Heuristics in judgment and decision-making
Heuristics are simple strategies or mental processes that humans, animals, organizations and machines use to quickly form judgments, make decisions, and find solutions to complex problems. This happens when an individual focuses on the most relevant aspects of a problem or situation to formulate a solution.
Some heuristics are more applicable and useful than others depending on the situation. Heuristic processes are used to find the answers and solutions that are most likely to work or be correct. However, heuristics are not always right or the most accurate. While they can differ from answers given by logic and probability, judgments and decisions based on a heuristic can be good enough to satisfy a need. They are meant to serve as quick mental references for everyday experiences and decisions. In situations of uncertainty, where information is incomplete, heuristics allow for the less-is-more effect, in which less information leads to more accuracy.
History
formulated one of the first models of heuristics, known as satisficing. His more general research program posed the question of how humans make decisions when the conditions for rational choice theory are not met, that is how people decide under uncertainty. Simon is also known as the father of bounded rationality, which he understood as the study of the match between heuristics and decision environments. This program was later extended into the study of ecological rationality.In the early 1970s, psychologists Amos Tversky and Daniel Kahneman took a different approach, linking heuristics to cognitive biases. Their typical experimental setup consisted of a rule of logic or probability, embedded in a verbal description of a judgement problem, and demonstrated that people's intuitive judgement deviated from the rule. The "Linda problem" below gives an example. The deviation is then explained by a heuristic. This research, called the heuristics-and-biases program, challenged the idea that human beings are rational actors and first gained worldwide attention in 1974 with the Science paper "Judgment Under Uncertainty: Heuristics and Biases" and although the originally proposed heuristics have been refined over time, this research program has changed the field by permanently setting the research questions.
The original ideas by Herbert Simon were taken up in the 1990s by Gerd Gigerenzer and others. According to their perspective, the study of heuristics requires formal models that allow predictions of behavior to be made ex ante. Their program has three aspects:
- What are the heuristics humans use?
- Under what conditions should humans rely on a given heuristic?
- How to design heuristic decision aids that are easy to understand and execute?
These two different research programs have led to two kinds of models of heuristics, formal models and informal ones. Formal models describe the decision process in terms of an algorithm, which allows for mathematical proofs and computer simulations. In contrast, informal models are verbal descriptions.
Formal models of heuristics
Simon's satisficing strategy
Herbert Simon's satisficing heuristic can be used to choose one alternative from a set of alternatives in situations of uncertainty. Here, uncertainty means that the total set of alternatives and their consequences is not known or knowable. For instance, professional real-estate entrepreneurs rely on satisficing to decide in which location to invest to develop new commercial areas: "If I believe I can get at least x return within y years, then I take the option." In general, satisficing is defined as:- Step 1: Set an aspiration level α
- Step 2: Choose the first alternative that satisfies α
- Step 3: If after time β no alternative has satisfied α, then decrease α by some amount δ and return to step 1.
Elimination by aspects
Unlike satisficing, Amos Tversky's elimination-by-aspect heuristic can be used when all alternatives are simultaneously available. The decision-maker gradually reduces the number of alternatives by eliminating alternatives that do not meet the aspiration level of a specific attribute.Recognition heuristic
The recognition heuristic exploits the basic psychological capacity for recognition in order to make inferences about unknown quantities in the world. For two alternatives, the heuristic is:If one of two alternatives is recognized and the other not, then infer that the recognized alternative has the higher value with respect to the criterion.For example, in the 2003 Wimbledon tennis tournament, Andy Roddick played Tommy Robredo. If one has heard of Roddick but not of Robredo, the recognition heuristic leads to the prediction that Roddick will win. The recognition heuristic exploits partial ignorance, if one has heard of both or no player, a different strategy is needed. Studies of Wimbledon 2003 and 2005 have shown that the recognition heuristic applied by semi-ignorant amateur players predicted the outcomes of all gentlemen single games as well and better than the seedings of the Wimbledon experts, as well as the ATP rankings. The recognition heuristic is ecologically rational when the recognition validity is substantially above chance. In the present case, recognition of players' names is highly correlated with their chances of winning.
