Variance decomposition of PANSS screening-to-baseline change reveals dominant subject-level variability and systematic rater effects in schizophrenia trials

Introduction

Changes in Positive and Negative Syndrome Scale (PANSS)[1] scores between screening and baseline in acute schizophrenia trials reflect a combination of factors, including measurement error, regression to the mean, and symptom fluctuation related to medication wash-out. When PANSS thresholds are used as inclusion criteria, scores at screening or baseline may be inflated at certain centers, potentially obscuring true illness severity and introducing downstream variability and bias. The extent to which short-interval PANSS changes reflect systematic differences at the country, site, or rater level—as opposed to predominantly subject-level variability—remains insufficiently characterized.

Methods

This study had two objectives: (1) to quantify the proportion of variance in PANSS screening-to-baseline change attributable to country, site, and rater across multiple trials; and (2) to identify countries, sites, and raters with atypical scoring patterns using empirical best linear unbiased predictions (BLUPs). Individual participant data were pooled from 15 multicenter acute schizophrenia trials (N = 6,259 subjects). Screening-to-baseline PANSS change was analyzed using linear mixed-effects models with restricted maximum likelihood estimation. Fixed effects included protocol and screening PANSS score, while random intercepts were specified for country, site, and rater nested within site to partition variance into hierarchical components. BLUPs were examined to detect entities with consistent deviations from overall patterns, indicating potential systematic scoring differences.

Results

The majority of variability in PANSS screening-to-baseline change was attributable to subject-level factors, with residual variance accounting for approximately 82% of the total. Among hierarchical components, between-rater variability contributed about 12% of the variance, while between-site variability accounted for approximately 5%. Between-country variability was negligible (<1%). A strong negative association was observed between screening PANSS score and subsequent change (−0.21 points per 1 additional PANSS point at screening), consistent with regression to the mean and potential baseline inflation. Notably, two-thirds of raters who participated in four or more protocols demonstrated consistent directional deviations in at least 75% of studies, suggesting stable rater-specific scoring tendencies. At the country level, the United States and Serbia showed significantly different predicted effects relative to the overall mean by about 0.5 PANSS point.

Discussion

These findings indicate that early PANSS change in acute schizophrenia trials is driven predominantly by subject-level variability, but with a meaningful and systematic contribution from rater effects and a smaller contribution from site effects. In contrast, country-level differences appear minimal once site and rater are accounted for, although BLUP estimates identified two countries (Serbia and the United States) with small but statistically reliable deviations from the overall mean screening-to-baseline change. The observed rater-level variance, along with consistent cross-study patterns in rater BLUPs, supports the presence of stable rater “styles,” including differences in administration or scoring practices. These effects are clinically relevant, as they can introduce measurement noise and reduce assay sensitivity. Accordingly, robust rater selection, training, calibration, and centralized monitoring remain critical for optimizing data quality in schizophrenia trials. Future work will extend this framework to post-randomization PANSS trajectories.

References

1. Kay, S. R.; Fiszbein, A.; Opler, L. A. (1987): The positive and negative syndrome scale (PANSS) for schizophrenia. In Schizophrenia Bulletin 13 (2), pp. 261–276.