combi-
nation that produces that yield at minimum total per-unit area cost. Apply-
ing the procedure of Eq. [17] and solving simultaneously (if needed) gives a
unique solution for the optimum of each variable.
80 NELSON, VOSS, & PESEK

III
II:

~
III
::t:
II:
III
a..
z
II:
~
~
III
II:
II:
o
~--------~----------------+-D
IG I
~--------~---------------~
X3 X4
RATE OF FERTILIZER
Fig. 3-5. Relationship between the response curve, OEA, the total fertilizer cost line, BFC,
and the optimal rate of fertilizer, X 3. Line BHD is the fixed cost of applying fertilizer. (Pe-
sek & Heady, 1958).

1. Graphical Methods ofEconomic Interpretation of Fertilizer Response

Data
The relationships described in the previous paragraphs may be
graphed as in Fig. 3-5. The conditions in Eq. [17] are defined graphically by
drawing a line parallel to the fertilizer cost line, BFC, (expressed in the
same units as Y) and tangent to the yield response curve, OEA. The
amount of fertilizer, X 3, needed to produce this magnitude of yield re-
sponse at this point of tangency is the optimal rate. The vertical distance,
EF, between the fertilizer cost line and the yield response curve, is the
profit and this distance is greatest at this fertilizer rate.
Economic interpretations for an experiment representing a single site-
year and involving two or more nutrients may be obtained readily by plot-
ting the yield contours and then on the same map drawing isoclines equal to
particular price ratios. In Fig. 3-6 is shown such a map for a NP fertilizer
on corn experiment. The isoquants are calculated by setting the values of
yield and fixing the levels of all but one of the nutrients and then solving for
the level of the other nutrient. There are several computer packages that
have contour plot procedures in which the computer does these calcula-
tions and then plots the contours. The slope of a tangent to each isoquant at
a given point represents the rate at which one input factor substitutes for or
replaces the other in maintaining output at a fixed level. If the isoquants
are curved, the rate at which one input factor substitutes for the other di-
minishes as the output is produced with more of the former and less of the
latter. The lines indicated by Pi in Fig. 3-6 are isoclines or expansion paths.
They denote the path of optimum (least cost) nutrient combination~ as
higher yield levels are obtained. They also connect points of equal tangen-
tial slopes on successively higher isoquants. Hence, they also denote points
on the isoquants that have equal marginal rates of substitution (i.e., the
ratios of marginal physical products for the two nutrients are equal). Liter-
ally, this is a measure of the relative productivity of the two inputs at the
intersection of an isocline with the isoquant.
EV ALVA TlON OF FERTILIZER RESPONSE 81

QUANTITY OF P

Fig. 3-6. Map showing isoquants and isoclines for graphical economic interpretation.

Use of the graphical procedure for situations where data from a series
of experiments are being combined makes it necessary to first substitute
values for the site-year variables and their interactions with the nutrients
into the general prediction equation. The resulting reduced general equa-
tion may then be used to produce isoquant plots as mentioned above.

2. Algebraic Procedures for Economic Interpretation of Fertilizer Response

Data
For a single nutrient model being fitted to data from a single site-year,
the optimal rate will be obtained by taking the derivative with respect to
the nutrient and setting it equal to the cost of fertilizer price of product ra-
tio (Eq. [17]). If one has a single experiment with two nutrients, two partial
derivatives will need to be taken, one with respect to each of the nutrients.
These will be set equal to the respective cost/price ratios. The solution of
the two equations will need to be done simultaneously if there is an interac-
tion term in the model. Otherwise, the two equations will be independent.
As an example, supposing the following two nutrient model pertains to a
set of data from a single site-year for crop yield

[19]

Because the model contains an NP term, the solution of the optimal rates
will need to be done simultaneously, i.e.,

[20]

where Pn and Pp are prices of Nand P, respectively, and Pc is the price of

the crop. These equations reduce to the following:
82 NELSON, VOSS, & PESEK

[21 ]

The set of two simultaneous equations may be readily solved for opti-
mal rates of Nand P using a computer assuming the following matrix rela-
tionship

[22]

where A is the matrix of regression coefficients given above,

[~]
is the vector of unknown optimal rates of Nand P fertilizer and (; is the
vector of price ratios minus the respective linear regression coefficients.

