Large responses to antidepressants or methodological artifacts? A secondary analysis of STAR∗D, a single-arm,...

To replicate Stone et al's (2022) finding that the distribution of response in clinical antidepressant trials is trimodal with large, medium-effect, and small subgroups.

To apply finite mixture modeling to pre-post Hamilton Depression Rating Scale (HDRS) differences (n = 2184) of STAR∗D study's level 1, a single-arm, open-label study. For a successful replication, the best fitting model had to be trimodal, with comparable components as in Stone et al. Secondary/sensitivity analyses repeated the analysis for different baseline levels of depression severity, imputed values, and patient-reported depression symptoms.

The best fitting models were either bimodal or trimodal but the trimodal solution did not meet criteria for replication. The bimodal model had 1 component with HDRS mean change of M = -13.0, SD = 6.7 and included 65.3% of patients, and another component with M = -1.8, SD = 5.1, 34.7%, respectively. For the trimodal model, the component with the largest change (M = -14.3, SD = 6.4) applied to 52% of patients, which differed substantially from the large effect component in Stone et al (M = -18.8, SD = 5.1), which applied to 7.2%. Secondary/sensitivity analyses arrived at similar conclusions, and for patient-reported depression symptoms the best fitting models were unimodal or bimodal.

This analysis failed to identify the trimodal distribution of response reported in Stone et al. In addition to being difficult to operationalize for regulatory purposes, results from mixture modeling are not sufficiently reliable to replace the more robust approach of comparing mean differences in depression rating scale scores between treatment arms.