Search tips
Search criteria 


Logo of nihpaAbout Author manuscriptsSubmit a manuscriptHHS Public Access; Author Manuscript; Accepted for publication in peer reviewed journal;
J Clin Exp Neuropsychol. Author manuscript; available in PMC 2014 March 1.
Published in final edited form as:
PMCID: PMC3600160

Parallel Psychometric and Cognitive Modeling Analyses of the Penn Face Memory Test in the Army Study to Assess Risk and Resilience in Service Members



The psychometric properties of the Penn Face Memory Test (PFMT; Gur et al., 1997) were investigated in a large sample (4,236 participants) of U.S. Army Soldiers undergoing computerized neurocognitive testing. Data were drawn from the initial phase of the Army Study to Assess Risk and Resilience in Servicemembers (Army STARRS), a large-scale study directed towards identifying risk and resilience factors for suicidal behavior and other stress-related disorders in Army Soldiers. In this paper we report parallel psychometric and cognitive modeling analyses of the PFMT to determine whether ability estimates derived from the measure are precise and valid indicators of memory in the Army STARRS sample.


Single-sample cross-validation methodology combined with exploratory factor and multidimensional item response theory techniques were used to explore the latent structure of the PFMT. To help resolve rotational indeterminacy of the exploratory solution, latent constructs were aligned with parameter estimates derived from an unequal-variance signal detection model.


Analyses suggest that the PFMT measures two distinct latent constructs, one associated with memory strength and one associated with response bias, and that test scores are generally precise indicators of ability for the majority of Army STARRS participants.


These findings support the use of the PFMT as a measure of major constructs related to recognition memory and have implications for further cognitive-psychometric model development.

Keywords: cognitive-psychometric modeling, item response theory, Penn Face Memory Test, Army STARRS

Psychometric analyses provide the bedrock for accurate test interpretation in clinical neuropsychological practice and research (Campbell & Fiske, 1959; Cronbach & Meehl, 1955; Loevinger, 1957). It is essential that studies of mental health provide clear evidence that test scores are reliable and valid indicators of psychological constructs. Modern psychometric techniques including multidimensional item response theory (MIRT; see Reckase, 2009) constitute the preeminent methodology for such analyses. In this paper we explore MIRT models of the Penn Face Memory Test (PFMT; Gur et al., 1993, 1994, 1997, 2001, 2010, 2011) and improve interpretations of parameter estimates using results from a parallel cognitive modeling analysis. These results are used to assess the measurement precision of the PFMT as a component of the neuropsychological battery being administered to U.S. Soldiers in the Army Study to Assess Risk and Resilience in Servicemembers (Army STARRS; Army STARRS, 2011). Army STARRS is a large-scale study directed towards the identification of biological and psychosocial risk and resilience factors for suicidal behavior and other stress-related disorders. The present study serves to establish the measurement properties of the PFMT and provides a basis for future development of predictive models.

Context of Analyses

Army STARRS is a collaborative research effort in response to the alarming increase in suicide rates among Army Soldiers over the past several years (US Army, 2010). The study, which is divided into several components, includes historical data, prospective data collection, and biological sample collection across a number of sub studies. The New Solider Study—a component of Army STARRS—involves the administration of computerized questionnaires, including psychiatric symptom inventories, psychological assessments, and neuropsychological tests to Army Soldiers at the onset of basic training. The final sample for the New Solider Study is expected to comprise tens of thousands of participants. Data collection for the project began in 2011 and was still underway at the time of writing this manuscript.

The neuropsychological battery being administrated to Soldiers establishes baseline cognitive functioning with the future aim of identifying markers of risk and resilience. Tests were designed to provide efficient assessment of neurocognitive domains that can be linked to specific brain systems (see Penn Computerized Neurocognitive Battery; Gur, Erwin, & Gur, 1992; Gur et al., 2010, 2012). Army STARRS test data will be linked with demographic data, measures of mental health, and genetic markers collected in the final sample. Therefore, it is important that these data are reliable and valid indicators of Army Soldiers’ neurocognitive functioning.

Model Development

Analyses of data drawn from psychological instruments rely on models that relate test items to latent abilities. Reliability coefficients from both classical test theory (e.g., Cronbach’s alpha) and common factor theory (e.g., omega) are based on specific assumptions about underlying measurement structures (e.g., tau-equivalent or common factor; see McDonald, 1999). Verification of such assumptions, including a proper understanding of the test space, pattern of loadings, and factor correlations, is needed to derive accurate reliability and validity coefficients. Unfortunately, the process of mapping relations between observed data and latent abilities in psychological assessment is complex. There is no guarantee, for example, that item success within a specific scale requires just a single cognitive ability; there is also no guarantee that success is determined equally by the same abilities for all test takers (see Haynes, Smith, & Hunsley, 2011). In neuropsychological assessment, this opacity is worsened when separate cognitive processes are involved in task performance (see Brandt, 2007), further hindering the goal of establishing the construct validity of test scores (see Strauss & Smith, 2009).

A combination of theory and empirical evidence is needed to overcome the difficulties associated with measuring latent abilities. Combining principles from correlational and experimental psychology in model development (cf. Cronbach, 1957) is exemplified in current methodology known as cognitive-psychometrics (see Batchelder, 2010). In analyses of the PFMT, cognitive-psychometric modeling can be used to explain unknown aspects of the test’s latent factor structure with experimentally-derived cognitive architecture. To do so, we start with a firm understanding of the PFMT’s theoretical bases.

The PFMT was designed to measure visual episodic memory for unstructured stimuli, a latent ability sensitive to right temporal lobe dysfunction (Gur et al., 1993; Gur, Jaggi, Shtasel, & Ragland, 1994). The task parallels aspects of commonly administrated measures of verbal episodic memory (e.g., Delis, Kramer, Kaplan, & Ober, 2000), requiring examinees to first memorize a list of faces, and then to identify test items as either targets (old faces) or distractors (new faces). Although there is strong evidence that neuropsychological tests can dissociate visual episodic memory from other cognitive abilities (Cabeza & Kingstone, 2006; Carroll, 1993; Lezak, Howieson, Loring, Hannay, & Fischer, 2004), it is more difficult to delineate specific sub-systems and/or levels of processing within the domain itself (e.g., Baddeley & Hitch, 1974; Cowan, 1988). Indeed, it is well known that episodic memory tasks activate several regions of the prefrontal cortex and medial temporal lobe in brain imaging studies (Cabeza & Kingstone, 2006; Gur et al., 1997; Kelley et al., 1998).

A parsimonious explanation of recognition memory test performance is offered by cognitive modeling studies based on signal detection theory (see Green & Swets, 1966; Wickens, 2002), which have suggested that differences between target and distractor responses can be attributed to two general constructs: 1) response bias; and 2) memory strength. Interplay between the constructs can be modeled using two unequal-variance Gaussian distributions (one for targets and one for distractors) and a set of decision criteria (Wixted, 2007). The free parameters of the model are d′ – distance between the means of the target and distractor distributions; σ2T – variance of the target distribution; and C – response bias (the distance between criterion placement and the mid-point of d′; see Snodgrass and Corwin, 1988; additional details provided in the Appendix). Memory strength is related to d′ and σ2T, where larger values of d′ and smaller values of σ2T equate with better accuracy. Response bias is represented by the C parameter, where higher values of C lead to conservative responding (favoring distractors) and lower values of C lead to liberal responding (favoring targets). Note that this definition of response bias is not synonymous with poor effort. Because the signal detection model has been widely discussed in the literature, specific formulas are not presented here. Importantly, parameters of the model can be conceptualized at the level of individual respondents (e.g., Ingham, 1970) so that differences in d′, σ2T, and C can be directly related to parameter values from psychometric models to better understand a test’s internal structure.

