OBJECTIVES: To identify groups of patients with schizophrenia with different compliance trajectories over three years and explore causes, correlates and consequences. METHODS: SOHO is an observational study of health outcomes in 10,000 patients with schizophrenia, from 10 countries, who initiated or changed their antipsychotic medication. Prospective data was available from seven measurement occasions over three years and comprising patient and clinician reported antipsychotic compliance, HRQL and side effects reported by patients, and clinician CGI ratings. Latent class (group-based) mixture modeling, following the methods developed by Nagin (1999) and Muthen (2004), was applied to identify the prevalence, and graphically characterize the profile, of compliance trajectories over time Model fit was evaluated using Lo-Mendell-Rubin Likelihood ratio tests, bootstrap p-values comparing n+1 versus n latent classes, and values of information criteria. Missing data were included under a MAR assumption using maximum likelihood estimation. RESULTS: Models with up to five classes were estimated and four classes interpreted. The modal class identified more than three quarters of the SOHO sample as largely compliant throughout the study; three further latent classes identified almost equal sized groups who increased or decreased their compliance over time, or who never complied. Cross tabulation of these latent compliance classes with classifications resulting from modeling of longitudinal clinical symptom and HRQL data revealed strong and statistically significant associations thereby illustrating co-evolution or linkage between parallel clinical, compliance and health outcome processes. CONCLUSIONS: Statistical summaries of longitudinal patterns offer important insights into the dynamics of compliance, and can be used to relate patterns of compliance with patterns of remission or relapse on clinical symptom scales, or change in HRQL. Mixture modelling of longitudinal data can be used to extend the propensity score approach to longitudinal measurements of clinical processes, which should characterize selection effects with higher precision and greater clinical validity.