Chapter 2

Key aspects of decision modelling

for economic evaluation

This chapter considers the basic elements of decision modelling for economic
All rights reserved. May not be reproduced in any form without permission from the publisher, except fair uses permitted under U.S. or applicable copyright law.

evaluation. It considers the key stages in developing a decision analytic model

and describes the cohort model, the main type of decision model used in the
field. The decision tree and Markov model are described in detail and exam-
ples provided of their use in economic evaluation.

2.1. The stages of developing a decision model

It is possible to identify a series of stages in developing a decision model for
economic evaluation. In part, this will involve some general choices concern-
ing the nature of the evaluation. This will include the measure of effect and
the time horizon, but also the perspective of the analysis; that is, whose costs
and effects we are interested in? Below is a list of the stages in the development
process which relate specifically to the decision modelling.

2.1.1. Specifying the decision problem

This involves clearly identifying the question to be addressed in the analysis.
This requires a definition of the recipient population and subpopulations. This
will typically be the relevant patients, but may include nonpatients (e.g. in the
case of screening and prevention programmes). This requires specific details
about the characteristics of individuals, but should also include information
about the locations (e.g. the UK NHS) and setting (e.g. secondary care) in
which the options are being delivered. The specific options being evaluated
also need to be detailed. These will usually be programmes or interventions,
but could include sequences of treatments with particular starting and
stopping rules.
Part of the definition of the decision problem relates to which institution(s)
is/are (assumed to be) making the relevant decision. In some cases this will be
explicitly stated – for example, in the case of a submission to a reimbursement
agency – but it will often have to be implied by the characteristics of the evalu-
ation, such as the sources of data used.
2006. OUP Oxford.

EBSCO Publishing: eBook Collection (EBSCOhost) printed on 10/30/2025 [Link] PM UTC via TULANE UNIVERSITY 1689059; Andrew Briggs, Mark
Sculpher, Karl Claxton; Decision Modelling for Health Economic Evaluation Account:ehost.
 16 KEY ASPECTS OF DECISION MODELLING FOR ECONOMIC EVALUATION

2.1.2. Defining the boundaries of the model

All models are simplifications of reality and it will never be possible for
a model to include all the possible ramifications of the particular option being
considered. Choices need to be taken, therefore, about which of the possible
consequences of the options under evaluation will be formally modelled.
For example, should the possible implications of antibiotic resistance be
assessed in all economic evaluations of interventions for infectious diseases?
Another example relates to whether or not to include changes in disease
transmission resulting from screening programmes for HIV. It has been shown
that including reductions in the horizontal transmission of HIV in such
models has a marked impact on the cost-effectiveness of screening (Sanders
et al. 2005).

2.1.3. Structuring a decision model

Given a stated decision problem and set of model boundaries, choices have to
be made about how to structure the possible consequences of the options
being evaluated. In part, this will be based on the nature of the interventions
themselves. For example, for an economic evaluation of alternative diagnostic
strategies for urinary tract infection in children, it was necessary to use a
complex decision tree to reflect the prior probability (prevalence) and diagnos-
tic accuracy (sensitivity and specificity) of the various single and sequential
screening tests (Downs 1999). In part, model structure will reflect what is
known about the natural history of a particular condition and the impact of
the options on that process; for example, the future risks faced by patients
surviving a myocardial infarction and the impact of options for secondary
prevention on those risks.
As a general approach to structuring a decision model, there is value in
having some sort of underlying biological or clinical process driving the
model. Examples of the former include the use of CD4 counts or viral load in
HIV models (Sanders et al. 2005). The latter approach is more common and
examples include the use of the Kurtzke Expanded Disability Status Scale in
multiple sclerosis (Chilcott et al. 2003), the Mini Mental State Examination in
Alzheimer’s disease (Neumann et al. 1999) and clinical events, such as myocar-
dial infarction and revascularization in coronary heart disease (Palmer et al.
2005). The cost-effectiveness of the relevant interventions can then be assessed
by attaching health-related quality-of-life weights and costs to states or path-
ways defined in this way. The advantage of using these biologically- or clini-
cally-defined states is that they should be well-understood. In particular, there
should be good evidence about the natural history of a disease in terms of

EBSCOhost: eBook Collection (EBSCOhost) printed on 10/30/2025 [Link] PM UTC via TULANE UNIVERSITY. All use subject to
[Link]
 THE STAGES OF DEVELOPING A DECISION MODEL 17

these definitions. This is particularly important when modelling a baseline

(e.g. disease progression without treatment) and in extrapolating beyond the
data from randomized trials.
There are no general rules about appropriate model structure in a given
situation. However, some of the features of a disease/technology that are likely
to influence choices about structure include:
◆ Whether the disease is acute or chronic and, if the latter, the number of
possible health-related events which could occur over time.
◆ Whether the risks of events change over time or can reasonably be assumed
to be constant.
◆ Whether the effectiveness of the intervention(s) (relative to some usual
care baseline) can be assumed constant over time or time-limited in
some way.
◆ If and when treatment is stopped, the appropriate assumptions about
the future profile of those changes in health that were achieved during
treatment. For example, would there be some sort of ‘rebound’ effect or
would the gains, relative to a comparator group, be maintained over time
(Drummond et al. 2005).
◆ Whether the probability of health-related events over time is dependent on
what has happened to ‘a patient’ in the past.

2.1.4. Identifying and synthesizing evidence

The process of populating a model involves bringing together all relevant
evidence, given a selected structure, and synthesizing it appropriately in terms
of input parameters in the model. Consistent with the general principles of
evidence-based medicine (Sackett et al. 1996), there needs to be a systematic
approach to identifying relevant evidence. Evidence synthesis is a key area of
clinical evaluation in its own right (Sutton et al. 2000) which is of importance
outside the requirements of economic evaluation. However, the requirements
of decision analytic models for economic evaluation have placed some
important demands on the methods of evidence synthesis. These include:
◆ The need to estimate the effectiveness of interventions despite the absence
of head-to-head randomized trials. This involves the use of indirect and
mixed treatment comparisons to create a network of evidence between
trials.
◆ The need to obtain probabilities of clinical events for models over a
standardized period of follow-up despite the fact that clinical reports
present these over varying follow-up times.

EBSCOhost: eBook Collection (EBSCOhost) printed on 10/30/2025 [Link] PM UTC via TULANE UNIVERSITY. All use subject to
[Link]
 18 KEY ASPECTS OF DECISION MODELLING FOR ECONOMIC EVALUATION

◆ The need for estimates of treatment effectiveness in terms of a common

endpoint although trials report various measures.
◆ The need to assess heterogeneity in measures between different
types of patients. Ideally this would be undertaken using individual
patient data, but metaregression can be used with summary data in some
situations.
These issues in evidence synthesis are being tackled by statisticians, often
within a Bayesian framework (Sutton and Abrams 2001; Ades 2003;
Spiegelhalter et al. 2004), and these are increasingly being used in decision
models for economic evaluation (Ades et al. 2006). An important area of
methodological research in the field relates to incorporating evidence synthesis
and decision modelling into the same analytic framework – ‘comprehensive
decision modelling’ (Parmigiani 2002; Cooper et al. 2004). This has the advan-
tage of facilitating a fuller expression of the uncertainty in the evidence base in
the economic evaluation.

2.1.5. Dealing with uncertainty and heterogeneity

Uncertainty and heterogeneity exist in all economic evaluations. This
is an area of economic evaluation methodology that has developed rapidly
in recent years (Briggs 2001), and its implications for decision modelling
represent an important element of this book. Chapters 4 and 5 provide more
detail about appropriate methods to handle uncertainty and heterogeneity.
Box 2.1 summarizes the key concepts, and these are further developed in
Chapter 4.

