Testes não paramétricos em análise de sobrevivência no contexto de inferência causal

Abstract

Research aimed at estimating causal effects of interventions on time-to-event outcomes in observational studies faces challenges, notably when the goal is to compare groups in the presence of confounding. Propensity score–based methodologies are well established in the literature and provide a robust framework for causal inference in observational studies. Weighting procedures have already been employed in nonparametric methods, such as the Kaplan-Meier estimator, allowing the construction of pseudo-populations using propensity scores, which yield adjusted, unbiased estimates in the presence of confounding. However, their integration with nonparametric hypothesis tests in survival analysis is still limited to the Log-Rank test and remains an area in need of methodological developments. In this dissertation, a generalization of nonparametric tests adjusted by the inverse probability of treatment weighting is developed, with a focus on comparing survival functions under different hazard patterns. The proposed methodology is assessed through simulation studies that consider alternative hypotheses involving proportional hazards, early and late separation, and crossing survival curves, as well as different censoring and sample size scenarios. The results show substantial gains in test power when the separation pattern is aligned with the weighting scheme used, suggesting that tests adjusted with appropriate weights can outperform classical approaches in the detection of causal effects.