Subjective confidence is usually thought of as the degree to which a person believes they are correct about a judgment and are willing to say so. Confidence can be important when there is no objective guide to accuracy; in these cases, decision makers will usually prefer to make the judgment in which they have the greatest confidence; confidence can drive further behaviors (Weber, BÃ¶ckenholt et al. 2000). Accordingly, there has been some concern that decision makers have appropriate levels of confidence in their judgments.
Subjective confidence can be measured in several ways. People are regularly asked to state their confidence verbally or numerically: “Just how sure are you of that?” Confidence in an outcome is also measured, perhaps more subtly, by asking how willing someone is to take a bet that the outcome they predict will actually occur. A person who is willing to take a bet that pays $10 if they are correct, but costs them $50 if they are wrong must be quite confident that they are correct!
An important measure of the accuracy of someone’s confidence judgments is their calibration. Calibration refers to the relationship between the numerical confidence than an event will occur and the actual frequency with which the event occurs. For example, if we followed a group of 100 patients presenting to an ER with chest pain for whom the attending physician had estimated a 30% chance (each) of a myocardial infarction, and 30 of those patients actually were diagnosed with an MI, the attending physician would be said to be perfectly calibrated in his judgments. That is, when he gives a 30% confidence of MI, it accurately reflects a 30% chance of MI.
Few people – physicians or patients – are well-calibrated in their confidence judgments around medical uncertainty. Overconfidence, in particularly, has been repeatedly demonstrated. Clinically promising approaches to improving calibration include discussion and searching for conflicting evidence. Koriat, et al. (1980) reduced overconfidence by asking people to consider conflicting evidence that would weigh against their initial belief. Arkes, et al. (1987) showed that group discussion with peers could improve calibration as well. Both of these techniques encourage deeper consideration of choice alternatives, and allow for the combination of multiple viewpoints (Weber, BÃ¶ckenholt et al. 2000; Armstrong 2001); moreover, collegial discussion of cases is a venerable tradition in medicine, and occurs naturally in most group practices.
Although we most often think of confidence as a statement about belief in the accuracy of a judgment, Weber et al. (2000) provided a convincing demonstration that confidence may instead reflect a lack of conflict about the decision. In their study, they asked 84 physicians to generate most-likely and second-most-likely diagnoses for cases, and to give their confidence in each diagnosis and in the proposition that the correct diagnosis was somewhere in their set. Confidence in the set should always be higher than confidence in any single diagnosis (and it was); moreover, if confidence reflects belief in accuracy of judgment, the confidence in the set should be higher when the top two diagnoses are both likely than when one is likely and the other considerably less likely. Instead, confidence was reduced when the top two diagnoses were both judged to be quite likely.
The authors conclude that expressions of confidence may actually be expressions of a lack of decision conflict. When there is only one likely diagnosis, the confidence in the set is high, because there is little conflict about which diagnosis is correct. When there are two likely diagnoses, however, there is much more conflict about which is the most likely diagnosis, and this is reflected in a lower overall confidence for the set. In support of this theory, physicians who mentioned a rival hypothesis when discussing their reasoning for selecting their most-likely diagnoses also tended to have lower confidence in that diagnosis (as well as in the set of diagnoses).