2.1.6. Assessing the value of additional research

The purpose of evaluative research, such as randomized control trials, is to
reduce uncertainty in decision making by measuring one or more parameters
(which may be specific to particular subgroups) with greater precision. This is
generally true in clinical research, but also in assessing cost-effectiveness.
Given limited resources, it is just as appropriate to use decision analytic
models to assess the value for money of additional research projects as to
assess alternative approaches to patient management. In quantifying the
decision uncertainty associated with a particular comparison, decision models
can provide a framework within which it is possible to begin an assessment of
the cost-effectiveness of additional research. This can be undertaken infor-
mally using simple sensitivity analysis by assessing the extent to which a
model’s conclusions are sensitive to the uncertainty in one (or a small number)

EBSCOhost: eBook Collection (EBSCOhost) printed on 10/30/2025 [Link] PM UTC via TULANE UNIVERSITY. All use subject to
[Link]
 THE STAGES OF DEVELOPING A DECISION MODEL 19

Box 2.1. Key concept in understanding uncertainty

and heterogeneity in decision models for
cost-effectiveness analysis
Variability: Individual patients will inevitably differ from one another
in terms, for example, of the clinical events that they experience and the
associated health-related quality of life. This variability cannot be reduced
through the collection of additional data.
Parameter uncertainty: The precision with which an input parameter
is estimated (e.g. the probability of an event, a mean cost or a mean utility).
The imprecision is a result of the fact that input parameters are estimated
for populations on the basis of limited available information. Hence uncer-
tainty can, in principle, be reduced through the acquisition of additional
evidence.
Decision uncertainty: The joint implications of parameter uncertainty in
a model result in a distribution of possible cost-effectiveness relating to the
options under comparison. There is a strong normative argument
for basing decisions, given available evidence, on the expectation of this
distribution. But the distribution can be used to indicate the probability
that the correct decision has been taken.
Heterogeneity: Heterogeneity relates to the extent to which it is possible
to explain a proportion of the interpatient variability in a particular meas-
urement on the basis of one or more patient characteristics. For example, a
particular clinical event may be more likely in men and in individuals aged
over 60 years. It will then be possible to estimate input parameters
(and cost-effectiveness and decision uncertainty) conditional on a
patient’s characteristics (subgroup estimates), although uncertainty in
those parameters will remain.

of parameters. Formal value-of-information methods are considered fully in

Chapters 6 and 7. These methods have the strength of reflecting the joint
uncertainty in all parameters. They also assess the extent to which reduction
in uncertainty through additional research would result in a change in deci-
sion about the use of a technology and, if there is a change, its value in terms
of improved health and/or reduced costs.
Each of these stages is crucial to the development of a decision model that is
fit for the purpose of informing real policy decisions.

EBSCOhost: eBook Collection (EBSCOhost) printed on 10/30/2025 [Link] PM UTC via TULANE UNIVERSITY. All use subject to
[Link]
 20 KEY ASPECTS OF DECISION MODELLING FOR ECONOMIC EVALUATION

2.2. Some introductory concepts in decision analysis

Decision analysis is based on some key ‘building blocks’ which are common to
all models. These are covered more fully in introductory texts (Weinstein and
Fineberg 1980; Hunink et al. 2001; Drummond et al. 2005), and are only
summarized here.

2.2.1. Probabilities
In decision analysis, a probability is taken as a number indicating the likeli-
hood of an event taking place in the future. As such, decision analysis shares
the same perspective as Bayesian statistics (O’Hagan and Luce 2003). This
concept of probability can be generalized to represent a strength of belief
which, for a given individual, is based on their previous knowledge and expe-
rience. This more ‘subjective’ conceptualization of probabilities is consistent
with the philosophy of decision analysis, which recognizes that decisions
cannot be avoided just because data are unavailable to inform them, and
‘expert judgement’ will frequently be necessary.
Specific probability concepts frequently used in decision analysis are:
◆ Joint probability. The probability of two events occurring concomitantly.
In terms of notation, the joint probability of events A and B occurring is
P(A and B).
◆ Conditional probability. The probability of an event A given that an event B
is known to have occurred. The notation is P(A|B).
◆ Independence. Events A and B are independent if the probability of event A,
P(A), is the same as the probability of P(A|B). When the events are inde-
pendent P(A and B) = P(A) × P(B).
◆ Joint and conditional probabilities are related in the following equation:
P(A and B) = P(A|B) × P(B). Sometimes information is available on the
joint probability, and the above expression can be manipulated to ‘condi-
tion out’ the probabilities.

2.2.2. Expected values

Central to the decision analytic approach to identifying a ‘preferred’ option
from those being compared under conditions of uncertainty is the concept of
expected value. If the options under comparison relate to alternative treatments
for a given patient (or an apparently homogeneous group of patients), then
the structure of the decision model will reflect the variability between patients
in the events that may occur with each of the treatments. The probabilities
will show the likelihood of those events for a given patient. On this basis, the

EBSCOhost: eBook Collection (EBSCOhost) printed on 10/30/2025 [Link] PM UTC via TULANE UNIVERSITY. All use subject to
[Link]
 SOME INTRODUCTORY CONCEPTS IN DECISION ANALYSIS 21

model will indicate a number of mutually exclusive ‘prognoses’ for a given

patient and option (more generally, these are alternative ‘states of the world’
that could possibly occur with a given option). Depending on the type
of model, these prognoses may be characterized, for example, as alternative
pathways or sequences of states. For a given option, the likelihood of each
possible prognosis can be quantified in terms of a probability, and their impli-
cations in terms of cost and/or some measure of outcome. The calculation of
an expected value is shown in Box 2.2 using the example of costs. It is derived
by adding together the cost of each of the possible prognoses weighted by the
probability of it occurring. This is analogous to a sample mean calculated on
the basis of patient-level data.

2.2.3. Payoffs
As described in the previous section, each possible ‘prognosis’ or ‘state of the
world’ can be given some sort of cost or outcome. These can be termed
‘payoffs’, and expected values of these measures are calculated. The origins of

Box 2.2. An illustration of the concept of expected

values using costs

Prognosis 1 Cost = 25
Probability = 0.4

Prognosis 2 Cost = 50
Probability = 0.2

Prognosis 3 Cost = 100

Probability = 0.1

Prognosis 4 Cost = 75
Probability = 0.3

Expected cost = (25 × 0.4) + (50 × 0.2) + (100 × 0.1) + (75 × 0.3) = 52.50

EBSCOhost: eBook Collection (EBSCOhost) printed on 10/30/2025 [Link] PM UTC via TULANE UNIVERSITY. All use subject to
[Link]
 22 KEY ASPECTS OF DECISION MODELLING FOR ECONOMIC EVALUATION

decision analysis are closely tied to those of expected utility theory (Raiffa
1968), so the standard payoff would have been a ‘utility’ as defined by von
Neumann-Morgenstern (von Neumann and Morgenstern 1944). In practice,
this would equate to a utility based on the standard gamble method of preference
elicitation (Torrance 1986). As used for economic evaluation in health care,
the payoffs in decision models have been more broadly defined. Costs would
typically be one form of payoff but, on the effects side, a range of outcomes
may be defined depending on the type of study (see Chapter 1). Increasingly,
quality-adjusted life-years would be one of the payoffs in a decision model for
cost-effectiveness analysis, which may or may not be based on utilities elicited
using the standard gamble.
The principle of identifying a preferred option on the basis of a decision
analytic model is on the basis of expected values. When payoffs are defined in
terms of ‘von Neumann-Morgenstern utilities’, this would equate with a
preferred option having the highest expected utility; this is consistent with
expected utility theory as a normative framework for decision making under
uncertainty. Although a wider set of payoffs are used in decision models for
economic evaluation, the focus on expected values as a basis for decision
making remains. This follows the normative theory presented by Arrow and
Lind (1970) arguing that public resource allocation decisions should exhibit
risk neutrality. For example, in cost-effectiveness analysis, the common incre-
mental cost-effectiveness ratio would be based on the differences between
options in terms of their expected costs and expected effects. However, the
uncertainty around expected values is also important for establishing the
value and design of future research, and this should also be quantified as part
of a decision analytic model. The methods for quantifying and presenting
uncertainty in models are described in Chapters 4 and 5, respectively; and the
uses of information on uncertainty for research prioritization are considered
in Chapters 6 and 7.

