Estimating Heterogenous Treatment Effects for Survival Data with Doubly Doubly Robust Estimator

TL;DR

This paper proposes a novel doubly doubly robust estimator for accurately estimating both average and heterogeneous treatment effects in complex survival data with truncation and censoring, leveraging deep learning techniques.

Contribution

It introduces a new estimator that addresses missing data issues in survival analysis, combining doubly robust properties with deep learning for improved causal effect estimation.

Findings

The proposed method outperforms traditional hazard ratio and Cox-based estimators in simulations.

It effectively handles missing counterfactuals and data truncation in survival analysis.

Demonstrates advantages in estimating heterogeneous treatment effects in complex survival data.

Abstract

In this paper, we introduce a doubly doubly robust estimator for the average and heterogeneous treatment effect for left-truncated-right-censored (LTRC) survival data. In causal inference for survival functions in LTRC survival data, two missing data issues are noteworthy: one is the missing data of counterfactuals for causal inference, and the other is the missing data due to truncation and censoring. Based on previous research on non-parametric deep learning estimation in survival analysis, this paper proposes an algorithm to obtain an efficient estimate of the average and heterogeneous causal effect. We simulate the data and compare our methods with the marginal hazard ratio estimation, the naive plug-in estimation, and the doubly robust causal with Cox Proportional Hazard estimation and illustrate the advantages and disadvantages of the model application.