|
Bayesian causal inference offers a principled approach to policy evaluation of proposed interventions on mediators. This approach frames scientific questions in terms of counterfactual contrasts, clarifying where cross-world counterfactual independence assumptions are needed. Parametric models are used for the outcome of interest and mediators being intervened upon.
Estimation is based on posterior predictive draws under two well-defined treatment regimes: one following the existing data generation process and one following the process under intervention. We can formulate contrasts such as the racial disparity in cancer survival remaining after an intervention that eliminates disparities in cancer stage.
Information from external data sources can be incorporated via informative priors. Prior information can be used for parameters not identified by the data, such as those related to unmeasured baseline or exposure-induced confounders. When priors arise from external data, this method is a coherent and transparent sensitivity analysis with a readymade mechanism for communicating uncertainty.
The method extends to models with multiple mediators or mediators measured longitudinally.
|
