Using survival information in truncation by death problems without the monotonicity assumption

Area of Science:

• Biostatistics
• Clinical Trials Methodology
• Causal Inference

Background:

• Randomized clinical trials (RCTs) can suffer from 'truncation by death,' where patient mortality before outcome measurement complicates causal effect estimation.
• Survivors may differ in prognostic variables, biasing comparisons between treatment and control groups.
• The well-defined causal effect is limited to the subgroup of patients who would always survive, but this subgroup is not directly observable.

Purpose of the Study:

• To address the challenge of estimating causal effects in the presence of truncation by death in RCTs.
• To develop methods for identifying and estimating causal effects within the subgroup of patients who would always survive.
• To improve the precision of causal effect bounds without relying on strong parametric assumptions.

Main Methods:

• Utilizing detailed survival information, both before and after the primary outcome measurement time point, to inform causal inference.
• Developing methods to sharpen the bounds of the subgroup causal effect by leveraging comprehensive survival data.
• Employing a copula model to relax the often-unrealistic monotonicity assumption regarding treatment's effect on survival time.

Main Results:

• The proposed methods enhance the precision of causal effect bounds for the always-surviving subgroup.
• Leveraging detailed survival data improves statistical inference without imposing strong parametric constraints.
• The copula model offers a more flexible approach to modeling treatment effects on survival compared to the monotonicity assumption.

Conclusions:

• Detailed survival information is a valuable resource for improving causal effect estimation in RCTs with potential truncation by death.
• The proposed approach provides more useful bounds for causal effects in challenging clinical trial settings.
• Relaxing the monotonicity assumption using copula models leads to more robust causal inference.