Neuropsychological testing assesses brain-behavior functioning, typically to examine whether an injury or illness has negatively affected specific cognitive functions. However, several factors – including fatigue, disengagement, distraction, unmotivated responding, and intentional underperformance – can interfere with the reliable and valid measurement of cognitive functioning. Gauging whether such factors have affected the measurement of cognitive functioning is a significant challenge. Over the past three decades, research to formally discern valid from invalid test performance has rapidly increased, leading to assessment tools termed performance validity tests (PVTs; Larrabee, Reference Larrabee2012). PVTs have provided substantial benefit to neuropsychological evaluations by providing evidence for or against credible cognitive performance.
Despite their clear utility, most PVT measures have limitations. Standalone PVTs typically require 10–15 min to administer, and many are still administered by paper and pencil. Such measures may not be suitable for “big data” acquisition, even though data integrity is a substantial concern in data collected at scale. Though several embedded PVTs exist, as described in Bilder & Reise (Reference Bilder and Reise2019), neurocognitive assessments would benefit from having measures of performance validity built into every test, which may be facilitated by examining aberrant patterns of responses or speed. PVTs also rely on cutoff scores, which prevent exploration of potential gradients in performance validity (Walters et al., Reference Walters, Berry, Rogers, Payne and Granacher2009) and impede flexibility for individuals utilizing the tests. For example, a clinician or researcher may value minimizing either false positives or false negatives, depending on population base rates of invalid performance or the evaluation context (McCormick et al., Reference McCormick, Yoash-Gantz, McDonald, Campbell and Tupler2013). Cutoff measures also preclude examination of potential variations in task engagement across a test battery. Relatedly, PVTs are typically used to predict insufficient engagement on separate, non-PVT measures, occasionally leading to challenging discrepancies (Loring & Goldstein, Reference Loring and Goldstein2019) or false positives (Lippa, Reference Lippa2018). In addition, PVTs usually do not factor in speed of responding, though using speed in addition to accuracy may increase the sensitivity of PVTs (Kanser et al., Reference Kanser, Rapport, Bashem and Hanks2019; Lupu et al., Reference Lupu, Elbaum, Wagner and Braw2018). Innovative methods for performance validity that leverage item-level psychometrics (e.g., response consistency) may be beneficial for acquisition of large samples over the internet or for tests administered through telehealth, which has become especially relevant during the COVID-19 pandemic.
Examining item-level responses has the potential to increase sensitivity to insufficient effort and task disengagement. Person-fit metrics are one valuable tool to detect problematic responding at the item level. Person-fit (Reise, Reference Reise1990; Tatsuoka & Linn, Reference Tatsuoka and Linn1983; Tellegen, Reference Tellegen1988) is a general term for psychometric methods designed to assess the consistency of a response pattern using some prior information (e.g., relative item difficulty). These methods examine whether an individual’s pattern of responding on a test fits with the individual’s overall ability level or the overall patterns of performance across all individuals. As a simple example, consider a 10-item test on which the first 5 items are of low difficulty and the last 5 items are of high difficulty. If an examinee answers the first 5 items correctly and last 5 items incorrectly, that pattern would fit the normative data well (high person-fit score) and suggest an ability level in between the difficulties of the easy items and the difficulties of the hard items. However, if an examinee answered the first 5 items incorrectly and last 5 items correctly, that pattern would not fit the normative data well – correct responses on hard items suggest that the examinee is of very high ability, whereas incorrect responses on easy items suggest that the examinee is of very low ability. This examinee would receive a low person-fit score (indicating person misfit). We propose that this lack of fit may be one useful indicator of invalid neurocognitive performance (e.g., reflecting haphazard responses or insufficient effort).
Here, we describe and apply novel methods for developing data-driven, embedded performance validity measures from item-level responses in the Penn Computerized Neurocognitive Battery (PennCNB; Gur et al., Reference Gur, Ragland, Moberg, Bilker, Kohler, Siegel and Gur2001, Reference Gur, Richard, Calkins, Chiavacci, Hansen, Bilker and Gur2012, Reference Gur, Richard, Hughett, Calkins, Macy, Bilker and Gur2010), a neurocognitive battery that has been applied worldwide in numerous studies. First, we describe and present data from simulations that model invalid responding using a multivariate combination of metrics. Second, we examine associations of these simulation-derived metrics with established validity rules in “real-world” data from two large, independent samples. Finally, we examine associations of these performance validity metrics with psychopathology in these samples to examine whether individuals with elevated psychopathology are likely to be identified as false positives on these metrics.
