Disentangled Representation for Causal Mediation Analysis

Estimating direct and indirect causal effects from observational data is crucial to understanding the mechanisms predicting behaviour under different interventions. Causal mediation analysis a method that often used reveal effects. Deep learning shows promise in analysis, but current methods only assume latent confounders affect treatment, mediator outcome simultaneously, fail identify types of (e.g., or outcome). Furthermore, are based on sequential ignorability assumption, which not feasible for dealing with multiple confounders. This work aims circumvent assumption applies piecemeal deconfounding as an alternative. We propose Disentangled Mediation Analysis Variational AutoEncoder (DMAVAE), disentangles representations into three accurately estimate natural effect, effect total effect. Experimental results show proposed outperforms existing has strong generalisation ability. further apply real-world dataset its potential application.