[~]
may then be solved by the following relationship derived from Eq. [22]:

[ ~] = A -1(: [23]

Solution of

[~]
could be readily obtained by using Procedure Matrix described in the man-
ual for the Statistical Analysis System (SAS Inst., 1982) or procedures
from comparable statistical software packages.

B. Inductive Inferences From the Series of Experiments to Decisions Made

by an Individual Farmer

If the response function is a general one developed for a series of site-

years, it would contain terms for site-years and site-years x controlled fer-
tilizer variables. For one of the site-years studied, or for an individual farm
that was not in the set of experiments studied, one would substitute the
levels of the various site-year variables for that particular site into the gen-
eral equation, which will leave an equation containing only constants and
terms for individual controlled variables. This would then be solved for the
optimum combination of added fertilizer nutrients, assuming a given set of
cost/price ratios and using the procedures given above.

C. Discounting (Cost of Money)

A decision-maker might not use the optimum rate of fertilizer, or even

some high rational rate, because he might discount future returns that are
EV ALVA TlON OF FERTILIZER RESPONSE 83

subject to uncertainty to a greater degree than the return on a "sure" in-

vestment. This discounting characteristic varies among individuals but en-
ters into decision making, although it may not be recognized.
Usually discounting is expressed as a percentage just as in the case of
interest. Suppose that the total discounting rate (including return on alter-
native investments) for an individual is 20%. Under these conditions, the
rate of fertilizer that this individual should use is given by
3 Y/8X = 1.2 Pxl P y. [24)

Pesek et a1. (1961) presented examples of how discounting affects fertilizer

use in the presence and absence of residual fertilizer effects and given one
or more cropping seasons. If compounding returns is considered, and the
period of investment in crop growth exceeds the length of one compound-
ing period, then the standard compounding formula must be used to deter-
mine the coefficient of the inverse price ratio in Eq. [24). The formula is

C = e(l + r)t = Px [25]

where C is the compounded value over time; e is the original cost; r is the
interest rate; and t is, time in years for each crop return period.

D. Equating Marginal Returns and Limited Capital

Frequently, the resources for buying fertilizer are limited to the extent
that not all fields can be fertilized to the level designated in Eq. [17) or even
to the level provided for in Eq. [24). The problem then becomes one of how
to allocate the investment in fertilizer among various crops and fields to
maximize net return. This goal is achieved when the marginal returns from
fertilizer use to all crops or all fields are the same. If the fertilizer response
equations for three fields or three crops are given as f( Y 1)' f( Y 2)' and
f( Y 3) and the costs of fertilizer and crops are appropriately designated, the
maximum revenue will occur when the following holds:

and the total of XI, X 2, and X 3 multiplied by the respective prices does not
exceed the limited resources. These equal marginal returns must be
greater than the returns in alternative investments.
If resources for fertilizer use are extremely limited, the maximum eco-
nomic return per unit of fertilizer applied plus the application cost is a rele-
vant consideration. The rate that produces this maximum economic return
is designated the minimum rate.
In Fig. 3-5 the return to fertilizer plus application cost is the ratio of
the difference between the response curve OEA and the total fertilizer cost
(line BFC) to the total fertilizer cost (line BFC). The relative return rises
rapidly to a maximum (frequently at the first or second increment of fertil-
izer) and then declines to zero at the point where the variable fertilizer cost
line crosses the response curve.
84 NELSON, VOSS, & PESEK

Application of this concept to a fertilizer response curve for a crop that

may be produced profitably without any fertilizer follows. The fertilizer re-
sponse curve is written

[26]

and the variable fertilizer cost line is:

C = m + rX, [27]

where m is fixed cost of fertilizer per unit area and r is the cost of the fertil-
izer in units of response.
The quantity to maximize is the difference between these two equa-
tions divided by the latter. This maximum is reached at the level of X given
in the following equation:

X' = [28]
rb l1

In many cases, some fertilizer is required to produce a crop profitably.

