[Witness Simulation Software Free Download.75
Hanne Rylaarsdam
One of the most unfortunate consequences of a legal system governed by human judgment is the alarming occurrence of false eyewitness identifications; too many individuals have been incarcerated for crimes that they did not commit. Many factors contribute to false convictions, but the principal one is faulty eyewitness identification, which contributed to the convictions in 75% of the DNA exonerations (see www.innocenceproject.org/).
A major contributor to faulty eyewitness identification is thought to be an overreliance on relative judgments (Wells, 1984). A relative judgment can involve making a selection of the lineup member who most resembles the perpetrator relative to the other lineup members, which can arise when comparisons are made across lineup members. Obviously, this is problematic if the police have an innocent suspect who is compared to foils that poorly resemble the perpetrator. On the other hand, an absolute judgment involves choosing the best-matching lineup member if, and only if, the degree of match to memory is above a criterion value. It is believed that a reliance on absolute judgments would enhance the accuracy of eyewitness identifications. This is the rationale for the recommendation that lineups be conducted sequentially (with lineup members being presented one at a time; e.g., Wells et al., 1998; for a review, see Gronlund, Andersen, & Perry, 2013).
Although absolute judgments are purported to yield better performance in lineup identifications, theoretical support for this proposal has been lacking. However, Clark, Erickson, and Breneman (2011) recently provided theoretical support by showing that a computational model (the WITNESS model, Clark, 2003) predicted that absolute judgments produced better performance than relative judgments in some circumstances.
We will begin by briefly introducing the WITNESS model. Following that, we will describe a restricted version of WITNESS (which, for pedagogical reasons, we refer to as WITNESSR) that we will compare to the original, unrestricted WITNESS model. We will show that the values taken by the relative/absolute decision weight parameters in WITNESS covary with the values taken by the decision criterion, making the decision weight parameters difficult to identify. That is, in most circumstances, WITNESSR is able to fit data as well with any value of the decision weights simply by adjusting the value of the criterion. This raises questions about the theoretical rationale for the superiority of absolute judgments, and the role of the relative/absolute judgment distinction in eyewitness decision-making.
The WITNESS model (Clark, 2003) is a direct-access matching model (for an overview of this type of model, see Clark & Gronlund, 1996) that has been adapted for eyewitness situations. WITNESS uses numerical representations of features as items in the matching process. The details of the WITNESS model are beyond the scope of this article. Instead, we highlight the aspects of the model necessary for our analysis. WITNESS is implemented in R and available as a package to download from -project.org/.
The evaluation of WITNESS versus WITNESSR proceeded as follows: First, we examined the identifiability of the w a parameter. A model is identifiable if different values of a parameter generate different predicted values (see Bamber & van Santen, 2000). We then conducted a parameter recovery simulation. Particular parameter values were used to generate response proportions, and then the resulting response proportions were fit by freely estimating the remaining parameters in order to determine the extent to which the initial generating parameter values could be recovered (see Lewandowsky & Farrell, 2011).
We will show that the WITNESS model is only partially identifiable, because the decision weight parameters cannot be uniquely specified. We will also show that the WITNESS model has a high degree of parameter variability involving w a and c, meaning that the parameter values recovered by a fitting algorithm vary greatly across iterations. Both of these analyses reveal that the values of the decision weight parameters are poorly identified and not well constrained by existing data. This makes any theoretical interpretation of the values or rank order of these parameters problematic.
We investigated the behavior of the decision weights at the bivariate level using contour plots. Contour plots allow one to plot one of the variables as a color, thus showing the relationship between three variables without having to resort to three-dimensional plots. For example, suppose that we want to investigate how a and SFS interact to create a particular RMSE value.Footnote 1 Using a contour plot, we can plot a on the x-axis and SFS on the y-axis, with RMSE ranging from light to dark throughout the plot. Darker areas of the plot show the combinations of a and SFS values that produce the best fit (i.e., the closest approximation between model and data). If two parameters are poorly identified at the bivariate level, then we should see large areas of the figure where nearly equal fits are obtained (i.e., large dark areas). Conversely, if two parameters are identifiable at the bivariate level, then we should see only a small dark area in the plot, signifying a tight area of best fit centered on the parameter values that generated the predicted response proportions.