Method
Neurocognitive battery
The PennCNB is a publicly available collection of tests assessing a range of neurocognitive functions (Gur et al., Reference Gur, Richard, Hughett, Calkins, Macy, Bilker and Gur2010; Moore et al., Reference Moore, Reise, Gur, Hakonarson and Gur2015). The 14 core PennCNB tests take approximately one hour to complete (see eTable 1). Numerous studies support the PennCNB’s validity and reliability, including sensitivity to individual differences (Gur et al., Reference Gur, Richard, Calkins, Chiavacci, Hansen, Bilker and Gur2012) and expected neurobehavioral profiles in clinical conditions (e.g., Aliyu et al., Reference Aliyu, Calkins, Swanson, Lyons, Savage and May2006; Hartung et al., Reference Hartung, Kim, Laney, Hooper, Radcliffe, Port and Furth2016; Roalf et al., Reference Roalf, Gur, Almasy, Richard, Gallagher, Prasad and Gur2013). The battery has been applied in large-scale genomic studies (Greenwood et al., Reference Greenwood, Lazzeroni, Maihofer, Swerdlow, Calkins, Freedman and Braff2019; Gulsuner et al., Reference Gulsuner, Stein, Susser, Sibeko, Pretorius, Walsh and McClellan2020; Scott et al., Reference Scott, Moore, Stein, Pretorius, Zingela, Nagdee and Gur2021), intervention studies (Bhatia et al., Reference Bhatia, Mazumdar, Wood, He, Gur, Gur and Deshpande2017; Scott et al., Reference Scott, Lynch, Cenkner, Kehle-Forbes, Polusny, Gur and Oslin2021), and even in space flight (Basner et al., Reference Basner, Savitt, Moore, Port, McGuire, Ecker and Gur2015; Garrett-Bakelman et al., Reference Garrett-Bakelman, Darshi, Green, Gur, Lin, Macias and Turek2019).
To create performance validity measures, we used six PennCNB tests conducive for psychometric characterization of item-level responses. Specifically, for these methods, tests need to be self-paced (no rapid signal-detection tests), include core measures of accuracy, and comprise fixed sets of administered items. Based on these criteria, we selected the following tests. Episodic Memory: The Penn Word Memory Test (CPW), Penn Face Memory Test (CPF) (Thomas et al., Reference Thomas, Brown, Gur, Hansen, Nock, Heeringa and Stein2013) and Visual Object Learning Test (VOLT) (Glahn et al., Reference Glahn, Gur, Ragland, Censits and Gur1997) have similar formats for presentation and recall. The CPW/CPF presents 20 words or faces to remember, respectively, and the recall portion shows these targets mixed with 20 non-targets. The VOLT shows participants a series of 10 three-dimensional Euclidean shapes, and the recall portion shows these targets mixed with 10 new shapes. Verbal Reasoning: The Penn Verbal Reasoning Test (PVRT) measures language-mediated reasoning ability using a series of analogy problems patterned after the Educational Testing Service factor-referenced test kit. Nonverbal Reasoning: The Penn Matrix Analysis Test (PMAT) measures nonverbal reasoning ability using problems similar to those used in the Raven’s Progressive Matrices Test (Raven, Reference Raven1989) and WAIS-IV Matrix Reasoning (Wechsler, Reference Wechsler2008). Emotion Identification: The Penn Emotion Recognition Test (ER40) measures the social cognition domain of emotion identification. Participants are shown 40 individual faces and must determine whether the emotion expressed by the actor’s face is happiness, sadness, anger, fear, or none at all. There are 4 female and 4 male faces for each emotion (4 × 2 × 5 = 40).
Participants
To generate parameters for the simulations described below, we used item-level accuracy and response time data from the Philadelphia Neurodevelopmental Cohort (PNC), a large (n = 9498), community-based sample of youth between ages 8 and 22. Please see Calkins et al., (Reference Calkins, Merikangas, Moore, Burstein, Behr, Satterthwaite and Gur2015) and Supplementary Methods for more extensive information on PNC recruitment, enrollment, and procedures. The Institutional Review Boards at the University of Pennsylvania (Penn) and CHOP approved this study.