The psychometric models used in our analyses are all based on the multidimensional extension of an item response theory graded response model (M-GRM; de Ayala, 1994; Muraki & Carlson, 1995; Samejima, 1997). The M-GRM is a psychometric framework for test data comprising ordered categorical responses. For the PFMT, examinees must choose from four responses for both target and distractor items indicating their confidence in seeing the test item previously: “Definitely Yes”, “Probably Yes”, “Probably No”, or “Definitely No”. The M-GRM first relies on a logistic distribution to model the probability that examinee i (N total) will respond with the kth or higher category (K total) to item j (J total) as a function of the item category threshold (djk), item discrimination (ajm), and examinee ability1 (θim) parameters:

equation M1

where aj and θi are vectors with M elements (one per factor) and uij is the realization of the response variate. From these values, the probabilities of responding in each of the categorical response options are then given by:

equation M2

equation M3

equation M4

As can be seen, the probability of responding with the first categorical response option is simply one minus the probability of responding with any other option. The probability of responding with an intermediate categorical response option is the difference of being in k or higher and k + 1 or higher for each category. Finally, the probability of responding with the last categorical response option is just the probability of responding with this option alone. Equations 2, 3, and 4 rely on the restriction that the item threshold parameters (djk) are strictly ordered from largest to smallest from the first to last categorical response options respectively. This assures that the response probabilities for a fixed ability level decrease from one category to the next (i.e., that correct PFMT response options are more difficult than incorrect PFMT response options).

The item threshold (djk) parameter is inversely related to item difficulty (i.e., higher values indicate lower difficulty). The item discrimination (aj) parameter can be interpreted similarly to a factor loading or an item-by-total biserial correlation. In a graphical representation of the IRT model, the djk and aj parameters determine the locations and slopes of the items’ category response functions. An example is illustrated in the top panel of Figure 1 for a hypothetical PFMT item. The plot shows that the probability of responding in the kth of K categories (y-axis) is a function of ability (x-axis). The exact shapes of the curves depend on the item parameters. The θ parameter represents examinees’ standings on the cognitive ability assumed to influence individual differences in item responses. The model represents item response variance as a function of examinee ability, item threshold (inverse of difficulty), and item discrimination.

Figure 1
Category response curves for hypothetical unidimensional (top) and multidimensional (bottom) Penn Face Memory Test items. DY = “Definitely Yes”, PY = “Probably Yes”, PN = “Probably No”, or DN = “Definitely ...

The vector notations aj and θi used in Equations 1, 23, and 4 imply that multiple cognitive abilities can influence responses. In the case where all items can discriminate between distinct levels of all abilities, the model is said to have within-item multidimensional structure. When items can discriminate just one of two or more abilities, the model is said to have between-item multidimensional structure. And, when all items can discriminate just a single ability, the model is said to have unidimensional structure. The within-item multidimensional model is the most general and the unidimensional model is the most restrictive. These three nested variants are depicted in Figure 2. Examples of within-item multidimensional item category response functions are provided in the bottom panel of Figure 1. Note that the hypothetical item can discriminate between levels of multiple abilities (i.e., the curvatures of the category response functions change with level of ability along both axes). Several sources can be consulted for further description of item response theory models in psychological and clinical assessment (Embretson & Reise, 2000; Reise & Waller, 2009; Thomas, 2011).

Figure 2
Unidimensional (left), between-item multidimensional (center), and within-item multidimensional (right) models.

Measurement Precision

Item response theory (IRT) offers a robust and comprehensive alternative to classical test theory (CTT). Both IRT and CTT provide descriptions of and formulas for measurement precision; however, there are important differences among them. In CTT, standard error (SEtot) is derived from test reliability by

equation M5

where σtot is the population standard deviation of the total number-correct true-score and ρ is the population reliability coefficient (Magnusson, 1966). Standard error in IRT is derived from item information—the instantaneous change in the probability of a correct response on a particular item at a particular ability level normalized by item variance at the ability (Lord, 1980). An item where a small change from a particular ability value produces a large normalized change in the probability of a correct response is informative at that ability. This occurs when item difficulty is closely matched with ability and when item discrimination is high (Baker & Kim, 2004). Test information at a particular ability (TIθ) is simply the sum of item information. From this, standard error as a function of ability, SEθ, is given by

equation M6

Whereas standard expositions of CTT provide a single estimate of total score measurement error based on reliability, IRT provides distinct estimates of measurement error that can vary depending on ability; that is, potentially unique values of SEθ can be assumed for unique values of θ.

It has been shown that CTT could be extended to include the concept of test information, and that IRT could be extended to include the concept of reliability (e.g., Mellenbergh, 1996). For example, reliability could be calculated in IRT across a range of ability estimates using the average standard error (Andrich, 1988). Yet, in practice, information and reliability are rarely both computed from a particular test model. CTT estimates of measurement precision are almost always reported as true score reliability, and IRT estimates of measurement precision are almost reported as information. Furthermore, because CTT estimates of reliability are for particular items administered to particular groups of examinees, standard error is often reported as a single value that confounds examinee and test properties (Embretson & Reise, 2000). Within the IRT framework, item properties, such as item difficulty, are modeled in addition to ability, facilitating estimates of a test’s psychometric properties that are independent from examinee properties (Embretson & Reise, 2000).

As a practical tool for researchers, IRT could easily offer both SEθ and ρ for evaluating measurement precision. Evaluations of SEθ involve comparisons among differing levels of ability. Usually, graphical representations of the SEθ function are used to determine ability locations with relatively better or worse measurement precision. By doing so, a test evaluator can determine whether a test is optimized for measurement within a specific population and a specific range of ability. Cutoffs for acceptable SEθ, however, are less commonly offered, which can sometimes lead to ambiguous interpretations of data. Unlike SEθ, acceptable cutoffs for ρ are commonly offered (.60 to .80 depending on the specific type of coefficient; see Haynes, Smith, & Hunsley, 2011). Typically, we should expect that a favorable ρ should accompany a favorable SEθ function with respect to a specific population. That is, reliability should be high when areas of low measurement error overlap with areas of high density in ability. However, because reliability is a reflection of average standard error, implying that it is possible to find equivalent reliabilities for tests with very different standard error functions, both indices of measurement precision should be evaluated.

Study Goals

The first goal of this study was to find an appropriate measurement model for the PFMT. This was accomplished by fitting nested measurement structures based on the M-GRM to a training data set, and then cross-validating the best fitting model in a validation data set. To improve interpretability, the psychometric results were then compared to cognitive modeling parameters derived from the unequal-variance signal detection model. Our second goal was to evaluate the measurement precision of PFMT latent ability scores. To do so, we compared SEθ values and evaluated ρ over the range of ability observed in Army STARRS.



Army Soldiers were recruited to volunteer without compensation for the Army STARRS New Solider Study at the start of basic training. The current sample comprises participants from three Army bases in the United States tested between February and June of 2011. All Soldiers were asked to provide informed, written consent prior to participation in research. Army commanders provided sufficient time to complete all surveys and tests which were administered in a group format using laptop computers. Research proctors monitored the testing environment and assisted with questions and technical difficulties. Surveys and tests were administered in a fixed order in 90 minute sessions over two days of testing. The PFMT was administered on the second day of testing.

