Many causal panel estimators construct counterfactuals using weights or latent structure estimated directly from observed outcomes, making their estimates sensitive to observation noise. Bayesian hierarchical models can account for this noise and pool information across related observations. Existing Bayesian approaches to causal panel analysis, however, typically specify counterfactual imputation within the probabilistic model, so changing estimators may require a new formulation. We develop a framework that learns the latent untreated panel with a probabilistic model and applies established causal estimators to its posterior draws. The model accommodates non-Gaussian observations and structured pooling, while re-estimation across draws propagates uncertainty through counterfactual imputation. We establish conditions for posterior consistency and characterize how pooling affects imputation risk. Under correct posterior specification, we show that the excess Bayes risk of our approach is bounded by the causal estimator's expected squared imputation error when given the true latent outcomes in observed untreated cells. We also develop a fast cross-validation approach for model selection and calibration of predictive uncertainty for treatment effects. We use simulations to examine when pooling helps or harms imputation and cross-validation to assess whether probabilistic modeling improves accuracy in multiple empirical applications.
翻译:暂无翻译