1. Main findings?
2. Implications?
3. Future research direction?
4. Conclusions?

Accuracy, Confidence, and Juror Perceptions in Eyewitness Identification Gary L. Wells, R. C. L, Lindsay, and Tamara J. Ferguson University of Alberta, Edmonton, Canada Subjects of both sexes individually witnessed the staged theft of a calculator. The 127 witnesses were then given the opportunity to identify the thief from a six-person picture array; from this sample, 24 accurate-identification witnesses and 18 inaccurate-identification witnesses were cross-examined with either leading or nonleading questions. Jurors were unable to distinguish accurate from inaccurate witnesses across the 42 cross-examination sessions (d' = .02). However, jurors in the leading-questions conditions were significantly more likely to believe accurate than inaccurate witnesses ( .05). However, because of the significant main effect for crossexamination style on the correctness measure, we calculated separate d's within cross-examination types. Within the leading-questions condition, the proportions of hits (.84) and false alarms (.73) produced evidence that jurors could reliably detect accurate from inaccurate witnesses (d' = .35, z = 3.98, p < .01). Surprisingly, when d' is calculated for the nonleading questions condition, the resultant value is negative (d' .39, z = 4.53, p < .01) owing to the fact that false alarms exceeded hits (.86 vs. .76, respectively). Jurors' confidence that they made a correct decision was unaffected by cross-examination type or witness accuracy. However, jurors in the leading-questions conditions felt significantly less confident that they could generally tell from a witness' testimony whether the witness was accurate than did jurors in the nonleading-questions conditions, M =3.83 and 5.32, respectively; F(l , 159) = 9.83, p < .01. Also, jurors in the leading-questions conditions expected a significantly lower percentage of witnesses on the average to be accurate than was expected by jurors in the nonleadingquestions condition, M = 54.6% and 62.9%, respectively; F(l, 159) = 8.62, p < .01. No other main effects or interactions were significant on the jurors' questions. Principal Inter correlations Five variables seemed particularly fruitful for further analysis regarding their relationships to one another: jurors' decisions to believe or disbelieve the witness; jurors' confidence in their decision to believe or disbelieve the witness; jurors' correctness in their decision to believe or disbelieve the witness; jurors' attributions of confidence to the witness; and witnesses' self-rated confidence. An intercorrelation matrix was constructed for these five variables and is presented in Table 3. The results of this analysis reveal that jurors' correctness in their decisions to believe or disbelieve the witness was unrelated to any of these variables. Although the correctness variable was unrelated to any other variable, the remaining four variables were all significantly related to one another. Perhaps most impressive is the fact that jurors' ratings of the witnesses' confidence accounted for nearly 50% of the variance in jurors' decisions of whether or not to believe the witness. Discussion We began our investigation with the question of whether or not a witness who makes a false identification of a defendant will have as much courtroom impact (as defined by the proportion of jurors who believe the witness) as a witness who makes an accurate identification. Our overall d' analysis indicated that jurors' belief decisions were unrelated to whether or not the witness that they observed had made an accurate identification. This independence between witnesses' accuracy and jurors' belief inferences is less distressing, however, when the d' analyses are calculated within cross-examination types. Specifically, there was a significant positive d' in the leading cross-examination condition, indicating that jurors could discriminate beTable 3 Principal Intercorrelations Variable 1. Jurors' believe/not believe decisions 2. Jurors' confidence in their decisions 3. Jurors' correctness in their decisions 4. Jurors' attributions of confidence to the witness 5. Witnesses' self-rated confidence .344* .051 .19 .706** .589** .126 .527** .348* .194 .549** Note. All correlations used the mean of jurors' responses for each session and thus are based on n *p <.05. **p < .01. 42.446 G. WELLS, R. LINDSAY, AND T. FERGUSON tween accurate and inaccurate witnesses. Unexpectedly, however, the nonleading crossexamination condition yielded a significant negative d' value. The negative d' value also indicates that jurors could discriminate between accurate and inaccurate witnesses, but jurors' decisions to believe appear to be governed by a significantly incorrect decision rule. In psychophysical terms, a negative d' seems to indicate that the subject responder has reversed his or her interpretation of noise only versus signal plus noise. Even though jurors' estimates of the witnesses' confidence were strongly related to jurors' decisions to believe or disbelieve the witness, it appears that the confidence of the witness was not the signal that allowed jurors to make accurate discriminations in the leading cross-examination conditions. Nor did the confidence of the witness serve as a faulty signal that could account for jurors' inaccurate discriminations in the nonleading crossexamination conditions. The evidence for this is contained in the fact that witnesses' confidence, whether measured by the witnesses themselves or through jurors' estimates, was unrelated to witnesses' accuracy for the 42 cross-examined witnesses. This was true of both the leading and nonleading cross-examination conditions. In fact, the only witness measure that was significantly related to witnesses' accuracy was witnesses' responses to the question, "What percentage of witnesses would be accurate under these conditions?" The fact that this question produced a significant effect for accuracy and no interaction with cross-examination style suggests that it may have fruitful validity in helping to distinguish between accurate and inaccurate witnesses. We also found no evidence to support the differential-attrition hypothesis. In fact, the data show a nonsignificant reversal effect wherein two of our accurate-identification