A total of 5,551 Army Soldiers recruited into the study participated in the New Solider Survey during the specified time frame. Of these, approximately 15% did not complete the PFMT (e.g., some did not show for the second day of testing) and another 8% were removed due to technical problems or potentially invalid data. Specifically, during one administration a subset of participants experienced a testing software malfunction. The software anomalies were corrected and incomplete response profiles flagged as errant. Also, soldiers who provided the same response to 12 or more consecutive items were flagged as errant. This automated rule for identifying invalid profiles on the PFMT was developed based on previous large-scale studies using the Penn Computerized Neurocognitive Battery (Gur et al., 2012). Profiles flagged as errant were not used for psychometric modeling. Automated quality control is required due the complexities of validating thousands of individual response profiles in large-scale studies. Other large-scale studies have reported similar exclusion rates due to technical problems (Hoerger, Quirk, & Weed, 2011).

These procedures resulted in a final sample of 4,236 participants for psychometric modeling. Demographic characteristics of participants with PFMT data and participants without PFMT data are shown in Table 1. No group differences were found for age and handedness; however, there tended to be fewer males, better educational attainment, and less ethnic and racial diversity in the group of participants with valid PFMT data.

Table 1
Demographic Characteristics of Participants with Penn Face Memory Test Data Present vs. Absent or Excluded


The PFMT was administered as part of a neurocognitive battery designed for efficient computerized assessment (Gur et al., 2001, 2010). The test begins by showing examinees 20 faces that they will be asked to identify later. Faces are shown in succession for an encoding period of 5 seconds each. After this initial learning period, examinees are immediately shown a series of 40 faces—20 targets and 20 distractors—and asked to decide if they have seen each face before by choosing 1 of 4 ordered categorical response options: 1 = “Definitely Yes”; 2 = “Probably Yes”; 3 = “Probably No”; or 4 = “Definitely No”. For consistency with IRT’s assumption of a monotonically increasing response functions, we recoded responses prior to analyses so that higher valued responses always implied more accurate responses (i.e., 1 is the least correct answer and 4 is the most correct answer for all items). Stimuli consist of black and white photographs of faces presented on a black background. All faces were rated as having neutral expressions and were balanced for gender and age (Gur et al., 1993, 2001). Examinees responses and response times are recorded during test administration; however, there are no time limits during recognition testing or explicit instructions to work quickly. The PFMT takes approximately 4 minutes to administer.

Typical PFMT scoring procedures produce the number of correctly recognized targets and correctly rejected distractors, and median response times for correct responses. There are no cutoffs for qualitative interpretations of performance. In previously published studies, the reliability of accuracy scores was .75 and the reliability of time scores was .87 (Gur et al., 2010). PFMT scores have also demonstrated expected associations with demographic variables such as gender (women perform better), age (younger examinees perform better), and education (more educated examinees perform better), and its strongest intercorrelations are with measures of word memory and spatial memory (Gur et al., 2010).

Model Analysis

Data were analyzed using a single-sample cross-validation. The sample was first randomly split into a training set (N = 2,118) and a validation set (N = 2,118). Exploratory psychometric analyses were used to determine the number of latent factors present in the data by examining eigenvalues (including a parallel analysis that compared observed eigenvalues to 10 randomly simulated eigenvalue sets) and factor loadings for multiple varimax-rotated (orthogonal) solutions (polychoric correlations were analyzed in both techniques). We next fit exploratory between-item multidimensional and within-item multidimensional graded response models to the data (both variations of Equation 1; see Figure 2) and compared the relative fit of each using Akaike information criterion (AIC) and Bayesian information criterion (BIC) values. AIC and BIC values penalize over-parameterized models and become smaller with better fit (for the use of AIC and BIC in IRT, see de Ayala, 2009).

The best fitting exploratory model from the training data was then cross-validated in the validation data to determine if the measurement structure would generalize across samples. Item fit was evaluated by assessing the amount of local dependence remaining between items (Chen & Thissen, 1997). Large values of local dependence (residual association [ϕc2] > 0.2) imply poor model fit (see de Ayala, 2009). This is because local dependencies—unaccounted for common variance between items—can distort parameter estimates and lead to incorrect conclusions regarding the reliability and validity of test scores. After verifying fit of the chosen model in the validation data, the entire sample of 4,236 participants was pooled in order to estimate final item and examinee parameter values. Lastly, to better understand the latent dimensions themselves, we compared fitted values to individual parameter estimates derived from a signal detection theory analysis.

All analyses were conducted using the R language (R Development Core Team, 2011). Eigenvalues and factor loadings were estimated using the ‘psych’ package for R (Revelle, 2011). Exploratory factor analyses were based on maximum likelihood estimation with orthogonal (varimax) rotations. MIRT model parameters were estimated using the ‘mirt’ package for R (Chalmers, 2011), which uses a stochastic approximation algorithm (see Cai, 2010) to fit exploratory parameter values to data. Data missing at random are accommodated in the package within a Metropolis-Hastings Robbins-Monro algorithm. Parameter estimates are based on the logistic model with the added multiplicative constant 1.702 which produces results that are nearly identically to the normal ogive model (see Haberman, 1974). Procedures and software for estimating individual differences in the unequal-variance signal detection model are less well established. Therefore, we used R to program a Metropolis-Hastings-within-Gibbs sampler for a hierarchical Bayesian representation of the model, and used approximated expected a posteriori (EAP) values as the final parameter estimates. Additional details about this procedure are provided in the Appendix.


Exploratory Analysis

Training Data

The training data were analyzed using exploratory psychometric techniques. The results reveal that the first 10 eigenvalues are all greater than one (8.03, 4.56, 1.76, 1.42, 1.31, 1.23, 1.21, 1.11, 1.08, and 1.04), which is sometimes considered to be the criterion for retention. In addition, when the eigenvalues were compared against values based on random simulations (parallel analysis), a total of 13 were sufficient to retain. Despite this, the first two eigenvalues appear to be dominant, and the amount of variance explained quickly diminishes to a nearly constant value by the third factor (20%, 11%, 4%, 4%, 3%, 3%, 3%, 3%, 3%, and 3%).

The loadings for 1-, 2-, and 3-factor exploratory solutions are presented in Table 2. The results appear to favor a 2-factor solution. Specifically, 2 factors maximized the number of items for which there is just one large, primary loading per item. For ease of interpretation, items in Table 2 are ordered by targets (first 20 rows; item numbers 3, 4, …, 40) and then distractors (second 20 rows; item numbers 1, 2, …, 39). As can be seen, this provides a clear distinction in the 2-factor solution: factor 1 is primarily related to correct rejections and factor 2 is primarily related to hits. For clarity, the first of these is referred to as the distractor-loaded factor and the second is referred to as the target-loaded factor. The 1-factor solution is difficult to interpret, as it appears to create one amalgam of factors with no consistent pattern of loadings. The 3-factor solution retains the distractor-loaded factor, but splits the target-loaded factor into two sub-facets, with items tending to load onto the first of these early in the test, and onto the second late in the test. The results of 4 or more-factor solutions (not reported here) suggest a high number of difficult to interpret single-item factors or factors with no dominant loadings at all. Because the amount of variance accounted for by the first two factors could be readily distinguished from higher factor solutions, coupled with the interpretability of the two-factor solution and its parsimony, we decided to proceed with the analysis using the 2-factor model. We return to this issue in the discussion.

