A1809
Title: Inference in high-dimensional linear mediation models under proportional asymptotics
Authors: Rajarshi Mukherjee - Harvard T.H. Chan School of Public Health (United States) [presenting]
Abstract: Mediation analysis in a linear structural equation model is considered under a proportional asymptotic regime where both the number of mediators and confounders are allowed to diverge proportionally to the sample size. In this regime, consistent and asymptotically normal estimators of direct, indirect, and suitable path-specific effects are provided through a carefully devised scheme without imposing any sparsity structure on the problem. Typical assumptions used in debiasing methodology under proportional asymptotic regimes do not apply to the mediation analysis problem, and a partial ridge-based debiasing scheme is employed to address the relevant subtleties. The effect of sample splitting and cross-fitting is additionally considered, and detailed numerical experiments are performed to validate the results in finite samples.