2.3. Cohort models

The overall purpose of a model structure is to characterize the consequences
of alternative options in a way that is appropriate for the stated decision prob-
lem and the boundaries of the model. The structure should also be consistent
with the key features of the economic evaluation, such as the perspective, time
horizon and measure of outcome. There are several mathematical approaches
to decision modelling from which the analyst can choose. One important
consideration is whether the model should seek to characterize the experience
of the ‘average’ patient from a population sharing the same characteristics, or

EBSCOhost: eBook Collection (EBSCOhost) printed on 10/30/2025 [Link] PM UTC via TULANE UNIVERSITY. All use subject to
[Link]
 COHORT MODELS 23

should explicitly consider the individual patient and allow for variability
between patients. As described previously, the focus of economic evaluation is
on expected costs and effects, and uncertainty in those expected values. This
has resulted in most decision models focusing on the average patient experi-
ence – these are referred to as cohort models. In certain circumstances, a more
appropriate way of estimating expected values may be to move away from the
cohort model, to models focused on characterizing variability between
patients. These ‘micro simulation’ models are discussed in Chapter 3, but the
focus of the remainder of this chapter is on cohort models.
The two most common forms of cohort model used in decision analysis for
economic evaluation are the decision tree and the Markov model. These are
considered below.

2.3.1. The decision tree

The decision tree is probably the simplest form of decision model. Box 2.3
provides a brief revision of the key concepts using a simple example from the
management of migraine (Evans et al. 1997); the decision tree has been
described in more detail elsewhere (Hunink et al. 2001; Drummond et al.
2005). The key features of a decision tree approach are:
◆ A square decision node – typically at the start of a tree – indicates a decision
point between alternative options.
◆ A circular chance node shows a point where two or more alternative events
for a patient are possible; these are shown as branches coming out of the
node. For an individual patient, which event they experience is uncertain.
◆ Pathways are mutually exclusive sequences of events and are the routes
through the tree.
◆ Probabilities show the likelihood of a particular event occurring at a
chance node (or the proportion of a cohort of apparently homogeneous
patients expected to experience the event). Moving left to right, the first
probabilities in the tree show the probability of an event. Subsequent prob-
abilities are conditional; that is, the probability of an event given that an
earlier event has or has not occurred. Multiplying probabilities along path-
ways estimates the pathway probability which is a joint probability (as
discussed previously).
Expected costs and outcomes (utilities in Box 2.3) are based on the princi-
ples in Box 2.2. Expected values are based on the summation of the pathway
values weighted by the pathway probabilities.
A somewhat more complicated decision tree model comes from a
cost-effectiveness analysis of alternative pharmaceutical therapies for

EBSCOhost: eBook Collection (EBSCOhost) printed on 10/30/2025 [Link] PM UTC via TULANE UNIVERSITY. All use subject to
[Link]
 24 KEY ASPECTS OF DECISION MODELLING FOR ECONOMIC EVALUATION

Box 2.3. Example of a decision tree based on Evans

et al. (1997)

No recurrence (0.594) Pathway

A
Relief (0.558)
Recurrence (0.406)
B
Sumatriptan
Endures attack (0.92)
C
No relief (0.442) Relief (0.998)
D
ER (0.08)
Migraine Hospitalisation (0.002)
attack E

No recurrence (0.703)
F
Relief (0.379)
Caffeine/ Recurrence (0.297)
ergotamine G
Endures attack (0.92)
H
No relief (0.621) Relief (0.998)
I
ER (0.08)
Hospitalisation (0.002)
J
Pathway
Sumatriptan Probability Cost Expected cost Utility Expected utility
A 0.331 16.10 5.34 1.00 0.33
B 0.227 32.20 7.29 0.90 0.20
C 0.407 16.10 6.55 −0.30 −0.12
D 0.035 79.26 2.77 0.10 0.0035
E 0.0001 1172.00 0.11 −0.30 −0.00003
Total 1.0000 22.06 0.41
Caffeine/erogotamine
F 0.266 1.32 0.35 1.00 0.27
G 0.113 2.64 0.30 0.90 0.10
H 0.571 1.32 0.75 −0.30 −0.17
I 0.050 64.45 3.22 0.10 0.0050
J 0.0001 1157.00 0.11 −0.30 −0.00003
Total 1.0000 4.73 0.20

gastro-oesophageal reflux disease (GORD) (Goeree et al. 1999). This model is

described in some detail here, both to ensure an understanding of the decision
tree, and to set up the case study used in Chapter 5 where the same model is
used to demonstrate the appropriate analysis of a probabilistic model.
As shown in Fig. 2.1, six treatment options are considered in the form
of strategies, as they define sequences of treatments rather than individual
therapies:
◆ A: Intermittent proton-pump inhibitor (PPI). Patients would be given PPI and,
if this heals the GORD, they would taken off therapy. If they experience

EBSCOhost: eBook Collection (EBSCOhost) printed on 10/30/2025 [Link] PM UTC via TULANE UNIVERSITY. All use subject to
[Link]
 Healed No Prescription Therapy
Strategy A: PPI
Not Healed
Step up therapy (Table 1)

Healed Maintenance PPI

Strategy B: PPI

Not Healed
Step up therapy (Table 1)
0-6 MONTHS 6-12 MONTHS

Healed Maintenance H2RA Recurrence Recurrence

Strategy C: H2RA (step up therapy – see (step up therapy – see
Table 1) Table 1)
Not Healed
Step up therapy (Table 1)
Endoscopically
Proven Erosive
Esophagitis Healed Maintenance Low Dose P.A.
No Recurrence No Recurrence
Strategy D: P.A

Not Healed
Step up therapy (Table 1)

Healed Maintenance H2RA

Strategy E: PPI
Not Healed
Step up therapy (Table 1)

Healed Maintenance Low Dose PPI

COHORT MODELS
Strategy F: PPI
Not Healed
Step up therapy (Table 1)

Fig. 2.1 Decision tree used to evaluate six treatment strategies for gastro-
oesophageal reflux disease (adapted from Goeree et al. (1999)). PPP, proton pump
inhibitor; H2RA: H2 receptor antagonists; PA: prokinetic agent. Table 1 refers to the
table from the original article.

25
EBSCOhost: eBook Collection (EBSCOhost) printed on 10/30/2025 [Link] PM UTC via TULANE UNIVERSITY. All use subject to [Link]
 26 KEY ASPECTS OF DECISION MODELLING FOR ECONOMIC EVALUATION

a recurrence of the condition, they would be returned to the PPI regimen.

