Identifying strategies to improve PANSS based dimensional models in schizophrenia: Accounting for multilevel structure, Bayesian model and clinical staging

Dimensional approaches can decompose a construct in a set of continuous variables, improving the characterization of complex phenotypes, such as schizophrenia. However, the five-factor model of the Positive and Negative Syndrome Scale (PANSS), the most used instrument in schizophrenia research, yielded poor fits in most confirmatory factor analysis (CFA) studies, raising concerns about its applications. Thus, we aimed to identify dimensional PANSS CFA models with good psychometric properties by comparing the traditional CFA with three methodological approaches: Bayesian CFA, multilevel modeling, and Multiple Indicators Multiple Causes (MIMIC) modeling. Clinical data of 700 schizophrenia patients from four centers were analyzed. We first performed a traditional CFA. Next, we tested the three techniques: 1) a Bayesian CFA; 2) a multilevel analysis using the centers as level; and 3) a MIMIC modeling to evaluate the impact of clinical staging on PANSS factors and items. CFA and Bayesian CFA produced poor fit models. However, when adding a multilevel structure to the CFA model, a good fit model emerged. MIMIC modeling yielded significant differences in the factor structure between the clinical stages of schizophrenia. Sex, age, age of onset, and duration of illness did not significantly affect the model fit. Our comparison of different CFA methods highlights the need for multilevel structure to achieve a good fit model and the potential utility of staging models (rather than the duration of illness) to deal with clinical heterogeneity in schizophrenia. Large prospective samples with biological data should help to understand the interplay between psychometrics concerns and neurobiology research.