Comparison of Time-to-First Event and Recurrent-Event Methods in Randomized Clinical Trials

Background: Most phase-3 trials feature time-to-first event end points for their primary and secondary analyses. In chronic diseases, where a clinical event can occur >1 time, recurrent-event methods have been proposed to more fully capture disease burden and have been assumed to improve statistical precision and power compared with conventional time-to-first methods. Methods: To better characterize factors that influence statistical properties of recurrent-event and time-to-first methods in the evaluation of randomized therapy, we repeatedly simulated trials with 1:1 randomization of 4000 patients to active versus control therapy, with true patient-level risk reduction of 20% (ie, relative risk=0.80). For patients who discontinued active therapy after a first event, we assumed their risk reverted subsequently to their original placebo-level risk. Through simulation, we varied the degree of between-patient heterogeneity of risk and the extent of treatment discontinuation. Findings were compared with those from actual randomized clinical trials. Results: As the degree of between-patient heterogeneity of risk increased, both time-to-first and recurrent-event methods lost statistical power to detect a true risk reduction and confidence intervals widened. The recurrent-event analyses continued to estimate the true relative risk (0.80) as heterogeneity increased, whereas the Cox model produced attenuated estimates. The power of recurrent-event methods declined as the rate of study drug discontinuation postevent increased. Recurrent-event methods provided greater power than time-to-first methods in scenarios where drug discontinuation was ≤30% after a first event, lesser power with drug discontinuation rates of ≥60%, and comparable power otherwise. We confirmed in several actual trials of chronic heart failure that treatment effect estimates were attenuated when estimated via the Cox model and that increased statistical power from recurrent-event methods was most pronounced in trials with lower treatment discontinuation rates. Conclusions: We find that the statistical power of both recurrent-events and time-to-first methods are reduced by increasing heterogeneity of patient risk, a parameter not included in conventional power and sample size formulas. Data from real clinical trials are consistent with simulation studies, confirming that the greatest statistical gains from use of recurrent-events methods occur in the presence of high patient heterogeneity and low rates of study drug discontinuation.