Heuristic decision making in medicine.

2012·1.2 MB·English

by  Marewski, Julian N.

PAGES_12_AG_1006_BA.qxd:DCNS#52 10/03/12 12:46 Page 77
C l i n i c a l r e s e a r c h
Heuristic decision making in medicine
Julian N. Marewski, PhD; Gerd Gigerenzer, PhD
Introduction
A
ccompanied by his anxious wife, a middle-aged
male patient arrives at a rural Michigan hospital. He suf-
fers from serious chest pain. The physician in charge, a
compassionate-looking woman, suspects acute ischemic
heart disease, but is not entirely sure. Should she assign
the patient to a regular nursing bed for monitoring? If it
really is acute ischemic heart disease, however, the patient
Can less information be more helpful when it comes to needs to be rushed immediately to the coronary care unit.
making medical decisions? Contrary to the common On the other hand, unwarrantedly sending the patient to
intuition that more information is always better, the the care unit is not only expensive, but can also decrease
use of heuristics can help both physicians and patients the quality of care for those patients who need it, while
to make sound decisions. Heuristics are simple decision those who do not are exposed to the risk of catching a
strategies that ignore part of the available information, potentially harmful, hospital-transmitted infection.
basing decisions on only a few relevant predictors. We How humans can solve this, and related complex deci-
discuss: (i) how doctors and patients use heuristics; and sion-making dilemmas in the medical world, is the cen-
(ii) when heuristics outperform information-greedy tral topic of this review article. For the emergency-room
methods, such as regressions in medical diagnosis. situation outlined above, there are different approaches
Furthermore, we outline those features of heuristics to tackling the problem, rooted in different traditions in
that make them useful in health care settings. These the decision sciences.
features include their surprising accuracy, transparency,
Keywords: medical decision making; fast-and-frugal heuristics; decision aids;
and wide accessibility, as well as the low costs and little biases; ecological rationality; bounded rationality
time required to employ them. We close by explaining
Author affiliations:University of Lausanne, Lausanne, Switzerland (Julian N.
one of the statistical reasons why heuristics are accu- Marewski); Max Planck Institute for Human Development, Berlin, Germany
rate, and by pointing to psychiatry as one area for (Gerd Gigerenzer)
future research on heuristics in health care. Address for correspondence:Julian N. Marewski, University of Lausanne, Faculty of
© 2012, LLS SAS Dialogues Clin Neurosci.2012;14:77-89. Business and Economics, Department of Organizational Behavior, Quartier UNIL-
Dorigny, Bâtiment Internef, Office 601, 1015 Lausanne, Switzerland
(e-mail: [email protected])
Copyright © 2012 LLS SAS. All rights reserved 77 www.dialogues-cns.org
PAGES_12_AG_1006_BA.qxd:DCNS#52 10/03/12 12:46 Page 78
C l i n i c a l r e s e a r c h
The first one is to leave all responsibility to the doctors. They introduced the Heart Disease Predictive Instrument,
Yet, in an actual rural Michigan hospital under study, which consists of a chart with some 50 probabilities and
doctors sent 90% of patients with severe chest pain to a logistic regression that enables the physician, with the
the coronary care unit; as a consequence, it became over- help of a pocket calculator, to compute the probability
crowded, quality of care decreased, and costs went up. that the patient should be admitted to the coronary care
The second approach is to try to solve the complex prob- unit. However, few physicians understand logistic regres-
lem with a complex algorithm. This is what a team of sions, and charts and calculators tend to be dropped the
medical researchers from the University of Michigan did. moment the researchers leave the hospital.
The third approach consists of teaching physicians effec-
tive heuristics. A heuristic is a simple decision strategy that
ignores part of the available information and focuses on
the few relevant predictors. Green and Mehr1developed
ST-segment
changes? one such heuristic for treatment allocation. This so-called
fast-and-frugal treeignores all probabilities and asks only
a few yes-or-no questions (Figure 1). Specifically, if a cer-
no yes tain anomaly appears in the patient's electrocardiogram
(ie, an ST-segment change), the patient is immediately
sent to the coronary care unit. No other information is
Chief complaint of Coronary considered. If there is no anomaly, a second variable is
care unit
chest pain? taken into account, namely whether the patient’s primary
complaint is chest pain. If not, the patient is classified as
low risk, and assigned to a regular nursing bed. Again, no
yes no additional information is considered. If the answer is yes,
a third and final question is asked to classify the patient.
