Error in Confidence Judgments

摘要

People are sometimes overconfident in their decisions, at least in laboratory settings. Or are they? Erev, Wallsten, and Budescu (1994) provided a demonstration that error could produce an overconfidence effect, depending on how data are analyzed. Justin, Winman, and Olsson's (2000) position was stronger; they suggested that there was "little or no evidence for an information-processing [overconfidence] bias in human judgment" (p. 388). While we agree with both Erev et al. and Juslin et al. that error can produce overconfidence effects, our paper outlined a problem with Juslin et al.'s technical approach to separating error from "real" overconfidence: such a procedure, when applied to confidence judgment data, will happily remove overconfidence effects regardless of their source (random error or "real" overconfidence due to systematic biases). Furthermore, their procedure gives unrealistically large estimates of the extent of such error. The fact that the procedure can lead to erroneous conclusions means it cannot be used to support such arguments as there being "little or no evidence for an information-processing bias in human judgment." Olsson et al.'s reply to our paper touches on three issues: (1) our error estimates are implausible and inconsistent with other findings; (2) our demonstrations notwithstanding, measurements of bias can only be taken seriously when error is accounted for; and (3) we misrepresent their position. We discuss each issue in turn below.