Table 2
Loadings for Exploratory 1-, 2-, and 3-Factor Solutions

To explore the need for non-primary ability loadings, 2-factor M-GRMs were fit to the data to determine whether items are better modeled by within- or between-item multidimensionality (i.e., if the non-primary ability loadings could be eliminated; center panel of Figure 2 vs. right panel of Figure 2). Both within- and between-item models converged without difficulty. The AIC and BIC values for the within-item model (162,433.30 & 163,564.90) were consistently lower than the AIC and BIC values for the between-item model (163,104.50 & 164,009.80). This suggests that within-item multidimensional structure provides a better fit in comparison to between-item multidimensional structure, even after accounting for the former’s increased complexity due to a higher number of parameters that must be estimated.

Cross Validation

The training data suggest that the latent structure of the PFMT is adequately fit by a 2-factor within-item M-GRM. To confirm this, the model was next applied to the validation data and evaluated for item fit. Local dependence values are shown in the top panel of Figure 3. As can be seen, the majority of values are less than .10, and none are greater than .20. The average value was .08, which implies overall good fit. Specifically, it appears as though all major sources of common variance between items can be attributed to just two latent variables without great ill effect on the final parameter estimates.

Figure 3
Top: Histogram of local dependence (ϕc2) between items after fitting the within-item multidimensional graded response model to the data. Bottom: Heat map of local dependence (ϕc2) between items. Darker colors imply greater local dependence. ...

A correlation plot (heat map) showing specific item-pair local dependencies is shown in the bottom panel of Figure 3. Areas of concentration (darker colors) are indicative of systematic model fit issues. As was done in Table 2, the items are grouped by targets (item numbers 3, 4, …, 40) and then distractors (item numbers 1, 2, …, 39). Again, it can be seen that most item-pair associations were very low. However, there does appear to be a tendency for items administrated later in the PFMT task (i.e., items 30 through 40) to show higher local dependence values in comparison to other item pairings.

Comparison to Cognitive Model

Item parameter estimates derived from the entire sample are reported in Table 3 and graphical representations in the form of item vectors are shown in Figure 4 (for generation of these plots see Reckase, 2009). Panel A of Figure 4 shows item vectors plotted in a coordinate system comprised of orthogonal axes (θ1 and θ2) which we have called the distractor- and target-loaded factors respectively. The angles formed between the vectors and axes convey the directions of association between the items and factors (i.e., smaller angles imply greater relative discrimination). The overall vector lengths, which are equal to the items’ multidimensional discrimination parameters, convey the magnitudes of association. The origins of the vectors are related to the items’ thresholds; smaller valued coordinates imply less difficulty. Because each item is associated with three thresholds, each has three vectors as well.

Figure 4
Unrotated (A & B) and rotated (C & D) solutions to the within-item multidimensional graded response model. A: unrotated orientation of the axes (θ1 and θ2) plotted with item vectors; B: unrotated orientation plotted with ...
Table 3
Parameter Estimates for the Within-Item Multidimensional Graded Response Model

The item threshold parameters (d1, d2, and d3) are generally positive, indicating that most items were of easy to moderate difficulty (recall that difficulty and threshold are inversely related). This is also conveyed by the high prevalence of item vectors originating in the upper-left and lower-left quadrants of panel A in Figure 4. Yet, while nearly all of the d3 parameters for targets in Table 3 are positive (d3 item vector coordinates less than 0), d3 parameters for distractors are mostly negative (d3 item vector coordinates greater than 0).

The item discrimination parameter estimates (a1 and a2) indicate that the first latent ability (θ1) is more strongly associated with distractors (i.e., distractor-loaded factor) and that the second latent ability (θ2) is more strongly associated with targets (i.e., target-loaded factor). This can also be seen in panel A of Figure 4 where the light gray item vectors (distractors) tend to align with θ1 and the dark gray item vectors (targets) tend to align with θ2.Unfortunately, this manifestation of the model is complicated by rotational indeterminacy. That is, due to the exploratory nature of the M-GRM, there is no correct orientation to the x- and y-axes (θ1 and θ2). Any consistent orthogonal rotation of the vector and axis values will produce the same model fit, with just the item weights changing. A proper rotation of the axes cannot be determined by the observed data alone, but instead must be “resolved” using a rotation procedure based on a theory-based criterion.

Our approach is to resolve this rotational indeterminacy using cognitive-psychometric modeling. Specifically, results from the M-GRM were correlated with individual parameter estimates derived from a specific application of the signal detection theory of recognition memory—the unequal-variance model. Correlations between the unrotated M-GRM dimensions (θ1 and θ2) and the estimated signal detection parameters are depicted in panel B of Figure 4. The vectors for d′, σ2T, and C are plotted at angles and lengths2 which convey the direction and magnitude of association, respectively, between each parameter and the unrotated θ1 and θ2 dimensions (smaller angles and longer vectors indicate stronger association). The unrotated M-GRM results provide no more than an ambiguous interpretation of the distractor- and target-loaded abilities with respect to the signal detection parameters. θ1 is positively correlated with both d′ and C. θ2 is positively correlated with d′ but negatively correlated with C. Neither ability is strongly correlated with σ2T, which could reflect poor estimation of the parameter. The unrotated solution was unexpected from the perspective of signal detection models of recognition memory. In particular, the latent constructs, which ought to align with memory strength and response bias, should manifest themselves equally in target responses and distractor responses.

To improve clarity on this issue, we next orthogonally rotated the solution to align θ1 with d′ and θ2 with C (an angle of approximately 26°). Specifically, by first averaging the angle of separation between θ1 and d′ with the angle of separation between θ2 and C, and then consistently rotating all parameter estimates by this amount, results from the M-GRM model were brought into alignment with results from the signal detection model. This is similar to classic targeted rotation procedures (see Mulaik, 2010), except that the angle of rotation was determined directly from associations among estimates of ability and the signal detection parameters rather than a targeted matrix for the item parameters. The resulting dimensions after rotation along with the signal detection parameter vectors are shown in panel C of Figure 4. It can be seen that the rotation improves interpretation of the M-GRM latent dimensions considerably. The first ability (θ1*) can now be regarded as an indicator of memory strength (i.e., closely aligned with d′), and the second ability (θ2*) can now be regarded as an indicator of response bias (i.e., closely aligned with C, although in the opposite direction). Because θ1 and θ2 were rotated by the same angle (i.e., orthogonal rotation) the dimensions remain orthogonal to one another.

The rotated dimensions θ1* and θ2* along with the PFMT item vectors are shown in panel D of Figure 4. Note that θ1* bisects the average angle between distractor and target item vectors. This suggests that θ1* reflects a general ability, where higher values of θ1* are associated with higher probabilities of correct response for both targets and distractors. Also note that θ2* is positively related to targets and negatively related to distractors. This suggests that θ2* reflects a tradeoff, where higher values of θ2* are associated with higher probabilities of correct response to targets but lower probabilities of correct response to distractors, and that lower values of θ2* are associated with lower probabilities of correct response to targets but higher probabilities of correct response to distractors.

The rotated discrimination parameters are reported in Table 3 (a1* and a2*). Importantly, the items’ overall multidimensional discriminating powers (and communalities) have not changed. Discrimination parameters for the first rotated dimension (a1*) are all positive, and generally moderate in size. Better memory strength leads to a higher number of hits and correct rejections. Discrimination parameters for the second rotated dimension (a2*) are of the opposite sign for targets and distractors. That is, high θ2* implies bias towards targets and low θ2* implies bias towards distractors. The average absolute value of a1* is larger than the average absolute value of a2*, which is more true of targets (|Ma1*| = .50 vs. |Ma2*| = .28) than of distractors (|Ma1*| = .60 vs. |Ma2*| = .50).