Take-the-best
The take-the-best heuristic exploits the basic psychological capacity for retrieving cues from memory in the order of their validity. Based on the cue values, it infers which of two alternatives has a higher value on a criterion. Unlike the recognition heuristic, it requires that all alternatives are recognized, and it thus can be applied when the recognition heuristic cannot. For binary cues, the heuristic is defined as:Search rule: Search cues in the order of their validity v.
Stopping rule: Stop search on finding the first cue that discriminates between the two alternatives.
Decision rule: Infer that the alternative with the positive cue value.
The validity vi of a cue i is defined as the proportion of correct decisions ci:
vi = ci / ti
where ti is the number of cases the values of the two alternatives differ on cue i. The validity of each cue can be estimated from samples of observation.
Take-the-best has remarkable properties. In comparison with complex machine learning models, it has been shown that it can often predict better than regression models, classification-and-regression trees, neural networks, and support vector machines.
Similarly, psychological studies have shown that in situations where take-the-best is ecologically rational, a large proportion of people tend to rely on it. This includes decision making by airport custom officers, professional burglars and police officers and student populations. The conditions under which take-the-best is ecologically rational are mostly known. Take-the-best shows that the previous view that ignoring part of the information would be generally irrational is incorrect. Less can be more.
Fast-and-frugal trees
A fast-and-frugal tree is a heuristic that allows to make a classifications, such as whether a patient with severe chest pain is likely to have a heart attack or not, or whether a car approaching a checkpoint is likely to be a terrorist or a civilian. It is called “fast and frugal” because, just like take-the-best, it allows for quick decisions with only few cues or attributes. It is called a “tree” because it can be represented like a decision tree in which one asks a sequence of questions. Unlike a full decision tree, however, it is an incomplete tree – to save time and reduce the danger of overfitting.Figure 1 shows a fast-and-frugal tree used for screening for HIV. Just like take-the-best, the tree has a search rule, stopping rule, and decision rule:
Search rule: Search through cues in a specified order.
Stopping rule: Stop search if an exit is reached.
Decision rule: Classify the person according to the exit.
In the HIV tree, an ELISA test is conducted first. If the outcome is negative, then the testing procedure stops and the client is informed of the good news, that is, “no HIV.” If, however, the result is positive, a second ELISA test is performed, preferably from a different manufacturer. If the second ELISA is negative, then the procedure stops and the client is informed of having “no HIV.” However, if the result is positive, a final test, the Western blot, is conducted.
In general, for n binary cues, a fast-and-frugal tree has exactly n + 1 exits – one for each cue and two for the final cue. A full decision tree, in contrast, requires 2n exits. The order of cues in a fast-and-frugal tree is determined by the sensitivity and specificity of the cues, or by other considerations such as the costs of the tests. In the case of the HIV tree, the ELISA is ranked first because it produces fewer misses than the Western blot test, and also is less expensive. The Western blot test, in contrast, produces fewer false alarms. In a full tree, in contrast, order does not matter for the accuracy of the classifications.
Fast-and-frugal trees are descriptive or prescriptive models of decision making under uncertainty. For instance, an analysis or court decisions reported that the best model of how London magistrates make bail decisions is a fast and frugal tree. The HIV tree is both prescriptive– physicians are taught the procedure – and a descriptive model, that is, most physicians actually follow the procedure.
