Background: Bipolar disorder (BD) is a chronic and severe mental illness. It requires a non-discontinued pharmacological treatment to prevent mood recurrences but nonadherence to medication is frequent. To this date, medication adherence in BD has been mostly evaluated in cross-sectional studies and often considered as a stable trait. We aimed to study medication adherence using a prospective person-oriented approach. Methods: 1627 BD patients were followed on a 2 years period and assessed every 6 months. Medication adherence was evaluated at each visit with the Medication Adherence Rating Scale (MARS). A latent class mixed model (LCMM) was used to identify trajectory classes of adherence over time. Regression analyses and linear mixed model were used to search for predictors and covariables of the trajectories. Results: Three distinct and robust trajectories of medication adherence have been identified: one that starts poorly and keeps deteriorating (4.8%), one that starts poorly but improves (9%) and one that starts well and keeps improving (86.2%). A good tolerance to psychotropic medications, low depressive symptoms, the absence of comorbid eating disorders and anticonvulsant medication were associated to a better prognosis of adherence. Along the follow-up, the lower were the depressive symptoms, the better was the medication adherence (p <.001) Limitations: The use of a single measure of medication adherence although it is a validated instrument and a possible positive selection bias that might limit the generalization of our findings. Conclusions: This study demonstrates that medication adherence in BD patients is a heterogeneous and potentially variable phenomenon.< Réduire