Mediation Analysis for Censored Survival Data under an Accelerated Failure Time Model

TL;DR

This paper investigates the differences between the 'difference' and 'product' methods for causal mediation analysis in censored survival data under AFT models, revealing conditions under which they produce biased or consistent estimates.

Contribution

It formally demonstrates that the two methods may diverge in estimates under censoring and model misspecification, providing guidance on their valid application in survival analysis.

Findings

The product method remains valid under independent censoring.

The difference method can be biased with non-collapsible error distributions.

Both methods are consistent in normal mediator-normal outcome cases.

Abstract

Recent advances in causal mediation analysis have formalized conditions for estimating direct and indirect effects in various contexts. These approaches have been extended to a number of models for survival outcomes including accelerated failure time (AFT) models which are widely used in a broad range of health applications given their intuitive interpretation. In this setting, it has been suggested that under standard assumptions, the "difference" and "product" methods produce equivalent estimates of the indirect effect of exposure on the survival outcome. We formally show that these two methods may produce substantially different estimates in the presence of censoring or truncation, due to a form of model misspecification. Specifically, we establish that while the product method remains valid under standard assumptions in the presence of independent censoring, the difference method can be biased in the presence of such censoring whenever the error distribution of the AFT model fails to be collapsible upon marginalizing over the mediator. This will invariably be the case for most choices of mediator and outcome error distributions. A notable exception arises in case of normal mediator-normal outcome where we show consistency of both difference and product estimators in the presence of independent censoring. These results are confirmed in simulation studies and two data applications.