PURPOSE: Pharmacovigilance monitors the safety of drugs after their approval and marketing. Timely detection of adverse effects is important. The true relationship between time-varying drug use and the adverse event risk is typically unknown. Yet, most current pharmacovigilance studies rely on arbitrarily chosen exposure metrics such as current exposure or use in the past 3 months. The authors used simulations to assess the impact of a misspecified exposure model on the timeliness of adverse effect detection. METHODS: Prospective pharmacovigilance studies were simulated assuming different true relationships between time-varying drug use and the adverse event hazard. Simulated data were analyzed by fitting conventional parametric and more complex spline-based estimation models at multiple, pre-specified testing times. The 'signal' was generated on the basis of the corrected model-specific p-value selected to ensure a 5% probability of incorrectly rejecting the null hypothesis of no association. RESULTS: Results indicated that use of an estimation model that diverged substantially from the true underlying association-reduced sensitivity and increased the time to detection of a clinically important association. CONCLUSIONS: Time to signal detection in pharmacovigilance may depend strongly on the method chosen to model the exposure. No single estimation model performed optimally across different simulated scenarios, suggesting the need for data-dependent criteria to select the model most appropriate for a given study.
PURPOSE: Pharmacovigilance monitors the safety of drugs after their approval and marketing. Timely detection of adverse effects is important. The true relationship between time-varying drug use and the adverse event risk is typically unknown. Yet, most current pharmacovigilance studies rely on arbitrarily chosen exposure metrics such as current exposure or use in the past 3 months. The authors used simulations to assess the impact of a misspecified exposure model on the timeliness of adverse effect detection. METHODS: Prospective pharmacovigilance studies were simulated assuming different true relationships between time-varying drug use and the adverse event hazard. Simulated data were analyzed by fitting conventional parametric and more complex spline-based estimation models at multiple, pre-specified testing times. The 'signal' was generated on the basis of the corrected model-specific p-value selected to ensure a 5% probability of incorrectly rejecting the null hypothesis of no association. RESULTS: Results indicated that use of an estimation model that diverged substantially from the true underlying association-reduced sensitivity and increased the time to detection of a clinically important association. CONCLUSIONS: Time to signal detection in pharmacovigilance may depend strongly on the method chosen to model the exposure. No single estimation model performed optimally across different simulated scenarios, suggesting the need for data-dependent criteria to select the model most appropriate for a given study.