Abstract
Mediation analysis is a powerful technique to assess how exposure affects the outcome of interest
mediated through an intermediated variable (mediator). By incorporating with the counterfactual
model, causal mediation analysis can further yield substantial insight into the causal mechanism
through the assessment of natural direct and indirect effects.
In this talk, I will show two critical problems for causal mediation analysis and propose solutions.
The first problem arises from the assumptions regarding no unmeasured mediator–outcome
confounding and no intermediate mediator–outcome confounding. The conventional methodology
of causal mediation analysis is invalid if these assumptions cannot be satisfied. However, checking
these assumptions presents practical challenges. To address this problem, a novel instrumental
blocker, a novel quasi-instrumental variable, is introduced to relax both of these assumptions. A
multiply robust estimation method is derived to mitigate the model misspecification problem. As
an illustration, we apply the proposed method to genomic datasets of lung cancer to investigate the
potential role of the epidermal growth factor receptor in the treatment of lung cancer. The second
problem discussed in this talk is about death truncation. In longitudinal studies, the problem of
truncation by death arises when individuals die between follow-up visits. Some variables may not
be well-defined for dead individuals, and in such a condition, the conventional mediation method
cannot be applicable. A novel approach is proposed to redefining natural direct and indirect effects,
which are generalized forms of conventional causal effects for survival outcomes.
Furthermore, three statistical methods are developed for the binary outcome of survival status and
formulated a Cox model for survival time. The proposed method is applied to explore the effect of
hepatitis C virus infection on mortality, as mediated through hepatitis B viral load.