Applicability of regression models for the analysis of ordinal PANSS data in schizophrenia: a cohort study

BACKGROUND: There is ongoing debate about the appropriate measurement level for symptom severity scores derived from clinical rating scales. The use of statistically inappropriate analytical methods for such data may distort results and lead to misinterpretations. AIM: To study the applicability of linear, beta, beta-binomial, and ordinal regression models for assessing changes in schizophrenia symptoms over time using the PANSS. METHODS: The study cohort comprised patients diagnosed with schizophrenia and schizoaffective disorder. Symptom severity was quantified using the Positive and Negative Syndrome Scale (PANSS). Observation period, sex, brexpiprazole prescription upon admission, brexpiprazole monotherapy, and dropout from the study prior to completion of follow-up were used as covariates. Bayesian mixed-effects regression models were fitted to the obtained dataset: Model 1 (normal distribution), Model 2 (ordered beta distribution), Model 3 (beta-binomial distribution), and Model 4 (ordinal regression). The applicability of the models was assessed using the fit index γ, the proportion of predicted values that corresponded with any possible schizophrenia symptom score on the PANSS. A model was considered consistent at a γ=1. Additionally, the 95% Highest Density Interval (HDI) was calculated for γ. RESULTS: The study enrolled 24 patients with schizophrenia (75% of whom were men) aged 20 to 45 years. The fit index γ was 0.00 (95% HDI 0.00–0.00) for Model 1; 0.17 (95% HDI 0.14–0.21) for Model 2; 1.00 (95% HDI 1.00–1.00) for Model 3; and 1.00 (95% HDI 1.00–1.00) for Model 4. Model 1 allows results that fall outside the range of the PANSS. Models 1 and 2 can produce fractional values. CONCLUSION: Statistical models designed to analyze continuous variables (linear and beta regression) are inapplicable for ordinal variables and, in particular, changes in schizophrenia symptoms over time on the PANSS. Beta-binomial and ordinal regression models are recommended for rating scores.