When Reconstruction Is Not Enough: Doubly Robust Counterfactual and Interventional Mediation Analysis with Learned Multimodal Representations
Not provided in the abstract
Abstract
The paper discusses the limitations of learned representations in high-dimensional multimodal health records for causal identification and proposes a new workflow for improving counterfactual and interventional mediation analysis.
Reality Card
The authors demonstrate that traditional covariate-only training objectives fail to ensure sufficient representation for causal identification, leading to significant biases in estimations.
In simulations, discarding a rare confounder resulted in indirect-effect intervals with only 2% coverage at a sample size of n = 3,000.
The reliance on reconstruction loss for diagnostics can miss critical issues, indicating a need for more robust predictive diagnostics.
Paper to code
Verified implementation resources so builders can test the paper’s claims instead of stopping at the abstract.