UCLA

The Positive and Negative Syndrome Scale (PANSS) is frequently described with five latent factors, yet published factor models consistently fail to replicate across samples and related disorders. We hypothesize that (1) a subset of the PANSS, instead of the entire PANSS scale, would produce the most replicable five-factor models across samples, and that (2) the PANSS factor structure may be different depending on the treatment phase, influenced by the responsiveness of the positive symptoms to treatment. Using exploratory factor analysis, confirmatory factor analysis and cross validation on baseline and post-treatment observations from 3647 schizophrenia patients, we show that five-factor models fit best across samples when substantial subsets of the PANSS items are removed. The optimal model at baseline (five factors) omits 12 items: Motor Retardation, Grandiosity, Somatic Concern, Lack of Judgment and Insight, Difficulty in Abstract Thinking, Mannerisms and Posturing, Disturbance of Volition, Preoccupation, Disorientation, Excitement, Guilt Feelings and Depression. The PANSS factor models fit differently before and after patients have been treated. Patients with larger treatment response in positive symptoms have larger variations in factor structure across treatment stage than the less responsive patients. Negative symptom scores better predict the positive symptoms scores after treatment than before treatment. We conclude that sparse factor models replicate better on new samples, and the underlying disease structure of Schizophrenia changes upon treatment.