Informal models of heuristics
In their initial research, Tversky and Kahneman proposed three heuristics—availability, representativeness, and anchoring and adjustment. Subsequent work has identified many more. Heuristics that underlie judgment are called "judgment heuristics". Another type, called "evaluation heuristics", are used to judge the desirability of possible choices.Availability
In psychology, availability is the ease with which a particular idea can be brought to mind. When people estimate how likely or how frequent an event is on the basis of its availability, they are using the availability heuristic. When an infrequent event can be brought easily and vividly to mind, this heuristic overestimates its likelihood. For example, people overestimate their likelihood of dying in a dramatic event such as a tornado or terrorism. Dramatic, violent deaths are usually more highly publicised and therefore have a higher availability. On the other hand, common but mundane events are hard to bring to mind, so their likelihoods tend to be underestimated. These include deaths from suicides, strokes, and diabetes. This heuristic is one of the reasons why people are more easily swayed by a single, vivid story than by a large body of statistical evidence. It may also play a role in the appeal of lotteries: to someone buying a ticket, the well-publicised, jubilant winners are more available than the millions of people who have won nothing.When people judge whether more English words begin with T or with K, the availability heuristic gives a quick way to answer the question. Words that begin with T come more readily to mind, and so subjects give a correct answer without counting out large numbers of words. However, this heuristic can also produce errors. When people are asked whether there are more English words with K in the first position or with K in the third position, they use the same process. It is easy to think of words that begin with K, such as kangaroo, kitchen, or kept. It is harder to think of words with K as the third letter, such as lake, or acknowledge, although objectively these are three times more common. This leads people to the incorrect conclusion that K is more common at the start of words. In another experiment, subjects heard the names of many celebrities, roughly equal numbers of whom were male and female. The subjects were then asked whether the list of names included more men or more women. When the men in the list were more famous, a great majority of subjects incorrectly thought there were more of them, and vice versa for women. Tversky and Kahneman's interpretation of these results is that judgments of proportion are based on availability, which is higher for the names of better-known people.
In one experiment that occurred before the 1976 U.S. Presidential election, some participants were asked to imagine Gerald Ford winning, while others did the same for a Jimmy Carter victory. Each group subsequently viewed their allocated candidate as significantly more likely to win. The researchers found a similar effect when students imagined a good or a bad season for a college football team. The effect of imagination on subjective likelihood has been replicated by several other researchers.
A concept's availability can be affected by how recently and how frequently it has been brought to mind. In one study, subjects were given partial sentences to complete. The words were selected to activate the concept either of hostility or of kindness: a process known as priming. They then had to interpret the behavior of a man described in a short, ambiguous story. Their interpretation was biased towards the emotion they had been primed with: the more priming, the greater the effect. A greater interval between the initial task and the judgment decreased the effect.
Tversky and Kahneman offered the availability heuristic as an explanation for illusory correlations in which people wrongly judge two events to be associated with each other. They explained that people judge correlation on the basis of the ease of imagining or recalling the two events together.
Representativeness
The representativeness heuristic is seen when people use categories, for example when deciding whether or not a person is a criminal. An individual thing has a high representativeness for a category if it is very similar to a prototype of that category. When people categorise things on the basis of representativeness, they are using the representativeness heuristic. "Representative" is here meant in two different senses: the prototype used for comparison is representative of its category, and representativeness is also a relation between that prototype and the thing being categorised. While it is effective for some problems, this heuristic involves attending to the particular characteristics of the individual, ignoring how common those categories are in the population. Thus, people can overestimate the likelihood that something has a very rare property, or underestimate the likelihood of a very common property. This is called the base rate fallacy. Representativeness explains this and several other ways in which human judgments break the laws of probability.The representativeness heuristic is also an explanation of how people judge cause and effect: when they make these judgements on the basis of similarity, they are also said to be using the representativeness heuristic. This can lead to a bias, incorrectly finding causal relationships between things that resemble one another and missing them when the cause and effect are very different. Examples of this include both the belief that "emotionally relevant events ought to have emotionally relevant causes", and magical associative thinking.