To examine the utility of validity metrics derived from these simulations, we applied these metrics to PNC and Marine Resiliency Study-II (MRS-II) data for external validation. MRS-II participants (n = 1444) have been described in detail previously (Acheson et al., Reference Acheson, Geyer, Baker, Nievergelt, Yurgil and Risbrough2015), with more information available in Supplementary Methods. The institutional review boards of the University of California San Diego, Penn, VA San Diego Research Service, and Naval Health Research Center approved this study.
Procedures
The PennCNB was administered by assessors who undergo extensive training by experienced assessors, including hands-on instruction, observation of mock sessions, feedback after practice assessments, and standardized certification procedures. For both samples, participants were administered the PennCNB on laptops by proctors trained by experienced Penn staff. However, the PNC used a one-to-one proctor-examinee setup, and the MRS-II administered tests with multiple service members and one proctor.
After test administration, assessors assigned a code to rate the quality of data and took detailed notes to indicate when an examinee exhibited disengagement from a task. In addition, algorithmic validity rules for PennCNB tests have been developed specific to each test to detect certain subject-related problems (e.g., extended periods of inattention, random or habitual responding). These rules include flagging of repeated instances of impossible response times (e.g., < 200 msec), excessive outliers in performance, and unusual response patterns (e.g., choosing the same option several times in a row). See Supplemental Methods for greater detail on these algorithmic validity rules. Experienced CNB data validators supervised by neuropsychologists make ultimate decisions to designate a test as invalid by integrating algorithmic rules, assessor comments, and visual data inspection. Though helpful for quality assurance, these validation procedures require significant manual intervention, are infeasible for collecting data at scale, and are not “gold standard” criteria, as they are less likely to detect more subtle task disengagement or insufficient effort than typical PVTs. Nonetheless, associations between our performance validity metrics and these validation rules provide a useful initial test of criterion validity.
Performance validity metrics
We used a combination of three methods to generate performance validity metrics.
Person-fit metric
Person-fit is a measure of how feasible an individual’s response pattern is given item characteristics (e.g., difficulty). Many person-fit indicators have been described – see Karabatsos (Reference Karabatsos2003) and Meijer & Sijtsma (Reference Meijer and Sijtsma2001). We selected two indicators that are least correlated with each other: the point-biserialFootnote 1 method (Donlon & Fischer, Reference Donlon and Fischer1968) and Kane & Brennan (Reference Kane and Brennan1980) dependability method. Both measures indicate the degree to which a person’s response pattern conforms to sample-level response patterns.Footnote 2 See Supplemental Methods for more details on these specific methods.
Response timeFootnote 3 outlier metric
This metric reflects the proportion of items on which the response time (RT) is at least two standard deviations from the expected RT, given that the response is correct or incorrect. Each item’s response time is modeled using the following regression equation:
$$Response\;Time = intercept + \;{\beta _1}\left( {Response} \right) + error$$
where “Response” is a dichotomous variable indicating whether the response is correct or incorrect. In this model, each item has an expected response time (given a correct/incorrect response). If the examinee’s response time is > 2.0 SD above or below the expected response time according to the model, then it is considered an outlier. The proportion of responses that are outliers is subtracted from 1.0 to arrive at a score for this metric, with higher values indicating greater validity. For example, if 10% of an individual’s responses were outliers, the score for this metric would be 1.00–0.10 = 0.90 (90% valid/non-outliers).
“Easiest Items” metric
Similar to most standalone PVTs, this metric reflects items that are answered correctly by an overwhelming majority of individuals. This metric determines the proportion correct on the three easiest items on each measure, where “easiest” is indicated by the highest proportion correct in the full sample. Typically, these items were answered correctly by >95% of the reference sample. This metric has 4 possible values (0%, 33%, 67%, 100%). Note that this method will inevitably share some variance with person-fit methods but may be particularly sensitive to intentional underperformance, where a common putative strategy is to feign memory loss (Tan et al., Reference Tan, Slick, Strauss and Hultsch2002), resulting in easy items answered incorrectly.
