ORIGINAL ARTICLE
Risk Preferences: Consequences for Test and
Treatment Thresholds and Optimal Cutoffs
Stefan Felder, PhD, Thomas Mayrhofer, PhD
Risk attitudes include risk aversion as well as higher-order
risk preferences such as prudence and temperance. This
article analyzes the effects of such preferences on medical
test and treatment decisions, represented either by test
and treatment thresholds or—when the test result is not
given—by optimal cutoff values for diagnostic tests. For
a risk-averse decision maker, effective treatment is
a risk-reducing strategy since it prevents the low health
outcome of forgoing treatment in the sick state. Compared
with risk neutrality, risk aversion thus lowers both the test
and the treatment threshold and decreases the optimal
INTRODUCTION
In 1975, Stephen G. Pauker and Jerome P. Kassirer1
introduced the concept of diagnostic risk, which
refers to uncertainty over the true health state of the
patient (definitions of terms in italics are given in
the supplemental glossary) and its impact on medical
treatment decisions. They showed how a decision
Received 17 October 2011 from the Faculty of Business and Econom-
ics, University of Basel, Basel, Switzerland (SF), and German Health
Economics Research Center CINCH, University of Duisburg-Essen,
Essen, Germany (SF, TM). Financial support for this study was provided
entirely by (employment) contracts with the University of Duisburg-
Essen, Germany, and the University of Basel, Switzerland. The funding
agreement ensured the authors’ independence in designing the study,
interpreting the data, writing, and publishing the report. The following
authors are employed by the sponsor: Stefan Felder (University of Basel)
and Thomas Mayrhofer (University of Duisburg-Essen). Revision
accepted for publication 16 May 2013.
Supplementary material for this article is available on the Medical
Decision Making Web site at http://mdm.sagepub.com/supplemental.
Address correspondence to Stefan Felder, University of Basel, Faculty
of Business and Economics, Peter Merian-Weg 6, PO Box, 4002 Basel,
Switzerland; e-mail: stefan.felder@unibas.ch.
Ó The Author(s) 2013
Reprints and permission:
http://www.sagepub.com/journalsPermissions.nav
DOI: 10.1177/0272989X13493969
MEDICAL DECISION MAKING/JANUARY 2014
Downloaded from mdm.sagepub.com by guest on May 6, 2015
test cutoff value. Risk vulnerability, which combines risk
aversion, prudence, and temperance, is relevant if there
is a comorbidity risk: thresholds and optimal cutoff values
decrease even more. Since common utility functions imply
risk vulnerability, our findings suggest that diagnostics in
low prevalence settings (e.g., screening) may be consid-
ered more beneficial when risk preferences are taken
into account. Key words: decision analysis; value of
information; provider decision making; decision rules;
patient decision making. (Med Decis Making
2014;34:33–41)
maker should decide between ‘‘treatment’’ and ‘‘no
treatment’’ based on the a priori probability of illness
in the absence of a diagnostic test. Five years later, the
same authors2 developed a threshold analysis, extend-
ing their earlier work to include the possibility of
diagnostic testing. They proposed a method for valu-
ing a diagnostic test and a rule, also dependent on
the a priori probability of illness, for when to use the
test.
Pauker and Kassirer’s analyses1,2 are rooted in the
expected utility theory, according to which the deci-
sion maker (DM) takes actions based on probabilities
of the possible outcomes, on one hand, and on a value
function over the outcomes, on the other hand. In the
empirical examples provided by Pauker and Kassirer,
probabilities take the form of survival and mortality
rates—treatment increases the probability of survival
for sick patients and reduces it for healthy patients—
where the states of survival and death are implicitly
valued at 1 and 0. In this case, the DM is risk neutral.
This article brings the von Neumann-Morgenstern
hypothesis to its full use and considers risk preferen-
ces comprehensively (i.e., risk aversion, prudence,3
and temperance4), which respectively concern the
second, third, and fourth derivatives of the DM’s util-
ity function. We connect prudence and temperance
using the concept of risk vulnerability,5 which
requires that the DM is mixed risk averse (i.e., the
derivates of the utility function alternate in signs).
Most of the commonly used utility functions (e.g.,
33
 FELDER, MAYRHOFER
logarithmic and root functions) imply mixed risk
aversion.6,7 Only when preferences are risk vulnera-
ble will a DM be more ‘‘risk averse’’ in an overall
sense. We analyze how these risk preferences affect
the thresholds in the a priori probability of illness
that define the prevalence range within which testing
is indicated. Furthermore, we demonstrate how
higher-order risk preferences influence the decider’s
choice of the positivity criterion, which is the optimal
cutoff value of a diagnostic test when the test charac-
teristics are not exogenously given.
The concept of risk aversion in decisions under
uncertainty is well established in economics. Gener-
ally speaking, risk aversion can be explained by the
decreasing marginal utility of income. A risk-neutral
individual is indifferent between a lottery and its (cer-
tain) expected value (i.e., the individual will accept
a fair bet). A risk-averse individual, however, requires
compensation (the risk premium) to accept the bet. In
the medical setting, risk aversion refers to the decreas-
ing marginal utility of health (i.e., a marginal improve-
ment of health is valued more at low health levels than
at high levels). Risk-averse patients would thus be pre-
pared to forego some of their health if they can thereby
reduce uncertainty over their health state.
However, risk preferences are only partially cap-
tured by risk aversion. Higher-order risk attitudes
such as prudence3 and temperance4 also appear to
be relevant in decision making under uncertainty.
