REVIEW

Users’ Guide to Medical Decision Analysis

Claudia C. Dobler, MD, PhD; Gordon H. Guyatt, MD, MSc; Zhen Wang, PhD;
and M. Hassan Murad, MD, MPH

Abstract

Clinicians regularly have to trade benefits and harms to choose between testing and treatment stra-
tegies. This process is often done by making global and implicit judgments. A decision analysis is an
analytic method that makes this process more explicit, reproducible, and evidence-based. While cli-
nicians are unlikely to conduct their own decision analysis, they will read publications of such ana-
lyses or use guidelines based on them. This review outlines the anatomy of a decision tree and
provides clinicians with the tools to critically appraise a decision analysis and apply its results to
medical decision making. Clinicians reading about a decision analysis can make two judgments. The
first judgment is about the credibility of the methods, such as whether the decision analysis addressed
a relevant clinical question, included all important outcomes, used the current best evidence to derive
variables in the model, and adopted the appropriate time horizon. The second judgment is about
rating confidence in the preferred course of action by determining the certainty in the model variables,
whether the results are robust in sensitivity analyses and if the results are applicable to a specific
patient. Results from a valid and robust decision analysis can inform both guideline panels and the
patient-clinician dyad engaged in shared decision-making.
ª 2021 Mayo Foundation for Medical Education and Research n Mayo Clin Proc. 2021;96(8):2205-2217

CLINICAL SCENARIO outweigh the risks unless individuals/pa-

Y
ou are a family physician who sees a tients also belong to the groups mentioned
54-year old man, born in India, who in the above recommendations [including From the Institute for Evi-
dence-Based Healthcare,
has lived in the United States for the immigrants from high burden countries].” Faculty of Health Sciences
past 3 years and was recently diagnosed with As your patient is an immigrant from a & Medicine, Bond Univer-
sity, Gold Coast, Queens-
diabetes mellitus. You are aware that India high-TB-burden country and also has dia-
land, Australia (C.C.D.,
has a high incidence of tuberculosis and there- betes, you conclude that your patient might Z.W., M.H.M.); Evidence-
fore your patient could be infected with Myco- potentially benefit from latent TB testing Based Practice Center,
Robert D. and Patricia E.
bacterium tuberculosis (latent tuberculosis and treatment. However, because the evi- Kern Center for the Sci-
[TB] infection). Moreover, you are aware dence is low quality, you explore further ence of Health Care De-
that patients with diabetes have a two- to and find a decision analysis with cost- livery, Mayo Clinic,
Rochester, MN (C.C.D.);
four-fold increased risk of developing active effectiveness measures addressing latent TB and the Department of
tuberculosis.1 You wonder if you should test testing and treatment in residents born Medicine and Department
of Health Research
your patient for latent TB using an outside the United States with and without Methods, Evidence &
interferon-gamma release assay test (IGRA) medical comorbidities.3 You ascertain that Impact, McMaster Univer-
or a tuberculin skin test (TST) and prescribe there is reasonable match between the deci- sity, Hamilton, Ontario,
Canada (G.H.G.).
preventive treatment if the test is positive. sion problem examined in the analysis and
To inform your discussion, you search your clinical question. How should you use
first for an evidence-based recommendation this decision analysis/cost-effectiveness anal-
and find a guideline indicating that there is ysis to help guide you and your patient in
only low-quality evidence addressing latent deciding whether to pursue the screening
TB screening and treatment in immigrants option?
from high-TB-burden countries.2 The same
guideline concludes that the benefits of sys- INTRODUCTION
tematic and routine latent TB testing and Medical decisions are often complex, and cli-
treatment in patients with diabetes “do not nicians must regularly weigh the pros and

Mayo Clin Proc. n August 2021;96(8):2205-2217 n https://doi.org/10.1016/j.mayocp.2021.02.003 2205

www.mayoclinicproceedings.org n ª 2021 Mayo Foundation for Medical Education and Research
 MAYO CLINIC PROCEEDINGS

The purpose of this users’ guide is to

ARTICLE HIGHLIGHTS introduce the fundamental concepts of deci-
sion analysis and cost-effectiveness analysis
d This guide provides clinicians with the tools to critically appraise
and provide users with the tools to appraise
a decision analysis and apply its results to medical decision
studies using these analyses. In-depth de-
making. scriptions of decision modeling techniques
d We describe how clinicians can judge the credibility of the can be found elsewhere.5-11 In 1995,
methods of a decision analysis and rate the confidence in the Richardson et al12,13 developed a users’ guide
preferred course of action. (in two parts) on how to use a “clinical deci-
d The outlined principles can also be applied to appraise a cost- sion analysis” and presented a framework for
critical appraisal of clinical decision ana-
effectiveness analysis.
lyses.12,13 Since then, there has been a prolif-
d We use a clinical scenario and analyze a published decision eration of medical decision analysis studies,
analysis to demonstrate the practical application of this guide. in particular cost-effectiveness analyses.
Markov or state-transition models, not
explained in the original users’ guide, are
cons of interventions in individual patients. now included in most decision analyses.
Even treatments that have been found to be This current guide presents a contemporary
effective in large trials are associated with conceptualization and update on decision
adverse events, and it may not always be analysis. It is structured using the following
apparent whether the benefits (in this case section headings: “Are the Results Valid?”
reduced risk of developing TB with treat- “What Are the Results?” and “How Do I
ment of latent TB) outweigh the potential Apply the Results to My Patient?” Before
harms (here, receiving treatment based on explicating the criteria for using a decision
a false-positive test result, or having an analysis, we will comment on when such
adverse effect, such as drug-induced hepati- an analysis might be helpful and present its
tis, from latent TB treatment). anatomy.
Clinicians often assess such risk-benefit Branch and node decision trees and state-
decisions based on their intuition and often transition models, which are explained in
make decisions based on mental shortcuts, this guide, are “basic” decision model tech-
also called heuristics.4 This form of intuitive niques. There are numerous other modeling
decision-making, although often helpful, can techniques that are used to predict the
lead to inaccurate assessment of the likely events that occur after a clinical or policy de-
harms and benefits of an intervention. cision is made.
Decision analysis addresses medical
decision-making by providing a system-
atic approach to evaluating alternative WHEN MIGHT MEDICAL DECISION ANAL-
strategies using objective evidence and YSIS BE HELPFUL?
careful, model-based evaluation. High- In clinical practice it may not always be clear
quality medical decision analyses and whether the benefits of an intervention
cost-effectiveness analyses provide deci- outweigh the potential harms. Therefore,
sion makers d both clinicians and their there is often uncertainty regarding the
patients as well as those involved in best course of action, even when randomized
health policy d with guidance regarding controlled trials have shown an interven-
complex decision problems. Although cli- tion’s effectiveness. For example, prophylac-
nicians will seldom construct their own tic anticoagulation in a hospitalized patient
decision trees when faced with a clinical reduces the risk of venous thromboembo-
conundrum, the insights they gain from lism but increases the risk of a major
understanding medical decision analysis bleeding event. The best course of action in
will help them to critically reflect on their this situation will depend on patients’ risk
own decision making. factors for venous thromboembolism,
n n
2206 Mayo Clin Proc. August 2021;96(8):2205-2217 https://doi.org/10.1016/j.mayocp.2021.02.003
www.mayoclinicproceedings.org
MEDICAL DECISION ANALYSIS

