A connection between survival multistate models and causal inference for external treatment interruptions

Area of Science:

• Biostatistics
• Clinical Trial Methodology
• Causal Inference

Background:

• Treatment interruptions in clinical trials, such as FDA-imposed holds, complicate survival outcome analysis.
• Multistate models offer a flexible framework for analyzing time-dependent covariates and treatment interruptions.

Purpose of the Study:

• To develop and validate novel methods (causal censoring and filtering) for estimating the true treatment effect on overall survival, unaffected by clinical trial holds.
• To establish a link between multistate modeling and causal inference for handling treatment interruptions.

Main Methods:

• Application of a multistate model framework to data from the START (Stimulating Targeted Antigenic Response To non-small-cell cancer) trial.
• Development and implementation of causal censoring and filtering techniques to adjust for external trial interruptions.
• Simulation studies to compare the performance of proposed methods against a naive approach.

Main Results:

• Causal censoring and filtering methods successfully estimate the initial treatment effect on overall survival in a hypothetical scenario without clinical holds.
• The proposed methods demonstrate favorable performance compared to a naive approach that ignores external impacts.
• The study provides theoretical justification for the causal interpretation derived from censoring and filtering in multistate models.

Conclusions:

• Causal censoring and filtering are robust methods for analyzing clinical trial data with treatment interruptions.
• These methods can be generalized for applications in other settings, including the COVID-19 pandemic.
• The findings support the use of these advanced statistical techniques for more accurate clinical trial outcome assessment.