The total yield curve and the total cost of production line including the var-
iable fertilizer cost must be considered and the quantity to be maximized is
the difference between the yield curve

[29]

and the total cost line

n = p + rX, [30]

where b o is the yield without fertilizer and p is the sum of the total cost of
production and of fertilizer per unit area.
The minimum rate of fertilizer which should be applied is given by the
solution for X' in the following equation:

X' = [31 ]
rb ll

This is the rate that maximizes return to investment per unit area in crop
production.
The interpretation ofthe value for X' in Eq. [28] is that all similar units
of area should receive this minimum rate of fertilizer if resources permit. If
there is not enough fertilizer to cover all units at this level, fertilizer should
be applied at this level as far as it will reach, and the other units should be
left unfertilized. In the case of Eq. [31], X' is interpreted as the rate to ap-
EV ALVA nON OF FERTILIZER RESPONSE 85

ply to all similar units, but if there is not enough to treat all units at this
level, these excess units should be left out of production. Any other action
would lead to anticipated reduction in revenue. Applications of these con-
cepts were made by Pesek and Heady (1958) and more recently by Voss
(1975).

E. Minimizing Losses

In some cases, projected yields, costs, and returns may strongly indi-
cate a loss regardless of any action. If, in spite of these projections, a crop is
to be produced, the fertilizer rate that will minimize losses is the economic
optimum rate, which does not consider costs other than the fertilizer. The
correctness of this rate may be visualized from Fig. 3-5. If the fixed cost
line BHD were elevated above the response curve OEA, the shortest dis-
tance (minimum loss) between BHD and OEA would be a tangent to OEA
parallel to BFC, which is the economic optimum.

F. The Opportunity Cost Principle

If there are some enterprises that compete for the farmer's capital,
one should compare the marginal returns from the various investments be-
fore carrying the rates of fertilizer to the point of most profit per hectare.
Perhaps putting part of the money intended for fertilizer purchase into a
poultry (chicken, Gallus gallus damesticus) operation would give a higher
incremental return/incremental cost ratio. There are some difficult ques-
tions that need to be asked. For example, if there is a cattle (beef, Bas
taurus) feeding enterprise, would it pay to apply less N fertilizer to the corn
produced on the farm and then buy corn from off of the farm to offset the
loss in potential corn production due to limited fertilizer? These decisions
require the determination of marginal returns for each course of action and
equating them. This is possible only if all production functions are known.
Thus, one must always consider the cost of lost opportunities in not placing
capital in some alternative enterprise when making decisions about opti-
mal fertilizer strategies. If it were not for risk and uncertainty, the choice of
an "optimum mix" of input allocations to various enterprises on the farm
would be a rather straightforward procedure that could be handled by lin-
ear programming methods. Unfortunately, there are many factors that
have an influence on the relative profitability of various enterprises on the
farm, and it is impossible to predict how important they will be during a
particular growing season. Some of these will be discussed in the following
section.

G. Dealing With Risk and Uncertainty

The rational economic range within which the choices of fertilizer

rates must be made has been described. This range is conditioned by the
expected (or average) response, the expected price structure, availability
86 NELSON, VOSS, & PESEK

