CausalFlow-T

CausalFlow-T was the only causal estimator that satisfied all five reliability criteria across synthetic benchmarks. CausalFlow-T had a mean rank of 1.83, outperforming NF without DAG, GNN-CVAE, CVAE, and TARNet. CausalFlow-T maximizes exact likelihood with the change-of-variables formula instead of using an ELBO approximation. CausalFlow-T uses a causal masked autoregressive normalizing flow conditioned on an LSTM-encoded patient history. CausalFlow-T computes counterfactuals through exact abduction-action-prediction by inverting observed variables, intervening on treatment, and decoding descendants in DAG order. CausalFlow-T constrains autoregressive factorization to a topological ordering of the expert DAG. CausalFlow-T uses dequantization for binary outcomes, which creates calibration costs on survival endpoints. CausalFlow-T estimates causal counterfactuals from completed longitudinal data.