Propensity Score Weighting Analysis of Survival Outcomes Using Pseudo-observations

TL;DR

This paper introduces a novel propensity score weighting method using pseudo-observations for causal inference in survival analysis, providing valid inference and demonstrating efficiency and applicability to multiple treatments.

Contribution

It develops a new class of propensity score weighting estimators based on pseudo-observations, with explicit variance formulas and efficiency properties for survival outcomes.

Findings

The method provides valid, resampling-free inference for survival causal effects.

Overlap weights are shown to be optimally efficient among balancing weights.

Simulations demonstrate favorable performance compared to existing methods.

Abstract

Survival outcomes are common in comparative effectiveness studies and require unique handling because they are usually incompletely observed due to right-censoring. A ``once for all'' approach for causal inference with survival outcomes constructs pseudo-observations and allows standard methods such as propensity score weighting to proceed as if the outcomes are completely observed. For a general class of model-free causal estimands with survival outcomes on user-specified target populations, we develop corresponding propensity score weighting estimators based on the pseudo-observations and establish their asymptotic properties. In particular, utilizing the functional delta-method and the von Mises expansion, we derive a new closed-form variance of the weighting estimator that takes into account the uncertainty due to both pseudo-observation calculation and propensity score estimation. This allows valid and computationally efficient inference without resampling. We also prove the optimal efficiency property of the overlap weights within the class of balancing weights for survival outcomes. The proposed methods are applicable to both binary and multiple treatments. Extensive simulations are conducted to explore the operating characteristics of the proposed method versus other commonly used alternatives. We apply the proposed method to compare the causal effects of three popular treatment approaches for prostate cancer patients.