Measurement Precision for Army STARRS

The SEθ function for the PFMT is plotted in the top panel of Figure 5. Curvature in the SEθ surface shows that precision in measurement varies as a function of ability. The function reaches a low point around one standard deviation below average memory strength and average response bias (the coordinate θ1* = −1.0, θ2* = −1.3). Notably, there is one “valley” of SEθ running diagonally3 from about (θ1* = −4.0, θ2* = −4.0) to (θ1* = 4.0, θ2* = 4.0). High points in the SEθ surface fall in regions where memory strength is very high and response bias is very low (e.g., θ1* = 4.0, θ2* = −4.0) or where memory strength is very low and response bias is very high (e.g., θ1* = −4.0, θ2* = 4.0)

Figure 5
Top panel: standard error function for the rotated solution to the within-item multidimensional graded response model. Bottom panel: multidimensional distribution of memory strength and response bias in the Army STARRS sample.

The bottom panel of Figure 5 plots the bivariate distribution of ability estimates in the Army STARRS sample. The highest density of ability occurs at (θ1* = 0, θ2* = 0), with both memory strength and response bias symmetrically distributed about the point. By visually lining-up the ability distribution in Figure 5 with the SEθ function in Figure 5, it can be seen that there is a close match. That is, low values of SEθ tend to occur near dense ability regions. However, the SEθ function does not appear to be optimized for this specific sample. The low point in the SEθ function suggests that ability estimates are most precise for slightly less able individuals. Indeed, estimated reliability in the current sample is .61; however, if we consider only those Soldiers with ability estimates falling within one standard deviation of the low point in the SEθ surface (θ1* = −1.0 & θ2* = −1.3), reliability jumps to .71. If we consider only those Soldiers with ability estimates falling within one standard deviation of an arbitrary high point in the SEθ surface (say θ1* = 1.0 & θ2* = −1.0), reliability drops to .42, and would be much lower had we chosen more extreme values.


Parallel psychometric and cognitive modeling analyses were used to better understand the measurement properties of the PFMT as a component of the neuropsychological battery being administrated to Soldiers in the Army Study to Assess Risk and Resilience in Servicemembers. Item response theory (M-GRM model) was used to confirm the number of latent cognitive abilities measured by the instrument. Then, cognitive theory (unequal-variance signal detection model) was used to apply psychological meaning to those abilities. Specifically, these analyses supported the construct validity of independent neurocognitive markers of memory strength (d′) and response bias (C). We then investigated the measurement precision of PFMT test scores and found adequate precision overall, but higher precision in low-average ability regions. Thus, PFMT scores were most precise for examinees with low-average memory strength. The results suggest that although the PFMT can provide useful measurements across a wide range of memory abilities, it is particularly useful for identifying differences among impaired individuals.

Cognitive-psychometric methodology helped explain unknown aspects of the test’s latent factor structure using experimentally-derived cognitive architecture. The approach provided a synthesis of statistical modeling and cognitive theory. Although the modeling procedures were complex in comparison to conventional approaches, the final measurements had a priori psychological meaning and did not require a posteriori adjustments (e.g., multiple regression or pattern analysis) to be made interpretable. In addition, the modeling procedures allowed us to examine standard error as a function of cognitive ability. This facilitated a more nuanced inspection of the PFMT’s psychometric properties, and allowed us to determine specific levels of memory associated with relatively better or worse measurement precision.

Penn Face Memory Test Model Structure

The within-item M-GRM (bottom panel of Figure 1 and right panel of Figure 2) is considered to be a relatively complex measurement structure. Within-item multidimensional models require greater effort to estimate and to interpret in comparison to between-item models (McDonald, 2000). A common solution to such complexity is to force between-item (simple) structure by allowing factors to correlate with each other. In panel A of Figure 4, for example, it is possible to obliquely rotate the x- and y-axes so that the target and distractor items primarily align with just one dimension. That is, it is possible to avoid models where items are indicators of multiple cognitive constructs by assuming that the constructs are highly correlated. Although such a solution is methodologically simpler, it fails to agree with cognitive theories suggesting that target and distractor responses should be similarly determined by two underlying constructs. A more psychologically meaningful, albeit complex, solution was to interpret the psychometric model (within-item M-GRM) from a cognitive perspective (i.e., targeted rotation using unequal-variance signal detection model parameters). This methodology produced clear psychological interpretations for the latent abilities—one representing memory strength and one representing response bias.

Aligning the exploratory MIRT results with a well-validated cognitive model helps to establish the PFMT’s construct representation; that is, how test scores (item responses) are determined by the properties of items, tasks, and underlying cognitive processes (see Embretson, 1983; Embretson, 2010). In our analyses, for example, the function of the first rotated ability (θ1*) was shown to be consistent with the construct representation of memory strength and the function of the second rotated ability (θ2*) was shown to be consistent with the construct representation of response bias. This stringent form of validity is supported by cognitive-psychometric modeling methodology (see Batchelder, 2010). Long recognized as the ideal synthesis of theories from experimental and correlational psychology (Cronbach, 1957), this methodology is rapidly expanding its applications (e.g., Brandt, 2007; Brown, Turner, Mano, Bolden, & Thomas, 2012; Embretson & Gorin, 2001; Townsend & Neufeld, 2010) and has the potential to improve neuropsychological assessment, particularly in an age of computerized measurement.

The model applied in the current study is regarded as a preliminary attempt to build a cognitive-psychometric architecture for neurocognitive measures of episodic memory. Although the model was made consistent with basic results from signal-detection theory (a descriptive cognitive model), it remains unclear how specific cognitive processes, as well as other commonly studied cognitive phenomena (e.g., list-length), can be accommodated by the measurement structure. Whether the cognitive-psychometric model described in this paper needs to be replaced by a synthesis of psychometric theory with a more explanatory cognitive theory of episodic memory (see Clark & Gronlund, 1996 for examples), or can be amended to handle the inherent complexity in episodic memory testing is unknown. Obvious limitations in the applied model are that changing the number of study items (faces) and the time between study and test phases will affect item difficulty. Also, the process of face recognition itself could affect memory strength for other stored but not yet tested faces. Consistent with these concerns, our results suggest that in addition to the two dominant factors underlying test performance (memory strength and response bias) nearly a dozen smaller factors were present in the data. Our analysis of local dependence (Figure 3) suggests that most of this residual common variance is found among items administrated later in the PFMT task. We assume that this variance reflects the impact of unaccounted for cognitive processes (e.g., learning-to-learn, network activation, etc.), but this hypothesis was not tested.

If our assumption is accurate, a more comprehensive model comprising structures for both cognitive abilities and cognitive processes is required to accurately account for episodic memory task performance. For example, there is some agreement in the field of cognitive modeling that memory strength can be further deconstructed into the additive (see Wixted, 2007) and/or complementary (see Onyper, Zhang, & Howard, 2010; Yonelinas, 2002) effects of familiarity and recollection. Whereas familiarity is characterized by the weak sense of having experienced some past event, recollection conveys clear and unequivocal memory probably related to memory search. Imaging studies have even suggested that these processes rely on distinct neuroanatomical structures (Davachi & Wagner, 2002; Diana, Yonelinas & Ranganath, 2007; Kim & Cabeza, 2007; Wheeler & Buckner, 2003), though perhaps not in a clear and unambiguous manner (Wixted & Squire, 2011).