of resources for fertilizer purchase, and alternative uses for capital that
might be expended for fertilizer. We considered the solutions as if the yield
functions were known and the expected outcome was accurately predicted,
or that the average of all functions over time is known and the producer will
operate long enough to realize this outcome. Hence, matters were viewed
in a risk setting.
Actual functions are not the same each year, depending on timeliness
of operations and natural hazards such as weather, disease, pests, etc.,
none of which are known at the beginning of production. Expressed in an-
other way, the variability of yields that will be produced from what is con-
sidered to be an optimum rate of fertilizer will be great from farm to farm
or from year to year on the same farm. This implies that the calculations
used in linear programming for choosing an optimal enterprise mix will not
necessarily be reliable for a given farm in a particular season. Even though
the results may be viewed from the point of view of risk over a period of
years, each year's outcome is uncertain. Therefore, we must consider deci-
sion making under uncertainty. Decision making under uncertainty is
quite different from decision making under risk. The behavior of an indi-
vidual regarding uncertainty is probably affected by age, tenure, family sit-
uation, equity, and psychological traits. Concepts for dealing with uncer-
tainty follow in the next section.

H. Game Models in Decision Making

Fertilizer recommendation strategies need to be tailored to various

probabilities of natural disasters occurring. The crop producer can choose
recommendations accordingly to his aversion to risk.
Walker et al. (1960, 1964) applied game theory models to fertilizer
and other crop-producing decisions. There are at least four models that
were explored in these studies; those of Laplace, Wald, Savage, and
Hurwicz.
The Laplace model is a special case of a simple risk model. In it, it is
assumed that each previously observed outcome has an equal chance of oc-
curring the next season, or that the average condition will occur. The alter-
native chosen is the one that is expected to give the highest return. Because
the "best" is always expected, the Laplace solution is an optimistic one and
provides no protection against the occurrence of low returns because of the
alternative chosen.
In contrast, the Wald model lends to conservative action. It is set up as
a game against nature with states of nature and alternatives making up the
choice matrix. The decision maker playing against nature assumes that nat-
ure "tries to do its worst," and therefore, always selects an alternative that
maximizes returns under the worst that nature presents. The fallacy is that
nature, unlike a living opponent, does not take conscious action.
A Savage minimum regret model is less conservative than the Waldo
Its operation starts with the Wald "states-alternatives" choice matrix, and
a new regret matrix is generated by subtracting the highest outcome alter-
EVALUATION OF FERTILIZER RESPONSE 87

native within each "state" from each other alternative. This new matrix is
one of ex post facto opportunity losses. The strategy is to minimize the op-
portunity loss for a given season or crop.
The Hurwicz criterion depends on the assignment of an optimism-
pessimism index a to an operator. This index lies between 0, optimistic,
and I, pessimistic; under the latter, the strategy in a given game is the same
as the Wald solution gives. If m is the minimum and M is the maximum
outcome under one alternative over all "states," and I-a is the individ-
ual's belief that M will occur, then am + (I-a) M is the a index for that
alternative. The alternative with the highest a is the preferred one.
Even more favorable outcomes may be determined by using more
than one alternative, e.g., part of the crop may be fertilized at one level
and another part at a higher or lower level. Simulation models can be used
to produce weather data assuming a specified underlying model. Simulated
weather data for a number of years are available almost instantly. These
data permit the testing of various fertilization strategies used in connection
with a given weather outcome and evaluation of the resulting outcome. Re-
search results are lacking at the present time to develop fertilization strate-
gies for individual farm situations.

