A Randomization-Based Method for Evaluating Time-Varying Treatment Effects

Abstract

Tests for paired censored outcomes have been extensively studied, with some justified in the context of randomization-based inference. These tests are primarily designed to detect an overall treatment effect across the entire follow-up period, providing limited insight into when the effect manifests and how it changes over time. In this article, we introduce new randomization-based tests for paired censored outcomes that enable both time-specific and long-term analysis of a treatment effect. The tests utilize time-specific scores, quantifying each individual's impact on sample survival at a fixed time, obtained via pseudo-observations. Moreover, we develop corresponding sensitivity analysis methods to address potential unmeasured confounding in observational studies where randomization often lacks support. To illustrate how our methods can provide a fuller analysis of a time-varying treatment effect, we apply them to a matched cohort study using data from the Korean Longitudinal Study of Aging (KLoSA), focusing on the effect of social engagement on survival.