Personalized and optimal treatment estimated by a value search method in a randomized clinical trial with psychodynamic versus cognitive behavioral therapy for depression.

Abstract

Objective: To estimate an optimal treatment selection in data from a recent RCT comparing short-term psychodynamic (STPP) and cognitive behavioral therapy (CBT) for depression.

Method: An estimator from causal inference with a $\mathrm{value}\mathrm{search}$ optimization was applied to estimate the optimal treatment choice for 100 patients diagnosed with major depressive disorder. Cross-validated prediction was performed, and the population mean outcome under optimal treatment was estimated.

Results: Outcomes (psychosocial functioning and depressive symptoms) for estimated optimal treatment were stratified by treatments received and compared for concordant versus discordant assignments. Large effect sizes were observed (range 0.41-2.07). "Positive beliefs about rumination" was found to be an important predictor, with higher baseline levels favoring CBT. An improvement of more than 30% relative to randomized treatment was considered clinically significant. For most patients (75%), CBT was identified as the optimal treatment, whereas patients in the optimal STPP group showed the greatest benefit for concordant assignment.

Discussion: The $\mathrm{value}\mathrm{search}$ method showed effective personalization, with greater improvement than typically reported in other methods of personalization with similar therapies and outcomes. A simple parametric user-specified treatment regime is easy to interpret, and robustness to model misspecification is appealing. Limited external validity is expected (small sample), and prospective validation is needed to guide clinical treatment selection. Head-to-head comparisons with other methods are required to draw conclusions about relative performance.