REFERENCES
Abraham, T. P., and Rao, V. Y. 1965. An investigation on functional models for fertilizer
response surfaces. Indian Soc. Agric. Stat. J. 18:45-61.
Anderson, R. L., and L. A. Nelson. 1971. Some problems in the estimation of single nutrient
functions. p. 203-222. In Bulletin of the International Statistical Institute, 44, Part 1.
International Statistical Institute, The Hague, Netherlands.
Anderson, R. L., and L. A. Nelson. 1975. A family of models involving intersecting straight
lines and concomitant experimental designs useful in evaluating response to fertilizer
nutrients. Biometrics 31:303-318.
Baird, B. L., and J. W. Fitts. 1957. An agronomic procedure involving the use of a central
composite design for determining fertilizer response surfaces. p. 135-143. In E. L.
Baum et al. (ed.) Fertilizer innovations and resource use. Iowa State College Press,
Ames,IA.
Balmukand, B. H. 1928. Studies in crop variation. V. The relation between yield and soil
nutrients. J. Agric. Sci. 18:602-627.
Bartlett, M. S. 1937. Properties of sufficiency and statistical tests. Proc. R. Soc. (London)
AI60:268-282.
Baule, B. 1918. Zu Mitscherlichs gesetz der physiologischen beziehungen. Landwirtsch.
Jahrb.51:363-385.
Baum, E. L., E. O. Heady, and J. Blackmore (ed.). 1956. Methodological procedures in the
use of fertilizer use data. Iowa State University Press, Ames, IA.
Baum, E. L., E. O. Heady, J. T. Pesek, and C. G. Hildreth (ed.). 1957. Economic and tech-
nical analysis of fertilizer innovations and resource use. Iowa State University Press,
Ames,IA.
Box, G. E. P. 1954. Exploration and exploitation ofresponse surfaces. Biometrics 10: 16-60.
Box, G. E. P., and J. S. Hunter. 1957. Multifactor experimental designs for exploring re-
sponse surfaces. Ann. Math. Stat. 28:195-241.
Boyd, D. A. 1972. Some recent ideas on fertilizer response curves. Potassium Symp. 9:461-
473.
Boyd, D. A., L. T. K. Yuen, and P. Needham. 1976. Nitrogen requirement of cereals. I. Re-
sponse curves. J. Agric. Sci. 87:149-162.
Cady, F. B., and D. M. Allen. 1972. Combining experiments to predict future yield data.
Agron. J. 64:211-214.
88 NELSON, VOSS, & PESEK