bleeding risk, and possible contraindications

TABLE 1. The Process of Conducting a Decision Analysis
to anticoagulation.14,15
Medical decision analysis helps to deter- 1. Formulate the question for a specific group of patients using an actionable
intervention and a clear alternative course of action considering outcomes that are
mine the best course of action when there
relevant for the decision at hand (ie, participants, intervention, comparator,
is a trade-off between different management
outcome, aka the PICO process).
strategies.
2. Construct a decision tree with decision branches for potential courses of action (at
Although the risk-benefit trade-off least two) followed by chance nodesa representing the occurrence or non-
commonly refers to clinical benefits and occurrence of relevant outcomes
harms, it can also include the trade-off be- 3. Determine if a Markov modelb should be added to the decision tree, which is
tween greater effectiveness and higher cost. desirable if the decision problem involves a risk that continuous over time.
In this case, the appropriate decision model 4. If using a Markov model,b choose an appropriate model time horizon and Markov
is a cost-effectiveness analysis that compares cycle length.
the relative costs and values (often expressed 5. For a cost-effectiveness analysis, choose the desired analysis perspective (eg, from a
as per quality-adjusted life year [QALY]) of health care provider view, societal view, etc).
different courses of action). Cost- 6. Identify the probabilities associated with chance nodesa and the utilities or values
effectiveness analyses that address patient- associated with the outcomes.
important effectiveness outcomes can inform 7. Conduct a literature search that inform the probabilities and utilities.
clinical decisions as well as health policy 8. Determine the preferred course of action by “folding back”c the decision tree.
decisions. 9. Conduct a sensitivity analysis that informs the preferred course of action.
a
Chance nodes are typically represented by a circle; chance nodes are steps in the decision tree
THE ANATOMY OF A DECISION ANALYSIS that involve uncertainties. The possible outcomes of these probabilistic events are not under the
BASED ON A DECISION TREE control of the decision maker and are shown as paths leading away from the node toward the
right.
Table 1 summarizes the typical process of a b
The Markov model is a stochastic model that shows uncertain events as transitions between
decision analysis based on a decision tree, defined health states and the rate of transitions and probabilities between the states. Markov
and Figure 1A shows an example of a deci- models are useful to model risk over time, timing of events, and events that may happen more
sion tree that lies at the heart of decision than once.
c
“Folding back” the decision tree analysis method indicates working backwards. One begins at a far
analysis. Decision trees include a decision right chance node of the decision tree. At every chance node the utilities from all chance branches
node (represented by a square) and a num- to the right are summed up by multiplying the probability assigned to every chance branch
ber of decision branches (Figure 1A). Every (providing the weight to derive the expected value) with its associated utility and adding the
decision branch is followed by further results from all branches.

branches representing the occurrence or

non-occurrence of relevant events that may
follow from the management decision (eg, a risk that continuous over time and has out-
developing drug-induced hepatitis from comes that occur at different points in time
latent TB treatment, dying from drug- (eg, ongoing risk over a lifetime of devel-
induced hepatitis, or dying from an unre- oping TB in a patient with latent TB), and
lated cause). These events are represented when outcomes of interest may happen
by chance nodes (circles) with the associated more than once (eg, development of a repeat
probability for alternative outcomes at that episode of TB).7
node. Terminal nodes (triangles) show the A state-transition model consists of cy-
final outcome of a decision path. cles of equal increments of time (eg, 1 month
Most outcomes occur over varying time or 12 months). One assumption of the
frames, and patients may experience more model is that in every cycle the patient re-
than one outcome over time. An optimal de- sides in a finite number of "health states."7
cision analysis will be able to take the time Thus, in the model, a patient may be alive
factor and the probability of a patient experi- and TB free, develop TB, be cured from TB,
encing more than one event over time into die from TB, or die from an unrelated cause
account. The methodology for this process (Figure 1B). When moving through cycles,
involves what is called a Markov or state- patients transition between these different
transition model (Figure 1B), which allows outcomes (eg, after developing active TB a
modeling a decision problem that involves patient might recover from TB or die of
Mayo Clin Proc. n August 2021;96(8):2205-2217 n https://doi.org/10.1016/j.mayocp.2021.02.003 2207
www.mayoclinicproceedings.org
 MAYO CLINIC PROCEEDINGS

Die from
Drug-induced hepatitis
hepatitis Markov-subtree
Latent TB Survive risk of TB
M
treatment
Markov-subtree See figure 1B
IGRA true risk of TB
positive No hepatitis
M