Regular Can following such a simple heuristic enable doctors to
Any one other factor? nursing make good allocation decisions? Figure 2shows the per-
(NTG, MI, ST , ST , T) bed
formance of all three approaches in their ability to pre-
dict heart attacks in the Michigan hospital. As can be
seen, the heuristic approach resulted in a larger sensi-
no yes
tivity (proportion of patients correctly assigned to the
coronary care unit) and a lower false-positive rate (pro-
Regular Coronary portion of patients incorrectly assigned to the coronary
nursing care unit care unit) than both the Heart Disease Predictive
bed
Instrument and the physicians. The heuristic approach
achieved this surprising level of performance by consid-
ering only a fraction of the information that the Heart
Figure 1. A simple heuristic for deciding whether a patient should be
Disease Predictive Instrument used.
assigned to the coronary care unit or to a regular nursing bed.
If there is a certain anomaly in the electrocardiogram (the so-
called ST segment) the patient is immediately sent to the Views on rationality: from unbounded
coronary care unit. Otherwise a second predictor is consid-
rationality and irrationality to ecological
ered, namely whether the patient’s chief complaint is chest
rationality
pain. If not, a third question is asked. This third question is
a composite one: whether any of five other predictors is pre-
sent. This type of heuristic is also called a fast-and-frugal tree. What to diagnose, whom to treat, what to eat, or which
Fast-and-frugal trees assume that decision makers follow a
stocks to invest in—our days are filled with decisions, yet
series of sequential steps prior to reaching a decision.
Abbreviations: NTG, nitroglycerin; MI, myocardial infarction; how do we make them, and how should we make them?
T, T-waves with peaking or inversion In the decision sciences and beyond, the answer to these
Adapted from ref 58 (based on Green and Mehr)1: Gigerenzer G. Gut two questions depends on one’s view of human ratio-
feelings: the Intelligence of the Unconscious. New York, NY: Viking
Press; 2007. Copyright © Viking Press 2007 nality. There are at least three views.
78
PAGES_12_AG_1006_BA.qxd:DCNS#52 10/03/12 12:46 Page 79
Heuristic decision making in medicine - Marewski and Gigerenzer Dialogues in Clinical Neuroscience - Vol 14 .No. 1 .2012
Unbounded rationality: optimization unboundedly rational optimization, refers to models that
do not assume full knowledge but take into account con-
The study of unbounded rationalityasks the question, if straints, such as information costs. Optimization models
people were omniscient, that is, if they could compute are common in fields such as economics or computer sci-
the future from what they know, how would they behave ence. The spirit of optimization is also reflected in the
and how should they behave? Optimization models such workings of the Heart Disease Predictive Instrument,
as Bayesian inference and the maximization of subjec- which is a linear regression model that computes opti-
tive expected utility take this view.2When judging, for mal beta weights.
instance, whom to treat, these models assume that deci-
sion makers will collect and evaluate all information, Irrationality: cognitive illusions and biases
weight each piece of it according to some criterion, and
then combine the pieces to maximize the chances of According to the second view, human reasoning is not
attaining their goals (eg, treating the needy while saving characterized by optimization but by systematic devia-
costs). Optimization under constraints, a sub-branch of tions from optimization, also called cognitive illusions,
errors, or simply irrationality. The heuristics-and-biases
framework3 proposes that humans commit systematic
nit 1 errors when judging probabilities and making decisions.
u
y care .9 Amilzthaotiuognh v tihewis, firt asmtilel wtaokreks doipffteimrs itzhaetiroenin— fsruocmh tahse m oapxtii--
ar
on .8 mization of expected utility—as the normative yardstick
or against which to evaluate human decision making.
c
he .7 Decisions that deviate from this standard can be expli-
d to t .6 cated by assuming that people suffer from cognitive lim-
e itations, such as a suboptimal information processing
n
g
assi .5 capacity or insufficient knowledge. Following this view,
y one might argue that the physicians’ large false-positive
ectl .4 rate and below-chance performance in making alloca-
corr tion decisions (Figure 2) reflect the workings of their
nts .3 limited cognitive abilities.
