A1068
Title: General targeted machine learning for modern causal mediation analysis
Authors: Ivan Diaz - NYU Langone Health (United States) [presenting]
Abstract: The literature on the non-parametric definition and identification of mediational effects has grown significantly in recent years, with important progress in addressing challenges in interpreting and identifying such effects. However, statistical methodology for non-parametric estimation has lagged, with few or no methods available for tackling non-parametric estimation in continuous or high-dimensional mediators. It is shown that the identification formulas for six of the most widely known non-parametric approaches to mediation analysis proposed in recent years (natural direct and indirect effects, randomized interventional effects, separable effects, organic direct and indirect effects, recanting twin effects, and decision-theoretic effects) can be recovered from just two statistical estimands. An all-purpose, one-step estimation algorithm that can be coupled with machine learning in any mediation study that uses any of these definitions of mediation is proposed. The estimators rely on a re-parameterization of the identification formulas in terms of sequential regressions and on first-order non-parametric von Mises approximations of the first bias of a plug-in estimator to construct estimators with desirable properties, such as asymptotic normality. The one-step estimator requires the estimation of complex density ratios on the potentially high-dimensional mediators, a challenge solved using recent advancements in so-called Riesz learning.