No latent TB See figure 1B

treatment

Die from
IGRA positive Drug-induced hepatitis
Latent TB hepatitis
Survive
treatment
IGRA false No hepatitis
positive
IGRA No latent TB
treatment

IGRA true
negative
Latent TB
IGRA negative Markob-subtree
testing IGRA false risk of TB
negative
M

Markob-subtree See figure 1B

Latent TB risk of TB
present
M
No testing
Latent TB not See figure 1B
present

This is a simplified decision tree that depicts two courses of action: No testing/treatment and interferon-gamma release essay (IGRA) and
treatment for positive IGRA that impact on outcomes associated with the presence or absence of latent TB
A

FIGURE 1. A, Example of decision tree structure. B, Markov sub-tree. A simplified decision tree depicts two courses of action: no
testing/treatment and interferon-gamma release assay (IGRA) and treatment for positive IGRA that impact outcomes associated with
the presence or absence of latent tuberculosis (TB).

TB). Another assumption is that future DID THE DECISION ANALYSIS ADDRESS A
outcome only depends on current outcome, RELEVANT CLINICAL QUESTION
not past outcome. INCLUDING ALL IMPORTANT OUTCOMES?
Beware that figures accompanying ana- Medical decision analyses should have a
lyses in published papers are highly simpli- clear focus 1) addressing a clinical question
fied schematics. They will usually not allow for a specific group of patients; 2) using an
the reader to understand the detailed model actionable intervention and one or multiple
structure. clear alternative course(s) of action; and 3)
We will now move to our criteria for using outcomes that are relevant for clinical
assessing the credibility and applying a deci- decision making (ie, participants, interven-
sion analysis (Table 2). tion, comparator, and outcome, aka the
PICO process). The decision tree on which
the decision analysis is based must include
FIRST JUDGMENT: WAS THE METHODOL- all choice alternatives that are relevant to
OGY OF THE DECISION ANALYSIS clinical decision making. Omitting manage-
CREDIBLE? ment alternatives can provide very
Table 3 summarizes all components of judg- misleading results in a cost-effectiveness
ment 1 d determining the credibility of the analysis (when an effective but less expen-
methods of a decision analysis. sive alternative is not included). In our
n n
2208 Mayo Clin Proc. August 2021;96(8):2205-2217 https://doi.org/10.1016/j.mayocp.2021.02.003
www.mayoclinicproceedings.org
MEDICAL DECISION ANALYSIS

Die from other

causes
Other death
Alive TB-free continue
TB free
Alive TB-free
Continue
Develop
TB
Active TB
Die from other
causes
Unrelated death
Active TB die from
TB
Markov TB death
model Continue
Recover
risk of TB
M from TB
Cured TB
Die from other
causes
Cured TB Other death
Continue
Cured TB
TB death

Other death

Unrelated
death

Legend:
Decision node: Indicates a choice facing the decision maker
Chance node: Indicates chance events
M Markov node, followed by Markov sub-tree (shown in figure 2.3)
Terminal node: Indicates final outcomes (absorbing states=death) or outcomes that indicate
the state in which the patient commences the next Markov cycle.

FIGURE 1. Continued

clinical example, relevant alternative strate- analysis of latent TB treatment with isoni-
gies include a TST (and treating if there is azid in close contacts of patients with infec-
evidence of infection), or an IGRA (and tious TB did not include any adverse events
treating if there is evidence of infection), of preventive TB treatment and might there-
and no testing and treating. fore have overestimated the benefit of pre-
A medical decision analysis should ventive TB treatment compared with no
include all outcomes that are important to treatment.17
patients (eg, mortality, quality of life, and
functional status) rather than surrogate out- IS THE ANALYTIC TIME HORIZON OF THE
comes (eg, result of a laboratory or imaging DECISION ANALYSIS APPROPRIATE?
test). In their efforts to simplify the decision The time frame of the model, often referred
model, decision analysts should not omit to as the time horizon of the analysis, should
rare but important outcomes such as death; be long enough to capture any costs and
any such omission reduces the trustworthi- benefits that occur because of the interven-
ness of the result. Including harms of an tion. For some decisions, choosing a lifetime
intervention is as important as including horizon is appropriate, for example, when
benefits. For example, a cost-effectiveness modeling a latent infection that can develop
Mayo Clin Proc. n August 2021;96(8):2205-2217 n https://doi.org/10.1016/j.mayocp.2021.02.003 2209
www.mayoclinicproceedings.org
 MAYO CLINIC PROCEEDINGS