Representativeness of base rates
A 1973 experiment used a psychological profile of Tom W., a fictional graduate student. One group of subjects had to rate Tom's similarity to a typical student in each of nine academic areas. Another group had to rate how likely it is that Tom specialised in each area. If these ratings of likelihood are governed by probability, then they should resemble the base rates, i.e. the proportion of students in each of the nine areas. If people based their judgments on probability, they would say that Tom is more likely to study Humanities than Library Science, because there are many more Humanities students, and the additional information in the profile is vague and unreliable. Instead, the ratings of likelihood matched the ratings of similarity almost perfectly, both in this study and a similar one where subjects judged the likelihood of a fictional woman taking different careers. This suggests that rather than estimating probability using base rates, subjects had substituted the more accessible attribute of similarity.Conjunction fallacy
When people rely on representativeness, they can fall into an error which breaks a fundamental law of probability. Tversky and Kahneman gave subjects a short character sketch of a woman called Linda, describing her as, "31 years old, single, outspoken, and very bright. She majored in philosophy. As a student, she was deeply concerned with issues of discrimination and social justice, and also participated in anti-nuclear demonstrations". People reading this description then ranked the likelihood of different statements about Linda. Amongst others, these included "Linda is a bank teller", and, "Linda is a bank teller and is active in the feminist movement". People showed a strong tendency to rate the latter, more specific statement as more likely, even though a conjunction of the form "Linda is both X and Y" can never be more probable than the more general statement "Linda is X". The explanation in terms of heuristics is that the judgment was distorted because, for the readers, the character sketch was representative of the sort of person who might be an active feminist but not of someone who works in a bank. A similar exercise concerned Bill, described as "intelligent but unimaginative". A great majority of people reading this character sketch rated "Bill is an accountant who plays jazz for a hobby", as more likely than "Bill plays jazz for a hobby".Without success, Tversky and Kahneman used what they described as "a series of increasingly desperate manipulations" to get their subjects to recognise the logical error. In one variation, subjects had to choose between a logical explanation of why "Linda is a bank teller" is more likely, and a deliberately illogical argument which said that "Linda is a feminist bank teller" is more likely "because she resembles an active feminist more than she resembles a bank teller". Sixty-five percent of subjects found the illogical argument more convincing.
Other researchers also carried out variations of this study, exploring the possibility that people had misunderstood the question. They did not eliminate the error. It has been shown that individuals with high CRT scores are significantly less likely to be subject to the conjunction fallacy. The error disappears when the question is posed in terms of frequencies. Everyone in these versions of the study recognised that out of 100 people fitting an outline description, the conjunction statement cannot apply to more people than the general statement.
Ignorance of sample size
Tversky and Kahneman asked subjects to consider a problem about random variation. Imagining for simplicity that exactly half of the babies born in a hospital are male, the ratio will not be exactly half in every time period. On some days, more girls will be born and on others, more boys. The question was, does the likelihood of deviating from exactly half depend on whether there are many or few births per day? It is a well-established consequence of sampling theory that proportions will vary much more day-to-day when the typical number of births per day is small. However, people's answers to the problem do not reflect this fact. They typically reply that the number of births in the hospital makes no difference to the likelihood of more than 60% male babies in one day. The explanation in terms of the heuristic is that people consider only how representative the figure of 60% is of the previously given average of 50%.Dilution effect
and colleagues suggest that representativeness explains the dilution effect, in which irrelevant information weakens the effect of a stereotype. Subjects in one study were asked whether "Paul" or "Susan" was more likely to be assertive, given no other information than their first names. They rated Paul as more assertive, apparently basing their judgment on a gender stereotype. Another group, told that Paul's and Susan's mothers each commute to work in a bank, did not show this stereotype effect; they rated Paul and Susan as equally assertive. The explanation is that the additional information about Paul and Susan made them less representative of men or women in general, and so the subjects' expectations about men and women had a weaker effect. This means unrelated and non-diagnostic information about certain issue can make relative information less powerful to the issue when people understand the phenomenon.Misperception of randomness