If patients fail to heal with their initial dose of PPI, the dose of the
therapy would be doubled (DD PPI) and, once healed, patients would be
maintained on standard dose PPI. If the GORD recurs, they would be given
DD PPI.
◆ B: Maintenance PPI. Patients would initially be treated with PPI and, once
healed, they would be maintained on PPI. If they fail to heal with their
initial dose, or if the GORD recurs subsequent to healing, they would be
given DD PPI.
◆ C: Maintenance H 2 receptor antagonists (H 2RA). Patients are initially
treated with H2RA and, if they heal, they are maintained on that drug. If
their GORD subsequently recurs, they are placed on double-dose H2RA
(DD H2RA). If patients fail to heal on the initial dose, they are given PPI
and, if they then heal, are maintained with H2RA. If the GORD subse-
quently recurs, they are given PPI to heal. If patients fail to heal initially
with PPI, they are moved to DD PPI and, if they then heal, are maintained
with PPI. If the GORD subsequently recurs on maintenance PPI, they are
given DD PPI to heal.
◆ D: Step-down maintenance prokinetic agent (PA). Patients would be given
PA for initial healing and, if this is successful, maintained on low dose (LD)
PA; if their GORD subsequently recurs, they would be put on PA to heal
again. If patients fail their initial healing dose of PA, they would be moved
to PPI to heal and, if successful, maintained on LD PA. If their GORD
subsequently recurs, they would be treated with PPI for healing. If patients
fail their initial healing dose of PPI, they would be moved to DD PPI and, if
successful, maintained on PPI. If the GORD subsequently recurs, they
would receive DD PPI to heal.
◆ E: Step-down maintenance H2RA. Patients would initially receive PPI to
heal and, if this is successful, they would be maintained on H2RA. If they
subsequently recur, they would be given PPI to heal. Patients who initially
fail on PPI would be given DD PPI and, if this heals the GORD, they would
be maintained on PPI. If their GORD subsequently recurs, healing would
be attempted with DD PPI.
◆ F: Step-down maintenance PPI. Patients would initially be treated with
PPI and, if this heals the GORD, would move to LD PPI. In the case of a
subsequent recurrence, patients would be given PPI to heal. Patients who
fail on their initial dose of PPI would be given DD PPI and, if successful,
maintained on PPI. If their GORD recurs, they would be given DD PPI
to heal.

EBSCOhost: eBook Collection (EBSCOhost) printed on 10/30/2025 [Link] PM UTC via TULANE UNIVERSITY. All use subject to
[Link]
 COHORT MODELS 27

The structure of the decision tree used in the study is shown in Fig. 2.1. For
each strategy, the initial pathway shows whether their GORD initially heals
and, if so, it indicates the maintenance therapy a patient will move to. If they
do not heal, they move to step up therapy as defined for each of the five strate-
gies. The figure shows that, for each pathway on the tree, there is a probability
of GORD recurrence during two periods: 0–6 months and 6–12 months.
Should this happen, step-up therapy is used as defined above. It should be
noted that the tree contains decision nodes to the right of chance nodes.
However, this indicates a treatment decision defined by the strategies rather
than a point in the tree where alternative courses of action are being compared.
To populate the model, the authors undertook a meta-analysis of random-
ized trials to estimate, for each drug, the proportion of patients healed at
different time points. They also used available trial data to calculate the
proportions of patients who recur with GORD over the two time periods.
Routine evidence sources and clinical opinion were used to estimate the cost
of therapies and of recurrence.
The decision tree was evaluated over a time horizon of 12 months. Costs
were considered from the perspective of the health system and outcomes were
expressed in terms of the expected number of weeks (out of 52) during which
a patient was free of GORD. Table 2.1 shows the base-case results of the analy-
sis. For each strategy over 1 year, it shows the expected costs and time with
(and without) GORD symptoms. The table shows the options that are domi-
nated or subject to extended dominance (Johannesson and Weinstein 1993),
and the incremental cost per week of GORD symptoms avoided are shown for
the remaining options. Figure 2.2 shows the base-case cost-effectiveness
results on the cost-effectiveness plane (Black 1990; Johannesson and Weinstein
1993). It shows that Option D is dominated as it is more costly and less effec-
tive than Options C, A and E. It also shows that Option F is subject to extended
dominance. That is, it lies to the left of the efficiency frontier defined by non-
dominated options. This means that it would be possible to give a proportion
of patients Option E and a proportion Option B and the combined costs and
effects of this mixed option would dominate Option F (see the discussion of
cost-effectiveness decision rules in Chapter 1).

2.3.2. Markov models

The advantages of Markov models
Although various aspects of the GORD case study, such as the choice of
outcome measure, can be criticized, the study provides a good example of a
cost-effectiveness model based around a decision tree structure. Some of the

EBSCOhost: eBook Collection (EBSCOhost) printed on 10/30/2025 [Link] PM UTC via TULANE UNIVERSITY. All use subject to
[Link]
 28
KEY ASPECTS OF DECISION MODELLING FOR ECONOMIC EVALUATION
Table 2.1 Base-case results from the gastro-oesophageal reflux disease case study (Goeree et al. 1999)

Strategy Expected Expected weeks Incremental Incremental effects DC/DE ($/GORD

1-year cost with (without) costs (DC) (DE) GORD weeks week averted)
per patient GORD per patient $ averted
$ in 1 year
C: Maintenance H2RA 657 10.41 (41.59) – – –

A: Intermittent PPI 678 7.78 (44.22) 21 2.63 8

E: Step-down maintenance H2RA 748 6.17 (45.83) 70* 1.61* 44*
B: Maintenance PPI 1093 4.82 (47.18) 345† 1.35† 256†
D: Step-down maintenance PA 805 12.60 (39.40) Dominated by A,C,E
F: Step-down maintenance PPI 955 5.54 (46.46) Dominated by extended
dominance

* Relative to strategy A
†Relative to strategy E

EBSCOhost: eBook Collection (EBSCOhost) printed on 10/30/2025 [Link] PM UTC via TULANE UNIVERSITY. All use subject to [Link]
 COHORT MODELS 29

+500

Incremental Costs (Additional Dollars)

B-Main PPI
(5.59,436)

F-Step Down Main PPI

(4.87,298)

D-Step Down [Link]

(−2.19,148)
[Link] E-Step Down Main. H2RA
(2.63,21) (4.24,91)
C-Main.H2RA
(0,0)
−6 Incremental Benefits (Gord Weeks Averted) +6

−500
Fig. 2.2 Base-case results of the gastro-oesophageal reflux disease case study shown
on the cost-effectiveness plane (taken from Goeree et al. (1999)). The line joining
Strategies C, A, E and B is the ‘efficiency frontier’.

potential limitations of the decision tree are also evident in the case study. The
first is that the elapse of time is not explicit in decision trees. GORD relapse
between 0 and 6 months and between 6 and 12 months has to be separately
built into the model as no element of the structure explicitly relates to failure
rate over time.
A second limitation of decision trees made evident in the GORD example is
the speed with which the tree format can become unwieldy. In the GORD case
study, only three consequences of interventions are directly modelled: initial
healing, relapse between 0 and 6 months and relapse between 6 and
12 months. GORD is a chronic condition and it can be argued that for this
analysis, a lifetime time horizon may have been more appropriate than one of
12 months. If a longer time horizon had been adopted, several further features
of the model structure would have been necessary. The first is the need to
reflect the continuing risk of GORD recurrence (and hence the need for step-up
therapy) over time. The second is the requirement to allow for the competing
risk of death as the cohort ages. The third is the consideration of other clinical
developments, such as the possible occurrence of oesophageal cancer in
patients experiencing recurrent GORD over a period of years. This pattern of
recurring-remitting disease over a period of many years and of competing