Cady, F. B., and R. J. Laird. 1973. Treatment design for fertilizer use experimentation.
CIMMYT Res. Bull. 26. International Maize and Wheat Improvement Center, Mexico
City, Mexico.
Cochran, W. G., and G. M. Cox. 1957. Experimental designs, 2nd ed. John Wiley & Sons,
Inc., New York.
Colwell, J. D. 1967. Calibration and assessment of soil test for estimating fertilizer require-
ments. I. Statistical models and tests ofsignificance. Aust. J. Soil Res. 5:275-293.
Colwell, J. D. 1968. Calibration and assessment of soil test for estimating fertilizer require-
ments. II. Fertilizer requirements and an evaluation of soil testing. Aust. J. Soil Res.
6:93-103.
Colwell, J. D. 1970. A comparison of soil test calibrations for the estimation of phosphorus
fertilizer requirements. Aust. J. Exp. Agric. Anim. Husb. 10:774-782.
Colwell, J. D. 1978. Computations for studies of soil fertility and fertilizer requirement.
Commonwealth Agricultural Bureaux, Farnham House, Farnham Royal, Slough, U. K.
Colwell, J. D. 1979. National soil fertility project, Vol. 2. Commonwealth Scientific and In-
dustrial Research Organization, E. Melbourne, Victoria, Australia.
Denmead, O. T., and R. H. Shaw. 1962. Availability of soil water to plants as affected by soil
moisture content and meteorological conditions. Agron. J. 54:385-390.
Draper, N. R., andH. Smith. 1981. Applied regression analysis, 2nd ed. John Wiley & Sons,
Inc., New York.
Escobar, G. J. A. 1967. Consideraciones sobre la comparici6n de disen6s de tratamientos,
M. S. Thesis. Escuela Nacional de Agricultura, S. A. G. Chapingo, Mexico.
Federer, W. T. 1955. Experimentai design. Macmillan Publishing Co., New York.
Fisher, R. A. 1924. The influence of rainfall on yields of wheat at Rothamsted. Philos. Trans.
R. Soc. London Ser. B:213:89-142.
Gomez, K. A. 1972. Techniques for field experiments with rice. The International Rice Re-
search Institute, Los Banos, Philippines.
Hader, R. J., M. E. Harward, D. D. Mason, and D. P. Moore. 1957. An investigation of
some relationships between copper, iron and molybdenum in the growth and nutrition
of lettuce: I. Experimental design and statistical methods for characterizing the re-
sponse surface. Soil Sci. Soc. Am. Proc. 21:59-74.
Hartley, H. O. 1950. The maximum F-ratio as a short-cut test for heterogeneity of variance.
Biometrika 37:308-312.
Hatheway, W. H., and E. J. Williams. 1958. Efficient estimation of the relationship between
plot size and the variability of crop yields. Biometrics 14:207-222.
Heady, E. 0., and J. L. Dillon. 1961. Agricultural production functions. Iowa State Univer-
sity Press, Ames, IA.
Heady, E. 0., andJ. T. Pesek. 1957. Some methodological considerations in the Iowa-TVA
Research Project of Fertilizer Use. p. 144-167. In E. L. Baum et al. (ed.) Economic and
technical analysis of fertilizer innovations and resource use. Iowa State University
Press,Ames,IA.
Heady, E. 0., J. T. Pesek, and W. G. Brown. 1955. Crop response surfaces and economic
optima in fertilizer use. Iowa Agric. Exp. Stn. Res. Bull. 424.
Heady, E. 0., J. T. Pesek, W. G. Brown, and J. P. Doll. 1961. Crop response surfaces and
economic optima in fertilizer use. p. 475-525. In E. O. Heady and J. L. Dillon (eds.)
Agricultural production functions. Iowa State University Press, Ames, IA.
Hendricks, W. A., and J. C. Scholl. 1943. Joint effects of temperature and rainfall on corn
yields. North Carolina Agric. Exp. Stn. Tech. Bull. 74.
Hildreth, C. G. 1957. Possible models for agro-economic research. p. 176-186. In E. L.
Baum et al. (eds.) Economic and technical analysis of fertilizer innovations and resource
use. Iowa State University Press, Ames, IA.
J6nsson, L. 1974. On the choice of a production function model for nitrogen fertilizer on
small grains in Sweden. Swed. J. Agric. Res. 4:87-97.
Laird, R. J., and F. B. Cady. 1969. Combined analysis of yield data from fertilizer experi-
ments. Agron. J. 61:829-834.
Mason, D. D., and D. Cooper. 1958. Characterization and utilization of climatic data in a
multiple covariance model for studying the relationships between yield response to fer-
tilizers and soil nutrients. In Determining yield response surfaces and economically opti-
mum fertilizer rates for corn under various soil and climatic conditions in North Caro-
lina. TVA Rep. T58-2AE. Tennessee Valley Authority, Muscle Shoals, AL.
EVALUATION OF FERTILIZER RESPONSE 89