Statistical analyses
All analyses were performed in R v3.6.1 (R Core Team, 2020).
Simulations
One set of simulations was performed for each test. First, item responses and response times were estimated using parameters generated from the PNC. Next, simulations were conducted using these parameters for n = 2,000 simulated participants. Item-level response patterns and response time estimates were used to generate varying proportions of invalid responses for a small percentage of simulated participants. Next, the three performance validity metrics were calculated for each measure for each simulated examinee. Finally, a model was built and cross-validated to predict invalid responses in the simulations. More specific steps for performing simulations were as follows:
1. To simulate item responses, we first estimated population parameters (e.g., discriminations, difficulties) of item-level data from the PNC, using the 2-parameter logistic IRT model (2PLM). For reference, eTable 2 shows the item-wise proportions correct for data on which simulations were based, and eTable 3 shows the 2PLM item parameter estimates used in simulations. In rare cases where a negative discrimination parameter was estimated in real data, it was fixed to zero before simulations. Of note, although responses from simulated examinees are unrelated to those from real examinees, simulated item responses have the exact same parameters (discrimination, difficulty) as items from the PNC. These item parameter estimates were used as population parameters for simulations in #3 below. See Supplementary Methods for details.
2. Response time distributions for correct and incorrect responses were estimated separately for each item using data from the PNC, using the JohnsonFit() function in the SuppDists package (Wheeler, Reference Wheeler2016). These measures provide item-level response time estimates from a distribution of responses most similar to a target distribution. For each item of each test, the program generated two sets of four estimates (one set each for correct and incorrect responses), used in simulations described in #4 below.
3. Response patterns were then simulated for 2,000 examinees, using population parameters estimated from #1 above and the simdata() function in the mirt package (Chalmers, Reference Chalmers2012). N = 2,000 was chosen because it is small enough to ensure some variability due to sample size – for example, two simulations of N = 2,000 will vary in their proportions correct (especially of the N = 100 invalid responders) – yet not so large that the phenomena demonstrated here may be limited to extremely large sample sizes, casting doubt on their utility.
4. For each simulated correct/incorrect item response from #3, response time was simulated by drawing randomly from a distribution with the same moments as those estimated in #2, using the rJohnson() function in the SuppDists package (Wheeler, Reference Wheeler2016). Therefore, if real-world examinees tended to have highly skewed response times for Item X on Test Y, response times for simulated examinees for Item X on Test Y were drawn from a similarly skewed distribution.
5. For a subset of the 2,000 simulated examinees, a percentage of their responses (see #8 below) were changed to be “invalid.” For these simulated low-validity examinees, the probability of a correct response on a percentage of responses was set at chance level (i.e., unrelated to discrimination and difficulty item parameters), and response times for these chance-level responses were drawn from a random uniform distribution between 200 and 3000 milliseconds. This randomness was meant to simulate careless responding insofar as response times and accuracy were unrelated to the discrimination or difficulty of the item (analogous to random or unmotivated responding). The “invalid” nature of response patterns of these randomly selected examinees was indicated (value=1) in an additional column. Other examinees received a 0 in this column.
Primary analyses simulated invalid performance for 100 (5%) of the 2,000 examinees, which was selected as a reasonable base rate of non-credible performance for non-clinical/community samples (Martin & Schroeder, Reference Martin and Schroeder2020; Ross et al., Reference Ross, Poston, Rein, Salvatore, Wills and York2016). However, given research in healthy undergraduates that found higher rates of invalid performance (Roye et al., Reference Roye, Calamia, Bernstein, De Vito and Hill2019), we also modeled invalid performance in 200 (10%) simulated participants as a sensitivity test.
6. With valid and invalid response patterns simulated, performance validity metrics were calculated for each measure for each simulated examinee. Person-fit metrics were calculated using r.pbis() and E.KB() functions in the PerFit package (Tendeiro et al., Reference Tendeiro, Meijer and Niessen2016).