decision model determines the output of the

TABLE 2. Guide for Appraising and Applying the Results of a Decision Analysis
model in terms of the best course of action
First Judgment: Evaluate the Credibility of the Methods of a Decision Analysis that maximizes utility.21,22 On a utility scale,
Did the decision analysis address a relevant clinical question including all important
perfect health is typically assigned a value of
outcomes?
1 and being dead a value of 0. Many health
Is the time horizona of the decision analysis appropriate?
Are variables in the model (probabilities and utilities) based on current best states in a decision model are associated
evidence? with decrements in quality of life (eg, a pa-
For cost-effectiveness analysis: what was the analysis perspective? tient experiencing an adverse effect of a TB
Second Judgment: Rate the Confidence in the Preferred Course of Action medication has a health state utility of less
What are the sources of the probabilities and utilities in the decision tree? than 1). Health state utilities enable calcula-
How robust are the decision analysis results? tion of QALYs as an outcome. If, for
Are the results applicable to my patient? example, a patient lives for 10 years in a
a
All Markov models include a time horizon, whereas simple decision trees do not. A Markov health state that is associated with a utility
model should therefore be used when the timing of events is important (eg, the risk of preventive of 0.7, this is equal to 7 QALYs.
treatment occurs now, but the benefit is expected in the future).
A well-conducted decision analysis
should present the sources of all parameters
(probabilities, utilities, and cost in a cost-
effectiveness analysis; see below) used in
into active disease at any point during the
the decision tree. Most published decision
patient’s life.18 A study of the risk of devel-
analyses report the sources of the parameter
oping TB in young adults with latent TB
estimates and some include a systematic re-
that only has a time horizon of 20 years19
view and meta-analysis at least as source
will therefore not capture the cumulative
for the effectiveness of the intervention(s)
lifetime risk. In other situations it may be
under consideration.23
appropriate to choose a time frame over a
A critical look at the sources of the prob-
few years, for example, when the decision
abilities and utilities is perhaps the easiest
model focuses on interventions to prevent
and most important critical judgement for
the recurrence of cancer.20
the clinician user of a decision analysis.
Ideally, all important parameters that materi-
ARE VARIABLES IN THE MODEL (PROBA- ally affect the results are based on systematic
BILITIES AND UTILITIES) BASED ON CUR- reviews. However, not all evidence derived
RENT BEST EVIDENCE? from systematic reviews is high quality
A decision tree is populated with events and depending on the quality of the included
health states, each with their associated studies and the review methodology. To
probabilities (eg, the likelihood of devel- the extent that evidence is low quality (that
oping active TB) and utilities (eg, a year of is, associated with high uncertainty), results
life with active TB may be judged as having of the decision analysis will be less robust.
only half the value of a year of life in full
health). FOR COST-EFFECTIVENESS ANALYSIS:
Every branch in a decision tree is WHAT WAS THE ANALYSIS PERSPECTIVE?
assigned a probability (Figure 2A). Together Decision analysis can include the expected
the probabilities of all branches following a costs of decision alternatives, referred to
chance node add up to 1. Probabilities that as cost-effectiveness analysis. Cost-
reflect the risk that a certain event will effectiveness analysis is commonly
happen are often modified by patient charac- performed to compare two or more decision
teristics such as age or gender. It is therefore alternatives where one offers an improved
necessary to specify the population charac- health outcome but at increased cost.24
teristics to which the results of the decision Cost-effectiveness analyses focus on the in-
analysis apply. cremental cost of an intervention for a
The combination of probabilities derived particular difference in outcome d for
from the literature and values (utilities) in a instance, the additional cost of a
n n
2210 Mayo Clin Proc. August 2021;96(8):2205-2217 https://doi.org/10.1016/j.mayocp.2021.02.003
www.mayoclinicproceedings.org
MEDICAL DECISION ANALYSIS

TABLE 3. Using the Guide: Judgment 1, Determining the Credibility of the Methods of a Decision Analysis (Cost-Effectiveness Analysis for
Latent TB), applied to the study of Tasillo et al.a
The decision analysis addressed a relevant clinical question of a specific group of patients (non-US born residents with and without medical
comorbidities). The five evaluated courses of action were clear and actionable in a clinical setting where both tests (tuberculin skin test [TST] and
interferon-gamma release assay [IGRA] test) are available. The main outcome quality-adjusted life-years (QALYs) was a patient-important outcome
relevant for clinical decision making.
The clinically most important possible outcomes (even when rare) were included: developing active TB, being cured from TB, dying from TB,
experiencing drug-induced hepatitis from latent TB treatment, dying from drug-induced hepatitis, dying from causes other than TB.
The decision analysis used a Markov model with a time horizon of a lifetime which allowed for appropriately modelling the risk of TB re-activation over
a person’s lifetime.
The decision analysis authors performed a meta-analysis for the diagnostic performance of the TST and IGRA d they did not specify that they followed
the procedures of a trustworthy systematic review.16 For the other variable values in the model the authors chose one or two studies to provide
estimates. The representativeness of these studies is uncertain. The effectiveness estimate of latent TB treatment (relative risk of developing TB with
treatment) was sourced from a randomized trial and the probability of drug-induced hepatitis was sourced from two randomized trials. The
probabilities of the risk of developing TB in patients with diabetes and latent TB the case-fatality rate of TB and the drug- case-fatality of drug-induced
hepatitis were sourced from one observational study respectively.
To support their choice of utilities the authors cite two studies, one addressing utilities in liver disease and the other in latent and active TB. The rigor
and representativeness of these studies remains uncertain.
The study was a cost-effectiveness analysis from the health care sector perspective. While the cost estimates are thus not suitable for individual decision
making, the effectiveness outcome of the model (QALYs) can be used for decision making in individual patients.
The authors provided most of the information needed to address confidence in the preferred course of action. They described the sources of the
probabilities and utilities in the decision tree although as we have noted the sources have weaknesses. The authors included a general statement that
“In general the base case conclusions were robust to changes in core model parameters.” A probabilistic sensitivity analysis using Monte Carlo
simulation was conducted for the prevalence of latent TB test characteristics and the rate of developing TB but not for any utility estimates). The
model seemed to be robust for supporting “testing and treatment” (based on assessment of QALYs) but not the choice of testing (see Table 5).
The sensitivity analysis results were provided for the cost-effectiveness (incremental cost-effectiveness ratio ¼ cost/QALY) but not for QALYs alone.
a
Tasillo et al.3

management strategy to delay a single death QALY, which reflects the quantity and qual-
from TB d referred to as an incremental ity of life. The use of a utility (perceived
cost-effectiveness ratio (ICER).25,26 Thus, health-related quality of life for a health
the ICER quantifies the trade-offs between state) allows comparisons across different
improved health outcomes and resources diseases and health sectors by using a com-
spent.27 mon unit of measure (cost/QALY gained).29
An intervention that is both less effective Benefits, and particularly costs, are likely to
and more expensive than an alternative differ depending on the perspective of the
intervention is considered to be “strictly target audience be it the patient, a hospital, a
dominated” and is clearly inferior.28 If there health maintenance organization, an employer,
are multiple interventions analyzed, inter- a government payer, or society. The perspective
ventions are listed according to cost, with will determine which costs and health out-
the cheapest and least effective intervention comes are included in the model and how
forming the baseline comparison. The ICER they are valued. A cost-effectiveness analysis
for the next-most-expensive intervention is of an infectious disease from a societal perspec-
then calculated by dividing the difference tive, for example, would include modeling of
in cost by the difference in effectiveness. In disease transmission and associated health out-
this way the analysts can calculate the comes in affected people. The most common
ICER for all remaining options (Table 4). perspective of cost-effective analyses is the
Dominated interventions are listed but not payer perspective, which is not recommended
included in the ICER calculation. by experts.25,26
The unit of effectiveness in a cost- If the effectiveness results of a cost-
effectiveness analysis is commonly the effectiveness analysis are used for clinical