EBSCOhost: eBook Collection (EBSCOhost) printed on 10/30/2025 [Link] PM UTC via TULANE UNIVERSITY. All use subject to
[Link]
 30 KEY ASPECTS OF DECISION MODELLING FOR ECONOMIC EVALUATION

clinical risks is characteristic of many chronic diseases such as diabetes,

ischaemic heart disease and some forms of cancer. In such situations, the need
to reflect a large number of possible consequences over time would result in
the decision tree becoming very ‘bushy’ and, therefore, difficult to program
and to present. As such, a Markov framework was used to further develop the
GORD model described above (Goeree et al. 2002).
The Markov model is a commonly used approach in decision analysis to
handle the added complexity of modelling options with a multiplicity of
possible consequences. Such models have been used in the evaluation of
screening programmes (Sanders et al. 2005), diagnostic technologies (Kuntz
et al. 1999) and therapeutic interventions (Sculpher et al. 1996). The added
flexibility of the Markov model relates to the fact that it is structured around
mutually exclusive disease states, representing the possible consequences of
the options under evaluation. Instead of possible consequences over time
being modelled as a large number of possible pathways as in a decision tree,
a more complex prognosis is reflected as a set of possible transitions between
the disease states over a series of discrete time periods (cycles). Costs and
effects are typically incorporated into these models as a mean value per state
per cycle, and expected values are calculated by adding the costs and outcomes
across the states and weighting according to the time the patient is expected to
be in each state.

A case study in HIV

The details of the Markov model can be illustrated using a case study. This is a
cost-effectiveness analysis of zidovudine monotherapy compared with zidovu-
dine plus lamivudine (combination) therapy in patients with HIV infection
(Chancellor et al. 1997). This example has been used for didactic purposes
before (Drummond et al. 2005), but is further developed here, and in Chapter 4
for purposes of probabilistic analysis.
The structure of the Markov model is shown in Fig. 2.3. This model charac-
terizes a patient’s prognosis in terms of four states. Two of these are based on
CD4 count: 200–500 cells/mm3 (the least severe disease state – State A) and
less than 200 cells/mm3 (State B). The third state is AIDS (State C) and the
final state is death (State D). The arrows on the Markov diagram indicate the
transitions patients can make in the model. The key structural assumption in
this early HIV model (now clinically doubtful, at least in developed countries)
is that patients can only remain in the same state or progress; it is not feasible
for them to move back to a less severe state. More recent models have allowed
patients to move back from an AIDS state to non-AIDS states and, through
therapy, to experience an increase in CD4 count. These models have also

EBSCOhost: eBook Collection (EBSCOhost) printed on 10/30/2025 [Link] PM UTC via TULANE UNIVERSITY. All use subject to
[Link]
 COHORT MODELS 31

State A
200 < cd4 < 500
cells/mm3

State B
cd4 < 200
cells/mm3
Death

State C
AIDS

Fig. 2.3 The structure of the Markov model used in the case study (Chancellor
et al. 1997).

allowed for the fact that prognosis for these patients is now understood in
terms of viral load as well as CD4 count (Sanders et al. 2005).
The transition probabilities governing the direction and speed of transitions
between disease states in the model are shown in Table 2.2 where a cycle is
taken as 1 year. For monotherapy, these ‘baseline’ (i.e. control group) proba-
bilities are taken from a longitudinal cohort study where data were collected
prior to any use of combination therapy. The zeros indicate that backwards
transitions are assumed not to be feasible. The transition probabilities for
combination therapy were based on an adjustment to the baseline values
according to the treatment effect of combination therapy relative to monother-
apy. This treatment effect took the form of a relative risk (0.509) which was
derived from a meta-analysis of trials. Although the treatment effect in the
trials was something rather different, it was assumed that the relative risk
worked to reduce the transition probabilities from one state to any worse state.
The calculation of these revised (combination) transition probabilities is
shown in Table 2.2. Any probability relating to the movement to a worse state
is multiplied by 0.509, and the probability of remaining in a state is corre-
spondingly increased. The separation of baseline probabilities from a relative

EBSCOhost: eBook Collection (EBSCOhost) printed on 10/30/2025 [Link] PM UTC via TULANE UNIVERSITY. All use subject to
[Link]
 32 KEY ASPECTS OF DECISION MODELLING FOR ECONOMIC EVALUATION

Table 2.2 Transition probabilities and costs for the HIV Markov model used in the
case study (Chancellor et al. 1997)

State at start of cycle State at end of cycle

1. Annual transition probabilities
(a) Monotherapy
State A State B State C State D
State A 0.721 0.202 0.067 0.010
State B 0.000 0.581 0.407 0.012
State C 0.000 0.000 0.750 0.250
State D 0.000 0.000 0.000 0.000

(b) Combination therapy

State A State B State C State D
State A 0.858 0.103 0.034 0.005
(1 – sum) (0.202 × RR) (0.067 × RR) (0.010 × RR)
State B 0.000 0.787 0.207 0.006
(1 – sum) (0.407 × RR) (0.012 × RR)
State C 0.000 0.000 0.873 0.127
(1 – sum) (0.25 × RR)
State D 0.000 0.000 0.000 1.000

2. Annual costs
Direct medical £1701 £1774 £6948 –
Community £1055 £1278 £2059 –
Total £2756 £3052 £9007 –

RR, relative risk of disease progression. Estimated as 0.509 in a meta-analysis.

The drug costs were £2278 (zidovudine) and £2086 (lamivudine).

treatment effect is a common feature of many decision models used for cost-
effectiveness. One advantage of this approach relates to the important task of
estimating cost-effectiveness for a particular location and population
subgroup. Often it is assumed that baseline event probabilities should be as
specific as possible to the location(s) and subgroup(s) of interest, but that the
relative treatment effect is assumed fixed.
It can be seen that all the transition probabilities are fixed with respect to
time. That is, the baseline annual probability of progressing from, for example,
State A to State B is 0.202, and this is the case 1 year after start of therapy and
it is also the case, for those remaining in State A, after 10 years. When these
time invariant probabilities are used, this is sometimes referred to as a Markov
Chain.
Table 2.2 also shows the annual costs associated with the different states.
These are assumed identical for both treatment options being compared –
excluding the costs of the drugs being evaluated. The drug costs were

EBSCOhost: eBook Collection (EBSCOhost) printed on 10/30/2025 [Link] PM UTC via TULANE UNIVERSITY. All use subject to
[Link]
 COHORT MODELS 33

£2278 (zidovudine) and £2086 (lamivudine). Outcomes were assessed

in terms of changes in mean survival duration so no health-related quality-
of-life weights (utilities) were included.

Cohort simulation
As explained previously, the calculation of expected costs and outcomes for all
cohort models involves summing the costs and outcomes of all possible conse-
quences weighted by the probability of the consequence. In the case of Markov
models, this involves calculating how long patients would spend in a given
disease state. This can be achieved using matrix algebra, but is more commonly
undertaken using a process known as cohort simulation. This is illustrated, for
the first two cycles for monotherapy, in Fig. 2.4. Exactly the same process
would be used for combination therapy based on the adjusted transition
probabilities in Table 2.2. This example uses a cohort size of 1000, but
this number is arbitrary and, for a cohort model, the same answer will emerge
for any size of starting cohort. Cohort simulation simply involves multiplying
the proportion of the cohort ending in one state by the relevant transition
probability to derive the proportion starting in another state. In a spreadsheet,
this is achieved by setting up the formulae for one cycle and then copying
down for subsequent cycles.
Calculating expected costs for a cohort model simply involves, for each
cycle, adding the costs of each state weighted by the proportion in the state,
and then adding across cycles. This is shown in Table 2.3 for monotherapy
based on the costs shown in Table 2.2 and annual drug costs for zidovudine of
£2278. The process of discounting to a present value is very straightforward in
a cohort simulation and this is also shown in Table 2.3.
Table 2.4 shows the calculation of expected survival duration (life
expectancy) for the monotherapy group. At its simplest, this involves adding
the proportion of the living cohort for each cycle, and adding across the cycles.
Of course, the units in which survival duration is estimated will depend on the
cycle length. In the case study, all transitions are assumed to take place at the
start of the cycle, so the simple approach to calculating life expectancy
assumes those dying during a cycle do not survive for any proportion of the
cycle. Strictly, an unbiased estimate of life expectancy would assume
that deaths occur halfway through a cycle. It is possible, therefore, to employ
a half-cycle correction as shown in the final column of Table 2.3. The half-
cycle correction is shown here just for life expectancy calculation. In principle,
however, it also applies to the calculation of expected costs. In the context of a
cost-effectiveness analysis where the focus is on the incremental costs and
outcomes of alternative options, it is unlikely that the half-cycle correction