Integrative frameworks for such findings are offered by computational global memory models (for a review see Clark & Gronlund, 1996). Most of these assume that memories, which are represented parametrically as associative networks of item and context information, are accessed through a process combining elements of global matching and interactive-cueing. Recollection equates to a specific match between a test item and a stored memory; familiarity equates to the marginal associative strength between a test item and the complete stored network of memories (e.g., Gronlund & Shiffrin, 1986; Raaijmakers & Shiffrin, 1980). These computational models are unique in that they are designed to explain underlying memory structures and processes at a level of specificity not feasible with more universally applicable models (e.g., generalized linear model). Further research is needed to determine if computational models can be given a cognitive-psychometric framework.

Measurement Properties of the PFMT and Implications for Army STARRS

Standard error is determined by item and examinee characteristics. Specifically, measurement precision is optimized when item difficulty is closely matched with ability and when item discrimination is high. This can be visualized in panel D of Figure 4. The area of the two-dimensional ability space (θ1* by θ2*) covered by the dark and light grey arrows (item vectors representing combined item difficulty and discrimination characteristics) is the area of ability where precise measurement occurs. This implies that the PFMT should be most precise when measuring low-average levels of memory strength (θ1*) and response bias (θ2*), which can also be observed in the standard error function displayed in the top panel of Figure 5. It is noteworthy that PFMT distractor items and PFMT target items contribute towards precision of measurement at complementary angles within the multidimensional ability space. That is, hits and correct rejections contribute unique value to the assessment of episodic memory, and should not be regarded as interchangeable data.

Most of the PFMT items were estimated to have low difficulties (inverse of d1, d2, and d3), which resulted in the test’s standard error function reaching a minimum (maximum measurement precision) in the low-average range of ability. In addition, the difficulties of specific PFMT response options (i.e., “Definitely Yes”, “Probably Yes”, “Probably No”, or “Definitely No”) were slightly disproportionate for targets and distractors. Targets were easier than distractors. This implies that distractors contributed relatively greater measurement precision in the high-average range of ability. All things being equal, adding distractor items to the PFMT should improve measurement precision for non-impaired examinees, while adding target items should improve measurement precision for impaired examinees, as the adjustment should make the test more or less difficult respectively.

Most of the PFMT items were estimated to have moderately sized discrimination parameters, but memory strength values (a1*) were generally larger than response bias values (a2*). Thus, memory strength was measured more precisely than response bias. As a result, variation in memory strength had a greater effect on standard error than did variation in response bias. Although somewhat difficult to discern in Figure 5, marginal variation in the memory strength axis is greater than marginal variation in the response bias axis. The surface does not form a consistent “U” shape. An implication is that, because the variances of both latent dimensions were standardized, memory strength was a greater relative determinant of response performance than was response bias, which could reflect relatively less true variance in Army Soldiers’ response biases in comparison to Army Soldiers’ memory strengths. Practically, PFMT users can consider the instrument a better indicator of memory strength than response bias, which is consistent with the purpose of the test.

Overall, analyses of the PFMT suggest that standard error values were low for the majority of Army STARRS participants’ ability estimates. The reliability of all estimates in the sample was .61. Thus, it can be argued that the PFMT approaches the general standard for adequate measurement precision for the Army STARRS sample as a whole. However, the reliability is much higher (.71) if we consider only those estimates in the low-average range of ability. This confirms that the PFMT’s measurement precision is optimized for examinees with low-average ability. Item difficulty would have to be raised (item thresholds lowered) by approximately one standard deviation in order for measurement precision to peak in the center of the Army STARRS ability distribution. However, doing so may not be desirable in a study of psychopathology, where the impaired end of the ability spectrum is the primary focus of predictive modeling. Because Army STARRS seeks to identify potential risk factors, such as below average cognitive functioning, the results generally support maintaining the test in its current instantiation. The capability to determine this is a benefit of modeling both SEθ and reliability.

Two aspects of the PFMT likely contribute to its favorable measurement precision: 1) items are generally of moderate difficulty; and 2) examinees are asked to respond using ordered categorical response options. Items with moderate difficulties are optimal for assessment of normative samples. PFMT items, which are somewhat geared towards clinical populations, can be characterized as being easy, but not too easy. Developers of fixed-length tests, particularly tests administered in time-limited settings, must choose a range of ability for which items will provide maximum precision. For the PFMT, even though ability measurements are most precise for examinees in the low-average range, examinees in the above average range are challenged enough by the content to avoid dramatic ceiling effects in the Army STARRS sample. Standard error values are also lowered by the PFMT’s rating scale format. Modeling intermediate response options (i.e., “Probably”) tends to minimize standard error in the bodies of ability distributions; extreme response options (i.e., “Definitely”) tend to minimize standard error near the tails. The PFMT’s polytomous format (ordered categorical response options) has the effect of creating a pseudo-continuous response scale, which helps to overcome some of the limitations due to categorical response data.

It is important to note that an accurate characterization of standard error goes beyond evaluations of a test’s measurement properties. Standard error can also be used to improve predictive modeling, as is planned for risk and resilience variables in Army STARRS. A key tenant of structural equation modeling (see Bollen, 1989) is that a properly specified measurement model can alleviate some of the problems that arise from regression with latent variables. Most prominently, regression coefficients become attenuated (shrink towards zero) as measurement error grows larger. Because standard error is a property of ability score estimates, it should be examined in predictive modeling to ensure that poorly measured examinees will not disproportionately impact analyses of a test’s predictive value.


A limitation of all latent variable studies is the inability to unequivocally define measured constructs. Although the rotated dimensions appear to represent memory strength and response bias, there is no guarantee that the PFMT measures these two latent variables. Evidence of nomothetic span (convergent and discriminant validity), much of which has been collected for the PFMT (e.g., Gur et al., 1997, 2001, 2010, 2012), further supports the test’s construct validity. In addition, our choice to align latent ability estimates from the M-GRM with parameter values from the unequal-variance signal detection model was based on an a priori chosen theoretical framework for interpreting response data. As with all models, it is not possible to prove that the rotation is correct or is a true reflection of reality. Indeed, any rotation of the parameter space (e.g., based on simple structure, targeted item parameters, etc.) would result in the same model fit. By showing how the rotation solution can be linked to a particular theory, however, we refocus the discussion of factor rotation strategy from mathematical criteria (varimax, promax, etc.) to criteria for adequate theory construction.

We noted that the response profiles of approximately 8% of Army Soldiers in the initial sample were flagged as errant and therefore not included in the analyses. The exclusions may have eliminated some valid response profiles resulting in overestimates of the PFMT’s model fit and measurement precision. It was also noted that demographic characteristics of Soldiers with data present were not identical to demographic characteristics of Soldiers with data absent or excluded. Disparities in gender, education, and race and ethnicity, though small, were observed between the two groups. These differences threaten the validity of findings from this study. However, perhaps more concerning than difference between Soldiers tested and Soldiers not tested, is the more general possibility of group biases. Specifically, it is possible that differential item functioning between demographically diverse groups of examinees, even within the same calibration sample, led to poor estimation of some parameter values. A potential solution is to divide the current sample into subsets based on demographic characteristics, and then investigate differential item functioning between groups. Such analyses, while valuable, would add considerable length to the current paper. They will be pursued in future studies.