Mayo Clin Proc. n August 2021;96(8):2205-2217 n https://doi.org/10.1016/j.mayocp.2021.02.003 2211

www.mayoclinicproceedings.org
 MAYO CLINIC PROCEEDINGS

Fatal hepatitis
u=0
P=.02
Drug-induced Survive, active Survive
u=1
hepatitis TB P=.99

P=.01 P=0.98x(0.1x0.9) Fatal TB

=.0882
P=.01
u=0

Latent TB Survive, no TB
treatment P=1–(0.02+0.0882) u=1
=.8918

Survive
u=1
Active TB P=.99

P=0.1x0.6=.06 Fatal TB
No hepatitis u=0
0.01
Latent TB P=.99
No TB
P=1–0.06=.94
u=1

Survive
Active TB P=.99
u=1
No latent TB P=.1 Fatal TB
treatment u=0
P=.01
No TB
u=1
P=.9
p: probability
u: utility
The following model assumptions were made:
• Probability (p) of developing active TB:
• Without latent TB treatment the baseline risk of developing TB is 10%
• With latent TB treatment the relative risk (compared to no treatment) is 0.6
• With incomplete latent TB treatment (in case of severe drug induced hepatitis)
the relative risk (compared to no treatment) is 0.9
• The probability of death if active TB develops is 1%
• The probability of drug-induced hepatitis from latent TB treatment is 1%
• The probability of death from severe drug-induced hepatitis is 2
Utilities were either 0 (death) or 1.0 (full health)
A

FIGURE 2. A, Simplified decision tree. B, Simplified “folded back” decision tree. This simplified decision
tree (without Markov model, for illustrative purposes only) shows the process of “folding back the tree” to
determine the best course of action in somebody who has latent tuberculosis (TB). In this case, the best
case of action is to give latent TB treatment with the “no latent TB treatment” alternative marked as the
rejected alternative. Total utility for latent TB treatment is 0.9992 versus 0.999 for no latent TB treatment.
The outcomes at chance nodes (green) are mutually exclusive, and their combined probability must sum
to 1.

decision making in individual patients, the the overall cost estimates, as they are from
model should not include any components a health sector rather than a patient perspec-
that are not relevant to the patient perspec- tive, and focus on the health outcomes for
tive. To the extent that the perspective dif- decision making in your patient including
fers from the one that is relevant to you, QALYs and the number of persons needed
you should take that into account. For to test and treat to prevent one case of TB.
instance, the individual clinician and patient You might want to add any information on
may be uninterested in system costs. Com- costs that are relevant to the patient, if avail-
ing back to the cost-effectiveness analysis able. Importantly, as secondary cases from
in our clinical scenario3: you would ignore TB transmission are included in this

n n
2212 Mayo Clin Proc. August 2021;96(8):2205-2217 https://doi.org/10.1016/j.mayocp.2021.02.003
www.mayoclinicproceedings.org
MEDICAL DECISION ANALYSIS

(0.0882x0.99)+ Fatal hepatitis

P=.02 u=0
0.8918=0.9791
Survive
Drug-induced Survive, active u=1
hepatitis P=.99
TB
(0.01x0.9791)+(0.99x
P=.01 P=.0882
0.9994)=0.9992 Fatal TB
u=0
P=.01

Latent TB Survive, no TB 0.99x1=0.99

treatment P=.8918 U=1

Survive
(0.06x0.99)+(0.94 u=1
x1)=0.9994 Active TB P=.99

P=0.1x0.6=.06 Fatal TB
No hepatitis u=0
0.01
P=.99
Latent TB
No TB 0.99x1=0.99
u=1
P=1–0.06=.94
0.99x1=0.99
Survive
u=1
P=.99
Active TB
No latent TB P=.1 Fatal TB
treatment u=0
P=.01

(0.1x0.99)+(0.9x No TB
u=1
1)=0.999 P=.9

This simplified decision tree (without Markov model, for illustrative purposes only)
demonstrates the process of “folding back the tree” to determine the best course of action in
somebody who has latent TB. In this case the best case of action is to give latent TB
treatment with the “no latent TB treatment” alternative marked as the rejected alternative.
Total utility for latent TB treatment is 0.9992 versus 0.999 for no latent TB treatment.

The outcomes at chance nodes (green) are mutually exclusive, and their combined
probability must sum to 1.
B

FIGURE 2. Continued

analysis, you would tell your patient that the branch (providing the weight to derive the
outcomes also include how other people expected value) with its associated utility
close to them would be affected by the cho- and adding the results from all branches.
sen course. When arriving back at the decision node,
the model will present the total utility asso-
SECOND JUDGMENT: WHAT IS THE CONFI- ciated with each option d the one with
DENCE IN THE PREFERRED COURSE OF the highest utility represents the best choice
ACTION? (unless, in a cost-effectiveness analysis, it
To determine the preferred course of action proves prohibitively expensive). Figure 2B
from a decision tree, one works backwards gives an example of the process of folding
(ie, from right to left, that is from the termi- back a decision tree.
nal nodes to the initial decision node) by
“folding back” the tree. At every chance HOW ROBUST ARE THE DECISION ANAL-
node the utilities from all chance branches YSIS RESULTS?
to the right are summed up by multiplying To test the impact of uncertainties on the
the probability assigned to every chance model outcome, the analyst should perform
Mayo Clin Proc. n August 2021;96(8):2205-2217 n https://doi.org/10.1016/j.mayocp.2021.02.003 2213
www.mayoclinicproceedings.org
 MAYO CLINIC PROCEEDINGS