Sometimes prudence and temperance are called
third- and fourth-order risk aversion, indicating their
complementarity to (second-order) risk aversion. Eeck-
houdt and Schlesinger8 characterize prudence and
temperance preferences over particular classes of lot-
tery pairs. Prudence is defined as a preference for disag-
gregating a sure loss and a zero-mean background risk.
This means that a DM is more willing to accept an extra
risk when his or her level of health is higher. Temper-
ance implies a preference for disaggregating two (inde-
pendent) zero-mean background risks. So, intuitively,
a DM facing one unavoidable background risk will
seek to reduce exposure to another background risk.
In the medical context, the background risk can be
a comorbidity risk, a disease or disorder that is pres-
ent in addition to the primary health states. We
assume that this risk is exogenous (i.e., neither its
existence nor its extent can be influenced by the treat-
ment decision). We develop a model with and with-
out considering the comorbidity risk. We will show
that a risk-averse DM will test and treat at lower a pri-
ori probabilities of illness than the risk-neutral DM.
Taking comorbidity risk into account, the risk-vul-
nerable DM will test and treat even earlier.
34  MEDICAL DECISION MAKING/JANUARY 2014
Downloaded from mdm.sagepub.com by guest on May 6, 2015
Like Pauker and Kassirer,1,2 we take a normative
stance and derive test and treatment thresholds that
indicate the DM’s best choice depending on the a pri-
ori probability of illness. However, the model may
also be used for a positive analysis and can possibly
explain why we observe testing in low probability
ranges, which we would not expect if the DM acted
in a risk-neutral manner.
The remainder of the article is organized as fol-
lows. We start the second section with a basic model
that captures risk neutrality and risk aversion. We
then add a comorbidity risk, which allows us to
address prudence and temperance under the concept
of risk vulnerability. The different risk preferences
can be compared with the ratio of the utility gain
from treatment in the sick state and the utility loss
from treatment in the healthy state, representing the
tradeoff a DM faces when deciding for or against treat-
ment under uncertainty. The third section investigates
the treatment decision without diagnostic testing and
analyzes how it changes when moving from risk neu-
trality to risk vulnerability. The fourth section adds
diagnostic testing and presents the effects of risk pref-
erences on the test and treatment thresholds. The fifth
section discusses the analysis for endogenous test
results. The sixth section presents a clinical example,
and seventh section discusses the findings and limita-
tions of the model and concludes.
RISK AVERSION, PRUDENCE, TEMPERANCE,
AND RISK VULNERABILITY
Our focus here is on the diagnostic risk inherent in
the decision on medical treatment. In our analysis,
we do not differentiate between physicians and
patients as DMs but rather assume that physicians
decide purely in the interest of their patients. The
patient’s health state i is unknown; he or she can be
sick or healthy, i ¼ s; h, where p describes the proba-
bility of the sick state. The DM must decide whether
to treat the patient (j ¼ þ) or not (j ¼ ). The DM can
also use diagnostics and make the treatment decision
dependent on the test result (i.e., will treat if the test
is positive and will not treat after a negative test
result). Therefore, we use j, j ¼ þ; , for the test result
as well as the treatment decision. We assume that the
DM knows the possible consequences of his or her
decision for the patient’s health. In general, the con-
sequences can be described by a vector of health
attributes with different levels. For simplicity, we
use a single interval-scaled variable H for the
patient’s health state. It follows then that Hh
 RISK PREFERENCES AND TEST AND TREATMENT THRESHOLDS

indicates a true-negative and Hhþ a false-positive test

U (H )
result, while Hsþ stands for a true-positive and Hs Risk-neutral DM
for a false-negative test result. UN (H −
h ) =U (H
A
−
h ) Risk-averse DM
LA
U (H +
)
‘‘No treatment’’ is associated with the extreme A h
LN
health outcomes, that is, the best outcome in the U (H
N
+
h )
healthy state (i ¼ h) and the worst outcome in the U (H
A
+
s )
sick state (i ¼ s). The intermediate outcomes are asso- U N ( H s+ )
ciated with ‘‘treatment.’’ The 4 possible health out- GA
GN
comes Hij (i ¼ h; s, j ¼ þ; ) can be ranked as U N ( H s− ) = U A ( H s− )
follows: Hh > Hhþ ; Hsþ > Hs , where the ranking of
the intermediate outcomes need not be determi-
0 H s− H s+ H h+ H h− H
ned—that is, Hhþ > Hsþ and Hhþ \Hsþ are both possi-
ble. The DM’s elementary utility function is given
Figure 1 The health utility functions for risk-neutral and risk-
by UðHij Þ, which is assumed to have a positive slope:
averse decision makers. DM, decision maker; A, risk averse; N,
U 0 ðHij Þ[qUðHij Þ=qHij > 0. Given the rank order of risk neutral; G, utility gain; L, utility loss; U, utility; H, health sta-
health states
 above,
 positive
  marginal
  utility
 implies tus; h, healthy; s, sick; 1 , treated; –, not treated.
that U Hh > U Hhþ ; U Hsþ > U Hs . The utility
function is assumed to be interval scaled, an assump- of the utility function. Without loss of generality, we
tion also required for the quality-adjusted life year can therefore assume that Hs is valued equally by N
(QALY) approach. and A and that they also equally value Hh . Due to
A positive test result followed by treatment will be the concavity of DM A’s utility function, his or her
beneficial in the sick state and harmful in the healthy utility gain from treatment in the sick state G is higher,
state. We define while his or her utility loss from treatment in the
        healthy state L is lower than DM N’s (see Figure 1).
G ¼ U Hsþ  U Hs > 0 and L ¼ U Hh  U Hhþ > 0 From these considerations, we can derive
ð1Þ GA > GN , LA \LN , or, given (1), GA =LA > GN =LN .