patie .2 HPheyasritc iDainssease Predictive Instrument
of Ecological rationality: fast and frugal heuristics
on .1 Fast-and-frugal tree
orti There is, however, an alternative to optimization and
op .0
Pr .0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 irrationality. A couple of thousand journal articles (and
Proportion of patients incorrectly assigned to the coronary care unit several years) before the heuristics-and-biases tradition
became popular, Herbert Simon, the father of what is
known as the bounded rationalityview, stressed that opti-
Figure 2. The performance of a decision tree for coronary care unit allo-
cations, compared with that of the Heart Disease Predictive mization is rarely possible in the real world, and thus a
Instrument, and physicians’ judgments. The x-axis represents theory of rationality needs to study how people make
the proportion of patients who were incorrectly assigned to
decisions when optimization is out of reach.4Instead of
the coronary care unit (false positive rate), and the y-axis
shows the proportion of patients who were correctly assigned relying on unrealistic optimization models and striving
to the coronary care unit (sensitivity). The diagonal line rep- to compute optimal solutions for a given task, so he
resents chance level, the area to the left of the diagonal bet- argued, people use simple strategies, seeking solutions
ter-than-chance. Note that the Heart Disease Predictive
that are good enough with respect to an organism’s
Instrument’s allocation decisions depend on how sensitivity is
traded off against the false-positive rate. This is why several goals. He also stressed that behavior and performance
data points are shown for this instrument. result from both cognition and an organism’s environ-
Adapted from ref 58 (based on Green and Mehr)1: Gigerenzer G. Gut ment (Box 1): “Human rational behavior … is shaped by
feelings: the Intelligence of the Unconscious. New York, NY: Viking
Press; 2007. Copyright © Viking Press 2007 a scissors whose two blades are the structure of task
79
PAGES_12_AG_1006_BA.qxd:DCNS#52 10/03/12 12:46 Page 80
C l i n i c a l r e s e a r c h
environments and the computational capabilities of the even at the risk of causing unnecessary harm. This is
actor” (p 7).5 exactly what a vast majority of US physicians seem to
Embracing this emphasis on simple decision strategies do: 93% of over 800 surgeons, obstetricians, and other
and their fit to the environment, the fast-and-frugal specialists at high risk of litigation reported practices
heuristics framework6,7has developed an ecological view of recommending a diagnostic test or treatment that is
of rationality through which it tries to understand how not the best option for the patient, but one that pro-
and whenpeople’s reliance on simple decision heuristics tects the physician against the patient as a potential
can result in smart behavior. In this view, heuristics can plaintiff, including, for instance, unnecessary CT scans,
be ecologically rationalwith respect to the environment biopsies, and MRIs, and more antibiotics than med-
and the goals of the actor. Here, being rational means ically indicated.10Similarly, in the rural Michigan hos-
that a heuristic is successful with regard to some outside pital discussed above, of about 90% of the patients
criterion, such as making a decision accurately and who were referred to the coronary care unit, only
quickly when a patient is rushed into the emergency roughly 25% actually had a myocardial infarction. In
room. Hammond8 called such outside criteria corre- environments where risk of being sued is high if a
spondence criteria, as opposed to coherence criteria, patient is mistakenly diagnosed and/or treated as
which are based on unboundedly rational optimization healthy and where physicians seek to avoid potential
models as a normative yardstick for rationality. lawsuits, it is ecologically rational for them to follow
For instance, while physicians’ decisions in Figure 2 the defensive heuristic “err on the safe side,” being
appear to be systematically biased towards mistakenly overcautious and prescribing more diagnostic tests and
assigning healthy patients to the coronary care unit, treatments than necessary. This defensive heuristic is
these decisions might in fact be viewed as ecologically not the same as an irrational reasoning error or a cog-
rational, as the following court trial illustrates. In 2003, nitive illusion, caused by people’s mental limitations.
Daniel Merenstein,9 a family physician in Virginia, But precisely because of this, as we will discuss next,
USA, was sued because he had informed a patient there is room for change: by changing the environ-
about the pros and cons of PSA (prostate-specific anti- ment, physicians can be led to rely on heuristics that
gen) tests, instead of just ordering one. Given that are more beneficial to the patient.
there is no evidence that the test does more good than
harm, he had followed the recommendations of lead- The science of fast-and-frugal heuristics
ing medical organizations and informed his patient,
upon which the man declined to take the test. The Doctors and other humans cannot foresee the future,
patient later developed an incurable form of prostate and cannot know if a diagnosis is correct for certain, or
cancer, and Merenstein was sued. The jury at the court if a treatment will cure a patient for certain. Rather, they
exonerated him, but found his residency liable for $1 have to make decisions under uncertainty and often
million. After that, Merenstein felt he had no choice under the constraints of limited time. According to the
other than to overdiagnose and overtreat patients, fast-and-frugal heuristics research program, these deci-
sions can nevertheless be made successfully, because
Box 1 people can rely on a large repertoire of heuristics—an
In the literature, a connection between the heuristics- adaptive toolbox—with each heuristic (ie, each tool)
and-biases view and Simon’s concept of bounded being adapted to a specific decision-making environ-
rationality is often invoked. However, although ment. By relying on a heuristic that is well adapted to a
Kahneman et al3 credited Simon in the preface to particular environment, a person can make sound deci-
their anthology (“Judgment under uncertainty: sions, often based on very little information in little time
heuristics and biases”), their major early papers, (hence “fast-and-frugal”).