TABLE 4. Cost-Effectiveness League Tablea

Incremental cost Incremental Incremental
Course of action Cost, $ (per person), $ Effectiveness (QALYs) effectiveness (QALYs) cost/effectiveness (ICER)
No testing 20 (assumed) 70 (assumed)
Confirm positive 67 47 70.0009 0.0009 52,200
TST 102 35 70.0011 0.0002 (175,000) dominated
IGRA 156 54 70.0016 0.0005 108,000
Confirm negative 203 47 70.0018 0.0002 235,000
a
In the league table, the possible courses of action are listed in ascending order of their cost. For each course of action the additional (incremental) cost and effectiveness is
compared to the previously listed course of action. The incremental cost-effectiveness ratio (ICER) for the tuberculin skin test (TST) strategy is higher than that of the next,
more effective, alternative–the interferon-gamma release assay (IGRA) strategy. It is therefore dominated by the combination of 2 alternatives and should not be used to
calculate appropriate ICERs. ICER, incremental cost effectiveness ratio; QALY, quality-adjusted life year.

sensitivity analysis d that is, varying proba- distribution reflecting the associated uncer-
bility and utility values through their plau- tainty. As we have noted, the analyst should
sible range d for all key variables in the base estimates on the best existing evidence
decision tree. For this purpose, each variable from systematic review of the relevant litera-
value (for probabilities and utilities) is ture. From such reviews will come the upper
assigned upper and lower bounds or a and lower bounds for the estimates (eg, 95%

Risk of developing active TB

Quality of life post-TB

Effectiveness of treatment for latent TB

Case fatility from TB

Case-fatality from drug-induced hepatitis

Axis for base case value (using point estimates)
Risk of developing drug-induced hepatitis

–0.1 0.1 0.3 0.5

Difference in quality-adjusted life years by treating LTBI

FIGURE 3. Tornado diagram. This is a fictitious example of a tornado diagram for illustrative purposes
only. The x-axis of the tornado diagram displays the main outcome (here, difference in quality adjusted
life-years by treating latent tuberculosis [TB]). The solid line represents the base case outcome/value, that
is, the outcome when the point estimates for all variables (probabilities and utilities) are used. Each variable
included in one-way sensitivity analysis has its own bar, and the width of the bar is indicative of the impact
that variation of the variable through its plausible range has. Values to the right of the vertical 0-line
indicate a net gain and those to the left indicate a net loss in QALYs with latent TB treatment. The
bars are arranged in descending order of width, so that the diagram is essentially a ranking of variables
based on their potential impact on the outcome of the decision analysis. Special attention should be given
to variables that cannot only significantly change the effect size of the outcome but the direction of the
overall effect. In the example, the two variables “Risk of developing active TB” and “Quality of life post-
TB” can result in different study conclusions (different effect direction) when varied through their plausible
range.

n n
2214 Mayo Clin Proc. August 2021;96(8):2205-2217 https://doi.org/10.1016/j.mayocp.2021.02.003
www.mayoclinicproceedings.org
MEDICAL DECISION ANALYSIS

TABLE 5. Using the Guide: Judgment 2, Determining the Confidence in the Preferred Course of Action (Cost-
Effectiveness Analysis for Latent Tuberculosis Infection), applied to the study of Tasillo et ala
What was the preferred course of action (focusing on effectiveness outcome only)?
Testing and treatment for latent tuberculosis (TB) always resulted in quality-adjusted life years (QALYs) gained
compared with no testing, independently of the testing algorithm used.
For non-US born patients with diabetes aged 57 years (comparable with our clinical scenario patient), the “confirm
negative” testing strategy in which a negative interferon-gamma release assay (IGRA) followed by a positive
tuberculin skin test (TST) (or an initial positive IGRA) was interpreted as indicative of latent TB was associated with
the highest gain in QALYs compared with no testing (incremental QALY ¼ 0.0018 ¼ 0.7 days) and the lowest
number needed to test and treat (n¼362). The “confirm positive” strategy, in which a positive TST was followed
by an IGRA and both had to be positive to indicate latent TB, had the lowest gain in QALYs compared with no
testing (incremental QALY ¼ 0.0009 ¼ 0.3 days) and the highest number needed to test and treat (n¼749).
How robust are the decision analysis results?
While the robustness of the results could potentially be compromised by concerns regarding sources of both
probabilities and utilities, the results were robust in sensitivity analyses. Sensitivity analysis of latent TB prevalence
indicated that the prevalence needed to be 4.5% in non-US born patients with diabetes (such as our patient in
the case scenario) in order for any testing to be the preferred intervention, a criterion that is very likely fulfilled in
the patient from India (see below).
While the preferred testing for latent TB varied in sensitivity analysis for TST specificity, the proportion of patients
returning for TST reading, age, and utility (quality of life) with latent TB and post-TB, some form of testing and
treatment (as opposed to no testing and treatment) was always the preferred option. Strategies including IGRA
were preferred in more than 60% of simulations for non-US born populations with diabetes. For utilities, while
certainty in the results was reduced based on the sources of most utilities, the model seemed to be robust for
supporting “testing and treatment” but not the choice of testing.
Do the results apply to my patient?
While the population assessed in the decision model comprised residents born outside the US with diabetes like our
patient, the latent TB prevalence estimate of 15.9% was based on the whole population of non-US born residents,
a very heterogeneous group with highly variable estimates for latent TB prevalence depending on the TB incidence
in their respective countries of birth. It is estimated that almost 40% of people living in India have latent TB,31 which
means that the pre-test probability in our patient in the case scenario is higher than that of the study population in
the decision model. Thus, the benefits of testing/treatment for latent TB compared with no testing are likely greater
in our patient compared with the result of the decision analysis. This increases the certainty in the preferred course
of action of testing and treating compared with no testing.
Confidence in the preferred course of action
The preferred course of action is consistently testing and treating for latent TB with the optimal testing strategy of
“confirm negative” showing a QALY gain of 0.7 days compared with no testing and a number needed to test and
treat of 362 in non-US born patients with diabetes aged 57 years (comparable to our patient). Based on the
robustness of the preferred course of action in sensitivity analysis, indicating that the results are not highly
determined by changes to parameter assumptions, we have moderate confidence in the results of the decision
analysis despite some concerns regarding sources of included parameters.
The differences in QALYs between different intervention options including no testing/treatment were minimal (less
than 1 day), indicating that this is a preference-sensitive decision and shared decision making, is essential. Your
patient indicates that he is keen to avoid developing active TB. He is not familiar with acute hepatitis; however, after
you describe the condition to him, he is not particularly concerned. He has a positive IGRA test and goes on to
receive latent TB treatment.
a
Tasillo et al.3