Higher-order risk preferences concern prudence,3
as the utility gain from treatment in the sick state and temperance,4 and their combination, called risk vul-
the utility loss from treatment in the healthy state. nerability.5 Although a positive third derivative of
the utility function, Ui ðHij Þ > 0, characterizes a DM
000
Note that the expressions gain and loss accentuate
the utility change from treatment more clearly than as prudent, a negative fourth derivative, Uiiv ðHij Þ\0,
the terms net benefit (B) and net cost (C) of treatment denotes a temperate DM. Now, mixed risk aversion
used by Pauker and Kassirer.1 (i.e., alternating signs of the derivatives of the pro-
First we distinguish between a risk-averse DM A gressive utility functions) implies that both U ð H Þ
and a risk-neutral DM N. The latter is indifferent and U 00 ð H Þ have a positive slope and are concave. If
between a certain outcome and an uncertain outcome the slope of U 00 ð H Þ decreases more—in relative
of the same expected value, while the former prefers terms—than the slope of U ð H Þ with increasing levels
the certain outcome to the uncertain prospect. of health, a DM’s temperance will exceed his or her
The 2 DMs’ different risk attitudes can be illus- risk aversion in absolute terms:
trated by the curvature of their utility functions.
N has a linear utility function; thus, its first derivative U iv ð H Þ U 00 ð H Þ
is constant (UN 0
ðHij Þ[qUðHij Þ=qHij ¼ const:), and T[  > A[  ; ð2Þ
U 000 ð H Þ U 0 ðHÞ
all higher-order derivatives equal zero (UN x
ðHij Þ[
x j j x
q UðHi Þ=ðqHi Þ ¼ 0 for all x  2). DM A has a concave where A denotes the index of absolute risk aversion
utility function with a negative second derivative, (also known as the Arrow-Pratt measure of absolute
that is, UA00 ðHij Þ\0. A values a marginal improvement risk aversion),9,10 and T is the index of absolute tem-
in health more at low levels than at high levels, while perance.5,11 Equation (2) is known to be a necessary
N is indifferent between marginal health improve- condition for risk vulnerability. Gollier and Pratt5
ments at different health levels. A can also be seen showed that a DM faced with a background risk will
as more inclined toward preventing the worst out- display more risk-averse behavior in an overall sense
come than N. only if his or her preferences are risk vulnerable. Tech-
Under the expected utility hypothesis, decisions nically, risk vulnerability requires not only mixed risk
are independent of positive linear transformations aversion but also that absolute temperance (and

ORIGINAL ARTICLE 35

Downloaded from mdm.sagepub.com by guest on May 6, 2015

 FELDER, MAYRHOFER
absolute prudence) exceeds absolute risk aversion. All
constant relative risk aversion (CRRA) utility functions
and all the common decreasing absolute risk aversion
(DARA)
pffiffiffiffiffi utility functions—for example, lnð H Þ and
H , which exhibit both DARA and CRRA—satisfy
the necessary conditions for risk vulnerability.5
Higher-order risk preferences become relevant if
a comorbidity risk m ~ exists ‘‘in the background’’ of
the health status. In this case, health status H itself
turns into a random variable, H þ m. ~ The comorbidity
risk is assumed to be exogenous—its outcome is inde-
pendent of the DM’s treatment decision. In the supple-
mental appendix, we show that risk vulnerability
implies EGV =ELV > GA =LA , where E stands for the
expected value and V indicates a risk-vulnerable DM.
A risk-vulnerable DM facing a background risk puts
even more weight on improving his or her level of
health in the sick state relative to health loss from treat-
ment in the healthy state than a merely risk-averse DM.
This leads to the following sequence of inequalities:
EGV =ELV > GA =LA > GN =LN :
When considering the utility gain from treatment in
the sick state and the utility loss from treatment in
the healthy state, the risk-neutral DM is indifferent
to marginal changes in the health levels in these 2
states. The risk-averse DM values a marginal increase
in health more in the sick state than in the healthy
state. When facing a background risk, the utility
gain-loss ratio of treatment increases even more for
a risk-vulnerable DM.
TREATMENT DECISION WITHOUT A DIAGNOS-
TIC TEST: THE THERAPEUTIC THRESHOLD
To estimate the value of information provided by
a diagnostic test, one must first consider the DM’s deci-
sion regarding whether to treat in a situation in which
no test is available. In this base case without comorbid-
ity risk, expected utility as a function of j, with j ¼ þ for
treatment and j ¼ for no treatment, becomes
 
 
EU j ð pÞ ¼ pU Hsj þ ð1  pÞU Hhj :
There is an a priori probability of illness p~ at which
the DM is indifferent between treatment and no treat-
ment. This treatment threshold, first derived by
Pauker and Kassirer,1 satisfies
L 1
p~ ¼ ¼ :
L þ G 1 þ G=L
36  MEDICAL DECISION MAKING/JANUARY 2014
Downloaded from mdm.sagepub.com by guest on May 6, 2015
For p\ p,~ the expected utility of no treatment exceeds
the expected utility from treatment, so that not treat-
ing is indicated. For p > p, ~ by comparison, the opti-
mal strategy is to treat.
The therapeutic threshold will depend on the
DM’s risk preferences. We classified the different
risk preferences with respect to the ratio of utility
gains and losses from treatment in the sick and
healthy states. The effects of risk neutrality, risk aver-
sion, and risk vulnerability on the therapeutic thresh-
old can be gleaned from (5) as p~ is decreasing in
ðG=LÞ. In other words, if either the utility gain from
treatment in the sick state increases or the loss from
treatment in the healthy state decreases, the thera-
peutic threshold decreases and treatment is indicated
at lower a priori probabilities. This also makes sense
intuitively: if either the potential benefit from treat-
ment increases or the potential harm decreases, the
treatment option becomes more favorable.
