Comparative effectiveness research with average hazard for censored time-to-event outcomes: simulation study and application to observational data

Abstract Background The average hazard is a summary measure of event time distributions with a given time window, $$[0,\tau ],$$ and allows intuitive interpretation as an average person-time incidence rate over the time window. This metric is calculated as the ratio of the cumulative incidence probability at $$\tau$$ to the restricted mean survival time at $$\tau$$ and can be estimated through non-parametric methods and thus robust. While previously proposed for randomized trials, its use in comparative effectiveness research remains underexplored. Methods We evaluate inference procedures for the difference and ratio of average hazards from two comparative groups, using six common confounding adjustment methods for survival functions, including direct standardization, inverse probability of treatment weighting (IPTW), propensity score matching, empirical likelihood, and augmented IPTW (AIPTW). Extensive simulation studies under varying model specification are conducted to assess bias, variance, coverage probability, and width of confidence interval. We apply the method to data from the preference cohort in the CANVAS study. Results All adjustment methods achieved satisfactory performance; AIPTW was notably robust under partial model misspecification. Conclusions Using difference in average hazards and ratio of average hazards as estimands, when combined with common confounding adjustment methods, is feasible and reliable for comparative effectiveness research. The average hazard-based analysis provides a practical alternative to the traditional hazard ratio approach for quantifying the magnitude of the intervention effect on survival outcomes.