Presentation (CTI)

OBJECTIVES: Group-based trajectory modeling (GBTM) is increasingly used to study longitudinal medication adherence, but design choices can introduce bias. We identify common design-related biases in claims-based GBTM adherence studies, and propose mitigation strategies.
METHODS: Drawing on published claims-based adherence trajectory studies and our research experience, we examined how bias arises when trajectories are used as the exposure versus the outcome and developed design recommendations.
RESULTS: When trajectories are used as the exposure (more common scenario), time-related biases include immortal time bias and reverse causality. Immortal time bias arises when individuals must survive and remain event-free throughout the trajectory period to be classified in a high-adherence trajectory, inducing spurious protective associations for high-adherence trajectories compared to trajectories showing lower adherence. Reverse causality occurs when outcomes or competing events during the trajectory period influence subsequent adherence, shaping trajectory membership rather than resulting from it. Allowing time-zero to begin during the trajectory period further exacerbates these biases. To mitigate time-related bias when trajectories serve as the exposure, we recommend:(1)temporally separating trajectory and outcome periods (landmark analyses), and (2)excluding competing events occurring during trajectory period. However, this approach may reduce generalizability and introduce selection bias, particularly when events are prevalent or populations are at high risk of death or disenrollment. Alternatively, allowing time-zero to begin during the trajectory period (i.e., overlapping trajectory and outcome periods) and censoring individuals at event occurrence reduces selection bias and increases generalizability, but exacerbates time-related biases. When trajectories are used as the outcome, the primary concern was selection (survival) bias related to continuous enrollment requirements during the trajectory period; time-related biases were less relevant.
CONCLUSIONS: There is no single optimal design; the choice must balance selection and time-related bias based on trajectory use (exposure vs outcome), competing event prevalence, population risk, and study objectives. Sensitivity analyses are essential to evaluate robustness across approaches.