Now, the utility gain from treatment increases for
a risk-averse DM compared with a risk-neutral DM,
ð3Þ GA > GN , while the utility loss from treatment
decreases, LA \LN . This leads to GA =LA > GN =LN
and thus to a lower therapeutic threshold for A than
for N, p~A \ p~N . A uses the treatment strategy as an
insurance device (it reduces the spread between the
possible health states) and thus opts for treatment at
lower prevalence rates than N.
In the situation with comorbidity risk, the third
and fourth derivatives of the utility function appear
to be relevant as well. Since EGV =ELV > GA =LA , it
follows that the therapeutic threshold is reduced
even further: p~V \ p~A . Although the background risk
is exogenous, the risk-vulnerable DM braces himself
or herself in the face of it and treats earlier than the
(merely) risk-averse DM.
Summing up, we conclude that (3) implies
p~V \ p~A \ p~N : ð6Þ
The therapeutic threshold will be lower for a risk-
averse DM than for a risk-neutral one. Facing comor-
bidity risk leads to a further reduction of the thresh-
old for a risk-vulnerable DM.
ð4Þ
TREATMENT DECISION WITH A DIAGNOSTIC
TEST: TEST AND TREATMENT THRESHOLDS
The availability of a diagnostic test increases the
DM’s options. Instead of 2 alternative actions (treat-
ing and not treating), he or she now has the additional
ð5Þ
option of using diagnostics. The performance of a test
 RISK PREFERENCES AND TEST AND TREATMENT THRESHOLDS
Sick (p)
Treatment
Healthy (1– p)
Sick (p)
No Treatment
Healthy (1– p)
Test (+); Se Treatment
Sick (p)
Test (–); 1–Se No Treatment
Test
Test (+); 1–Sp Treatment
Healthy (1–p)
Test (–); Sp No Treatment
Figure 2 Decision tree with a diagnostic test. p, probability of being
sick; U, utility; H, health status; h, healthy; s, sick; 1, treated/test
positive; –, not treated/test negative; Se, sensitivity; Sp, specificity.
depends on its ability to separate the sick from the
healthy patients, which is usually captured by the
positive and negative likelihood ratios connected to
positive and negative test results, respectively. If
the test is employed, we assume that a positive test
result will lead to treatment, while a negative out-
come will not. The decision situation for the DM in
the base model is shown in Figure 2.
In the base case, where we assume that the test
itself implies no harm, the expected utility of a diag-
nostic test, EU Dx ð pÞ, can be written as follows:
    
EU Dx ð pÞ ¼ p Se  U Hsþ þ ð1  SeÞU Hs þ ð1  pÞ
    
Sp  U Hh þ ð1  SpÞU Hhþ
        is treatment. A will prefer not to run the risk of
¼ p U Hs þ Se  G þ ð1  pÞ U Hhþ þ Sp  L ;
where Se is the sensitivity or true-positive rate, 1 – Se
is the false-negative rate, Sp is the specificity or true-
negative rate, and 1 – Sp is the false-positive rate of
the test. The value of diagnostic information is
defined as the additional expected utility resulting
from the use of the test. The reference situation is
the expected utility of not treating for p\ p~ and the
expected utility from treatment for p  p.~ Given equa-
tions (1), (4), and (7), we can solve for the value of
information of a diagnostic test and find
8
>
> EU Dx ð pÞ  EU  ð pÞ ¼
<
p  Se  G  ð1  pÞð1  SpÞL for 0  p\ p~;
VI ð pÞ ¼
>
> EU Dx ð pÞ  EU þ ð pÞ ¼
:
pð1  SeÞG þ ð1  pÞSp  L for p~  p  1:
ORIGINAL ARTICLE
Setting the equations of (8) to zero, we obtain the test
U ( H s+ ) and treatment thresholds of Pauker and Kassirer2:
U ( H h+ ) ð1  SpÞL 1
p~Dx ¼ ¼
Se  G þ ð1  SpÞL 1 þ LRþ G=L
U (H −
s )
Sp  L 1
and p~Rx ¼ ¼ ; ð9Þ
U ( H h− ) ð1  SeÞG þ Sp  L 1 þ LR G=L
U ( H s+ )
where LRþ ¼ Se=ð1  SpÞ denotes the positive and
LR ¼ ð1  SeÞ=Sp the negative likelihood ratio.
U ( H s− ) Like the therapeutic threshold, the test and treatment
thresholds mark probabilities at which the DM is
U ( H h+ )
indifferent between 2 actions. At the test threshold,
U ( H h− )
p~Dx , he or she is indifferent between not treating and test-
ing followed by treatment in the case of a positive test
result). At the treatment threshold, p~Rx , the DM is indif-
ferent between testing and treating without prior testing.
Since LRþ > 1 > LR > 0, p~Dx is located below the
therapeutic threshold and p~Rx above it, p~Dx \ p\ ~ p~Rx .
As the test and treatment thresholds are decreasing
in ðG=LÞ, (3) yields the following rank order for the
effect of risk preferences:
p~Dx ~Dx
V \p ~Dx
A \p ~Rx
N and p ~Rx
V \p ~Rx
A \pN : ð10Þ
The downward shift of the thresholds as we move
from a risk-neutral DM to a risk-vulnerable DM can
be explained intuitively as follows. Under diagnostic
risk, treatment is the risk-reducing strategy. A risk-
averse DM values a test more at low a priori probabil-
ities where the reference strategy in the situation
without a test is no treatment. Consequently, A will
test earlier than N, implying a lower test threshold.
