Efficient Inference for Time-to-Event Outcomes by Integrating Right-Censored and Current Status Data

TL;DR

This paper introduces a semiparametric data fusion method that efficiently combines right-censored and current status data to improve survival probability estimation, outperforming existing approaches.

Contribution

It develops a novel doubly robust and efficient estimator for survival analysis that leverages both data types, achieving semiparametric efficiency and demonstrating practical gains.

Findings

The proposed estimator attains the semiparametric efficiency bound.

Incorporating current status data improves estimation efficiency.

Simulation results confirm the method's superior performance.

Abstract

We propose a semiparametric data fusion framework for efficient inference on survival probabilities by integrating right-censored and current status data. Existing data fusion methods focus largely on fusing right-censored data only, while standard meta-analysis approaches are inadequate for combining right-censored and current status data, as estimators based on current status data alone typically converge at slower rates and have non-normal limiting distributions. In this work, we consider a semiparametric model under exchangeable event time distribution across data sources. We derive the canonical gradient of the survival probability at a given time, and develop one-step estimators along with the corresponding inference procedure. Specifically, we propose a doubly robust estimator and an efficient estimator that attains the semiparametric efficiency bound under mild conditions. Importantly, we show that incorporating current status data can lead to meaningful efficiency gains despite the slower convergence rate of current status-only estimators. We demonstrate the performance of our proposed method in simulations and discuss extensions to settings with covariate shift. We believe that this work has the potential to open new directions in data fusion methodology, particularly for settings involving mixed censoring types.