Representativeness explains systematic errors that people make when judging the probability of random events. For example, in a sequence of coin tosses, each of which comes up heads or tails, people reliably tend to judge a clearly patterned sequence such as HHHTTT as less likely than a less patterned sequence such as HTHTTH. These sequences have exactly the same probability, but people tend to see the more clearly patterned sequences as less representative of randomness, and so less likely to result from a random process. Tversky and Kahneman argued that this effect underlies the gambler's fallacy; a tendency to expect outcomes to even out over the short run, like expecting a roulette wheel to come up black because the last several throws came up red. They emphasised that even experts in statistics were susceptible to this illusion: in a 1971 survey of professional psychologists, they found that respondents expected samples to be overly representative of the population they were drawn from. As a result, the psychologists systematically overestimated the statistical power of their tests, and underestimated the sample size needed for a meaningful test of their hypotheses.Anchoring and adjustment
Anchoring and adjustment is a heuristic used in many situations where people estimate a number. According to Tversky and Kahneman's original description, it involves starting from a readily available number—the "anchor"—and shifting either up or down to reach an answer that seems plausible. In Tversky and Kahneman's experiments, people did not shift far enough away from the anchor. Hence the anchor contaminates the estimate, even if it is clearly irrelevant. In one experiment, subjects watched a number being selected from a spinning "wheel of fortune". They had to say whether a given quantity was larger or smaller than that number. For instance, they might be asked, "Is the percentage of African countries which are members of the United Nations larger or smaller than 65%?" They then tried to guess the true percentage. Their answers correlated well with the arbitrary number they had been given. Insufficient adjustment from an anchor is not the only explanation for this effect. An alternative theory is that people form their estimates on evidence which is selectively brought to mind by the anchor.The anchoring effect has been demonstrated by a wide variety of experiments both in laboratories and in the real world. It remains when the subjects are offered money as an incentive to be accurate, or when they are explicitly told not to base their judgment on the anchor. The effect is stronger when people have to make their judgments quickly. Subjects in these experiments lack introspective awareness of the heuristic, denying that the anchor affected their estimates.
Even when the anchor value is obviously random or extreme, it can still contaminate estimates. One experiment asked subjects to estimate the year of Albert Einstein's first visit to the United States. Anchors of 1215 and 1992 contaminated the answers just as much as more sensible anchor years. Other experiments asked subjects if the average temperature in San Francisco is more or less than 558 degrees, or whether there had been more or fewer than 100,025 top ten albums by The Beatles. These deliberately absurd anchors still affected estimates of the true numbers.
Anchoring results in a particularly strong bias when estimates are stated in the form of a confidence interval. An example is where people predict the value of a stock market index on a particular day by defining an upper and lower bound so that they are 98% confident the true value will fall in that range. A reliable finding is that people anchor their upper and lower bounds too close to their best estimate. This leads to an overconfidence effect. One much-replicated finding is that when people are 98% certain that a number is in a particular range, they are wrong about thirty to forty percent of the time.
Anchoring also causes particular difficulty when many numbers are combined into a composite judgment. Tversky and Kahneman demonstrated this by asking a group of people to rapidly estimate the product 8 x 7 x 6 x 5 x 4 x 3 x 2 x 1. Another group had to estimate the same product in reverse order; 1 x 2 x 3 x 4 x 5 x 6 x 7 x 8. Both groups underestimated the answer by a wide margin, but the latter group's average estimate was significantly smaller. The explanation in terms of anchoring is that people multiply the first few terms of each product and anchor on that figure. A less abstract task is to estimate the probability that an aircraft will crash, given that there are numerous possible faults each with a likelihood of one in a million. A common finding from studies of these tasks is that people anchor on the small component probabilities and so underestimate the total. A corresponding effect happens when people estimate the probability of multiple events happening in sequence, such as an accumulator bet in horse racing. For this kind of judgment, anchoring on the individual probabilities results in an overestimation of the combined probability.