Finally, unlike a true cross-validation, where independent samples are drawn from the same population, our analyses were based on independent samples drawn from the same sample. This implies that our results were cross-validated to the sample, but not population of Army Soldiers.


The PFMT was found to measure visual memory strength and response bias with adequate precision for the majority of Soldiers participating in Army STARRS. The results are promising, and suggest that the test has the potential to provide useful measurements of Army Soldiers’ cognitive functioning. Although the psychometric model applied to the PFMT in the current study likely does not account for all relevant sources of task and item variance, it does provide a basis for future cognitive-psychometric modeling of the instrument.


Army STARRS was sponsored by the Department of the Army and funded under cooperative agreement number U01MH087981 with the U.S. Department of Health and Human Services, National Institutes of Health, National Institute of Mental Health (NIH/NIMH). The contents are solely the responsibility of the authors and do not necessarily represent the views of the Department of Health and Human Services, NIMH, the Department of the Army, or the Department of Defense.


Because the densities of the ordered categorical response options were estimated to be slightly disproportionate for targets and distractors, an unequal-variance version of the signal-detection model was used to model the data. The unequal-variance signal-detection model is not identifiable from the observed data alone. Therefore, the mean of the distractor distribution was fixed to 0 and the variance was fixed to 1 for all examinees. The mean of the target distribution for each examinee was set equal to estimates of d′ and the variance (σ2T) was set equal to 1 plus estimates of σ2 (see below). The middle criterion for each examinee was found by adding estimates of bias (C) to the midpoint of estimates of d′. Individual differences in the placement of the upper and lower criteria were not uniquely estimated, but were instead fixed by adding and subtracting .5, respectively, to each examinee’s middle criterion estimate. The value of .5 was found to provide good fit for data aggregated at the group level. All free parameters (d′, σ2, and C) were estimated using a hierarchical Bayesian approach with a Metropolis-Hastings-within-Gibbs sampler (see Gelman, Carlin, Stern, & Rubin, 2004) written in R. The chain was thinned by every 10th draw and run for 3,000 iterations after burn-in. We used informative prior distributions to improve parameter recovery. The priors were specified as follows: d′ ~ Normal(1.5, .25); σ2 ~ Inv-Gamma(5,1); and C ~ Normal(0, .25). The means of the d′ and C prior distributions were based on preliminary analyses fitted at the group level. The shape and scale parameters of the inverse gamma prior results in a prior mean of .25 for σ2 and a prior mean of 1.25 for σ2T. This value is consistent with the commonly reported ratio of distractor to target variance of .8 (i.e., σ2F/σ2T = 1.00/1.25; see Mickes, Wixted, & Wais, 2007).


1The terms “ability”, “factor”, and “construct” are similarly used to indicate a latent variable.

2Because the angles and vector lengths were based on the sample data, rather than the population model, the values do not maintain orthogonal geometry within the axes. Therefore, we used the average angle between each dimension to plot the vectors.

3SEθ functions are difficult to summarize for within-item multidimensional models. Whereas changes in ability for unidimensional models can occur in just one dimension, changes in ability for multidimensional models can occur in multiple dimensions. Therefore, in order to determine how SEθ changes as a function of ability, a gradient describing the direction of movement within the ability plane must first be established. For simplicity, we aligned this gradient with the direction of steepest decent. In practice, observed SEθ values are likely to vary depending on changes in this angle.