At high a priori probabilities, the reference strategy
a false-negative test result and thus treat earlier.
ð7Þ If there is a comorbidity risk, the utility gain-loss
ratio of treatment for a risk-vulnerable DM increases
even further, leading to lower thresholds. A risk-
vulnerable DM will test and treat earlier—and there-
fore also treat more often, as the prevalence range
within which treatment is indicated expands—when
faced with a comorbidity risk. Since the comorbidity
risk is exogenous, the DM cannot address it directly
through his or her behavior. The DM will instead
reduce overall risk by choosing treatment, the strat-
egy that reduces his or her endogenous risk.
RISK PREFERENCES AND THE OPTIMAL CUTOFF
ð8Þ
Our analysis of the effect of risk preferences on test
and treatment behavior can be extended to include the
37
Downloaded from mdm.sagepub.com by guest on May 6, 2015
 FELDER, MAYRHOFER
situation where the DM has to determine a cutoff value
that distinguishes positive from negative test results. A
good example is the prostate-specific antigen (PSA)
test for the detection of prostate cancer. The analysis of
a blood sample results in a PSA value that the physician
declares as a positive or negative test result depending
on his or her chosen cutoff value. Correspondingly, the
test characteristics sensitivity and specificity are no lon-
ger given but determined by the chosen cutoff.
The test technology can then be plotted on the
receiver operating characteristics (ROC) curve, which
represents all the feasible maximal sensitivity-
specificity pairs. In the simplest case where there is
only 1 marker, such as the PSA value, each possible
cutoff value corresponds to 1 point on the ROC curve.
In general, a higher cutoff value for a test result increases
specificity at the expense of sensitivity. The DM’s task is
to choose the optimal cutoff value at which the patient’s
expected utility is maximized at a given a priori proba-
bility of illness. The optimal cutoff value corresponds to
a point along the ROC curve that satisfies the following
condition (see McNeil and others,12 Metz,13 and Felder
and others14):
dSe L=G
¼ :
dð1  SpÞ p=ð1  pÞ
According to this equation, the slope of the ROC
curve for the optimal cutoff value is equal to the prod-
uct of the inverse of the a priori odds and L/G. With an
increase in the prevalence rate p or in the utility gain
G, the slope of the ROC curve decreases. This implies
an upward move along the ROC curve, which in turn
increases sensitivity at the expense of specificity.
Provided that the marker is positively correlated
with the presence of the illness, this implies a lower
optimal cutoff value. This result is intuitive, since
the importance of detecting the sick increases in the
probability of illness or with a higher utility gain of
treatment in the sick state. The opposite applies for
an increase in the utility loss L, where a downward
move along the ROC curve is optimal, leading to
lower sensitivity and increased specificity.
As the slope of the ROC curve is decreasing in
ðG=LÞ, we can easily qualify the effect of risk preferen-
ces on the optimal point along the ROC. Denoting the
cutoff value with x—which refers to the diagnostic
marker, rather than to a point on the ROC curve—we
then derive the following rank order for the optimal
(*) cutoffs as a function of risk preferences:
xV \xA \xN :
38  MEDICAL DECISION MAKING/JANUARY 2014
Downloaded from mdm.sagepub.com by guest on May 6, 2015
If the test result shows a value below the optimal cut-
off value, the physician should not treat. By contrast,
a higher value should lead to treatment. A lower opti-
mal cutoff then implies earlier treatment. The intui-
tion for the rank order shown in (12) is the same as
in the last two sections: risk aversion increases the
utility gain in the sick state and decreases the utility
loss in the healthy state, leading to earlier treatment.
Given a background risk, risk vulnerability strength-
ens this effect, causing a further decrease in the opti-
mal cutoff value.
CLINICAL EXAMPLE
For illustrative purposes, we take a clinical exam-
ple from Sox and others.15 A physician examines
a 55-year-old man with a headache and progressive
unilateral weakness. The physician suspects a brain
tumor, which can be treated with brain surgery. For
simplicity, we consider life expectancy LE to be the
only outcome parameter and ignore quality of life
(which was considered in the original example).
This does not lessen the significance of the analysis,
as we can assume decreasing marginal utility of
ð11Þ LE.16 If he is healthy, the patient can expect to live
another 21 years. Unnecessary brain surgery will
reduce his remaining LE to 20 years. If he has a brain
tumor and remains untreated, he can expect to live for
only 2 more years, while brain surgery in this case
will increase his remaining LE to 11 years.
First, we calculate the therapeutic threshold for
DMs with different risk preferences. The utility func-
tion of a risk-neutral DM is linear. To compare utility
values for risk-neutral and risk-averse (risk vulnera-
ble) DMs, we divide LE by 21. The utility gain
G and utility loss L from treatment can then be
expressed as GN ¼ ð11  2Þ=21 ¼ 0:429 and LN ¼
ð21  20Þ=21 ¼ 0:048. Therefore, the therapeutic
threshold for a risk-neutral DM will be p~N ¼ LN =
ðGN þ LN Þ ¼10:00%. Thus, in this situation, without
further diagnostics, brain surgery is the optimal strat-
egy for a risk-neutral DM, provided that the a priori
probability of the patient having a brain tumor is at
least 10%.