Examples
People's valuation of goods, and the quantities they buy, respond to anchoring effects. In one experiment, people wrote down the last two digits of their social security numbers. They were then asked to consider whether they would pay this number of dollars for items whose value they did not know, such as wine, chocolate, and computer equipment. They then entered an auction to bid for these items. Those with the highest two-digit numbers submitted bids that were many times higher than those with the lowest numbers. When a stack of soup cans in a supermarket was labelled, "Limit 12 per customer", the label influenced customers to buy more cans. In another experiment, real estate agents appraised the value of houses on the basis of a tour and extensive documentation. Different agents were shown different listing prices, and these affected their valuations. For one house, the appraised value ranged from US$114,204 to $128,754.Anchoring and adjustment has also been shown to affect grades given to students. In one experiment, 48 teachers were given bundles of student essays, each of which had to be graded and returned. They were also given a fictional list of the students' previous grades. The mean of these grades affected the grades that teachers awarded for the essay.
One study showed that anchoring affected the sentences in a fictional rape trial. The subjects were trial judges with, on average, more than fifteen years of experience. They read documents including witness testimony, expert statements, the relevant penal code, and the final pleas from the prosecution and defence. The two conditions of this experiment differed in just one respect: the prosecutor demanded a 34-month sentence in one condition and 12 months in the other; there was an eight-month difference between the average sentences handed out in these two conditions. In a similar mock trial, the subjects took the role of jurors in a civil case. They were either asked to award damages "in the range from $15 million to $50 million" or "in the range from $50 million to $150 million". Although the facts of the case were the same each time, jurors given the higher range decided on an award that was about three times higher. This happened even though the subjects were explicitly warned not to treat the requests as evidence.
Assessments can also be influenced by the stimuli provided. In one review, researchers found that if a stimulus is perceived to be important or carry "weight" to a situation, that people were more likely to attribute that stimulus as heavier physically.
Affect heuristic
"Affect", in this context, is a feeling such as fear, pleasure or surprise. It is shorter in duration than a mood, occurring rapidly and involuntarily in response to a stimulus. While reading the words "lung cancer" might generate an affect of dread, the words "mother's love" can create an affect of affection and comfort. When people use affect to judge benefits or risks, they are using the affect heuristic. The affect heuristic has been used to explain why messages framed to activate emotions are more persuasive than those framed in a purely factual way.Others
- Control heuristic
- Contagion heuristic
- Effort heuristic
- Familiarity heuristic
- Fluency heuristic
- Gaze heuristic
- Hot-hand fallacy
- Naive diversification
- Peak-end rule
- Recognition heuristic
- Scarcity heuristic
- Similarity heuristic
- Simulation heuristic
- Social proof
Theories
Cognitive laziness
An effort-reduction framework proposed by Anuj K. Shah and Daniel M. Oppenheimer states that people use a variety of techniques to reduce the effort of making decisions.Attribute substitution
In 2002 Daniel Kahneman and Shane Frederick proposed a process called attribute substitution which happens without conscious awareness. According to this theory, when somebody makes a judgment which is computationally complex, a rather more easily calculated heuristic attribute is substituted. In effect, a difficult problem is dealt with by answering a rather simpler problem, without the person being aware this is happening. This explains why individuals can be unaware of their own biases, and why biases persist even when the subject is made aware of them. It also explains why human judgments often fail to show regression toward the mean.This substitution is thought of as taking place in the automatic intuitive judgment system, rather than the more self-aware reflective system. Hence, when someone tries to answer a difficult question, they may actually answer a related but different question, without realizing that a substitution has taken place.
In 1975, psychologist Stanley Smith Stevens proposed that the strength of a stimulus is encoded by brain cells in a way that is independent of modality. Kahneman and Frederick built on this idea, arguing that the target attribute and heuristic attribute could be very different in nature.
Kahneman and Frederick propose three conditions for attribute substitution:
- The target attribute is relatively inaccessible.
- An associated attribute is highly accessible.
- The substitution is not detected and corrected by the reflective system.
Kahneman gives an example where some Americans were offered insurance against their own death in a terrorist attack while on a trip to Europe, while another group were offered insurance that would cover death of any kind on the trip. Even though "death of any kind" includes "death in a terrorist attack", the former group were willing to pay more than the latter. Kahneman suggests that the attribute of fear is being substituted for a calculation of the total risks of travel. Fear of terrorism for these subjects was stronger than a general fear of dying on a foreign trip.