Base-Rate Neglect

Source Results: Table 13.2, Study 2

The excerpt does not include the table of stimuli and responses.

Figure 13.3 plots median judgments of confidence as a function of Bayesian posterior probability for high (.90) and low (.10) prior probabilities. The figure shows that subjects were overconfident in the low base-rate condition and underconfident in the high base-rate condition.

The graph itself is not included in this excerpt.

These results are consistent with Grether’s (1980, 1992) studies concerning the role of the representativeness heuristic in judgments of posterior probability. Unlike the present study, in which both prior probabilities and data were presented in numeric form, Grether’s procedure involved random sampling of numbered balls from a bingo cage. He found that subjects overweighted the likelihood ratio relative to prior probability, as implied by representativeness, and that monetary incentives reduced but did not eliminate base-rate neglect. Grether’s results, like those found by Camerer (1990) in an extensive study of market trading, contradict the claim of Gigerenzer, Hell, and Blank (1988) that explicit random sampling eliminates base-rate neglect.

Our analysis implies that people are prone to overconfidence when the base rate is low and to underconfidence when the base rate is high. Dunning, Griffin, Milojkovic, and Ross (1990) observed this pattern in a study of social prediction. In their study, each subject interviewed a target person before making predictions about the target’s preferences and behavior (e.g., “If this person were offered a free subscription, which magazine would he choose: Playboy or the New York Review of Books?”). The authors presented each subject with the empirically derived estimates of the base-rate frequency of the responses in question (e.g., that 68% of prior respondents preferred Playboy). To investigate the effect of empirical base rates, Dunning and colleagues analyzed separately the predictions that agreed with the base rate (i.e., “high” base rate predictions) and the predictions that went against the base-rate (i.e., “low” base rate predictions). Overconfidence was much more pronounced when base-rates were low (confidence = 72%, accuracy = 49%) than when base rates were high (confidence = 79%, accuracy = 75%). Moreover, for items with base rates that exceeded 75%, subjects’ predictions were actually underconfident. This is exactly the pattern implied by the hypothesis that subjects evaluate the probability that a given person would prefer Playboy over the New York Review of Books on the basis of their impression of that person with little or no regard for the empirical base rate; that is, the relative popularity of the two magazines in the target population.