Medical Decision Support Systems

Biases in Human
Decision Making

.Yuval Shahar M.D., Ph.D

The Need to Assess Probabilities
• People need to make decisions constantly, such as
during diagnosis and therapy
• Thus, people need to assess probabilities to classify
objects or predict various values, such as the
probability of a disease given a set of symptoms
• People employ several types of heuristics to assess
probabilities
• However, these heuristics often lead to significant
biases in a consistent fashion
Three Major Human Probability-
Assessment Heuristics/Biases
(Tversky and Kahneman, 1974)
• Representativeness
– The more object X is similar to class Y, the
more likely we think X belongs to Y
• Availability
– The easier it is to consider instances of class Y,
the more frequent we think it is
• Anchoring
– Initial estimated values affect the final
estimates, even after considerable adjustments
 A Representativeness Example
• Consider the following description:
“Steve is very shy and withdrawn, invariably
helpful, but with little interest in people, or
in the world of reality. A meek and tidy
soul, he has a need for order and structure,
and a passion for detail.”
• Is Steve a farmer, a librarian, a physician,
an airline pilot, or a salesman?
 The Representativeness Heuristic
• We often judge whether object X belongs to
class Y by how representative X is of class Y
• For example, People order the potential
occupations by probability and by similarity in
exactly the same way
• The problem is that similarity ignores multiple
biases
 Representative Bias (1):
Insensitivity to Prior Probabilities
• The base rate of outcomes should be a major
factor in estimating their frequency
• However, people often ignore it (e.g., there are
more farmers than librarians)
– E.g., the lawyers vs. engineers experiment:
• Reversing the proportions in the group had no effect on
estimating the profession, given a description
• Giving worthless evidence caused the subjects to ignore the
odds and estimate the probability as 0.5
– Thus, prior probabilities of diseases are often ignored
when the patient seems to fit a rare-disease description
 Representative Bias (2):
Insensitivity to Sample Size
• The size of a sample withdrawn from a
population should greatly affect the
likelihood of obtaining certain results in it
• People, however, ignore sample size and
only use the superficial similarity measures
• For example, people ignore the fact that
larger samples are less likely to deviate
from the mean than smaller samples
 Representative Bias (3):
Misconception of Chance
• People expect random sequences to be “representatively
random” even locally
– E.g., they consider a coin-toss run of HTHTTH to be more
likely than HHHTTT or HHHHTH
• The Gambler’s Fallacy
– After a run of reds in a roulette, black will make the overall run
more representative (chance as a self-correcting process??)
• Even experienced research psychologists believe in a law
of small numbers (small samples are representative of
the population they are drawn from)
 Representative Bias (4):
Insensitivity to Predictability
• People predict future performance mainly by
similarity of description to future results
• For example, predicting future performance
as a teacher based on a single practice lesson
– Evaluation percentiles (of the quality of the
lesson) were identical to predicted percentiles of
5-year future standings as teachers
 Representative Bias (5):
The Illusion of Validity
• A good match between input information and output
classification or outcome often leads to unwarranted
confidence in the prediction
• Example: Use of clinical interviews for selection
• Internal consistency of input pattern increases
confidence
– a series of B’s seems more predictive of a final grade-point
average than a set of A’s and C’s
– Redundant, correlated data increases confidence
 Representative Bias (6):
Misconceptions of Regression
• People tend to ignore the phenomenon of
regression towards the mean
– E.g., correlation between parents’ and children’s heights
or IQ; performance on successive tests
• People expect predicted outcomes to be as
representative of the input as possible
• Failure to understand regression may lead to
overestimate the effects of punishments and
underestimate the effects of reward on future
performance (since a good performance is likely to
be followed by a worse one and vice versa)
 The Availability Heuristic
• The frequency of a class or event is often
assessed by the ease with which instances
of it can be brought to mind
• The problem is that this mental availability
might be affected by factors other than the
frequency of the class
 Availability Biases (1):
Ease of Retrievability
• Classes whose instances are more easily
retrievable will seem larger
– For example, judging if a list of names had
more men or women depends on the relative
frequency of famous names
• Salience affects retrievability
– E.g., watching a car accident increases
subjective assessment of traffic accidents
 Availability Biases (2):
Effectiveness of a Search Set
• We often form mental “search sets” to
estimate how frequent are members of some
class; the effectiveness of the search might
not relate directly to the class frequency
– Who is more prevalent: Words that start with r
or words where r is the 3rd letter?
– Are abstract words such as love more frequent
than concrete words such as door?
 Availability Biases (3):
Ease of Imaginability
• Instances often need to be constructed on the
fly using some rule; the difficulty of
imagining instances is used as an estimate of
their frequency
– E.g. number of combinations of 8 out of 10
people, versus 2 out of 10 people
– Imaginability might cause overestimation of
likelihood of vivid scenarios, and underestimation
of the likelihood of difficult-to-imagine ones
 Availability Biases (4):
Illusory Correlation
• People tended to overestimate co-occurrence of
diagnoses such as paranoia or suspiciousness
with features in persons drawn by hypothetical
mental patients, such as peculiar eyes
• Subjects might overestimate the correlation due
to easier association of suspicion with the eyes
than other body parts
 The Anchoring and Adjustment
Heuristic
• People often estimate by adjusting an initial
value until a final value is reached
• Initial values might be due to the problem
presentation or due to partial computations
• Adjustments are typically insufficient and
are biased towards initial values, the anchor
Anchoring and Adjustment Biases (1):
Insufficient Adjustment
• Anchoring occurs even when initial estimates (e.g.,
percentage of African nations in the UN) were explicitly
made at random by spinning a wheel!
• Anchoring may occur due to incomplete calculation, such
as estimating by two high-school student groups
– the expression 8x7x6x5x4x3x2x1 (median answer: 512)
– with the expression 1x2x3x4x5x6x7x8 (median answer: 2250)
• Anchoring occurs even with outrageously extreme anchors
(Quattrone et al., 1984)
• Anchoring occurs even when experts (real-estate agents)
estimate real-estate prices (Northcraft and Neale, 1987)
Anchoring and Adjustment Biases (2):
Evaluation of Conjunctive and Disjunctive Events

