witness.est: Estimate identification probabilities using WITNESS

In dustinfife/witness: This package fits eyewitness data using Clark's (2003) WITNESS model

Description

witness takes the parameters provided by the user then simulates the WITNESS procedure and outputs probabilities

Usage

  witness.est(parameters, guilty = NULL,
    perp.removed = NULL, sequential = NULL, data = NULL,
    ilen = 100, lsize = 6, meth = "WITC",
    suspectMatch = FALSE, seedit = TRUE,
    sample.size = 1000, optim = FALSE, goodFoil = FALSE)

Arguments

parameters	A parameters matrix. See details.
guilty	A vector of logical values (i.e, TRUE or FALSE). This vector indicates which rows in the dataset contain the guilty suspect.
perp.removed	A vector of logical values indicating which rows have the perp removed from the lineup.
sequential	A vector of logical values indicating which rows should be fitted with a sequential lineup.
data	Optional. Provide the dataset to be fit and this function will compute the objective function
ilen	The length of the features vector. Default is 100
lsize	Lineup size. Default is 6.
meth	Either "WITC" or "original". "WITC" will fix the wa/wr parameters, while "original" will allow them to vary.
suspectMatch	Logical. Should the foils be matched to the suspect?
seedit	Logical. Should the random seed be fixed? This should be set at true when using a fitting algorithm so that differences between iterations are due to difference in fit alone, not because of random noise.
sample.size	Defines how many iterations the WITNESS function performs.
optim	Logical. When TRUE, the function returns a single value (the fit of the model to the data). This must be specified when using a optimization routine (e.g., genetic algorithm).
goodFoil	Logical. Should one of the foils be a "dead ringer"?

Details

The WITNESS model (Clark, 2003) is a direct-access matching model that has been adapted for eyewitness situations. WITNESS uses numerical representations of features as items in the matching process. More specifically, each lineup member is represented as a vector of n arbitrary features (the features do not map to literal features of a person, such as eyes or a nose, however). The value n is typically fixed to some arbitrary constant such as 100. Each position in the vector (or each feature) is assigned a value between -1 and 1.

To begin, WITNESS generates a perpetrator vector (PERP) by randomly generating numbers on the interval -1, 1. This vector serves as the basis for all of the subsequent lineup members. The model next "encodes" the features in the PERP vector to a memory vector. The parameter a governs the degree to which the Memory (MEM) vector matches the PERP vector; a is the probability that each individual feature in PERP will be successfully encoded to MEM. For example, if a = .4, then each feature of PERP will have a 40% chance of being represented within MEM. Conversely, there is a 1-a (60%, in this case) chance that the feature will be replaced with another arbitrary random feature. Put differently, the correlation coefficient between PERP and MEM will be approximately a. This results in a noisy representation in memory, which accounts for phenomena such as the failure to encode details of the perpetrator and post-event interference.

WITNESS next creates the lineup members for comparison to memory. For target absent lineups (i.e., an innocent suspect replaces the guilty suspect), a new vector is created to represent the innocent suspect (SUSP). This vector is governed by the parameter SSP, or Similarity of the Suspect to the Perpetrator. SSP is the probability that each feature of SUSP will match the corresponding feature of PERP. As SSP approaches 1, SUSP will be more similar to PERP (and likewise, less similar as it approaches 0).

Next, WITNESS uses one of two more parameters to create the remaining lineup members (i.e., the foils), each to simulate a different method of foil selection. We refer to these vectors of foils as FOILS. Foils can be selected either because they match the description of the perpetrator (description-matched foils) or because they match the appearance of the suspect (suspect-matched foils). The parameter SFP, or Similarity of the Foils to the Perpetrator, simulates a description-matched lineup. SFP is the probability that a given feature of a given foil will match a corresponding feature in PERP (similar to SSP, where 0 yields no shared features and 1 replicates PERP). Note that this results in the same foils being used in target-present and target-absent lineups. SFS, Similarity of the Foils to the Suspect, is a similar parameter; however, it uses the suspect in each given lineup (PERP or SUSP) to generate the foils. Note that this results in different foils being used in target-present and target-absent lineups. In either case, the final result is a lineup constructed of vector representations of a suspect (guilty or innocent) and five foils, all of which share common features.

Following the construction of the lineup, the model must simulate the actual lineup procedure. WITNESS accomplishes this by comparing each lineup vector (i.e., PERP, SUSP, and FOILS) to MEM to create match values, or assessments of the degree to which two vectors overlap, for each lineup member. These match values are the dot products of each lineup vector to MEM divided by the total number of features. Larger dot products indicate a closer match to memory. After computing these match values, the values are used to execute the decision aspect of the lineup.

In order to model relative/absolute contributions, WITNESS employs two parameters: wa, the decision weight for the absolute contribution, and wr, the decision weight for the relative contribution. These parameters are proportionally complimentary in that they are constrained to sum to 1 (i.e., wa + wr = 1). WITNESS uses these weights to determine the contributions of the two lineup members with the largest match values. wa governs the contribution of the best match to MEM (BEST), whereas the contribution of the second best match (NEXT) is governed by wr . When making its decision, WITNESS will choose BEST if the evidence [EV = wa * BEST + wr * (BEST - NEXT)] exceeds c, the decision criterion for simultaneous lineups. Thus, if wa = 1 (and so wr = 0), the decision would be made entirely based on BEST's match to memory (absolute contribution only); if wr = 1, the decision would be made based only on the magnitude of the difference between BEST and NEXT (relative contribution only). Both relative and absolute judgments can contribute to identification decisions (meaning that wa could take any value within the interval of 0 to 1). If the resulting EV value does not exceed c, the lineup is rejected, meaning that no individual is selected from the lineup.

The user must specify the parameters of the model in matrix form. The columns of the matrix are as follows: e (encoding parameter), ssp (similarity of suspect to perp), sfs (similarity of foil to suspect), c (decision criteria), and wa (relative vs. absolute criteria). The rows correspond to each data row to be fit. See examples.

Examples

# create data matrix
dataMatrix = matrix(c(.471, .230, .350,
							.208, .137, .513,
							.396, .431, .242,
							.166, .081, .669), nrow=4, byrow=TRUE)
# create an parameter.form
params = matrix(c(rep("e", times=4),
		0, 0, 0, 0,
		"sfs1", "sfs1", "sfs2", "sfs2",
		"cr1", "cr2", "cr1", "cr2",
		1, 1, 1, 1), nrow=4)


est = witness.est(parameters=params, data=dataMatrix, meth="WITC", sample.size=100)
est

dustinfife/witness documentation built on May 15, 2019, 6:05 p.m.