  • Andrich D. Rasch models for measured. Newbury Park, CA: Sage; 1988.
  • Army STARRS. National Institute of Mental Health Strategic Plan. 2011.
  • Baddeley AD, Hitch G. Working memory. In: Bower GH, editor. The psychology of learning and motivation: Advances in research and theory. New York: Academic Press; 1974. pp. 47–89.
  • Baker FB, Kim S-H. Item response theory: Parameter estimation techniques. 2. New York: Dekker; 2004.
  • Batchelder WH. Cognitive psychometrics : Using multinomial processing tree models as measurement tools. In: Embretson SE, editor. Measuring psychological constructs : Advances in model-based approaches. Washington, DC: American Psychological Association; 2010. pp. 71–93.
  • Bollen KA. Structural equations with latent variables. New York: Wiley; 1989.
  • Brandt M. Bridging the gap between measurement models and theories of human memory. Journal of Psychology. 2007;215:72–85.
  • Brown GG, Turner T, Mano QR, Bolden K, Thomas ML. Experimental manipulation of working memory model parameters: An exercise in construct validity. 2012 Manuscript submitted for publication. [PubMed]
  • Cabeza R, Kingstone A, editors. Handbook of functional neuroimaging of cognition. 2. Cambridge, Mass: MIT Press; 2006.
  • Cai L. High-dimensional exploratory item factor analysis by a Metropolis-. Hastings Robbins-Monro algorithm. Psychometrika. 2010;75:33–57.
  • Campbell DT, Fiske DW. Convergent and discriminant validation by the multitrait-multi-method matrix. Psychological Bulletin. 1959;56:81–105. [PubMed]
  • Carroll JB. Human cognitive abilities: A survey of factor-analytic studies. New York: Cambridge University Press; 1993.
  • Chalmers P. mirt: Multidimensional Item Response Theory. 2011
  • Chen WH, Thissen D. Local dependence indexes for item pairs using item response theory. Journal of Educational and Behavioral Statistics. 1997;22:265–289.
  • Clark SE, Gronlund SD. Global matching models of recognition memory: How the models match the data. Psychonomic Bulletin & Review. 1996;3:37–60. [PubMed]
  • Cowan N. Evolving conceptions of memory storage, selective attention, and their mutual constraints within the human information-processing system. Psychological Bulletin. 1988;104:163–191. [PubMed]
  • Cronbach LJ. The two disciplines of scientific psychology. American Psychologist. 1957;12:671–684.
  • Cronbach LJ, Meehl PE. Construct validity in psychological tests. Psychological Bulletin. 1955;52:281–302. [PubMed]
  • Davachi L, Wagner A. Hippocampal contributions to episodic encoding: Insights from relational and item-based learning. Journal of Neurophysiology. 2002;88:982–990. [PubMed]
  • de Ayala RJ. The influence of dimensionality on the graded response model. Applied Psychological Measurement. 1994;18:155–170.
  • de Ayala RJ. The theory and practice of item response theory. New York, NY: Guilford Press; 2009.
  • Delis DC, Kramer JH, Kaplan E, Ober BA. The California Verbal Learning Test. 2. San Antonio: The Psychological Corporation; 2000.
  • Diana RA, Yonelinas AP, Ranganath C. Imaging recollection and familiarity in the medial temporal lobe: A three-component model. Trends in Cognitive Sciences. 2007;11:379–386. [PubMed]
  • Embretson SE. Construct validity: Construct representation versus nomothetic span. Psychological Bulletin. 1983;93:179–197.
  • Embretson SE. Cognitive design systems : a structural modeling approach applied to developing a spatial ability test. In: Embretson SE, editor. Measuring psychological constructs : Advances in model-based approaches. Washington, DC: American Psychological Association; 2010. pp. 274–273.
  • Embretson SE, Gorin J. Improving construct validity with cognitive psychology principles. Journal of Educational Measurement. 2001;38:343–368.
  • Embretson SE, Reise SP. Item response theory for psychologists. Mahwah, NJ, US: Lawrence Erlbaum Associates Publishers; 2000.
  • Gelman A, Carlin JB, Stern HS, Rubin DB. Bayesian Data Analysis. 2. Boca Raton: Chapman and Hall/CRC; 2004.
  • Green DM, Swets JA. Signal detection theory and psychophysics. New York: Wiley; 1966.
  • Gronlund SD, Shiffrin RM. Retrieval strategies in recall of natural categories and categorized lists. Journal of Experimental Psychology. 1986;12:550–561. [PubMed]
  • Gur RC, Erwin RJ, Gur RE. Neurobehavioral probes for physiologic neuroimaging studies. Archives of General Psychiatry. 1992;49:409–414. [PubMed]
  • Gur RC, Jaggi JL, Ragland JD, Resnick SM, Shtasel D, Muenz L, Gur RE. Effects of memory processing on regional brain activation: cerebral blood flow in normal subjects. The International journal of neuroscience. 1993;72:31–44. [PubMed]
  • Gur RE, Jaggi JL, Shtasel DL, Ragland JD. Cerebral blood flow in schizophrenia: Effects of memory processing on regional activation. Biological Psychiatry. 1994;35:3–15. [PubMed]
  • Gur RC, Ragland JD, Moberg PJ, Turner TH, Bilker WB, Kohler C, et al. Computerized neurocognitive scanning: I. methodology and validation in healthy people. Neuropsychopharmacology. 2001;25:766–776. [PubMed]
  • Gur RC, Ragland JD, Mozley LH, Mozley PD, Smith R, Alavi A, et al. Lateralized changes in regional cerebral blood flow during performance of verbal and facial recognition tasks: Correlations with performance and “effort” Brain and Cognition. 1997;33:388–414. [PubMed]
  • Gur RC, Richard J, Hughett P, Calkins ME, Macy L, Bilker WB, Brensinger C, Gur RE. A cognitive neuroscience-based computerized battery for efficient measurement of individual differences: Standardization and initial construct validation. Journal of Neuroscience Methods. 2010;187:254–262. [PMC free article] [PubMed]
  • Gur RC, Richard J, Calkins ME, Chiavacci R, Hansen JA, Bilker WB, Loughead J, Connolly JJ, Qiu H, Mentch FD, Abou-Sleiman PM, Hakonarson H, Gur RE. Age group and sex differences in performance on a computerized neurocognitive battery in children age 8–21. Neuropsychology. 2012 Epub ahead of print. [PMC free article] [PubMed]
  • Haberman SJ. The analysis of frequency data. Chicago: University of Chicago Press; 1974.
  • Haynes SN, Smith GT, Hunsley JD. Scientific foundations of clinical assessment. New York: Routledge; 2011.
  • Hoerger M, Quirk SW, Weed NC. Development and validation of the Delaying Gratification Inventory. Psychological Assessment. 2011;23:725–738. [PMC free article] [PubMed]
  • Ingham J. Individual differences in signal detection. Acta Psychologica. 1970;34:39–50. [PubMed]
  • Kelley W, Miezin F, McDermott K, Buckner R, Raichle M, Cohen N, et al. Hemispheric specialization in human dorsal frontal cortex and medial temporal lobe for verbal and nonverbal memory encoding. Neuron. 1998;20:927–936. [PubMed]
  • Kim H, Cabeza R. Differential contributions of prefrontal, medial temporal, and sensory-perceptual regions to true and false memory formation. Cerebral Cortex. 2007;17:2143–2150. [PubMed]
  • Lezak MD, Howieson DB, Loring DW, Hannay HJ, Fischer JS. Neuropsychological assessment. 4. New York, NY: Oxford University Press; 2004.
  • Loevinger J. Objective tests as instruments of psychological theory. Psychological Reports, Monograph Supplement. 1957;3:635–694.
  • Lord FM. Applications of item response theory to practical testing problems. Hillsdale, NJ: Lawrence Erlbaum; 1980.
  • Lord FM, Novick MR. Statistical theories of mental test scores. Reading MA: Addison-Wesley; 1968.
  • Magnusson D. Test theory. Reading MA: Addison-Wesley; 1966.
  • McDonald RP. Test theory: A unified treatment. Mahwah, NJ: Lawrence Erlbaum; 1999.
  • McDonald RP. A Basis for Multidimensional Item Response Theory. Applied Psychological Measurement. 2000;24:99–114.
  • Mellenbergh GJ. Measurement precision in test score and item response models. Psychological Methods. 1996;1:293–299.
  • Mickes L, Wixted JT, Wais PE. A direct test of the unequal-variance signal-detection model of recognition memory. Psychonomic Bulletin & Review. 2007;14:858–865. [PubMed]
  • Mulaik SA. Foundations of factor analysis. 2. Boca Raton, FL: Chapman & Hall/CRC; 2010.
  • Muraki E, Carlson JE. Full-information factor analysis for polytomous item responses. Applied Psychological Measurement Special Issue: Polytomous Item Response Theory. 1995;19:73–90.
  • Onyper SV, Zhang YX, Howard MW. Some-or-none recollection: Evidence from item and source memory. Journal of Experimental Psychology. 2010;139:341–364. [PMC free article] [PubMed]
  • R Development Core Team. R: A language and environment for statistical computing. R Foundation for Statistical Computing; Vienna, Austria: 2011. URL
  • Raaijmakers JGW, Shiffrin RM. SAM: A theory of probabilistic search of associative memory. In: Bower GH, editor. The Psychology of Learning and Motivation. New York: Academic Press; 1980. pp. 207–262.
  • Reckase MD. Multidimensional item response theory. New York: Springer; 2009.
  • Reise SP, Waller NG. Item response theory and clinical measurement. Annual Review of Clinical Psychology. 2009;5:25–46. [PubMed]
  • Revelle W. psych: Procedures for Personality and Psychological Research. 2011
  • Samejima F. The graded response model. In: van der Linden WJ, Hambleton RK, editors. Handbook of modern item response theory. New York, NY: Springer; 1997. pp. 85–100.
  • Snodgrass J, Corwin J. Pragmatics of measuring recognition memory - applications to dementia and amnesia. Journal of Experimental Psychology-General. 1988;117:34–50. [PubMed]
  • Strauss ME, Smith GT. Construct validity: Advances in theory and methodology. Annual review of clinical psychology. 2009;5:1–25. [PMC free article] [PubMed]
  • Thomas ML. The value of item response theory in clinical assessment: A review. Assessment. 2011;18:291–307. [PubMed]
  • Townsend JT, Neufeld RWJ. Introduction to special issue on contributions of mathematical psychology to clinical science and assessment. Journal of Mathematical Psychology. 2010;54:1–4.
  • US Army. Army health promotion, risk reduction, and suicide prevention. Report 2010. 2010 Retrieved October 11, 2011, from
  • Yonelinas AP. The nature of recollection and familiarity: A review of 30 years research. Journal Of Memory & Language. 2002;46:441–517.
  • Wheeler M, Buckner R. Functional dissociation among components of remembering: Control, perceived oldness, and content. Journal of Neuroscience. 2003;23:3869–3880. [PubMed]
  • Wickens TD. Elementary signal detection theory. New York: Oxford University Press; 2002.
  • Wixted JT. Dual-process theory and signal-detection theory of recognition memory. Psychological Review. 2007;114:152–176. [PubMed]
  • Wixted JT, Squire LR. The medial temporal lobe and the attributes of memory. Trends in Cognitive Sciences. 2011;15:210–217. [PMC free article] [PubMed]