Executive Summary

Research continues on the effects of “system“ and “estimator” variables on the accuracy of eyewitness identification. “System” variables are factors that are under the control of law enforcement; e.g., the number of persons in a lineup and the lineup instructions given. “Estimator” variables reflect conditions at the scene, such as levels of light, distance between the eyewitness and perpetrator, or presence of a weapon at the time of the incident.

Some of the key results from recent research include:

1. An eyewitness’ level of confidence in an identification and the speed in which the identification is made can be reliable predictors of accuracy under some conditions affected by system and/or estimator variables. Both are stronger predictors of accuracy when the eyewitness makes an identification from a lineup than when the eyewitness responds that the suspect is not in the lineup.
2. Simultaneous lineups appear to elicit higher overall performance than sequential lineups, although sequential lineups orient eyewitnesses to use a more conservative decision process.
3. Recent research on constructing lineups indicates that eyewitness identification performance appears superior when the filler faces in the lineup are dissimilar rather than similar to the suspect, as long as the faces all match the suspect’s description.

While receiver operator characteristic (ROC) curves are useful for comparing two procedures in light of another variable (e.g., expressed confidence level), traditional statistical methods such as logistic regression are more powerful for comparing the accuracies of two procedures (e.g., sequential versus simultaneous lineups; “blind” versus “non-blind” lineups) particularly when multiple factors are involved (e.g., weapon presence versus absence; good versus poor lighting).

Measures and displays that consider both positive predictive value and negative predictive value provide additional metrics of performance in proposed eyewitness identification procedures. Most analyses consider only positive predictive value (probability that the ID was correct), but negative predictive value is also critical (probability that exclusion was correct) – for purposes of both low false positive rates and for recognizing that the true perpetrator remains to be identified.

“Machine learning” algorithms such as random forests can be useful for predicting the likelihood of “choosing” (versus “not choosing”) a suspect from a lineup, and for estimating the accuracy of the decisions, based on conditions of the scenario (e.g., delay, presence of weapon).

With large studies of 1000+ participants, these newer approaches can model complex interactions among system and estimator variables.

Ecological validity of online or lab experiments, versus more realistic experimental conditions, needs to be evaluated.

The response by law enforcement agencies, lawmakers, and judges to research in the eyewitness area has been striking:

Change has been steady, and the pace of change has been more pronounced in the six years since the NAS report was released.

Several state courts have reconsidered their rulings and developed new approaches informed by eyewitness evidence research.  States have enacted laws requiring that police agencies adopt eyewitness identification policies meeting minimum criteria.  Other states have adopted model policies, adopted new jury instructions, or otherwise promoted reception of research in police policies. Policing organizations have adopted new policies.

Future work should solidify these gains, through investment in training and further dissemination of these best practices.

However, some aspects of law enforcement and judicial practice has been resistant to research recommendations, including videotaping lineups and limiting subsequent identifications and confidence statements in the courtroom. Reform efforts should focus on those challenges.