which appear in the same volume, do not cite Simon’s There are different sets of mechanisms that help people
work on bounded rationality. Thus, the connection to choose among the heuristics. The first depends on the
between heuristics-and-biases and bounded rational- workings of basic cognitive capacities, such as memory.11
ity was possibly made in hindsight.61 The interplay of these capacities with the environment
creates for each heuristic a cognitive nichein which it
80
PAGES_12_AG_1006_BA.qxd:DCNS#52 10/03/12 12:46 Page 81
Heuristic decision making in medicine - Marewski and Gigerenzer Dialogues in Clinical Neuroscience - Vol 14 .No. 1 .2012
can be applied. For instance, the frequency and recency Heuristics in health care?
with which we have encountered information in our
environment influences what information we remem- Although the science of fast-and-frugal heuristics has
ber, and how quickly we remember it. What information started to make an impact in the medical community,22
comes to the mental stage, and how quickly it arrives the heuristics-and-biases perspective still dominates as
there, in turn determines what heuristics are applicable of today.23For instance, Elstein24refers to heuristics as
to solve a given task. A second set of mechanisms for “mental shortcuts commonly used in decision making
selecting heuristics includes social and individual learn- that can lead to faulty reasoning or conclusions” (p 791),
ing processes that can make people more prone to citing them as a source of many errors in clinical rea-
choose one applicable heuristic over another.12 soning.
Importantly, by changing the environment people can Some medical researchers, however, recognize the
be led to rely on different heuristics. For instance, in potential of fast-and-frugal heuristics to improve deci-
environments with a lower risk of being sued, doctors sions. For example, as McDonald25writes, “admitting the
may rely on different medical heuristics. In Switzerland, role of heuristics confers no shame” (p 56). Rather, the
where litigation is less common, only 41% of general goal should be to formalize and understand heuristics so
practitioners and 43% of internists reported that they that their use can be effectively taught, which could lead
sometimes or often recommend a PSA test for legal to less practice variation and more efficient medical care.
reasons.13 Similarly, Elwyn et al26state that “The next frontier will
involve fast-and-frugal heuristics; rules for patients and
Past research on fast-and-frugal heuristics clinicians alike” (p 574). In what follows, we will discuss
different ways in which the study of heuristics can
The heuristics in the adaptive toolbox can be classified inform medical decision making.
along several nonexclusive categories. These categories
include: (i) how the heuristic processes information (eg, How practitioners and patients make decisions
assigning different importance to different predictor
variables by ordering them sequentially, as in Figure 1); In medical decision making and beyond, the science of
(ii) whether the heuristic is applicable to the social fast-and-frugal heuristics focuses on at least three main
domain (eg, to doctor-patient interactions or bargaining questions. The first question is descriptive: what heuris-
at the bazaar); (iii) whether the heuristic is a model of tics do doctors, patients, and other stakeholders use to
inductive inference about unknown quantities and make decisions? The second question is closely interre-
future events (eg, in medical diagnosis or weather fore- lated with the first one, and deals with ecological ratio-
casting); or (iv) whether the heuristic represents a model nality: to what environmental structures is a given
for decisions that are based exclusively on the contents heuristic adapted—that is, in which environments does
of one’s memories (eg, in quiz shows or under time pres- it perform well, and in which does it not? The third ques-
sure in a medical emergency). tion focuses on practical applications: how can the study
Corresponding models of heuristics have been studied of people’s repertoire of heuristics and their fit to envi-
in diverse domains, including applied ones, such as ronmental structures aid decision making?
enforcing proenvironmental behaviour or forecasting Let us begin with the descriptive question of how prac-
customers’ activities in business, as well as in the basic titioners and patients make decisions. Here, fast-and-
sciences, ranging from animal behavior to the law, frugal heuristics differ from traditional, information-
finance, or psychology.14,15At the same time, a number of greedy models of medical decision making, such as
heuristics for very different tasks have been proposed: expected utility maximization, Bayesian inference, or
heuristics for mate search,16inferences about politicians,17 logistic regression.