confidence intervals from a systematic review Analysts can present results of one-way sensi-
of primary studies). One-way sensitivity anal- tivity analyses in a tornado diagram, which
ysis d that is, considering the possible true provides information on the relative plausible
values of one variable at a time d is impact of the variables addressed in sensi-
commonly used to explore the influence of tivity analyses (see Figure 3, which shows a
plausible variation on the decision problem. fictitious example).18
Mayo Clin Proc. n August 2021;96(8):2205-2217 n https://doi.org/10.1016/j.mayocp.2021.02.003 2215
www.mayoclinicproceedings.org
 MAYO CLINIC PROCEEDINGS

In addition to variable values, the chosen decision analysis important to my patient for
complexity of a model (number of variables decision making?
included) and the structure of the model Discussing a patient’s preferences and
(eg, the type of defined health states in a values and ascertaining that they are consis-
Markov model) can potentially impact the tent with those assumed in the model is
result of a decision analysis. Sensitivity anal- essential. Focusing on elucidating patients’
ysis can be used to determine whether the preferences is particularly important if the
results are robust to changes in model study conclusion changed in the sensitivity
complexity and structure. analysis of utilities used in the decision
The robustness of the model results is analysis.
decreased when sensitivity analysis results
in important changes in the estimates of ef- CONCLUSION
fect, most seriously when it changes the di- Decision analyses provide a systematic
rection of the overall effect. If the study approach to assessing the risk-benefit bal-
conclusion (direction of the effect) changes ance of alternative strategies using objective
in the sensitivity analysis of utilities, the evidence and careful, model-based evalua-
preferred course of action is strongly influ- tion. The information obtained from a deci-
enced by patients’ preferences. sion analysis can be used for clinical
In addition to one-way sensitivity anal- decision making as well as decision making
ysis, analysts can conduct probabilistic sensi- at a health policy level. When applying the
tivity analysis using “Monte Carlo results of a decision analysis, we first judge
simulation.”30 Here, the analyst can examine the credibility of the methods of the decision
the impact of simultaneously varying results analysis and how confident we are in the
of several input variables across their prob- preferred course of action.
able range. This process is repeated in
numerous simulations (eg, 100,000), and
in each simulation the value for each variable SUPPLEMENTAL ONLINE MATERIAL
Supplemental material can be found online
is selected at random according to the plau-
at http://www.mayoclinicproceedings.org.
sible range or probability distribution. The
Supplemental material attached to journal
analyst can then present the proportion of
articles has not been edited, and the authors
simulations in which one decision option
take responsibility for the accuracy of all
or another was the preferred option.
data.
We will have more confidence in the
result if one decision option is the preferred
course in a high number of simulations (eg, Abbreviations and Acronyms: ICER = incremental cost-
effectiveness ratio; IGRA = interferon-gamma release assay
in 90% of simulations or more). test; QALY = quality-adjusted life year; TB = tuberculosis;
Returning to the clinical scenario TST = tuberculin skin test
(Table 5), you are reasonably confident in
the preferred course of action.
Potential Competing Interests: The authors report no po-
tential competing interests.

Grant Support: Dr Dobler is supported by a fellowship of

ARE THE RESULTS APPLICABLE TO MY
the Australian National Health and Medical Research Coun-
PATIENT? cil (grant# APP1123733). The fellowship sponsor had no
You can address the applicability of the deci- role in manuscript design, data interpretation, or writing of
sion analysis to a specific patient by asking the manuscript.
the following questions: 1) Is the population Correspondence: Address to Claudia C. Dobler, MD, PhD,
in the decision analysis similar to my patient? Institute for Evidence-Based Healthcare, Bond University,
2) Is the intervention (the intervention op- Gold Coast, Queensland, Australia ([email protected];
tions) in the decision analysis similar to the Twitter: @ClaudiaCDobler).
intervention(s) under assessment in the clin- ORCID
ical situation? 3) Is the main outcome of the Claudia C. Dobler: https://orcid.org/0000-0002-5460-

n n
2216 Mayo Clin Proc. August 2021;96(8):2205-2217 https://doi.org/10.1016/j.mayocp.2021.02.003
www.mayoclinicproceedings.org
MEDICAL DECISION ANALYSIS

