Average treatment effects tell us how an intervention shifts outcomes on average, but clinicians and policymakers often need to know what would have happened to a specific person under a different treatment. The new PO-Flow framework models treatment-specific outcome distributions as invertible transport maps, allowing it to generate potential outcomes, estimate conditional average treatment effects, and make factual-conditioned counterfactual predictions.
PO-Flow is built on continuous normalizing flows. To handle the treatment imbalance common in observational data, the authors add cross-fitted inverse-probability weighting to the flow-matching loss. They also prove sufficient conditions for recovering individual counterfactuals: under shared exogenous noise and a common encoded reference distribution, scalar recovery follows from increasing noise-to-latent maps, while multivariate outcomes require triangular or convex-gradient structure.
In experiments across four scalar benchmarks and seven multivariate or image settings with dimensions up to 784, PO-Flow achieved the lowest PEHE and potential-outcome RMSE on all scalar benchmarks and attained or tied the best mean in 26 of 35 multivariate comparisons. Controlled ablations showed that inverse-probability weighting improved all reported scalar counterfactual metrics. A conference version of the paper appears at ISIT 2026.