and choices between risky alternatives,18to name a few. How physicians make diagnostic decisions is potentially
In the applied world, heuristics have been used to pre- modelled by fast-and-frugal trees, a branch of heuristics
dict, for example, the performance of stocks,19the out- that assumes decision makers to follow a series of
comes of sports competitions,20or the results of political sequential steps prior to reaching a decision. Such trees
elections.21 ask only a few yes-or-no questions and allow for a deci-
81
PAGES_12_AG_1006_BA.qxd:DCNS#52 10/03/12 12:46 Page 82
C l i n i c a l r e s e a r c h
sion after each one. Like most other heuristics, fast-and- medication.29Fast-and-frugal trees, rather than full deci-
frugal trees are built around three rules; one that speci- sion trees, are also routinely used in HIV testing and
fies in what direction information search extends in the cancer screening,30and have been identified as descrip-
search space (search rule); one that specifies when infor- tive models of behavior in other areas beyond medicine,
mation search is stopped (stopping rule), and one that including the law.31
specifies how the final decision is made (decision rule). What about the patients? Even patients with higher edu-
In their general form, fast-and-frugal trees can be sum- cation often rely on a simple heuristic when it comes to
marized as follows: their own health, even when it contradicts their acade-
(cid:129)Search rule: Look up predictors in the order of their mic viewpoint. For instance, although most economists
importance. subscribe to neoclassical theories of unboundedly ratio-
(cid:129)Stopping rule: Stop search as soon as one predictor nal models and advocate weighing all pros and cons of
variable allows it. alternatives in their research, when surveyed about their
(cid:129)Decision rule: Classify according to this predictor vari- own real-life decisionsabout whether to participate in
able. PSA screening, 66% of more than 100 American econo-
Fast-and-frugal trees are characterized by the limited mists said that they had not weighed any pros and cons
number of exitsthey have; only a few predictors can be of PSA screening, but simply trusted their doctor’s
looked up, but they will always lead to a decision. For advice. They presumably followed the heuristic ‘‘If you
instance, the heuristic shown in Figure 1represents one see a white coat, trust it.” Another 7% indicated that
such fast-and-frugal tree with four exits. Specifically, a their wives or relatives had influenced their decision.32
fast-and-frugal tree has n+ 1 exits, where nis the num- The simple social heuristic “trust your doctor” is eco-
ber of binary predictor variables. In comparison, more logically rational in environments where physicians
information-greedy approaches have many more exits; understand health statistics, do not rely on defensive
Bayes’ rule, for example, can be represented as a tree decision heuristics for fear of litigation, and have no con-
with 2n exits. Contrary to more information-greedy flicts of interest, such as earning money, a free dinner, or
approaches, fast-and-frugal trees make themselves effi- another kind of gratification for prescribing certain med-
cient by introducing order—which predictors are the ications or for using certain diagnostic techniques. Yet,
most important ones?—making themselves efficient. in the American health care system, where none of these
A number of fast-and-frugal trees have been identified factors holds, reliance on this heuristic can become
as potential descriptive models of behavior. Dhami and potentially maladaptive.
Harries,27for example, compared a fast-and-frugal tree
to a regression model on general practitioners' decisions Saving lives by changing the environment
to prescribe lipid-lowering drugs for hypothetical
patients. Both models fitted the prescriptions equally Not only in the United States, but also in other countries,
well (but see Box 2). Similar results were obtained by can changing health care environments pay off, and
Backlund et al28for judgments regarding drug treatment sometimes even save lives. Consider the following exam-
of hyperlipidemia as well as for diagnosing heart failure, ple. Numerous Germans and Americans die each year
and by Smith and Gilhooly for describing antidepressant while waiting for an organ donor.33Even though expen-
sive advertising campaigns are conducted to promote
Box 2 organ donation, relatively few citizens sign a donor card:
A heuristic’s ability to account for behavioral data according to Johnson and Goldstein,34a study published
should not only be tested by assessing its fitto those in 2003, about 12% in Germany and 28% in the US. In
data, with fit meaning that relevant parameters can contrast, about 99.9% of the French are potential donors
be adjusted to the data. It should also be assessed how (Box 3). These dramatic differences among Western
well the heuristic predicts (ie, generalizes to) new countries can be explained by the interplay between the
data, with all relevant parameters being fixed and not legal environment and people’s reliance on the default
adjustable to these data.62Data fitting does not pro- heuristic. According to this social heuristic, a person
vide a good test; the real test is in prediction. should not act if a trustworthy institution has made an
implicit recommendation: “If there is a default, do noth-
82
PAGES_12_AG_1006_BA.qxd:DCNS#52 10/03/12 12:46 Page 83
Heuristic decision making in medicine - Marewski and Gigerenzer Dialogues in Clinical Neuroscience - Vol 14 .No. 1 .2012
ing about it.” By German law, no one is a donor without with v = C/(C +W) where C is the number of correct
their or their family's explicit consent. In France, in con- inferences when a predictor discriminates, and W the
trast, the default is that everyone is an organ donor number of wrong inferences.
unless they explicitly opt out. Depending on the legal (cid:129)Search rule: Search through predictors in order of their
environment, the same simple heuristic produces very validity.
different behavior, with very different outcomes for the (cid:129)Stopping rule: Stop on finding the first predictor that
general public and those who urgently need an organ.34 discriminates between the alternatives (eg, possible
In short, the descriptive study of practitioners’ and predictor values are 1 and 0).
patients’ use of heuristics as well as the fit between these (cid:129)Decision rule: Infer that the alternative with the posi-
heuristics and the environment can help in understand- tive predictor value (1) has the higher criterion value.