EBSCOhost: eBook Collection (EBSCOhost) printed on 10/30/2025 [Link] PM UTC via TULANE UNIVERSITY. All use subject to
[Link]
 34
KEY ASPECTS OF DECISION MODELLING FOR ECONOMIC EVALUATION
Cycle number State A State B State C State D

0 1000 0 0 0
0.
20 0.06
2= 7=6
0.721=721 20 7 0.010
2 =10

1 721 202 67 10
0.0
10=
7 0.
0. 25
0.721=520 06 =1

0.2

0.75=50
7= 7

02

1.000=10
48

=1
46 0.0

0.
0.581=117
12

40
=2

7=
83
2 520 263 181 36

Fig. 2.4 Illustration of the first two cycles of a cohort simulation for monotherapy for
the Markov model used in the case study.

EBSCOhost: eBook Collection (EBSCOhost) printed on 10/30/2025 [Link] PM UTC via TULANE UNIVERSITY. All use subject to [Link]
 COHORT MODELS 35

Table 2.3 Calculation of expected costs for monotherapy in the case study Markov
model based on the costs shown in Table 2.2, annual drug costs of £2278 and an
annual discount rate of 6%

Cycle Proportion of cohort in each state Costs (£)

(year) A B C D Undiscounted Discounted
0 1000
1 721 202 67 10 5463* 5153†
2 520 263 181 36 6060 5393
3 376 258 277 89 6394 5368
4 271 226 338 165 6381 5055
5 195 186 364 255 6077 4541
6 141 147 361 350 5574 3929
7 102 114 341 444 4963 3301
8 73 87 309 531 4316 2708†
9 53 65 272 610 3682 2179
10 38 49 234 679 3092 1727
11 28 36 198 739 2564 1350
12 20 26 165 789 2102 1045
13 14 19 136 830 1708 801
14 10 14 111 865 1377 609
15 7 10 90 893 1103 460
16 5 8 72 915 878 346
17 4 5 57 933 695 258
18 3 4 45 948 548 192
19 2 3 36 959 431 142
20 1 2 28 968 337 105

63 745 44 663

*{[721 × (2756 + 2278)] + [202 × (3052 + 2278)] + [67 × (9007 + 2278)] + [10 × 0]} / 1000
†5463 / [(1 + 0.06)1]

will make a lot of difference to results unless the cycle length is long as a
proportion of the model’s time horizon.
Of course, the cohort simulation and calculation of expected costs and
expected survival duration, shown in Fig. 2.4 and Tables 2.3 and 2.4 for the
monotherapy baseline, would also need to be undertaken for combination
therapy. The process would be exactly the same but would involve different
transition probabilities and costs.

EBSCOhost: eBook Collection (EBSCOhost) printed on 10/30/2025 [Link] PM UTC via TULANE UNIVERSITY. All use subject to
[Link]
 36 KEY ASPECTS OF DECISION MODELLING FOR ECONOMIC EVALUATION

Table 2.4 Calculation of life expectancy over 20 cycles for monotherapy for the case
study Markov model. Calculation with and without a half-cycle correction is shown

Cycle Proportion of cohort in each state Life years

(year) A B C D Die at start Die in middle
0 1000
1 721 202 67 10 0.990 0.995
2 520 263 181 36 0.964 0.977
3 376 258 277 89 0.911 0.937
4 271 226 338 165 0.835* 0.873
5 195 186 364 255 0.745 0.790
6 141 147 361 350 0.650 0.697
7 102 114 341 444 0.556 0.603†
8 73 87 309 531 0.469 0.513
9 53 65 272 610 0.390 0.429
10 38 49 234 679 0.321 0.355
11 28 36 198 739 0.261 0.291
12 20 26 165 789 0.211 0.236
13 14 19 136 830 0.170 0.190
14 10 14 111 865 0.135 0.152
15 7 10 90 893 0.107 0.121
16 5 8 72 915 0.085 0.096
17 4 5 57 933 0.067 0.076
18 3 4 45 948 0.052 0.059
19 2 3 36 959 0.041 0.047
20 1 2 28 968 0.032 0.036

7.996 8.475

*(271 + 226 + 338) / 1000

†{102 + 114 + 341 + [0.5 × (444 – 350)]}/1000

The Markov assumption

Although the Markov version of a cohort model provides greater flexibility
than a decision tree, it also has some important restrictions in the context of
structuring complex prognoses. The restriction relates to what is known as the
Markov assumption, or ‘memoryless’ feature of Markov models. This assump-
tion means that once a notional patient has moved from one state to another,

EBSCOhost: eBook Collection (EBSCOhost) printed on 10/30/2025 [Link] PM UTC via TULANE UNIVERSITY. All use subject to
[Link]
 SUMMARY 37

T = 3 yrs after
initial treatment
Local
recurrence
Remission Death

Regional Transition probabilities equal

T = 8 yrs after
recurrence regardless of timings or
initial treatment
directions of earlier transitions

Note: T = time since initial operation (i.e. since start of model)

Fig. 2.5 Example of a simple Markov model to illustrate the Markov assumption.

the Markov model will have ‘no memory’ regarding where the patient has
come from or the timing of that transition. This is illustrated in Fig. 2.5 using
a simple Markov model to characterize prognosis following treatment for
cancer. It shows that patients can have two forms of cancer recurrence: local
and regional. Once treated for their recurrence, patients then move into a
remission state but, once in that state, all patients are considered homoge-
neous regardless of where they have come from (what type of recurrence) or
the timing of the recurrence after initial treatment for the cancer. The implica-
tion of this is that it is not possible to have different probabilities of future
transitions (e.g. the probability of death in Fig. 2.5) according to the nature or
timing of the recurrence.
In general, the Markov assumption means that it may not be straightfor-
ward to build ‘history’ into this type of model. That is, to model a process
where future events depend on past events. This may be possible by adding
additional states to the model and incorporating time dependency into transi-
tion probabilities. Both of these extensions to the Markov model are consid-
ered in Chapter 3.

2.4. Summary
This chapter has described some of the key concepts for the use of decision
analysis for economic evaluation. The majority of cost-effectiveness studies
are based on cohort modelling. This provides a flexible framework which can
be programmed relatively easily (often in spreadsheets) and evaluated rapidly.
However, the cohort framework can be restricting. In the case of Markov
models, for example, the memoryless feature of these models can cause diffi-
culties in modelling a complicated prognosis.