7. To assess the performance of the validity metrics in detecting invalid responders in this simulation, we split the sample randomly into a training (75%) and testing (25%) set. In the training set, a logistic regression model was built to predict invalid responders using the new performance validity metrics. Then, the model (built in the training set) was used to predict invalid responses in the testing set, with area under the receiver operating characteristic (ROC) curves generated using each performance validity metric individually (e.g., only the person-fit metric) and using all metrics combined. Of note, to eliminate collinearity among performance validity metrics, they were regressed out of each other before building models that included them all.
8. Steps 3 through 7 were repeated for all possible percentages of invalid responses (#5 above) per test. For example, for the PMAT (24 items), a set of simulations was conducted for when 100 examinees had invalid responses to 24 items (24/24 = 100% invalid), 23 items (23/24 = 95.8% invalid), 22 items (22/24 = 91.7% invalid), etc., to only 1 item invalid. For exploratory purposes, a response pattern simulated to be invalid was considered invalid even if it had only one “random” response. Although examining a large range of invalid responses may not reflect distributions of invalid responses in real-world participants, this method avoids choosing an arbitrary cutoff and examines the methods’ utility across a distribution of levels of invalid responding. The ability of performance validity metrics to detect invalid response patterns in the test sample was recorded at each possible percentage of invalid responses. For each test and each percentage of invalid responses, 500 simulations (2000 examinees each) were run. In total, there were: 500 simulations × 179 total items (across 6 tests) = 89,500 simulations total.
Coefficients for each simulation were saved, and final hard-coded equations below represent the average across all simulations.
Application of validity estimates to developmental and adult neurocognitive data
Next, as an initial test of criterion validity, we examined whether the multivariate CNB validity estimates (MCVE) were associated with established CNB validation rules. We generated MCVEs for PNC and MRS-II participants and used these scores to predict previously established valid/invalid classifications for each test using logistic regressions and ROC curves. Specifically, equations created from “Simulations” above were used to calculate a weighted sum of the four individual validity metrics per person, providing out-of-sample validity estimates. These validity estimates were used to “predict” valid/invalid classification rules determined by a human team. As detailed above, although these are not gold standard PVTs, they serve as an important proof-of-concept test.
Finally, to assess whether the MCVE could be confounded by clinical conditions (depression, anxiety, etc.), we examined MCVE data for real PNC participants with and without six psychopathological diagnoses: depression, obsessive-compulsive disorder (OCD), post-traumatic stress disorder (PTSD), psychosis, attention deficit hyperactivity disorder (ADHD), and conduct disorder. See supplementary methods and Calkins et al. (Reference Calkins, Merikangas, Moore, Burstein, Behr, Satterthwaite and Gur2015) for greater details on psychopathology assessment and Calkins et al. (Reference Calkins, Moore, Merikangas, Burstein, Satterthwaite, Bilker and Gur2014) for clinical composition of the sample. Similar analyses were conducted for MRS-II; because clinical assessments differed, we compared participants with and without: depression, PTSD, anxiety, insomnia, alcohol problems, psychosomatic/physical problems, and marked lack of perspective-taking (low interpersonal reactivity). Specific instruments can be found in Supplemental Methods, and details of score generation for the seven MRS-II psychopathology scales and clinical composition can be found in Moore et al. (Reference Moore, Risbrough, Baker, Larson, Glenn, Nievergelt and Gur2017).
Results
Simulations
We first used simulations and the MCVE prediction model incorporating our three performance validity metrics to classify simulated examinees as valid versus invalid responders. Figure 1 shows the relationship between simulated effort (number of valid responses across items) and the model’s ability to classify response patterns as valid or invalid when simulating 5% of examinees as invalid. As expected, the more valid responses a simulated examinee gave (i.e., moving from left to right on the x-axis), the ability of the MCVE to detect invalid responding in these simulations decreased. Further, the individual metrics (e.g., person-fit) differed by test in how well they detected invalid responses. For example, for the ER40, individual metrics except the “easiest items” metric perform well (AUC > 0.80) up to 75% valid responding. By contrast, for the VOLT, the relative utility of the methods was less clear; the easiest items metric outperformed RT outlier and person-fit metrics between ∼30% and ∼90% valid responses, though all fell below AUC = 0.80 when 60% or more of responses were valid. Notably, for every test except PMAT, the combined MCVE performed marginally to substantially better than any individual metric, supporting the utility of our approach. See Supplemental Results for greater details on the shape of curves in Figure 1. As shown in eFigure 1, sensitivity analyses showed similar results when simulating 10% of examinees as invalid, supporting results above. Additionally, eFigure 2 shows that increases in the proportion of items invalid for each test generally results in decreased internal consistency (Cronbach’s α), supporting effects on validity.
