A machine-learning based causal mediation framework without the no-omitted-confounder assumption for latent variables and intensive longitudinal data

• Applicant: Professor Dr. Holger Brandt
• Subject Area: Personality Psychology, Clinical and Medical Psychology, Methodology
• Term: since 2023
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 529799871

In this proposal, I develop a new framework for the identification of mediator variables in treatment studies that does not rely on the assumption of no-omitted-confounders or sequential ignorability. Instead, the new methods allow the existence of unobserved confounders and still provide unbiased and consistent estimates under a wide range of conditions. I provide three important extensions of an existing basic model, the so-called rank preserving model: First, I will develop a (Bayesian) latent variable modeling approach that has the same flexibility as structural equation models (SEM). While SEM implicitly and automatically assume sequential ignorability if applied to mediation analyses, the new model does not need this implausible assumption. Second, I will extend the model for intensive longitudinal data via a dynamic approach that allows researchers to assess person- and time-related heterogeneity of causal mediation effects via random effects. Third, I will use a machine learning based method, the causal effect variational autoencoder, to extend the RPM for both high-dimensional problems with many covariates and non-randomized treatments if at least noisy proxy variables exist that can be used to account for potential confounding. In a final milestone, I will provide an over-arching comparison of different state-of-the-art machine learning methods for causal mediation models that were developed in the past few years. Besides the statistical developments, model performance will be investigated with simulation studies and illustrated with empirical data under each milestone. The output of the proposal is a flexible, widely usable approach to identify mediator variables in a causal framework even if relevant confounders are omitted from the analysis.

• DFG Programme: Research Grants