ORGANIZATIONAL BEHAVIOR AND HUMAN DECISION PROCESSES 49, 188-207 (1991) Overconfidence in Estimation: Testing the Anchoring-and-Adjustment Hypothesis RICHARD A. BLOCKANDDAVID R. HARPER Montana State University People are frequently overconfident in the accuracy of their estimates of uncertain quantities. The present study requested 50%- or 90%-confidence ranges. Overconfidence is shown when less than the target percentage of ranges include the true value. Tversky and Kahneman (1974) proposed that people use an anchoring-and-adjustment heuristic: They begin with a starting value, one supplied to them or generated by them, and insufficiently adjust their estimates around this anchor. The present data support the proposed anchoring process. If subjects receive another person’s point estimates, their own implicit point estimates are correlated with these values. However, an- choring-and-adjustment processes do not invariably produce overconfidence. Subjects who receive anchors are no more overconfident than are those who do not receive anchors. If subjects are required to produce a point estimate first, overconfidence decreases; processes involved in explicitly displaying the point estimate are implicated. Overconfidence may occur because people do not realistically assess their estimation ability. 8 1991 Academic press, IK. A person usually does not obtain all of the relevant information before making a judgment or decision. Even if the information is potentially available, such as in reference books, a person may not invest the time and energy needed to obtain it. Cognitive conceit-that is, overconfi- dence in personal cognitive attributes, such as knowledge and decision- making abilities-may also decrease the likelihood that someone will seek potentially important information. Instead, a person may simply estimate relevant quantities, such as numerical attributes of objects and probabil- ities of events. Difficulties may ensue if a person has only imprecise information available in memory concerning these objects or events. Alpert and Raiffa (1982) conducted an important study in which sub- jects estimated ten uncertain quantities, such as the total US egg produc- Address correspondence to Richard A. Block, Department of Psychology, Montana State University, Bozeman, MT 59717-0001. David R. Harper is now at Bowling Green State University. We thank the students who were involved in collecting and analyzing these data, especially Joseph Mangiantini and Win Turner. We also thank Robert Patterson, Wesley Lynch, and several anonymous referees for helpful comments on the manuscript. Some of these data were presented at the 1986 and 1987 meetings of the Psychonomic Society and at the Fourteenth Interdisciplinary Conference in 1989. 188 0749-5978191 $3.00 Copyright 0 1991 by Academic Press, Inc. All rights of reproduction in any form reserved.OVERCONFIDENCE IN ESTIMATION 189 tion in 1965. Alpert and Raiffa explained in detail how the subjects should assess probability distributions for these quantities. They used the method of direct fractile assessment; for example, a person gives a lower quartile estimate (i.e., a .25 fractile) and an upper quartile estimate (i.e., a .75 fractile). These two estimates define a 50%-confidence interval, or a range that the person thinks is as likely to include the true value as to exclude it. If a person estimates several quantities, the percentage of ranges that include the true value can be compared to the target percent- age, which in this example is 50%. If someone realistically assesses the extent of his or her knowledge concerning the quantities, in the long run exactly 50% of the ranges should contain the true value. Alpert and Raiffa found, instead, that only about 33% of the ranges did. This is a classic overconfidence effect, since people appear to be unjustifiably confident in the accuracy of their ranges (cf. Adams & Adams, 1961). Pitz (1974) obtained similar data showing what he called a hyperprecision effect.’ He also found that subjects are less overconfident when they estimate quan- tities about which they presumably have greater knowledge. Lichten- stein, Fischhoff, and Phillips (1982) concluded that “the most pervasive finding in recent research is that people are overconfident with general- knowledge items of moderate or extreme difficulty” (p. 314) and that “overconfidence is most extreme with tasks of great difficulty” (p. 315). Tversky and Kahneman (1974) suggested that a person performing this kind of task implicitly uses a heuristic called anchoring and adjustment. Along with two more well-known heuristics (representativeness and availability), people are assumed to use this heuristic in the process of making judgments under conditions of uncertainty. Tversky and Kahne- man reported an anchoring effect in an experiment in which subjects received arbitrary starting values. Each subject estimated various quan- tities, such as the percentage of African countries belonging to the United Nations. Before each estimate, the experimenter spun a “wheel of fortune” in the subject’s presence, and the subject first judged whether the resulting number was higher or lower than the true value of the quan- tity. Then the subject estimated the quantity by adjusting upward or downward from the starting value. Median estimates of the percentage in i Peterson and Pitz (1988) recently studied how subjects use information in a prediction task. They preferred the term uncertainty to the ones used in previous articles, hyperpre- cision (Pitz, 1974) and overconfidence (Peterson & Pitz, 1986). In this usage, the term uncertainty refers to a belief about possible values of a quantity, and confidence refers to a