Sox and others15 report the following utilities for
life expectancies, derived by the standard gamble
method: U ð21Þ ¼ 1; U ð2Þ ¼ 0:25; U ð0Þ ¼ 0. We
approximated these utility levels by the utility func-
tion U ðLEÞ ¼ LE0:6 6:212, which represents a mixed
risk-averse DM who is also risk vulnerable. The util-
ð12Þ ities of the remaining life expectancies become
 RISK PREFERENCES AND TEST AND TREATMENT THRESHOLDS
U ð21Þ ¼ 1:000, U ð20Þ ¼ 0:971, U ð11Þ ¼ 0:679, and
U ð2Þ ¼ 0:244 (note that U ð2Þ differs slightly from
0.25). The utility gain G and the utility loss L from
treatment change to GA ¼ 0:679  0:244 ¼ 0:435 and
LA ¼ 1:000  0:971 ¼ 0:029, leading to a therapeutic
threshold of 6.23% (p~A ¼ LA =ðGA þ LA Þ ¼ 0:029=
ð0:435 þ 0:029Þ). The risk-averse DM would therefore
prescribe brain surgery at a much lower a priori prob-
ability. He or she values the utility gain to utility loss
ratio G=L more highly than the risk-neutral DM by
a factor 1.67 (GA =LA ¼ 15:1; GN =LN ¼ 9:0).
Next, we assume that the patient’s health status
depends not only on the presence or absence of the
primary illness but also on the presence or absence
of a comorbidity. We add a symmetric zero-mean
comorbidity risk E½m ~  ¼ 0, where m ~ can take on the
values 1 2 and –2 with 50% probability each. Hence,
the patient’s life expectancy in the healthy states
(sick states) becomes 23 or 19 (4 or 0) with equal prob-
ability if he is not treated and 22 or 18 (13 or 9) if he is
treated.
For a risk-neutral DM, the additional comorbidity
risk changes nothing, as the expected utility gain G
and expected loss L of treatment remain at 0.429
(GN ¼ 0:5  ð13 þ 9  4  0Þ=21) and at 0.048
ðLN ¼ 0:5  ð23 þ 19  22  18Þ=21Þ. For the risk-
vulnerable DM,
 EGV increases  to 0.491
’ð0:5=6:212Þ  130:6 þ 90:6  40:6  00:6 and EGV
increases to 0.02893 ½’ð0:5=6:212Þ  230:6 þ 190:6 
220:6  180:6 Þ, which results in a higher utility gain
to utility loss ratio EGV =ELV ¼ 17:0 and thus in a fur-
ther decrease in the therapeutic threshold to
p~V ¼ 5:57%.
Finally, we take into account that the physician
can use a computed tomography (CT) scan with a sen-
sitivity of 95% and a specificity of 97% for further
diagnostics. The CT produces false-positive and
false-negative results, which must be considered
before using the test. The likelihood ratios for the
CT scan amount to LRþ ¼ 31:67 and LR ¼ 0:05.
Using for the test and treatment thresholds and the
utility gain to utility loss ratios from treatment
reported above, we obtain the following values:
p~Dx
N ¼ 0:35%, p ~Dx
A ¼ 0:21%, p ~Dx
V ¼ 0:19% and p ~Rx
N ¼
Rx Rx
68:31%, p~A ¼ 56:30%, p~V ¼ 53:34%. We note that
risk aversion in particular has a strong effect on the
therapeutic threshold as compared with risk neutral-
ity. Moreover, recent empirical studies show that
higher-order risk preferences can have a sizable
impact on decisions in general (see Ebert and Wie-
sen17) and on medical treatment decisions in particu-
lar (see Krieger and Mayrhofer18).
ORIGINAL ARTICLE
Downloaded from mdm.sagepub.com by guest on May 6, 2015
DISCUSSION AND CONCLUSION
In this article, we analyze the impact of risk prefer-
ences on decisions over medical testing and treat-
ment. In particular, we investigate the effect of
a comorbidity risk in the presence of higher-order
risk attitudes. A risk-averse DM values the health
gain from treatment in the sick state more and the
health loss from treatment in the healthy state less
than a risk-neutral DM. When facing an exogenous
comorbidity risk, the utility gain-loss ratio for
a risk-vulnerable DM will increase even further.
Risk attitudes affect the test and treatment thresh-
olds. Both thresholds decrease with an increasing
utility gain-loss ratio of treatment, leading to both
testing and treatment at lower a priori probabilities
of illness. A risk-averse DM will thus test and treat
at lower probabilities than a risk-neutral DM. Facing
a comorbidity risk, a risk-vulnerable DM will test and
treat even earlier.
Corresponding results apply for endogenous test
results. Since a risk-averse DM realizes a higher util-
ity gain-loss ratio of treatment, he or she will lower
the test’s cutoff value to increase sensitivity at the
expense of specificity. In the presence of a comorbid-
ity risk, risk vulnerability strengthens this effect,
causing an even further decrease in the optimal
cutoff.
The threshold analysis is particularly relevant in
the low-prevalence setting of screening programs. If
DMs are risk averse (and risk vulnerable), they will
realize a larger utility gain from treatment in the
sick state and a smaller utility loss in the healthy
state. They will therefore test at lower a priori proba-
bilities than if they were risk neutral. Risk aversion
and higher-order risk preferences might thus explain
some of the observed excess testing (e.g., for breast
cancer19 and prostate cancer20). Of course, prospect
theory21,22 and defensive medicine23 are other possi-
ble explanations.