Figure 1. Prediction accuracy (AUC) for the multivariate CNB performance validity estimate in predicting true (Simulated) careless responding in 5% of the simulated examinees, by proportions of valid responses. Note. For visual simplification, “Person-Fit Method” is the average of the person-fit indices. CPF = Penn Face Memory Test; CPW = Penn Word Memory Test; VOLT = Visual Object Learning Test; ER40 = Penn Emotion Recognition Test; PMAT = Penn Matrix Analysis Test; PVRT = Penn Verbal Reasoning Test.
The simulations summarized in Figure 1 were used to generate hard-coded equations for use in real data:
MCVEmem = 0.42 × Outlier_Score + 0.02 × Acc_3easy + 0.05 × PersonFit1 + 0.50 × PersonFit2
MCVEnon-mem = 0.34 × Outlier_Score + 0.00 × Acc_3easy + 0.22 × PersonFit1 + 0.44 × PersonFit2
where “MCVEmem” is for episodic memory tests, “MCVEnon-mem” is for all non-memory tests, “Outlier_Score” is the RT outlier validity metric, “Acc_3easy” is the easiest items metric, “PersonFit1” is the point-biserial metric, “PersonFit2” is the Kane-Brennan dependability metric, and coefficients reflect averages across each test type (memory and non-memory). Note that non-memory tests are combined because we intend for the MCVE to be generalizable across multiple tests. Furthermore, there is domain specificity in PVTs that might be lost when aggregating across memory and non-memory domains (Erdodi, Reference Erdodi2019). Supplementary file “MCVE_estimation.R” provides the R script for calculating validity estimates, where memory and non-memory versions can be specified by commenting out relevant lines indicated.
Application of validity metrics to two independent samples
Next, these equations were used to generate MCVE scores for participants in the PNC and MRS-II, which were used to predict valid/invalid classifications previously determined by the investigators’ quality assurance team for each test. eFigure 3 shows results of these ROC curves for the PNC. AUCs ranged from 0.61 for the PMAT to 0.75 for the CPW. Using an AUC > 0.70 cutoff, the MCVE shows acceptable classification accuracy for four of the six tests. eFigure 4 shows ROC results for the MRS-II. AUCs ranged from 0.65 for the PMAT to 0.94 for the ER40. In this sample, the MCVE achieved acceptable (or remarkable) classification accuracy for five of six tests. To facilitate decisions regarding use of MCVE thresholds, eFigures 5 and 6 and Supplementary File MCVE_sensitivity_specificity_coordinates.csv show how sensitivity and specificity vary as this threshold is varied. Overall, the table suggests that a specificity of at least 0.80 can be achieved with a threshold of approximately 0.90 ± 0.05 on the MCVE. Table 1 presents in-sample prediction statistics for the PNC and MRS-II. See Supplementary Results for more information.
Table 1. In-Sample prediction statistics for the six neurocognitive tests predicting human-determined validity, in the Philadelphia Neurodevelopmental cohort (PNC) and marine resiliency Study-II (MRS-II) samples
Note. PNC = Philadelphia Neurodevelopmental Cohort; MRS-II = Marine Resiliency Study-II; Thresh = threshold; Sens = sensitivity; Spec = specificity; AUC = area under the receiver operating characteristic curve; mean threshold across samples = 0.84.
Associations with psychopathology
eFigure 7 shows how psychiatric diagnostic groups differed on MCVE scores in the PNC. For OCD, PTSD, and conduct disorder, no significant differences in MCVE were found. Depression was associated with higher MCVE (i.e., more valid performance) on the PVRT (p < 0.005), CPW (p < 0.0005), and VOLT (p < 0.005). Psychosis was associated with lower MCVE on the ER40 (p < 0.05) but higher MCVE on the CPW (p < 0.0005). Finally, ADHD was associated with lower MCVE on all tests (p < 0.05 to p < 0.0005).
eFigure 8 shows how psychiatric diagnostic groups differed on MCVEs in the MRS-II. For PTSD, anxiety, insomnia, alcohol problems, and lack of perspective-taking, no significant differences in MCVE were found. Depression was associated with lower MCVE on the ER40 (p < 0.05), and psychosomatic/physical problems were associated with lower MCVE on the PVRT (p < 0.05) and VOLT (p < 0.0005).