witnesses refused to be cross-examined, whereas only one of our false-identification witnesses refused. Interestingly, subjects' verbal reports of to what extent they believe that they can tell whether a witness is accurate by listening to her or his testimony produced results that were opposite to what might be expected based on the d' analyses. Specifically, jurors in the leading-question conditions felt less competent at detecting whether a witness was accurate than did subjects in the nonleadingquestion conditions. This apparent discrepancy between subjects' verbal reports and their behavioral performance may parallel Evans and Wason's (1976) contention that people have little or no access to higher order cognitive processes. The data indicate that although jurors' decisions to believe the witness are highly related to their ratings of the witnesses' confidence, the confidence-accuracy relationship is very poor. Recent research in the area of probability calibration suggests that low accuracy-confidence relationships may be related to a lack of relevant "feedback experience." Specifically, recent research by Lichtenstein and Fischhoff (1977, Experiment 2) indicated that there was practically no relationship between accuracy and confidence on a task that was novel (e.g., identifying whether a handwriting specimen was American or European). However, after allowing subjects to study a different sample of handwriting specimens, correctly labeled as American or European, subjects' confidence was significantly related to accuracy of decision. Do people get the kind of feedback in daily encounters that should allow them to be well calibrated regarding confidence-iaccuracy relationships in eyewitness identification? Perhaps not. In fact, we argue that daily encounters and the assumptions on which they operate might serve to maintain the poor accuracy-confidence relationship and enhance overconfidence. First we note that with rare exceptions, the recognition of one's own friends is not the same task as that required in criminal identifications. Criminal identifications typically involve a single exposure followed in time by a recognition test. We often see someone in a hallway, on the street, or at a party who we believe we recognize, but, not knowing the person well, we simply smile or make some ambiguous gesture of recognition. Such gestures are almost always reciprocated and, thus, we consider our belief (i.e., that we made a correct recognition) to beEYEWITNESS IDENTIFICATION 447 confirmed. However, in signal detection terms, we have failed to distinguish between a "hit" and a "false alarm," with the verdict almost always being that it was a hit. Our perceptual set has led us to think of such events as confirmations. Of course, in some cases we do get feedback. For example, at a social gathering a third party might take initiative of the form, "Have you and John met?" However, the nature of social intercourse usually takes the form of a pause following such a question so that if one person says "yes," the other person will be able to avoid an embarrassing "no" response. Indeed, there seems to be a general belief that if one person says "yes," both individuals believe that they have previously met; the remainder of the conversation is to determine where or when they met. While this may provide the person who says "no" with a disconfirmation, the overall nature of daily experience is that if one believes that he or she recognizes another person, it is practically disconfirmable. Thus, people's daily experiences may lead them to believe that if they make a recognition, they are correct. Because they lack veridical feedback experience, people fail to be sensitive to factors leading to correct and incorrect recognitions which, in turn, leads to eyewitnesses' poor accuracy-confidence calibration. Perhaps, as in Lichtenstein and Fischhoff (1977, Experiment 2), eyewitnesses can be trained for accuracy-confidence calibration. However, the current evidence indicates that self-rated and overtly expressed confidence is largely irrelevant in determining the criminal-identification accuracy of an eyewitness. This relative lack of relationship between accuracy and confidence, in turn, hinders the opportunity for differential attrition among accurate versus false identification witnesses and makes the jurors' task less likely to yield veridical decisions. Finally, the significant effect for type of cross-examination on jurors' correctness in their decisions suggests that future research may delineate factors that can increase the veridicality of jurors' decisions. Generalizing to the Courtroom Although the implications of the current research for criminal justice are rather obvious, we must for the moment insert a limit on the extent to which we are willing to generalize. Specifically, it must be remembered that the confidence-accuracy data and the jurors' detection data were based on the use of a "constant crime" design. By constant crime, we mean that each witness saw approximately the same event for the same amount of time, and so on. Suppose, however, that the crime was different for each witness. It may be that accurate and inaccurate witnesses do not differ in their confidence within a given crime, but, given a sample of witnesses who observed different crimes, the relationship may exist. It seems only reasonable, for example, that witnesses who observe a criminal for 15 sec will be both less accurate and less confident than witnesses who observe a criminal for 15 min. Thus, although the confidence-accuracy relationship may be weak or nonexistent as measured between witnesses within an event, it may exist as measured between events. A similar generalization problem exists over whether jurors are "overbelieving" of witnesses or not. The current data indicated that among witnesses who made a choice, 74% were accurate; a value that is reasonably close to the 79.8% figure that represents the jurors' rate of belief. However, it is not at all clear what would happen to the jurors' rate of belief if the crime would have been of the type that produced a low rate of accuracy. It is possible that the jurors' rate of belief is around 80% irrespective of the actual rate of witness accuracy. Reference Note 1. Wells, G. L., et al. Guidelines for empiricall