Mead, R., and D. J. Pike. 1975. A review of response surface methodology from a biometric
viewpoint. Biometrics 31:803-851.
Mitscherlich, E. A. 1909. Das gesetz des minimums and das gesetz des abnehmenden bo-
denertrages. Landwirtsch Jahrb. 38:537-552.
Mitscherlich, E. A. 1930. Die bestimr.1Ung des dungerbedurfnisses des bodens, 3rd ed. West
Berlin.
Mombiela, F. A., and L. A. Nelson. 1980. Evaluation of fertilizer recommendation metho-
dologies using simulated corn P responses in North Carolina Norfolk-like soils with
varying soil P fertility. North Carolina Agric. Res. Service Tech. Bull. 268.
Mombiela, F., J. J. Nicholaides III, and L. A. Nelson. 1981. Method to determine the appro-
priate mathematical form for incorporating soil test levels in fertilizer response models
for recommendation purposes. Agron. J. 73:937-941.
National Academy of Sciences-National Research Council, Committee on Economics of
Fertilizer Use of the Agricultural Board. 1961. Status and methods of research in eco-
nomic and agronomic aspects of fertilizer response and use. NAS-NRC Pub. 918. Na-
tional Academy of Sciences-National Research CounciL Washington, DC.
Neider, J. A. 1966. Inverse polynomials, a useful group of multifactor response functions.
Biometrics 22: 128-141.
Nelson, L. A., and R. 1. McCracken. 1962. Properties of Norfolk and Portsmouth soils: Sta-
tistical summarization and influence on corn yields. Soil Sci. Soc. of Am. Proc. 26:497-
502.
Niklas, H., and M. Miller. 1927. Beitrage zur mathematischen Formulierung des Ertragge-
setzes. Z. Pflanzenernaehr. Dueng. Bodenkd. 8A:289-297.
Pesek, 1. T., and E. O. Heady. 1958. Derivation and application of a method for determin-
ing minimum recommended rates of fertilization. Soil Sci. Soc. Am. Proc. 22:419-423.
Pesek, 1. T., E. O. Heady, and L. C. Dumenil. 1961. Influence of residual fertilizer effects
and discounting on optimum fertilizer rates. Trans. lnt. Congr. Soil Sci. 7th 3:220-227.
Runge, E. C. A. 1968. Effects of rainfall and temperature interactions during the growing
season on corn yield. Agron. J. 60:503-507.
Runge, E. C. A., and R. T. Odell. 1958. The relation between precipitation, temperature
and the yield of corn on the Agronomy South Farm, Urbana, Illinois. Agron. J. 50:448-
454.
Ryan, 1. G., and R. K. Perrin. 1973. The estimation and use of a generalized response func-
tion for potatoes in the Sierra of Peru. North Carolina Agric. Exp. Stn. Tech. Bull. 214.
Schultz, E. F., Jr. 1955. Rules of thumb for determining expectations of mean squares in
analysis of variance. Biometrics 11: 123-135.
Smith, H. F. 1938. An empirical law describing heterogeneity in the yields of agricultural
crops. J. Agric. Sci. 28:1-23.
Sopher, C. D., and R. J. McCracken. 1973. Relationships between individual soil proper-
ties, management practices and corn yields on selected South Atlantic Coastal Plain
soils. Agron. 1. 65:595-599.
Sopher, C. c., R. J. McCracken, and D. D. Mason. 1973. Relationships between drought
and corn yields on selected South Atlantic Coastal Plain Soils. Agron. J. 65:351-354.
SAS Institute. Inc. 1982. SAS user's guide: Statistics. 1982 ed. SAS Institute, Inc., Cary, NC.
Steel. R. G. D., and J. H. Torrie. 1980. Principles and procedures of statistics. 2nd cd.
McGraw-Hill. Inc .. New York.
van Bavel, C. H. M. 1953. A drought criterion and its application in evaluating drought inci-
dence and hazard. Agron. J. 45:167-172.
von Boguslawski, E., and B. Schneider. 1962. Die dritte annaherung des ertragsgesetzes. 2.
Mitteilung. Z. Acker Pflanzenbau 116: 113-128.
Voss, R., and J. Pesek. 1967. Yield of corn grain as affected by fertilizer rates and environ-
mental factors. Agron. J. 59:567-572.
Voss, R. D. 1975. Fertilizer N: The key to profitable corn production with changing prices
and production costs. p. 215-229. In Proc. 30th Annu. Corn and Sorghum Res. Conf.,
Chicago. IL. December. 1975. American Seed Trade Association. Washington, DC.
Voss, R. E., J. J. Hanway, and W. A. Fuller. 1970. Influence of soil. management, and cli-
matic factors on the yield response by corn (Zea mays L.) to N, P, and K fertilizer.
Agron. J. 62:736-740.
Walker, O. L., E. O. Heady. and J. T. Pesek. 1964. Application of game theoretic models to
agricultural decision making. Agron. J. 56: 170-173.
90 NELSON, VOSS, & PESEK

Walker, O. L., E. O. Heady, L. G. Tweeten, and J. T. Pesek. 1960. Application of game

theory models to decisions on farm practices and resource use. Iowa Agric. Exp. Stn.
Res. Bull. 488.
Wood, C. L., and F. B. Cady. 1981. Intersite transfer of estimated response surfaces. Bio-
metrics 37: I-tO.