Multiply robust causal inference of the restricted mean survival time difference

Area of Science:

• Biostatistics
• Epidemiology
• Survival Analysis

Background:

• Hazard ratio (HR) is common for survival analysis but relies on proportional hazards.
• Restricted Mean Survival Time (RMST) difference is a robust alternative, especially when HR assumptions are violated.
• Current RMST estimation using inverse probability of treatment weighting (IPTW) can be biased if propensity score models are misspecified.

Purpose of the Study:

• To develop and evaluate a multiply robust estimator for the causal difference in RMST.
• To address limitations of existing IPTW methods for RMST estimation.
• To provide a more reliable method for comparative effectiveness research in survival analysis.

Main Methods:

• Proposed an empirical likelihood-based weighting approach for RMST estimation.
• Developed a multiply robust estimator consistent if the correct propensity score model is within a specified set.
• Evaluated the proposed method through simulations and applied it to real-world data.

Main Results:

• Simulation results demonstrated the robustness and improved efficiency of the proposed estimator compared to IPTW.
• The empirical likelihood approach showed less bias and higher efficiency in finite samples than IPTW from a correctly specified model.
• Direct application of machine learning for propensity scores led to biased results in simulations.

Conclusions:

• The proposed multiply robust empirical likelihood method provides a more reliable estimation of causal differences in RMST.
• This method offers advantages over traditional IPTW, particularly when propensity score models may be misspecified.
• The approach is applicable to real-world comparative effectiveness studies, such as evaluating antibiotic prophylaxis effects on childhood allergies.