EBSCOhost: eBook Collection (EBSCOhost) printed on 10/30/2025 [Link] PM UTC via TULANE UNIVERSITY. All use subject to
[Link]
 38 KEY ASPECTS OF DECISION MODELLING FOR ECONOMIC EVALUATION

2.5. Exercise: replicating a Markov model of HIV/AIDS

2.5.1. Overview
The aim of this exercise is to get you to replicate a previously published
Markov model by Chancellor and colleagues (1997). The model itself is some-
what out of date, in that it compares combination therapy (lamivudine and
AZT) with monotherapy (AZT alone). Nevertheless, the model is straightfor-
ward enough to serve as a simple example that it is possible to replicate in a
short period of time.
The basic structure of the model was given in Fig. 2.3, which shows that the
disease process is structured as chronic, such that patients can move to succes-
sively more serious disease states, but cannot recover.
The cycle length of the model is 1 year and is evaluated over 20 years (after
which more than 95 per cent of patients are expected to have died).
Use the data given below to populate the model and calculate the incremental
cost-effectiveness ratio for AZT monotherapy. You should be able to replicate
a figure for the ICER of £6276.
1. Transition probabilities. These were calculated from the counts of individuals
that were observed to move between the four health states each
year in a longitudinal data set from the Chelsea and Westminster Hospital
in London. These counts are given in Table 2.5 and you should be able
to verify that these counts give the transition probabilities presented in
Table 2.2 part 1(a).
2. State costs. These are given in the original article separately by state for
‘direct medical’ and ‘community care’ costs and these were reproduced in
Table 2.2 part 2.
3. Costs of drugs. The yearly cost of AZT monotheray is given as £2278 and
lamivudine is given as £2086.50.
4. Treatment effect. A meta-analysis of four trials is reported in the paper
giving a pooled relative risk of disease progression as 0.509. Note from
Table 2.2 part 1(b) how this treatment effect is applied to the monother-
apy transitions in the model.

Table 2.5 Counts of transition between the four model states per year

A B C D Total
A 1251 350 116 17 1734
B 0 731 512 15 1258
C 0 0 1312 437
D

EBSCOhost: eBook Collection (EBSCOhost) printed on 10/30/2025 [Link] PM UTC via TULANE UNIVERSITY. All use subject to
[Link]
 EXERCISE: REPLICATING A MARKOV MODEL OF HIV/AIDS 39

5. Discounting. The original analysis was based on discounting the costs at

6 per cent but not discounting the estimates of life years.

2.5.2. Step-by-step guide to constructing the model

Open the file ‘Exercise 2.5 – Template’ and start at the <Parameters> worksheet.
You will see that the cells to be completed are coloured yellow.
Using the figures from Table 2.5, complete columns C and D on the
<Parameters> worksheet for the transition probabilities (rows 9–17). Column
C should contain the number of events of interest and column D should
contain the complement, such that columns C plus D should equal the appro-
priate row total from Table 2.5 above. Now calculate the respective transition
probabilities in column B using the counts in columns C and D (the reason for
structuring the spreadsheet like this will become apparent in the exercise in
Chapter 4).
Enter the other information for the state costs, drug costs, treatment effect
and discounting given above in the appropriate place in the template.
Having entered the input parameters of the model, the task now is to
construct the Markov models for the treatment alternatives: combination and
monotherapy. Note that the parameters you have just entered are in named
cells, such that they can be referred to by the shortened name in column A of
the <Parameters> worksheet.
If you move to the Markov worksheet, you will see the structure of the
Markov model laid out for you for each treatment alternative. Start with the
monotherapy model.

1. Generating the Markov trace

The initial concern is with generating the Markov trace, that is, showing the
proportions of patients that are in any one state at any one time. This is to be
entered in columns C to F representing the four model states, and G provides
a check (as the sum across C to F must always equal the size of the original
cohort – which is set to 1 in cell C7 such that the Markov trace represents
proportions of patients).
The first step in building a Markov model is to define the transition matrix.
This is a matrix that shows the probability of transition from states
represented in the rows to states represented in the columns. To save you
time, a copy of the transition matrix is reproduced in Table 2.6 for the
monotherapy arm.
i. Start by making sure that you understand the transition matrix. In
particular, make sure you can replicate it from the information given
in the diagram of the model in Fig. 2.3.

EBSCOhost: eBook Collection (EBSCOhost) printed on 10/30/2025 [Link] PM UTC via TULANE UNIVERSITY. All use subject to
[Link]
 40 KEY ASPECTS OF DECISION MODELLING FOR ECONOMIC EVALUATION

Table 2.6 Transition matrix for the monotherapy arm

Transition matrix A B C D
A tpA2A tpA2B tpA2C tpA2D
B 0 tpB2B tpB2C tpB2D
C 0 0 tpC2C tpC2D
D 0 0 0 1

ii. Use the transition matrix to populate the Markov model. This will
involve representing the transitions between the different states repre-
sented in columns C to F.
Hints: Start with just one row at a time (e.g. the first – row 8). Once you
have this row correct you can simply copy it down to the other rows.
Although it is tempting to look across the rows of the transition matrix –
consider looking down the columns as this indicates, for a given state,
where the people entering that state come from. Indeed, the transition
matrix describes the proportion of patients in the other states in the previ-
ous cycle arriving in the given state for this cycle. For example, looking
down for ‘State B’ we find that tpA2B patients in ‘State A’ last period
arrive in State B this period and that tpB2B patients in ‘State B’ last
period stay in ‘State B’ this period. The formula for cell D8 should there-
fore read: =C7*tpA2B+D7*tpB2B
iii. When you get to the dead state, do not be tempted to make this the
remainder of the other cells in the row – the use of remainders in this
way will mean that any errors could go unnoticed. Instead, it is good
practice to complete all states as you think they should be then sum
across the states to make sure the total sums to one. Do this check in
column G and make sure it does!
iv. When you think you have the first row correct, copy this row down to
the 19 rows below. If your check in column G is still looking good then
most likely you have done it correctly.
By far the most tricky bit of building the Markov model in Excel is
now complete. Now that you have the Markov trace you can calculate
the cost and effects.

2. Estimating life years

In column H calculate the proportion of the cohort that is in one of the ‘alive’
states for each of the years in the model. In column I apply the standard
discount formula1 (although we have set this to zero in the base case). Finally,

EBSCOhost: eBook Collection (EBSCOhost) printed on 10/30/2025 [Link] PM UTC via TULANE UNIVERSITY. All use subject to
[Link]
 EXERCISE: REPLICATING A MARKOV MODEL OF HIV/AIDS 41

in row 29, sum the columns to give the expected life years both discounted and
undiscounted.

3. Estimating costs
In column K, calculate the costs for each time period by applying the state
costs, not forgetting that AZT is given for the whole time period. In column L,
apply the standard rate of discount for costs. Again, in Row 29, sum the
columns to give costs both undiscounted and discounted.

4. Adapting the model for combination therapy

You now need to repeat the steps above, but this time for combination therapy.
i. Start by copying the whole monotherapy model to the corresponding
combination therapy model. This will minimize repetition as tasks
now relate to adjusting the model for combination treatment. (Note:
remember to anchor the year (column A) in the discounting formula
before copying).
ii. Firstly, the treatment effect needs to be added in. In the original article,
the relative risk parameter was applied to all transitions. The corre-
sponding transition matrix for combination therapy is, therefore,
given in Table 2.7.
iii. Use this transition matrix to adjust the trace in rows 8 and 9 only (as
the base case assumption in the originally reported model was that
treatment effect was limited to 2 years).
Now add in the cost of lamivudine for these 2 years only in column V (as the
drug is assumed to be given for only 2 years)

1 Recall that the standard discount rate is given by 1/(1 + r)t where r is the discount rate and
t represents time (in years).