The commonly used utility functions with CRRA
(e.g., logarithmic and root functions) imply mixed
risk aversion—that is, have derivatives with progres-
sively alternating signs6,7 and feature risk vulnerabil-
ity. Constant absolute risk aversion (CARA) utility
functions (i.e., exponential utility functions) also
exhibit mixed risk aversion but do not show risk vul-
nerability, implying that a DM with a CARA utility
function does not react to the zero-mean comorbidity
risk. However, although exponential utility functions
are often more convenient analytically than
CRRA functions, they are usually considered less
39
 FELDER, MAYRHOFER
plausible.24 Interestingly, assuming that the comor-
bidity risk applies only to the sick state, the only
assumption necessary for EG=EL > G=L to hold is
that the DM be prudent—in this case, even DMs
with CARA utility functions will appear to be more
risk averse.25
We model health as a continuous variable,
although it clearly has multidimensional aspects.
The theory of multivariate risk aversion, which takes
into account that utility is a function of many varia-
bles (e.g., Duncan26), could be applied to the health
setting. However, since there are thousands of possi-
ble combinations of health variables, it is feasible to
think of overall H as being continuous.
Recent experimental evidence confirms risk-averse
and prudent behavior among individuals facing uncer-
tainty.17,27,28 Empirical results for temperance are
ambiguous so far. While Deck and Schlesinger27 find
(weakly) intemperate behavior, studies by Ebert and
Wiesen17 and Noussair and others29 point in the other
direction, in line with our theoretical assumptions.
Ebert and Wiesen17 also measure the intensity of
higher-order risk preferences and find that the com-
pensation subjects are willing to pay to avoid third-
order risk is significantly higher than for second-
order risk, emphasizing the importance of measuring
higher-order risk preferences in general.
Krieger and Mayrhofer18 study whether individu-
als in an experimental laboratory display risk-averse
and prudent behavior in medically framed treatment
decisions, where DMs’ payoffs are monetary. Their
theoretical approach is similar to the one presented
in this article. Their experiment provides evidence
for risk-averse and prudent behavior. Such risk pref-
erences have a considerable impact on the therapeu-
tic threshold, leading to its reduction by 41% relative
to the risk-neutral position. Risk aversion accounts
for three-fourths of this effect and prudent behavior
for one-fourth. Furthermore, they find that the medi-
cal framing of the decision as compared with a neutral
framing does not affect second-order risk preferences
but is associated with more frequent and stronger
prudent behavior. Since—assuming that the comor-
bidity risk applies only to the sick state—the impact
of higher-order risk preferences on test and treatment
thresholds as well as on optimal cutoffs is analogous
to the therapeutic threshold, the results of Krieger and
Mayrhofer illustrate the potential importance of mea-
suring risk preferences, particularly risk aversion and
prudence, for medical treatment recommendations.
This article has at least 3 limitations. First, our
model is rooted in the expected utility framework,
which has been criticized for its insufficient capacity
40  MEDICAL DECISION MAKING/JANUARY 2014
Downloaded from mdm.sagepub.com by guest on May 6, 2015
to predict individual behavior, especially with regard
to possible probability weighting. Alternatives to
expected utility theory, such as rank-dependent
choice models, have been shown to be more in line
with actual behavior.30,31 Nonetheless, recent stud-
ies, such as the one by List,32 challenge the validity
of rank-dependent theory. They claim that experi-
enced individuals behave largely in accordance
with expected utility theory. Since physicians rou-
tinely decide on medical tests and treatments, they
should be able to make unbiased estimations of a pri-
ori probabilities of illnesses and thus make coherent
treatment decisions. Moreover, this study’s nature
is normative, and ‘‘expected utility is the best theory
to determine which decisions to take.’’33(p697)
The second limitation is the restriction to diagnos-
tic risk. In general, DMs face not only diagnostic risk
but also therapeutic risk, which refers to uncertainty
over the treatment outcome. Eeckhoudt34 shows that
risk aversion increases the critical probability of suc-
cessful treatment in this case. The intuition here is
that treatment is a risk-increasing action, so a risk-
averse DM treats later than a risk-neutral one. There-
fore, the effects of higher-order risk preferences on the
thresholds may cancel each other out when diagnostic
and treatment risk are analyzed simultaneously.
Bleichrodt and others35 analyze the effect of
comorbidities on the treatment decision in the con-
text of therapeutic risk. They mirror the QALY model
and assume that treatment only affects quality of life,
that comorbidity only has an impact on life duration,
and that patients can influence treatment intensity.
They find that comorbidities do not affect the treat-
ment decision, since the prudence premium3 can be
interpreted as the DM’s risk premium for a longer
life. However, in contrast to Eeckhoudt’s34 and our
models, the background risk in Bleichrodt and others
is imposed on a second dimension (i.e., life duration)
rather than on the patient’s health level.
The third limitation concerns our rather abstract
clinical example, which should be regarded as
merely illustrative. Further empirical and experi-
mental research is needed to produce reliable infor-
mation on higher-order risk preferences regarding
medical decision making.
Pure health outcomes such as mortality or survival
rates are often used instead of utility measures. This
may lead to unreasonably high test and treatment
thresholds and cutoff values. The use of QALYs
instead would reduce this bias, as QALYs reflect
patient preferences over health statuses when they
are calculated by methods grounded in the expected
utility theory and using the standard gamble method,
 RISK PREFERENCES AND TEST AND TREATMENT THRESHOLDS
which reflects higher-order risk preferences if pres-
ent. However, often only the primary illness is taken
into account when eliciting QALYs but not the possi-
bility of comorbid conditions.
We conclude that medical decisions should be
based on the accurate measurement of preferences
in an extended standard gamble approach. Our
results are relevant for clinical guidelines. In particu-
lar, our findings suggest that diagnostic tests in low-
prevalence settings (e.g., screening) may appear
more beneficial if (higher-order) risk preferences are
taken into account.
