Learning Outcomes That Maximally Differentiate Psychiatric Treatments

To develop a statistical method that uncovers clinically meaningful differences between active psychiatric treatments, even when traditional rating scales fail to do so.

We introduce Supervised Varimax (SV), a novel algorithm that transforms individual items from clinical rating scales into a small set of optimized outcomes that maximally differentiate treatments. SV was applied to data from two large, multi-center, randomized controlled trials: CATIE (schizophrenia) and STAR*D (treatment-resistant depression).

SV identified significant differential treatment effects that were not evident in the original analyses. In CATIE Phase I, olanzapine was more effective than quetiapine and ziprasidone for hostility, and perphenazine outperformed ziprasidone for emotional dysregulation. In Level 2 of STAR*D, bupropion augmentation was more effective than buspirone augmentation for patients with increased appetite. These findings were validated using post-hoc permutation testing and matched to clinical subgroups using simple, symptom-based rules.

SV enables precision psychiatry by optimizing outcome definitions to enhance treatment differentiation in RCTs. This approach provides interpretable, clinically actionable insights using existing trial data, without requiring complex predictive modeling or additional biomarkers.

CATIE (NCT00014001), STAR*D (NCT00021528).