Longitudinal causal effect estimation is challenged by time‑varying confounding and treatment‑confounder feedback loops. Existing approaches typically require fitting dataset‑specific nuisance models and are often limited to pre‑specified treatment sequences or prediction horizons, incurring heavy computation at deployment. We introduce PICLE‑PFN (Policy In‑Context Longitudinal causal Effect estimation with Prior‑data Fitted Networks), the first foundation model that supports dynamic treatment policies. PICLE‑PFN encodes target policies as in‑context demonstrations of past decisions, providing a unified interface that covers a broad class of policies. The model is pretrained on synthetic data generated from longitudinal structural causal models, amortizing causal effect estimation across datasets, policies, and horizons. When presented with a new observational dataset and a target policy, it predicts conditional average potential outcomes (CAPOs) in a single forward pass, without any task‑specific training or hyper‑parameter tuning. Experiments on synthetic and semi‑synthetic longitudinal benchmarks show that PICLE‑PFN attains RMSE comparable to state‑of‑the‑art dedicated estimators, markedly improves uncertainty calibration, and reduces deployment runtime to a fraction of the baseline.
Review