Clark, S.E. (2003). A memory and decision model for eyewitness identification. Applied Cognitive Psychology, 17, 629-654.

Clark's article describes a computational model of memory called WITNESS. Clark designed this model to help understand the eyewitness identification process. Here are the main assumptions of WITNESS:

1. Information is represented by vectors made up of features (or elements). For simplicity, in the example below I created vectors where the features are 1, -1, or 0.

2. Assume you have a vector that represents the perpetrator's features and that vector is called P. The witness's memory for the perpetrator will be another vector call M. A person’s memory won’t exactly match the perpetrator because memory is imperfect. In other words, M will be similar to P, but won’t exactly match P.

3. How close M is to P depends on the probability that any individual feature will be stored correctly. The parameter c represents the probability that any feature in M will match the corresponding feature in P. Psychologically c is going to vary as a function of things like viewing conditions, retention interval, degree of attention and so on. In the example below I set c to .80. In Clark’s simulations c ended up being much lower than that.

4. The members of the lineup itself are also mentally represented as feature vectors. The foils are represented by vectors Fj and the suspect by S_g or S_i depending on whether the person is guilty or innocent. For simplicity Clark assumes that S_g = P. How similar the innocent suspect is to the perpetrator is given by the parameter S(S,P). This parameter gives the probability that any feature of P will also occur in S_i

5. How similar each foil is to the guilty suspect is given by S(F_j,P) and the similarity between each foil and the innocent suspect is given by S(F_j,S).

6. Recognition memory depends on determining how similar each lineup member is to what's stored in memory. Similarity is computed by multiplying each element of one vector by the corresponding element of the other vector, then adding all of those products and dividing by the number of non-zero elements (see the figure below for an example).

7. The example below provides a simplified situation of a target present lineup and foils that have been matched to the suspect where S(F_j,P) =.50

Absolute Versus Relative Judgments

WITNESS is designed to simulate both absolute and relative judgment processes. Absolute judgment processes are modeled as cases where the similarity of whichever lineup is the best match exceeds some a response criterion. Relative judgments are modeled as cases where the best match exceeds the next best match by some criterion. In WITNESS the lineup judgment is determined by a weighted mixture of the two processes. If this combined value exceeds the response criterion (C_ID ) then the person selects the best match. If all the lineup members are below the lower response criterion (C_REJ ) then the person says that the perpetrator is not in the lineup. If the value is someplace in between the witness will say “I don’t know.”

If w_ABEST + w_RDIFF > C_ID then select the best match

If BEST < C_REJ Reject the lineup

If neither of these things happens the person responds ‘don’t know’

BEST represents the similarity between the best match and what is stored in memory. DIFF is the difference between the top match and the next best match. The weights in this case indicate the degree to which absolute or relative judgments predominate with w_A + w_R = 1. You could simulate a pure absolute judgment process by setting w_A =1. You could simulate a pure relative judgment strategy by setting w_R = 1. What’s neat about this is that it assumes choices are based on both the absolute strength of the match as well as how strong that match is relative to other choices. 

Fits to Actual Data

Clark used simulations of his model to fit data from three different experiments that compared the match to suspect with match to description methods. The first fits were to data from Tunnicliff and Clark (2000). It worked pretty well for the suspect matched data (Figure 3 on page 638). It didn’t work quite as well for the description matched. In particular, the reject and don’t know responses didn’t match up very well.

Clark next fit data from Juslin et al. (1996). These fits weren’t quite as good and again the description match case was especially problematic (Figure 4 on page 639). Clark argues that this is because the false identification rate was substantially above chance suggesting that the innocent suspect was substantially more similar to the perpetrator than were the foils. When this assumption was built into the model Clark says the fits were “dead on”.

Lastly, Clark fit data from Wells et al. (1993). The model had real trouble matching this data, particularly because the lineup rejection rate was somewhat (although not significantly) higher for the target present lineup in the suspect matched case. This is a puzzling finding when you think about it. Why should people be more likely to reject a lineup that contains the suspect than one that doesn’t? The model predicts the opposite result. Clark also points out that one might predict the low correct identification rates in the suspect matched case by making c very low, but he points out that won’t work because identification accuracy is pretty good in the description matched case.

Clark tried to deal with these issues in two ways. First, because one of the problems appeared to be the lineup rejections he combined the rejections and 'don't know' responses and tried fitting that. The model still didn't do a very good job matching the data (Figure 5 on page 642). Clark also examined the possibility that the foils in the match to suspect case were similar not only to the innocent suspect but also to the actual perpetrator. The model did better if one made that assumption, but still had trouble with the lineup rejections and 'don't know' responses.

Conclusions

The WITNESS model fit many aspects of the data from the three experiments, but had trouble with other aspects. The similarity parameters were the main thing that differed across the experiments (Table 5 on page 646). Clark also talks about "Consistent Mispredictions of the Model". For instance, the data from the experiments consistently showed baseline conditional probabilities of selecting the innocent suspect at rates greater than chance in the description based lineups. The model also predicted larger differences between target present and absent lineups that occurred in the actual experiments. The model also didn't do a great job with the lineup rejections.

Clark acknowledges certain limitations of the current model, some of which involve the necessarily noisy data that he was trying to model, others have to do with simplifying assumptions that are necessary to make in early stages of model development. These include how to model the match to description process for choosing foils. In the present simulations this was done by creating foils that had a certain degree of similarity to the perpetrator [i.e. S(F,P)]. This captures the fact that in the match to description method, foils are not picked to match features of the suspect [i.e. S(F,S)] but rather to match features of the description, which itself is based on the witness's memory of the suspect. The model also makes the simplifying assumption, as do many models of memory, that memory is represented by a vector of features. Clark also examined some other ways of modeling the decision processes and they did not work as well as the mixed approach that he ended up adopting.

In this article Clark developed an interesting model of how basic memory processes are combined with decision rules specific to the eyewitness task and used the model to fit data from three experiments. The model did a good job accounting for many of the qualitative pattern of results from those experiments, and Clark has provided plausible accounts for the cases where the model didn't quite capture the data.