ing not only how health care decisions are made, but Brighton35,36showed that, across many data sets from dif-
how they can be improved. This leads us to the third— ferent real-world domains, it was the rule rather than the
the applied—question. exception that take-the-best outperformed sophisticated
computational machineries in predicting new (eg, yet
Can less be more? unknown) data. In the past years, a number of studies
have striven to make similar comparisons between
Heuristics have various general features that render heuristics and information-greedy tools in medical deci-
them especially suitable tools to improve applied med- sion making. One of the most recent of these attempts,
ical decision making. Let us point out just some of these. for example, focuses on fast-and-frugal trees for diag-
nosis of mental disorders such as depression.40
Accuracy
Transparency
As numerous studies have shown, when used in the cor-
rect environment, simple decision heuristics can surpass Because heuristics are simple, they are transparent and
the accuracy of more sophisticated, information-greedy generally easy to teach and to use in applied settings.
classification and prediction tools, including that of Consider, once more, the tree shown in Figure 1: in order
regression models or neural nets. Brighton,35,36for exam- to make an accurate decision quickly, the doctor has to
ple, compared the performance of heavy-weight com- ask at most three simple yes-or-no questions. The deci-
putational machineries such as classification and regres- sion-making process is completely transparent and can be
sion trees (CART37) or the decision tree induction easily communicated to a patient if needed. In contrast,
algorithm C4.538 to that of a heuristic called take-the- dealing with the various probabilities and symptoms cov-
best.39This heuristic resembles the fast-and-frugal tree ered by the Heart Disease Predictive Instrument is more
shown in Figure 1; it bases a decision on just one good cumbersome and complicated. As a result, the decision-
reason. Take-the-best simplifies decision making by making process seems less transparent and is likely more
searching sequentially through binary predictor vari- difficult to explain to a patient.
ables that can have positive values (1) or not (0) and by Teaching simple, transparent heuristics to doctors can
stopping after the first predictor that discriminates. In also help them to better understand health statistics, that
contrast to more complex (eg regression) models that is, the information on which informed medical diagnoses
assign optimal (eg, beta) weights to the various predic- and treatment decisions should be based. Unfortunately,
tor variables they integrate, take-the-best simply orders there is evidence that many doctors do not know how to
predictors unconditionally according to their validity v, correctly interpret such statistics. For instance,
Gigerenzer et al41gave 160 gynecologists the statistics
Box 3 needed for calculating that a woman with a positive
As of writing this article, the numbers reported by breast cancer screening mammogram actually has can-
Johnson and Goldstein34in 2003 have changed. For cer: a sensitivity of 90%, a false-positive rate of 9%, and
instance, in 2010 Germany had about 17% potential a prevalence of 1%. The physicians were asked what they
donors.63 would tell a woman who tested positive about her
chances of having breast cancer. The best answer is about
83
PAGES_12_AG_1006_BA.qxd:DCNS#52 10/03/12 12:46 Page 84
C l i n i c a l r e s e a r c h
1 out of 10 women; the results for the remaining 9 out of Simple Triage and Rapid Treatment, START,49a heuristic
10 are false alarms (false positives). As it turns out, 60% that can be categorized into the branch of fast-and-
of the gynecologists believed that 8 or 9 out of 10 women frugal trees,50allowed paramedics to rapidly split the vic-
who tested positive would have cancer, and 18% thought tims into main groups, including those who required
that the chances were 1 in 100. A similar lack of under- immediate medical treatment and those whose treat-
standing among physicians has been reported in diabetes ment was not as urgent.
prevention studies,42the evaluation of HIV tests,43and
other medical tests and treatments.44-48Making health sta- Accessibility and costs
tistics transparent can help doctors to understand them.
One very simple heuristic, for instance, is to change the Well-functioning heuristics can be made easily accessi-
mathematical format in which the relevant numbers are ble and help treatment and diagnosis even in situations
represented. To illustrate this, consider the case of mam- where access to technology is restricted. For instance, for
mography screening once more. It is easy to teach physi- macrolide prescription in young children with commu-
cians to translate the given probabilities into what is nity-acquired pneumonia, a tree with only two predictor
called natural frequencies, and to draw a corresponding variables—age and duration of fever—was developed as
tree to visualize the numbers. As Figure 3shows, all the a decision aid (Figure 4).51This frugal decision aid turned
physicians have to do is to think of 1000 women. Ten of out to be only slightly less accurate than a scoring sys-
these women are expected to have breast cancer (= 1% tem based on logistic regression (72% versus 75% sen-
prevalence). Of these 10 women, 9 will test positive (= sitivity), but using it does not require expensive tech-
90% sensitivity). Of the 990 women who do not have nology. As a result, this decision aid can be made easily
cancer, roughly 89 will still test positive (= 9% false pos- accessible to millions of children worldwide, even in
itive rate). When the format was changed to such natural poor countries.
frequencies, most of the gynecologists (87%) understood Simple heuristics can also aid in saving costs in rich,
that 9+89 = 98 will test positive. Of these 98, only 9 will developed countries, as the following example illustrates.
actually have breast cancer, equaling roughly 1 out of 10 In the US, there are about 2.6 million emergency room
(= 10%). visits each year for dizziness or vertigo.52 Emergency
Speed
Duration of fever ≤2 days?