• People tend to overestimate the probability of

conjunctive events (e.g., success of a plan that
requires success of multiple steps)
• People underestimate the probability of
disjunctive events (e.g. the Birthday Paradox)
• In both cases there is insufficient adjustment
from the probability of an individual event
Anchoring and Adjustment Biases (3):
Assessing Subjective Probability Distributions
• Estimating the 1st and 99th percentiles often leads to
too-narrow confidence intervals
– Estimates often start from median (50th percentile) values,
and adjustment is insufficient
• The degree of calibration depends on the elicitation
procedure
– state values given percentile: leads to extreme estimates
– state percentile given a value: leads to conservativeness
 A Special Type of Bias: Framing
• Risky prospects can be framed in different
ways- as gains or as losses
• Changing the description of a prospect
should not change decisions, but it does
• Prospect Theory (Kahneman and Tversky,
1979) predicts such anomalies due to the
fact that the negative effect of a loss is
larger than the positive effect of a gain
 Framing Experiment (I)
• Imagine the US is preparing for the
outbreak of an Asian disease, expected to
kill 600 people (N = 152 subjects):
– If program A is adopted, 200 people will be
saved (72% preference)
– If program B is adopted, there is one third
probability that 600 people will be saved and
two thirds probability that no people will be
saved (28% preference)
 Framing Experiment (II)
• Imagine the US is preparing for the
outbreak of an Asian disease, expected to
kill 600 people (N = 155 subjects):
– If program C is adopted, 400 people will die
(22% preference)
– If program D is adopted, there is one third
probability that nobody will be die and two
thirds probability that 600 people will die (78%
preference)
 Summary: Heuristics and Biases
• There are several common heuristics people employ
to estimate probabilities
– Representativeness of a class by an object
– Availability of instances as a frequency measure
– Adjustment from an initial anchoring value
• All heuristics are quite effective, usually, but lead to
predictable, systematic errors and biases
• Understanding biases might decrease their effect