Discussion
Here, we used novel, item-level psychometrics to establish data-driven, embedded performance validity metrics for individual tests in a widely used computerized neurocognitive battery. Our performance validity metrics detected patterns of invalid responding in data that simulated careless responding, even at subtle levels. Moreover, a multivariate combination of these metrics predicted previously established validity rules on most measures in independent developmental and adult datasets. Furthermore, there were few differences on this measure by clinical diagnostic groups, even in diagnoses associated with reduced cognitive performance in prior studies (e.g., Barzilay et al., Reference Barzilay, Calkins, Moore, Wolf, Satterthwaite, Cobb Scott and Gur2019; Gur et al., Reference Gur, Calkins, Satterthwaite, Ruparel, Bilker, Moore and Gur2014; Kaczkurkin et al., Reference Kaczkurkin, Sotiras, Baller, Barzilay, Calkins, Chand and Satterthwaite2020; Service et al., Reference Service, Vargas Upegui, Castaño Ramírez, Port, Moore, Munoz Umanes and Freimer2020). These results provide proof-of-concept evidence for the potential utility of data-driven performance validity metrics that leverage existing individual test data. Importantly, the MCVE also allows one to examine performance validity across a spectrum, providing greater flexibility for estimating the likelihood of invalid responses and offering more precise data for investigators interested in understanding their data at a fine-grained level. Though these methods were specifically applied to the PennCNB, they could be applied to any computerized test with accuracy and response time data.
Although robust measures exist to detect insufficient engagement on neurocognitive tests, such measures typically rely on cutoff scores, do not integrate response speed, do not allow one to examine variations in validity across a test battery, and are not scalable for large-scale data acquisition. Our metrics offer a method to examine gradients of performance validity, provide multiple measures across individual tests, and integrate response speed and item-level psychometrics. Similarly, item-level metrics from another computerized battery, the NIH Toolbox-Cognition, have recently shown promise predicting externally validated PVTs in a mild traumatic brain injury sample (Abeare et al., Reference Abeare, Erdodi, Messa, Terry, Panenka, Iverson and Silverberg2021). Such metrics may facilitate administration of neurocognitive tests through telehealth or other remote applications, as well as the integration of cognitive testing into “big data” efforts, including genomics. In addition, compared to standalone PVTs, the nature of these validity metrics may be less amenable to coaching, which can be problematic in forensic cases (Suhr & Gunstad, Reference Suhr and Gunstad2007). Examining variations in performance validity across a test battery may be especially informative in this regard, as sophisticated simulators of cognitive impairment may target specific tests rather than performing poorly across all measures (Lippa, Reference Lippa2018). Critically, we provide equations that PennCNB users can apply to examine performance validity in data already collected. However, these metrics should be integrated with other available data, where possible, including behavioral observations by test administrators and participant clinical history.
Importantly, performance validity metrics should be minimally associated with most clinical conditions, even those with established neurocognitive deficits. There were few associations between our validity metrics and diagnoses of depression, PTSD, anxiety, psychosis spectrum disorders, or externalizing disorders, supporting our approach. This feature is especially relevant considering neurocognitive performance differences in these conditions (e.g., Barzilay et al., Reference Barzilay, Calkins, Moore, Wolf, Satterthwaite, Cobb Scott and Gur2019; Gur et al., Reference Gur, Calkins, Satterthwaite, Ruparel, Bilker, Moore and Gur2014; Kaczkurkin et al., Reference Kaczkurkin, Sotiras, Baller, Barzilay, Calkins, Chand and Satterthwaite2020). However, there were diagnoses with reduced validity metrics that should be addressed. First, ADHD showed MCVE reductions across all tests. It is unclear whether these reductions are measuring a behavioral phenotype of ADHD or reflect high base rates of non-credible performance in ADHD (Suhr & Berry, Reference Suhr and Berry2017). Caution is warranted in applying the MCVE to individuals with ADHD before additional research is conducted. Second, youth with significant psychotic symptoms and Marines with depression showed slightly lower MCVE on ER40, so caution is warranted in interpreting ER40 MCVE data in these populations. In contrast, lower MCVE from the Marines with psychosomatic/physical problems appears consistent with research showing that individuals with higher psychosomatic complaints have elevated rates of performance validity concerns (Dandachi-FitzGerald et al., Reference Dandachi-FitzGerald, van Twillert, van de Sande, van Os and Ponds2016; Martin & Schroeder, Reference Martin and Schroeder2020).
