Causal DAGs

The article concludes that DAG constraints and exact inference solve different failure modes. The causal estimation assumptions require sequential ignorability given LSTM-encoded history and the expert DAG, plus correct DAG specification. Sensitivity analyses found that removing a unique pathway increased bias, while removing a redundant pathway increased cross-seed instability. DAG constraints make interventions propagate only through causal descendants. DAG constraints separate confounding and prevent anti-causal intervention propagation. Exact invertible inference preserves patient-specific exogenous noise and enables structural propagation through mediation paths. The fixed expert-specified DAG can cause systematic bias if edges, edge directions, or latent confounding are misspecified.