Abstract T P147: Estimating the Effect of Antidepressant Use on Stroke Risk in the Presence of Competing Risks: A Simulation Study

Abstract

Introduction: Research on determinants of stroke must account for competing risks (events, such as death, that preclude subsequent stroke). Fine and Gray competing risk regression models are increasingly used in place of traditional Cox proportional hazards models to account for competing risks, but may not always provide clinically relevant estimates. Hypothesis: Even when antidepressant use has no effect on stroke risk, both Fine and Gray and Cox models can yield spurious associations if antidepressants affect mortality. Methods: We simulated data (50 simulations) on 3000 individuals followed for up to ten years with biennial assessments of stroke status. We considered two scenarios. In both scenarios, we set antidepressant use to have no effect on stroke risk. In Scenario 1, we set antidepressant use to increase mortality (HR (hazard ratio)=2.00). In Scenario 2, we set antidepressant use to decrease mortality (HR=0.50). Results: In Scenario 1, the average stroke incidence rate was 16.5/1,000 person years and the average mortality rate was 33.9/1,000 person years. Even though antidepressants had no effect on stroke in the simulations, both Fine and Gray (average HR=0.69) and Cox (average HR=0.83) models estimated protective effects of antidepressants on stroke. Fine and Gray models incorrectly rejected the null hypothesis in 22% of simulations and Cox models incorrectly rejected the null hypothesis in 10% of simulations. In Scenario 2, the average incidence rate was 16.9/1,000 person years and the average mortality rate was 31.3/1,000 person years. Even though antidepressants had no effect on stroke in the simulations, both Fine and Gray (average HR=1.22) and Cox (average HR=1.09) models estimated protective effects of antidepressants on stroke. Fine and Gray models incorrectly rejected the null hypothesis in 20% of simulations and Cox models incorrectly rejected the null hypothesis in 8% of simulations. Conclusions: For exposures that affect mortality, both Cox models and Fine and Gray corrections can yield misleading results regarding effects on stroke risk. This is important for interpreting findings on controversial topics, such as the effect of antidepressants on stroke.