OBJECTIVE: To classify a cohort of depressed adolescents recruited to the UK IMPACT trial, according to trajectories of symptom change. We examined for predictors and compared the data-driven categories of patients with a priori operational definitions of treatment response. METHOD: Secondary data analysis using growth mixture modelling (GMM). Missing data were imputed. Trajectories of self-reported depressive symptoms were plotted using scores taken at six nominal time points over 86 weeks from randomisation in all 465 patients. RESULTS: A piecewise GMM categorised patients into two classes with initially similar and subsequently distinct trajectories. Both groups had a significant decline in depressive symptoms over the first 18 weeks. Eighty-four per cent (84.1%, n = 391) of patients were classed as 'continued-improvers' with symptoms reducing over the duration of the study. A further class of 15.9% (n = 74) of patients were termed 'halted-improvers' with higher baseline depression scores, faster early recovery but no further improvement after 18 weeks. Presence of baseline comorbidity somewhat increased membership to the halted-improvers class (OR = 1.40, CI: 1.00-1.96). By end of study, compared with classes, a clinical remission cut-off score (≤27) and a symptom reduction score (≥50%) indexing treatment response misclassified 15% and 31% of cases, respectively. CONCLUSIONS: A fast reduction in depressive symptoms in the first few weeks of treatment may not indicate a good prognosis. Halted improvement is only seen after 18 weeks of treatment. Longitudinal modelling may improve the precision of revealing differential responses to treatment. Improvement in depressive symptoms may be somewhat better in the year after treatment than previously considered.