Using Individual-Participant-Data Meta-Analysis and Predictive Causal Modeling to Identify Treatment Responders

Introduction: Individual-participant-data meta-analysis (IPD-MA) coupled with predictive causal modeling (PCM) can help 1) address inconsistencies among studies by standardizing outcomes, 2) account for differences among study samples, 3) compare interventions not directly tested in a head-to-head trial, and 4) allows for identification and description of treatment responders.

Methods: A PCM was fit using data from 3 randomized trials (two three-arm, one two-arm) of 3 HIV prevention approaches (emotion regulation, family-based, knowledge and skills) involving 1320 adolescents with mental health concerns (age_mean=15.32; 59% female, 40% Caucasian, 16% Hispanic) assessed across 9-12 months following treatment. The PCM used Bayesian regularized regression (BLASSO) with 13 baseline characteristics (demographics, substance-use, sexual risk, psychiatric impairment, and psychiatric diagnosis) to build prediction models for each of the treatment arms in each study. Models were then used to generate 200 simulations of how each participant would have responded to each treatment in each study. Estimates were pooled across study and each arm compared with all others. BLASSO was again used to identify which background characteristics predicted treatment response. The outcome used condomless sex and number of partners to identify low or decreased risk from baseline (0) versus high or increased risk (1).

Results: The accuracy of the predictive models was moderate to good (area under the curve= .68 to .83). Results suggested minimal differences among prevention approaches (risk-difference scores: -.02 to .07) and identified psychiatric diagnosis, history of substance-use, female gender, and African- American Race as prominent predictors of treatment response with the pattern of response differing for each prevention approach.

Discussion: IPD-MA of clinical trials coupled with PCM can help identify characteristics of participants who best respond to interventions. The approach is heavily dependent on the accuracy of the prediction models that generate the simulations and more work is needed to improve this accuracy.