Judgment under uncertainty: heuristics and biases - Kahneman & Tversky - 1974 - Article

People base their decision making often on the likelihood of a particular event, however, they estimate the likelihood of events with the help of heuristics. Heuristics can be useful because they make the assessment of probabilities easier but they can also lead to systematic errors. Judgements are often made based on information that is not completely valid. Subsequently, this information is processed according to heuristic rules which leads to errors in the estimation. An example is the estimation of distance based on clarity. An object that has a higher clarity is perceived as being closer which is partly true but not always. In this article three heuristics leading to several biases are described.

Representativeness

The representativeness heuristic is used “when probabilities are evaluated by the degree to which A is representative of B, that is, by the degree to which A resembles B’’ . This heuristic can lead to bias when people are asked to estimate the likelihood that a description of certain personal characteristics belongs to someone who is engaged in a particular profession. If this occurs there are several factors that are not taken into account.

Base-rates

People should take base-rates into account but they often show insensitivity to prior probability of outcomes. In some professions are more people engaged than in other professions. If probability estimates are based on representativeness, prior probabilities will not be considered. This has been studied in experiments in which the base-rats were manipulated after which the subjects had to estimate the likelihood that someone was engaged in a certain profession based on descriptions. For instance, in a group of 70 engineers and 30 lawyers, the odds of being an engineer are higher than those of being a lawyer. Results show that participants base their evaluation on whether the description was stereotypical (was representative of a certain group) instead of prior probabilities. However, they did use prior probabilities if no additional information was provided to them. When no specific evidence was provided, the base-rates are accurately used but when useless evidence was given, the base-rates were neglected.

Insensitivity for sample size

Also in estimating the probability of getting a particular result in a sample drawn from a specified population, the representativeness heuristic is often used. People assume that if in a random sample the average height of men is 180 centimetres, this will be the average height in the population as well. People do not take the sample size into account. An experiment in which they had to estimate the distribution of average height for various sample size found that people produce the same distribution for all sample sizes. This was even the case when it was emphasized to take sample sizes into account. In another study people were told about a large hospital and a small hospital and that about 50% of the babies born were boys. At some days the percentage of boys reaches 60% and it was asked whether this would happen more often in the small or large hospital or that it would be about the same. People think that it happens as often in small as in large hospitals but according to sampling theory, the expected amount of days on which more than 60% of the born babies are boys is much greater in the small hospitals because a large sample is less likely to deviate from 50%. In addition, people do not take posterior probability into account. Posterior probability refers to the probability that a sample has been drawn from one population rather than from another. Estimations are often only based on sample proportion and not by the size of the sample which is necessary to know in order to estimate the posterior odds, which are often being underestimated.

Misconceptions of chance

Misconceptions of chance refers to the expectation that the sequence of events generated by a random process will be representative even if the sequence is very short. For instance, people estimate it to be more likely to get the sequence H-T-H-T-T-H than the sequence H-H-H-T-T-T because the first looks more randomly chosen. Also, H-T-H-H-H-H is estimated to be more unlikely because the fairness of the coin is not shown in this sequence. It can be concluded that people expect the essential characteristics of a process to be reflected even in a very small sample, this is called locally representativeness. One study found that these false estimations are also present in experienced research psychologists, who, as a result, used to small samples and overestimated their results.

Insensitivity to predictability

Predictions are often made based on representativeness, for instance, if one is asked to estimate the future profit of a company based on a short description, this will be influenced by the favourability of the description. In these kind of situations, people often use only favourableness and do not consider the reliability of the descriptions.

Illusion of validity

People often make predictions based on the outcome that fits best with certain input, that is, the outcome that is most representative of the input. the extent to which one is confident about his prediction also depends heavily on representativeness but not on reliability, therefore it is called an illusion of validity. This occurs in selection interviews, although it has been shown that they are highly fallible, they are still frequently used. In addition, redundancy (expressing statements multiple times to increase reliability) in input decreases the accuracy while it increases confidence.

Misconceptions of regression

Often, regression to the mean can be observed. This refers to the fact that individuals scoring high on one test score somewhat lower if they take the same test again whereas individuals who score low the first time, are more likely to score higher on the second test. This kind of regression is often not taken into account if decisions have to be made in daily life, maybe because people have the idea that the outcome should be representative of the input. because of this misconception, the effectiveness of punishment is overestimated and the effectiveness of reward is underestimated.

Availability

People often estimate probabilities of frequencies based on the ease that something pops up in their minds. An example is how often a certain event has occurred among your friends. A problem is that availability Is not only dependent on frequency and probability, therefore, basing decisions on availability leads to biases. These biases can be the result of the retrievability of instances (how difficult is it to retrieve a similar situation). This is influenced by familiarity, since familiar events come easier to mind. salience, referring to certain concepts being activated, is another aspect that is of influence. People can also be biased because of the effectiveness of a search set. This can be illustrated with a task in which people are asked whether there are more words starting with an r or with an r as third letter. they start thinking of as many words as they can and it is just easier to search for words by their first letter. Another bias is the bias of imaginability which refers to people generating instances on which they base frequency estimates. How easy they can construct instances determines their estimation of frequency. This has been tested by the question how many different committees you can form of k members if there are 10 people. The answer is maximal 252 for k=5. If the answer is based on imaginability, one would think of much more small committees than of large committees which contrasts the correct bell-shaped function. Risk estimation based on what comes easily to mind can lead to under and overestimation. The occurrence of two events at the same time is also often misinterpreted. People perceive correlations between things that are actually not correlated. This is called the illusory correlation.

Adjustment and anchoring

People often base their estimates on a starting value that they adjust to get a final value. the adjustment of the starting value is often wrong because different starting values lead to different final estimates, this is called anchoring. Giving people arbitrary starting values influences the estimate despite rewarding for the right answer. In addition, anchoring can take place in case of incomplete computation. This occurs when people have to do a multiplication with higher numbers at the beginning, they will overestimate the outcome. There can also be biases in the evaluation of conjunctive and disjunctive events. This happens when people are asked to bet on a conjunctive or a simple event whereas the conjunctive event has a lower probability of success compared to the simple event. Because of anchoring, the overall probability will be overestimated in conjunctive problems and underestimated in disjunctive problems. This is also why people underestimate probabilities of failure in complex systems. In decision analysis, experts are asked to assess a quantity such as the average value of the Dow-Jones on a certain day in the form of a probability distribution. A judge is someone who is calibrated which means that he can assess a certain percentage of quantities correctly. Getting probability distributions for a certain quantity can be done in two ways: 1. Asking the subject to select values that correspond to specified percentiles of his probability distribution and 2. By asking for the probability that a true value will exceed some specified values. Nevertheless, these methods do not yield the same results. It can be concluded that the success of calibration goes together with the used procedure to obtain the results.

Discussion

Several judgemental heuristics are described in this article and it has become clear that everyone can be affected by the resulting biases, both laymen and experienced researchers. Interestingly, people do not discover the principles of sampling and regression themselves even though they have been exposed to them their entire life. This is mainly because people do not attend to the relevant instances in life. The main contribution of modern decision theory is that it gives us a subjective interpretation of probability that can be applied to all kinds of events while being embedded in a general theory of rational decision. Although there is no formal way to test the compatibility of probability judgements with all the beliefs of the judge, compatibility of his judgements with his own knowledge, heuristics and biases is the aim of a rational judge.