CRA-PFN: Treatment Effect Prediction with Mixed-Role Covariates via Prior-Data Fitted Networks

Conditional average treatment effect (CATE) estimation from observational data usually presumes that valid adjustment covariates are known. In practice, however, treatment and outcome are designated while the remaining covariates may mix pre-treatment causes with prognostic covariates, mediators, descendants, and colliders, potentially hindering CATE identification and estimation by existing methods. To address this, we introduce the Causal Role-Agnostic Prior-data Fitted Network (CRA-PFN), which is pretrained under a role-heterogeneous prior over structural causal models (SCMs) and predicts task-specific CATEs without causal-role labels or a designated adjustment set being supplied to the estimator. We formalize role-agnostic CATE identifiability and derive conditions for mean-square consistency of CATE prediction as the mixed-role observational data grow. We further characterize the population behavior of the training objective relative to the prior-induced Bayes estimator, and derive a CATE error decomposition that separates estimator error from residual uncertainty. Extensive experiments across diverse datasets and out-of-distribution SCMs demonstrate that CRA-PFN achieves strong performance and generalizes effectively across mechanism shifts.