ACKNOWLEDGMENTS
We are grateful to the associate editor and 2 anonymous ref-
erees for most helpful comments regarding not only spe-
cific points but also the organization and presentation of
the material. We also thank Miriam Krieger and the partic-
ipants of the 13th Biennial European Conference of the
Society for Medical Decision Making in Hall in Tirol, Aus-
tria (2010), for their remarks.
REFERENCES
1. Pauker SG, Kassirer JP. Therapeutic decision making: a cost
benefit analysis. N Engl J Med. 1975;293(5):229–34.
2. Pauker SG, Kassirer JP. The threshold approach to clinical deci-
sion making. N Engl J Med. 1980;302(20):1109–17.
3. Kimball MS. Precautionary saving in the small and in the large.
Econometrica. 1990;58(1):53–73.
4. Kimball MS. Precautionary motives for holding assets. In: New-
man P, Milgate M, Falwell J, eds. New Palgrave Dictionary of
Money and Finance. London: Macmillan; 1992. p. 158–61.
5. Gollier C, Pratt JW. Risk Vulnerability and the tempering effect
of background risk. Econometrica. 1996;64(5):1109–23.
6. Brocket PL, Golden LL. A class of utility functions containing
all the common utility functions. Manage Sci. 1987;33(8):955–64.
7. Caballé J, Pomansky A. Mixed risk aversion. J Econ Theory.
1996;71:485–513.
8. Eeckhoudt L, Schlesinger H. Putting risk in its proper place. Am
Econ Rev. 2006;96:280–9.
9. Arrow KJ. Essays in the Theory of Risk-bearing. London: North-
Holland; 1971.
10. Pratt J. Risk Aversion in the small and in the large. Econometr-
ica. 1964;32:122–36.
11. Eeckhoudt L, Gollier C, Schlesinger H. Changes in background
risk and risk taking behavior. Econometrica. 1996;64(3):683–9.
12. McNeil BJ, Keeler M, Adelstein SM. Primer on certain ele-
ments of medical decision making. N Engl J Med. 1975;293(5):
211–5.
13. Metz CE. Basic principles of ROC analysis. Semin Nucl Med.
1978;8(4):283–98.
ORIGINAL ARTICLE
Downloaded from mdm.sagepub.com by guest on May 6, 2015
14. Felder S, Werblow A, Robra BP. A-priori risk and optimal test
accuracy in prenatal diagnostics. Med Decis Making. 2003;23(5):
406–13.
15. Sox HS, Blatt MA, Higgins MC, Marton KI. Medical Decision
Making. Philadelphia, PA: American College of Physicians; 2007.
Reprint, Boston: Butterworth-Heinemann; 1988.
16. Pratt JW, Zeckhauser RJ. Willingness to pay and the distribu-
tion of risk and wealth. J Polit Econ. 1996;104(4):747–63.
17. Ebert S, Wiesen D. Joint measurement of risk aversion, pru-
dence and temperance. J Risk Uncertain. Forthcoming 2013.
18. Krieger M, Mayrhofer T. Patient Preferences and Treatment
Thresholds under Diagnostic Risk: An Economic Laboratory
Experiment. Ruhr Economic Papers No. 321. Essen, Germany:
RWI; 2012.
19. Breen N, Gentleman JF, Schiller JS. Update on mammography
trends: comparisons of rates in 2000, 2005, and 2008. Cancer. 2011;
117(10):2209–18.
20. Farwell WR, Linder JA, Jha AK. Trends in prostate-specific
antigen testing from 1995 through 2004. Arch Intern Med. 2007;
167(22):2497–502.
21. Kahneman D, Tversky A. Prospect theory: an analysis of deci-
sion under risk. Econometrica. 1979;47(2):263–91.
22. Tversky A, Kahneman D. Advances in prospect theory: cumu-
lative representation of uncertainty. J Risk Uncertainty. 1992;5(4):
297–323.
23. DeKay ML, Asch DA. Is the defensive use of diagnostic tests
good for patients, or bad? Med Decis Making. 1998;18(1):19–28.
24. Friend I, Blume ME. The demand for risky assets. Am Econ
Rev. 1975;65(5):900–922.
25. Felder S, Mayrhofer T. Medical Decision Making—A Health
Economic Primer. Berlin, Germany: Springer; 2011.
26. Duncan GT. A matrix measure of local multivariate risk aver-
sion. Econometrica. 1977;45(4):895–903.
27. Deck C, Schlesinger H. Exploring higher order risk effects. Rev
Econ Stud. 2010;77(4):1403–20.
28. Ebert S, Wiesen D. Testing for prudence and skewness seeking.
Manage Sci. 2010;57(7):1334–49.
29. Noussair CN, Trautmann ST, van de Kuilen G. Higher order
risk preferences, demographics, and financial decisions. 2011.
Available from: http://ideas.repec.org/p/dgr/kubcen/2011055.
html
30. Abdellaoui M. Parameter-free elicitation of utility and proba-
bility weighting functions. Manage Sci. 2000;46(11):1497–512.
31. Bleichrodt H, Pinto JL. A parameter-free elicitation of the prob-
ability weighting function in medical decision analysis. Manage
Sci. 2000;46(11):1485–96.
32. List J. Neoclassical theory versus prospect theory: evidence
from the market place. Econometrica. 2004;72(2):615–25.
33. Wakker PP. Lessons Learned by (from?) an economist working
in medical decision making. Med Decis Making. 2008;28(5):690–8.
34. Eeckhoudt L. Risk and Medical Decision Making. Boston:
Kluwer; 2002.
35. Bleichrodt H, Crainich D, Eeckhoudt L. The effect of comorbid-
ities on treatment decisions. J Health Econ. 2003;22:805–20.
41