Quick applicability is another important feature of well-
functioning heuristics, particularly in emergency situa- no yes
tions. After the attacks of September 11, 2001, the
Age ≤3 years? Mycoplasma
pneumoniae
1000 risk low
women no yes
Mycoplasma Mycoplasma
pneumoniae pneumoniae
10 990 risk high risk moderate
cancer no cancer
Figure 4. A fast-and-frugal tree for making decisions about macrolide
9 1 89 901 prescriptions, proposed by Fisher et al51 (see also
positive negative positive negative Katsikopoulos et al58for an in-depth discussion). Macrolides
are the first-line antibiotic treatment of community-acquired
pneumonia. The fast-and-frugal tree signals that first-line
Figure 3. A simple tree to represent probabilities as natural frequencies, macrolide treatment may be limited to individuals with com-
designed to help pyhsicians and patients understand health munity-acquired pneumonia who have had fever for more
statistics. than 2 days and who are older than 3 years.
84
PAGES_12_AG_1006_BA.qxd:DCNS#52 10/03/12 12:46 Page 85
Heuristic decision making in medicine - Marewski and Gigerenzer Dialogues in Clinical Neuroscience - Vol 14 .No. 1 .2012
room personnel need to detect the rare instances where Imagine two doctors. One doctor, let’s call him Professor
such dizziness is due to a dangerous brain stem or cere- Complexicus (PhD), is known for his scrutiny—he takes
bellar stroke. MRI with diffusion-weighted imaging can all information about a patient into account, including
help doctors to make this challenging diagnosis. Another the most minute details. His philosophy is that all infor-
diagnostic tool, a simple bedside exam, was developed mation is potentially relevant, and that considering as
by Kattah et al.52An alarm is raised if at least one of much information as possible benefits decisions. The
three simple tests indicates a stroke. other physician, Doctor Heuristicus, in contrast, relies
This bedside exam represents a tallyingheuristic. In con- only on a few pieces of information, perhaps those that
trast to fast-and-frugal trees and take-the-best, which she deems to be the most relevant ones. We can think of
assign more or less importance to specific predictor vari- the two doctors’ decision strategies as integration mod-
ables by ordering them, tallying treats all predictors els. One of Professor Complexicus’ models might read
equally, for example, by simply counting them. In its gen- like this: y= w x a1+ w x a2+ w x a3+ w x a4+ w x a5+
1 1 2 2 3 3 4 4 5 5
eral form, tallying can be described as follows. wx ai+ z. A simpler model of Doctor Heuristicus could
i i
(cid:129)Search rule: Search through predictors in any order. throw away some of the free parameters, w, a, and z, as
i i
(cid:129)Stopping rule: Stop search after mout of a total of M well as some of the predictor variables, x, such that y=
i
predictors (with 1 < m≤M). If the number of positive w x + z. The criterion both doctors wish to infer could
1 1
predictors is the same for both alternatives, search for be the number of days different patients will need to
another predictor. If no more predictors are found, recover from a medical condition, y. The predictor vari-
guess. ables, x, could be the type of condition the patients suf-
i
(cid:129)Decision rule: Decide for the alternative that is favored fer from, the patients’ overall physical constitution or
by more predictors. age, or the number of times loving family members have
As it turns out, Kattah et al’s52 simple bedside exam visited the patients in the hospital thus far.
yields a larger sensitivity than MRI, while the false-pos- In a formal, statistical analysis, a comparative evaluation
itive rate is only slightly larger than that of the MRI, of these two models would entail computing R2or some
which did not raise any false alarms. In contrast to the other goodness-of-fit index between the models’ esti-
MRI, which can take up to 5 to 10 minutes plus several mations and the observed number of days it took the
hours of waiting time, entails costs of more than $1000, patients to recover. Such measures are based on the dis-
and is not available everywhere, the bedside exam takes tance between a model’s estimate and the criterion y.
