Estimating the accuracy of neurocognitive effort measures in the absence of a “gold standard”.

Abstract

Psychologists frequently use symptom validity tests (SVTs) to help determine whether evaluees' test performance or reported symptoms accurately represent their true functioning and capability. Most studies evaluating the accuracy of SVTs have used either known-group comparisons or simulation designs, but these approaches have well-known limitations (potential misclassifications or lack of ecological validity). This study uses latent class modeling (LCM) implemented in a Bayesian framework to estimate SVT classification accuracy based on data obtained from real-life forensic evaluations. We obtained archival data from 1,301 outpatient evaluees who underwent testing with the Computerized Assessment of Response Bias (CARB), the Test of Memory Malingering (TOMM), and the Word Memory Test (WMT) in a forensic evaluation context. Under various data models, Markov chain Monte Carlo methods implemented via WinBUGS converged to target distributions that permitted Bayesian estimates of SVT accuracy. Under the most plausible model (conditional dependence in test results), classification accuracies (expressed as area under the "trapezoidal" receiver operating characteristic curve ± standard deviation) were as follows: CARB = 0.765 ± 0.030, WMT = 0.929 ± 0.020, and TOMM = 0.771 ± 0.034. At decision thresholds that hold false positive rates at 0.02, the SVTs would detect invalid responses (true positives) at rates of approximately 35%, 65%, and 49%, respectively, for the 3 tests. Though LCM methods have limitations, this study suggests that they offer an approach to SVT evaluation that avoids methodological pitfalls of known-group research designs while retaining ecological validity that is absent in simulation studies.