Prevalence estimation and validation of new instruments in psychiatric research: an application of latent class analysis and sensitivity analysis

Abstract

Prevalence and validation studies rely on imperfect reference standard (RS) diagnostic instruments that can bias prevalence and test characteristic estimates. The authors illustrate 2 methods to account for RS misclassification. Latent class analysis (LCA) combines information from multiple imperfect measures of an unmeasurable latent condition to estimate sensitivity (Se) and specificity (Sp) of each measure. With simple algebraic sensitivity analysis (SA), one uses researcher-specified RS misclassification rates to correct prevalence and test characteristic estimates and can succinctly summarize a range of scenarios with Monte Carlo simulation. The authors applied LCA to a validation study of a new substance use disorder (SUD) screener and a larger prevalence study. With a traditional validation study analysis in which an error-free RS (Structured Clinical Interview for DSM-IV Axis I Disorders [SCID]; M. H. First, R. L. Spitzer, M. Gibbon, & J. Williams, 1990) is assumed, the authors estimated the screener had 86% Se and 75% Sp. Validation study estimates from LCA were 91% Se, 81% Sp (screener), 73% Se, and 98% Sp (SCID). SA in the prevalence study suggested the prevalence of SUD was underestimated by 22% because SCID was assumed to be error-free. LCA and SA can assist investigators in relaxing the unrealistic assumption of perfect RSs in reporting prevalence and validation study results.