Limitations and future directions
Although initial results are promising, additional studies are needed for validation. As mentioned above, one limitation is the lack of a “gold standard” PVT criterion, which should be examined in future work. Future validation studies could take several forms, including comparison of the MCVE to multiple PVTs or known groups designs in which patients expected to display underperformance because of external incentives are compared to those without external incentives (see Larrabee, Reference Larrabee2012). Though these methods could be applied to detect intentional underperformance, they are likely detecting unmotivated responding in these samples given the lack of external incentive for underperforming.
A second limitation is that data were simulated from unidimensional IRT models, though many tests will produce multidimensional data. For example, memory tests often produce data showing two separate dimensions reflecting the ability or tendency to detect a target (e.g., knowing one has seen a face previously) and to reject a foil (e.g., knowing one has not seen a face previously). Future studies should examine whether incorporating multidimensional psychometric modeling improves these measures. Relatedly, methods of detecting aberrant response patterns have advanced considerably since the development of metrics used here, especially in clinical and personality research (Falk & Ju, Reference Falk and Ju2020; Lanning, Reference Lanning1991; Tellegen, Reference Tellegen1988). While our metrics were selected based on generalizability (e.g., no reliance on model estimation) and appropriateness for neurocognitive tests, future work may benefit from testing more advanced methods.
A third limitation of the present study is that the weighted composites (equations on pg. 13) are based on simulations in which the number of invalid responders is fixed at 5%. While we tested 10% invalid in additional simulations, the weights in our equations could change slightly if the base rates of invalid performance were lower (e.g., high-stakes test) or higher (e.g., with external incentives to malinger).
Ensuring the validity of assessments is critical for accurate interpretation in medicine and psychology. However, PVTs arouse considerable controversy, as there can be substantial consequences to classifying an assessment as invalid, including stigma, loss of benefits, and psychological distress. Thus, most PVTs have emphasized specificity over sensitivity, which should be considered in applying these metrics. Relatedly, researchers must currently use subjective judgment (including examination of the MCVE distribution in their sample) to decide what proportion of the sample to flag or remove. Even in cases where the MCVE distribution shows extreme outliers, the judgment is still subjective, which is a weakness of the metrics proposed here. MCVEs have not been extensively examined in clinical or medicolegal contexts, and further research is needed before they are used in such contexts. Furthermore, neuropsychologists using PVTs should integrate multiple sources of data to determine invalid performance, including behavioral observations and expected cognitive performance given the patient’s history (Heilbronner et al., Reference Heilbronner, Sweet, Morgan, Larrabee, Millis and Participants2009).
Conclusions
These limitations notwithstanding, we provide promising initial data on the development and application of novel, data-driven, dimensional performance validity metrics for individual tests in the PennCNB. We show that these metrics are sensitive to subtle patterns of invalid responding in simulated data, correspond well with established quality assurance metrics in two independent datasets, and are not associated with diagnoses of most psychiatric disorders. These methods may facilitate modeling of unmotivated, random, or disengaged responding in remote or large-scale cognitive data collection, enhancing the validity and precision of such assessments.
Acknowledgements
This work was supported by the NIH (HD095278; MH089983; MH117014, MH119219; T32MH019112) and the Penn-CHOP Lifespan Brain Institute. The views expressed in this article are those of the authors and do not necessarily reflect the position or policy of the University of Pennsylvania or the Department of Veterans Affairs. We thank the VISN4 MIRECC Writing Group for helpful comments on a draft of this manuscript. Portions of these data were presented at the 2021 Annual Meeting of the International Neuropsychological Society.
Conflicts of interest
All authors declare no potential conflicts of interest.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S1355617722000893
Open access