We chose a set of parameter values that produced response proportions for WITNESSR and WITNESS (see Table 1) that were similar to one another and typical of actual data. We varied each parameter between 0 and 1 (except for c, which never exceeded .2) in increments of .02, resulting in 50 different values for each parameter. We then crossed each of these values with 50 different values of w a (ranging from 0 to 1). For each of these parameter combinations, we iterated the WITNESS and WITNESSR models 10,000 times in order to estimate response proportions with stability, and then calculated RMSE. The result was four different 50 50 matrices of RMSE values for every combination of parameter values. These results were plotted as described above in a contour plot. Figure 1 depicts WITNESS, with w a on the x- axis and the other four parameters, one at a time, on the y-axis. Figure 2 depicts WITNESSR (with w a set to 1.0); the different plots show interactions between all of the pairs of parameters. The crosshairs in the plots indicate the parameter combinations used to generate the response proportions in Table 1.
We fit the response proportions from Step 1 using a steepest-descent algorithm, to see how well the models could recover the generating parameter values. If the fit was better than RMSE = .05, the parameter values were retained for later consideration. Otherwise, the parameter values were ignored, because the fitting algorithm failed to produce acceptable fits. Note that our intent was not to find the best parameter values. If this had been our intent, we would have used another algorithm less sensitive to local minima, such as a genetic algorithm or a simulated annealing procedure. Instead, our intent was to understand the degree of variability in parameter estimation under conditions that produced what we have observed to be typical fits.
Figure 5 shows three histograms. The two on the top correspond to the original WITNESS model; the left histogram shows the distribution of the recovered c parameters, and the right histogram shows the distribution of the recovered w a parameter. The bottom histogram corresponds to WITNESSR and shows the values of the c parameter.
Distribution of parameter recovery values across 1,000 iterations of a steepest-descent fitting algorithm. The bottom plot is the distribution of the c parameter for WITNESSR, and the top plots correspond to the c and w a parameters for the original WITNESS model
In WITNESS, the absolute match value is necessarily higher than any relative match value, because the relative match value is the absolute match value minus the next-best match value. Consequently, if WITNESS puts more weight on the higher evidence (the absolute match), it necessarily makes the total evidence have a higher value. But we have shown that an appropriate increase in the criterion generally can maintain the same predicted response proportions. As a consequence, the w a /w r parameters are difficult to identify within the current implementation of the WITNESS model. We believe that this undermines the theoretical support offered by Clark et al. (2011) for the superiority of absolute judgments in eyewitness identification decision-making. Moreover, the theoretical rationale for the relative/absolute conceptualization of eyewitness decision-making should be reassessed in light of these findings.
One alluring aspect of the relative-versus-absolute judgment concept is that it is introspectively intuitive. It is easy to conceptualize the experience of a relative decision process: We evaluate multiple options and seemingly make comparisons amongst them, and it seems logical for these comparisons to be incorporated into the decision process. As we previously mentioned, several studies have found relationships between the experiential self-reports of absolute or relative decisions and witness accuracy (Dunning & Stern, 1994; Kneller et al., 2001; Lindsay et al., 1991). However, experiential reports, intuition, or introspection do not validate psychological processes, and formal modeling allows researchers to investigate phenomena in a manner constrained so as to control natural human biases and errors of intuition (see Hintzman, 1991). For example, Hintzman (1986) demonstrated that a single-store exemplar model of memory accounted for perceived abstractions from episodic memory, despite conventional intuition that a separate abstraction process and memory store were necessary to explain this phenomenon.
795a8134c1