Baseline Symptom Severity is a Robust Predictor of Treatment Dropout Across Psychiatric Diagnoses: An Individual-Participant Analysis of Seven NIMH Trials

Objective: Treatment dropout undermines psychiatric trials and clinical care, and its predictors are usually studied within one diagnosis. Using harmonized individual participant data (IPD), we investigated which dropout predictors are robust across diagnoses and which are tied to study design. Method: IPD analysis of seven NIMH trials (N = 8543). A study-stratified Cox model (race-complete N = 6677) tested age, sex, race, and within-study z-scored severity, and each predictor was graded against proportional-hazards, equal-coefficient (follow-up tier), leave-one-study-out, random-effects, and severity-recoding checks. Results: Lower baseline severity robustly predicted treatment completion, significant and same-direction under every specification: z-score, percentile-rank and median-split codings (HR = 1.124, p < 0.001), leave-one-study-out, random-effects pooling, and both short and long follow-up trials. The study-specific severity estimates were directionally consistent in all seven trials, but the magnitude of the pooled association was driven mainly by the short-follow-up MDD tier (HR = 1.270, p < 0.001 vs. HR = 1.057, p = 0.026), and only STAR*D reached significance on its own. Age, sex, and race were informative but study-dependent, varying across follow-up horizon (equal-coefficient test rejected for age, sex, and severity) and between studies (I2 = 72%-84%); proportional hazards held for age and race but not severity or sex (χ2 = 18.1, df = 4, p = 0.001). Conclusion: Baseline symptom severity is a robust, diagnosis-spanning marker of dropout risk and a practical target for retention support, whereas age, sex, and race need study-specific interpretation. Grading pooled estimates with assumption checks distinguishes durable transdiagnostic signals from design-bound ones, a template for future IPD dropout research.