0189; Zhen Wang: https://orcid.org/0000-0002-9368- 14. Qaseem A, Chou R, Humphrey LL, Starkey M, Shekelle P.
6149; M. Hassan Murad: https://orcid.org/0000-0001- Venous thromboembolism prophylaxis in hospitalized patients:
5502-5975 a clinical practice guideline from the American College of Phy-
sicians. Ann Intern Med. 2011;155(9):625-632.
15. Grant PJ, Conlon A, Chopra V, Flanders SA. Use of venous
thromboembolism prophylaxis in hospitalized patients. JAMA
REFERENCES Intern Med. 2018;178(8):1122-1124.
1. Al-Rifai RH, Pearson F, Critchley JA, Abu-Raddad LJ. Associa- 16. Murad MH, Montori VM, Ioannidis JP, et al. How to read a systematic
tion between diabetes mellitus and active tuberculosis: a sys- review and meta-analysis and apply the results to patient care: users’
tematic review and meta-analysis. PLoS One. 2017;12(11): guides to the medical literature. JAMA. 2014;312(2):171-179.
e0187967. 17. Diel R, Nienhaus A, Schaberg T. Cost-effectiveness of isoniazid
2. Guidelines on the Management of Latent Tuberculosis Infection. chemoprevention in close contacts. Eur Respir J. 2005;26(3):
Geneva, Switzerland: World Health Organization; 2015. 465-473.
3. Tasillo A, Salomon JA, Trikalinos TA, Horsburgh CR Jr, 18. Dobler CC, Martin A, Marks GB. Benefit of treatment of latent
Marks SM, Linas BP. Cost-effectiveness of testing and treatment tuberculosis infection in individual patients. Eur Respir J. 2015;
for latent tuberculosis infection in residents born outside the 46(5):1397-1406.
United States with and without medical comorbidities in a 19. Taylor WC, Aronson MD, Delbanco TL. Should young adults
simulation model. JAMA Intern Med. 2017;177(12):1755-1764. with a positive tuberculin test take isoniazid? Ann Intern Med.
4. Tversky A, Kahneman D. Judgment under uncertainty: heuristics 1981;94(6):808-813.
and biases. Science. 1974;185(4157):1124-1131. 20. Younis T, Rayson D, Skedgel C. The cost-utility of adjuvant
5. Caro JJ, Briggs AH, Siebert U, Kuntz KM; Force I-SMGRPT. chemotherapy using docetaxel and cyclophosphamide
Modeling good research practicesdoverview: a report of the compared with doxorubicin and cyclophosphamide in breast
ISPOR-SMDM Modeling Good Research Practices Task cancer. Curr Oncol. 2011;18(6):e288-e296.
Forced1. Value Health. 2012;15:796-803. 21. Swales JD. Science in a health service. Lancet. 1997;349(9061):
6. Roberts M, Russell LB, Paltiel AD, et al. Conceptualizing a model: a 1319-1321.
report of the ISPOR-SMDM Modeling Good Research Practices 22. Hunink M, Glasziou P, Siegel J, et al. Decision Making in Health
Task Forced2. Value Health. 2012;15(6):804-811. and Medicine: Integrating Evidence and Values. New York, NY:
7. Siebert U, Alagoz O, Bayoumi AM, et al. State-transition modeling: a Cambridge University Press; 2001.
report of the ISPOR-SMDM Modeling Good Research Practices 23. Malhotra A, Wu X, Forman HP, Matouk CC, Gandhi D, Sanelli P.
Task Forced3. Value Health. 2012;15(6):812-820. Management of tiny unruptured intracranial aneurysms: a compar-
8. Karnon J, Stahl J, Brennan A, et al. Modeling using discrete event simu- ative effectiveness analysis. JAMA Neurol. 2018;75(1):27-34.
lation: a report of the ISPOR-SMDM Modeling Good Research 24. Ryder HF, McDonough C, Tosteson AN, Lurie JD. Decision
Practices Task Forced4. Value Health. 2012;15(6):821-827. Analysis and Cost-effectiveness Analysis. Semin Spine Surg.
9. Pitman R, Fisman D, Zaric GS, et al. Dynamic transmission modeling: 2009;21(4):216-222.
a report of the ISPOR-SMDM Modeling Good Research Practices 25. Sanders GD, Neumann PJ, Basu A, et al. Recommendations
Task Force–5. Value Health. 2012;15(6):828-834. for conduct, methodological practices, and reporting of
10. Briggs AH, Weinstein MC, Fenwick EAL, et al. Model parameter cost-effectiveness analyses: Second Panel on Cost-
estimation and uncertainty: a report of the ISPOR-SMDM Effectiveness in Health and Medicine. JAMA. 2016;316(10):
Modeling Good Research Practices Task Forced6. Value 1093-1103.
Health. 2012;15(6):835-842. 26. Neumann PJ, Sanders GD, Russell LB, Siegel JE, Ganiats TG, eds.
11. Eddy DM, Hollingworth W, Caro JJ, et al. Model transparency Cost-Effectiveness in Health and Medicine. 2nd Edition. New
and validation: a report of the ISPOR-SMDM Modeling Good York, NY: Oxford University Press; 2016.
Research Practices Task Forced7. Value Health. 2012;15(6): 27. Bang H, Zhao H. Median-based incremental cost-effectiveness
843-850. ratio (ICER). J Stat Theory Pract. 2012;6(3):428-442.
12. Richardson WS, Detsky AS. Users’ guides to the medical liter- 28. Cohen DJ, Reynolds MR. Interpreting the results of cost-
ature. VII. How to use a clinical decision analysis. A. Are the re- effectiveness studies. J Am Coll Cardiol. 2008;52(25):2119-
sults of the study valid? Evidence-Based Medicine Working 2126.
Group. JAMA. 1995;273(16):1292-1295. 29. Dobler CC. Screening strategies for active tuberculosis: focus on
13. Richardson WS, Detsky AS. Users’ guides to the medical liter- cost-effectiveness. Clinicoecon Outcomes Res. 2016;8:335-347.
ature. VII. How to use a clinical decision analysis. B. What are 30. Sonnenberg FA, Beck JR. Markov models in medical decision mak-
the results and will they help me in caring for my patients? Ev- ing: a practical guide. Med Decis Making. 1993;13(4):322-338.
idence Based Medicine Working Group. JAMA. 1995;273(20): 31. Chadha VK. Tuberculosis epidemiology in India: a review. Int J
1610-1613. Tuberc Lung Dis. 2005;9(10):1072-1082.

Mayo Clin Proc. n August 2021;96(8):2205-2217 n https://doi.org/10.1016/j.mayocp.2021.02.003 2217

www.mayoclinicproceedings.org