little time, is less cost-intensive, and can be conducted And indeed, fitting Professor Complexicus’ strategy of
anywhere. paying attention to more variables and weighting them
In short, relying on heuristics as a tool for medical deci- in an optimal way (ie, minimizing least squares) to obser-
sion making can help practitioners to make accurate, vations about past patients (ie, the ones where one
transparent, and quick decisions, often while depending already knows how many days they needed to recover),
on little technology and few financial resources. Less will always lead to a larger R2 than fitting Doctor
information, complexity, time, and technology can be Heuristicus’ simpler strategy to these observations. Put
more efficient, even when it comes to medical decision differently, when it comes to explaining past observa-
making. tions from hindsight, Professor Complexicus will do the
more convincing job. Given how well Professor
Why heuristics work Complexicus does in explaining the time patients
needed to recover in the past, it seems intuitive that his
One reason for the surprising performance of heuristics estimations should also fare better than those of Doctor
is that they ignore information. As we have explained Heuristicus when it comes to predicting future patients’
above, this makes them quicker to execute, easier to time to recover.
understand, and easier to communicate. Importantly, as Yet this is not necessarily the case. Goodness-of fit mea-
can be shown by means of mathematical analysis and sures alone cannot disentangle the variation in the
computer simulations,36,53 it is also this feature that dri- observations due to the relevant variables from the vari-
ves part of the predictive power of heuristics. Let us illus- ation due to random error, or noise. In fitting past obser-
trate this with a simplifying, fictional story. vations, models can end up taking into account such
85
PAGES_12_AG_1006_BA.qxd:DCNS#52 10/03/12 12:46 Page 86
C l i n i c a l r e s e a r c h
noise, thus mistakenly attributing meaning to mere is likely to hold true with respect to Professor
chance. As a result, a model can end up overfittingthese Complexicus’ and Doctor Heuristicus’ strategies:
observations. Professor Complexicus’ complex strategy is likely to be
Figure 5illustrates a corresponding situation in which more prone to overfitting past observations than Doctor
one model, Model A(thin line) overfits already existing, Heuristicus’ simple one. As a result, Dr. Heuristicus’
past observations (filled circles; eg, old patients) by chas- strategy is likely to be better able to predict new obser-
ing after noise in those observations. As can be seen, this vations than Professor Complexicus’ strategy.
model fits the past observations perfectly but does a rel- In short, when data are not completely free of noise,
atively poor job of predicting new observations (filled increased complexity (eg, integrating as much informa-
triangles; eg, new patients). Model B(thick line), while tion as possible) makes a model more likely to end up
not fitting the past observations as well as Model A, cap- overfitting past observations, while its ability to predict
tures the main trends in the data and ignores the noise. new ones decreases (although see Box 4). But what mat-
This makes it better equipped to predict new observa- ters in many applied medical settings is less the ability to
tions, as can be seen from the deviations between the explain (ie, fit) past observations than to make accurate
model’s predictions and the new observations, which are inferences about future, unknown observations, such as
indeed smaller than the deviations for Model A. about new, yet unseen patients.
Importantly, the degree to which a model is susceptible
to overfitting is related to the model’s complexity. One Summary and outlook for future research
factor that contributes to a model’s complexity is its
number of free parameters. As is illustrated in Figure 5, Rationality has many meanings. Most theories assume
the complex, information-greedy Model A overfits past that the future can be known with certainty, including
observations; Model B, in turn, which has fewer free the probabilities, for instance, for weighting different
parameters and which takes into account less informa- pieces of information, so that unboundedly rational opti-
tion, captures only the main trends in the past observa- mization methods can define rational choice. There are
tions, but better predicts the new observations. The same two variants of these: those that assume that people’s
behavior can actually be modeled by this form of
unboundedly rational optimization, and those that
Model A
assume that people’ behavior systematically deviates
Model B
Past observations from it, manifesting irrational cognitive illusions, biases,
New observations and errors. This article dealt with a third perspective,
which asks how people make decisions when the condi-
tions for optimization are not met. That is the case for
most real-world decisions, including in medicine. In
Box 4
Obviously, ignoring too much information and too
many parameters can also be detrimental. A well-
functioning model needs to achieve a balance
between both extremes. As is known in the model
selection literature, decreasing a model’s complexity
can eventually lead to underfitting; thus, in an uncer-
tain world, there is often an inversely U-shaped func-
Figure 5. Illustration of how two models fit past observations (filled cir- tion between model complexity and predictive
cles) and how they predict new observations (triangles). The
power.60Moreoever, besides the number of free para-
complex Model A (thin line) overfits the past observations and
is not as accurate in predicting the new observations as the meters a model has, other factors also contribute to
simple Model B (thick line). model complexity, such as a model’s functional form
Adapted from ref 60: Pitt MA, Myung IJ, Zhang S. Toward a method for
and the extension of the allowable parameter space.64
selecting among computational models for cognition. Psychol Rev.
2002;109:472–491. Copyright © American Psychological Association 2002
86