OBJECTIVE: To evaluate two hypotheses: that self-selection bias contributed to lack of medication advantage at the 36-month assessment of the Multimodal Treatment Study of Children With ADHD (MTA) and that overall improvement over time obscured treatment effects in subgroups with different outcome trajectories. METHOD: Propensity score analyses, using baseline characteristics and severity of attention-deficit/hyperactivity disorder symptoms at follow-up, established five subgroups (quintiles) based on tendency to take medication at the 36-month assessment. Growth mixture model (GMM) analyses were performed to identify subgroups (classes) with different patterns of outcome over time. RESULTS: All five propensity subgroups showed initial advantage of medication that disappeared by the 36-month assessment. GMM analyses identified heterogeneity of trajectories over time and three classes: class 1 (34% of the MTA sample) with initial small improvement followed by gradual improvement that produced significant medication effects; class 2 (52%) with initial large improvement maintained for 3 years and overrepresentation of cases treated with the MTA Medication Algorithm; and class 3 (14%) with initial large improvement followed by deterioration. CONCLUSIONS: We failed to confirm the self-selection hypothesis. We found suggestive evidence of residual but not current benefits of assigned medication in class 2 and small current benefits of actual treatment with medication in class 1.