Primary Submission Category: Mediation
Causal Mediation Analysis with Hidden Mediators
Authors: Laura Montelisciani, Eric Tchetgen Tchetgen,
Presenting Author: Laura Montelisciani*
Introduction: In causal mediation analysis, assumptions regarding the presence of a measurable mediator without measurement error are often impractical with observational data. A recent proximal causal inference framework enables the estimation of natural direct (NDE) and indirect (NIE) effects of the treatment on the outcome, even with measurement error or a hidden mediator.
Estimation: Two mediator proxies, Z and W, meeting specific conditional independence conditions, are required. Z and W must be directly caused by the mediator and associated with the outcome only via the mediator. For a continuous outcome, two linear models are fitted to estimate the NDE. The first determines the conditional expectation of W on treatment and Z, and the second, derives an unbiased estimator for the NDE from the regression of the outcome on the plugged in predicted conditional expectation of W and the treatment.
The estimate of NIE is then obtained by subtracting the estimate of NDE from the Average Treatment Effect, where the latter is estimated using the observed data. Estimation of NDE and NIE is possible even in the presence of interaction between the treatment and the hidden mediator.
Discussion: This estimation approach offers the benefit of accurately estimating both NDE and NIE in situations involving a hidden/mismeasured mediator, utilizing straightforward linear models and avoiding assumptions on measurement error.