Table 2.7 Transition matrix for the combination therapy arm

Transition matrix A B C D
A 1-tpA2B*RR- tpA2B*RR tpA2C*RR tpA2D*RR
tpA2C*RR-
tpA2D*RR

B 0 1-tpB2C*RR- tpB2C*RR tpB2D*RR

tpB2D*RR

C 0 0 1-tpC2D*RR tpC2D*RR
D 0 0 0 1

EBSCOhost: eBook Collection (EBSCOhost) printed on 10/30/2025 [Link] PM UTC via TULANE UNIVERSITY. All use subject to
[Link]
 42 KEY ASPECTS OF DECISION MODELLING FOR ECONOMIC EVALUATION

5. Cost-effectiveness estimates
The final task is simply to copy the corresponding results from the sums at
the bottom of the Markov model sheet (row 29) to the <Analysis> worksheet
(row 5) and to calculate the appropriate increments and ICER.
Congratulations, you have now replicated the Markov model. Compare
your result for the ICER to that given in the solution (£6276). Is any debugging
required? If it is, then you may want to compare your Markov trace and stage
costs for monotherapy against those reported in Table 2.3.

References
Ades, A. E. (2003) A chain of evidence with mixed comparisons: models for multi-parameter
evidence synthesis and consistency of evidence, Statistics in Medicine, 22: 2995–3016.
Ades, A. E., Sculpher, M. J., Sutton, A., Abrams, K., Cooper, N., Welton, N., et al. (2006)
‘Bayesian methods for evidence synthesis in cost-effectiveness analysis’,
PharmacoEconomics 24: 1–19.
Arrow, K. J. and Lind, R. C. (1970) ‘Risk and uncertainty: uncertainty and the evaluation of
public investment decisions’, American Economic Review, 60: 364–78.
Black, W. C. (1990) ‘The CE plane: a graphic representation of cost-effectiveness’, Medical
Decision Making, 10: 212–214.
Briggs, A. (2001) ‘Handling uncertainty in economic evaluation and presenting the results’
in M. Drummond and A. J. McGuire (eds) Economic evaluation in health care: merging
theory with practice. Oxford, Oxford University Press; pp. 172–214.
Chancellor, J. V., Hill, A. M., Sabin, C. A., Simpson, K. N. and Youle, M. (1997) ‘Modelling
the cost effectiveness of lamivudine/zidovudine combination therapy in HIV infection’,
PharmacoEconomics, 12: 1–13.
Chilcott, J., McCabe, C., Tappenden, P., O’Hagan, A., Cooper, N. J., Abrams, K., et al.
(2003) ‘Modelling the cost effectiveness of interferon beta and glatiramer acete in the
management of multiple sclerosis’, British Medical Journal, 326: 522.
Cooper, N. J., Sutton, A. J., Abrams, K. R., Turner, D. and Wailoo, A. (2004) ‘Comprehensive
decision analytical modelling in economic evaluation: A Bayesian approach’, Health
Economics, 13: 203–226.
Downs, S. (1999) ‘Technical report: urinary tract infections in febrile infants and young
children’, Pediatrics, 103: 54.
Drummond, M. F., Sculpher, M. J., Torrance, G. W., O’Brien, B. and Stoddart, G. L. (2005).
Methods for the economic evaluation of health care programmes. Oxford, Oxford
University Press.
Evans, K. W., Boan, J. A., Evans, J. L. and Shuaib, A. (1997) ‘Economic evaluation of oral
sumatriptan compared with oral caffeine/ergotamine for migraine’,
PharmacoEconomics, 12: 565–577.
Goeree, R., O’Brien, B. J., Hunt, R., Blackhouse, G., Willan, A. and Watson, J. (1999)
‘Economic evaluation of long term management strategies for erosive oesphagitis’,
PharmacoEconomics, 16: 679–697.
Goeree, R., O’Brien, B. J., Blackhouse, G., Marshall, J., Briggs, A. H. and Lad, R. (2002)
‘Cost-effectiveness and cost-utility of long-term management strategies for heartburn’,
Value in Health, 5: 312–328.

EBSCOhost: eBook Collection (EBSCOhost) printed on 10/30/2025 [Link] PM UTC via TULANE UNIVERSITY. All use subject to
[Link]
 REFERENCES 43

Hunink, M., Glaziou, P., Siegel, J., Weeks, J., Pliskin, J., Elstein, A., et al. (2001) Decision
making in health and medicine. Integrating evidence and values. Cambridge, Cambridge
University Press.
Johannesson, M. and Weinstein, S. (1993) ‘On the decision rules of cost-effectiveness
analysis’, Journal of Health Economics, 12: 459–467.
Kuntz, K. M., Fleishmann, K. E., Hunink, M. G. M. and Douglas, P. S. (1999) ‘Cost-
effectiveness of diagnostic strategies for patients with chest pain’, Annals of Internal
Medicine, 130: 709–718.
Neumann, P. J., Hermann, R. C. and Kuntz, K. M. (1999) ‘Cost-effectiveness of donepezil
in the treatment of mild or moderate Alzheimer’s disease’, Neurology, 52: 1138–45.
O’Hagan, A. and Luce, B. (2003) A primer on Bayesian statistics in health economics and
outcomes research. Bethesda, MD, Medtap International.
Palmer, S., Sculpher, M., Philips, Z., Robinson, M., Ginnelly, L., Bakhai, A., et al. (2005)
‘Management of non-ST-elevation acute coronary syndromes: how cost-effective are
glycoprotein IIb/IIIa antagonists in the UK National Health Service?’, International
Journal of Cardiology, 100: 229–240.
Parmigiani, G. (2002) Modeling in medical decision making: a Bayesian approach.
Chichester, Wiley.
Raiffa, H. (1968) Decision analysis: introductory lectures on choices under uncertainty.
Reading, MA, Addison-Wesley.
Sackett, D. L., Rosenberg, W. M. C., Gray, J. A. M., Haynes, R. B. and Richardson, W. S.
(1996) ‘Evidence-based medicine: what it is and what it isn’t’, British Medical Journal,
312: 71–72.
Sanders, G. D., Bayoumi, A. M., Sundaram, V., Bilir, S. P., Neukermans, C. P. and et al.
(2005) ‘Cost-effectiveness of screening for HIV in the era of highly active antiretroviral
therapy’, New England Journal of Medicine, 352: 570–585.
Sculpher, M., Michaels, J., McKenna, M. and Minor, J. (1996) ‘A cost-utility analysis of
laser-assisted angioplasty for peripheral arterial occlusions’, International Journal of
Technology Assessment in Health Care 12: 104–125.
Spiegelhalter, D. J., Abrams, K. R. and Myles, J. P. (2004). Bayesian approaches to clinical
trials and health-care evaluation. Chichester, Wiley.
Sutton, A. J. and Abrams, K. R. (2001) ‘Bayesian methods in meta-analysis and evidence
synthesis’, Statistical Methods in Medical Research, 10: 277–303.
Sutton, A. J., Abrams, K. R., Jones, D. R., Sheldon, T. A. and Song, T. A. (2000) Methods for
meta-analysis in medical research. Chichester, Wiley.
Torrance, G. W. (1986) ‘Measurement of health state utilities for economic appraisal –
a review’, Journal of Health Economics, 5: 1–30.
von Neumann, J. and Morgenstern, O. (1944) Theory of games and economic behavior.
New York, Princeton University Press.
Weinstein, M. C. and Fineberg, H. V. (1980) Clinical decision analysis. Philadelphia, PA,
WB Saunders Company.

EBSCOhost: eBook Collection (EBSCOhost) printed on 10/30/2025 [Link] PM UTC via TULANE UNIVERSITY. All use subject to
[Link]
 This page intentionally left blank

EBSCOhost: eBook Collection (EBSCOhost) printed on 10/30/2025 [Link] PM UTC via TULANE UNIVERSITY. All use subject to
[Link]