Generalized Bayes for Causal Inference

A team of researchers including Emil Javurek, Dennis Frauen, Yuxin Wang, and Stefan Feuerriegel has introduced a novel 'Generalized Bayes for Causal Inference' framework, addressing a core challenge in causal machine learning: principled uncertainty quantification. Standard Bayesian methods require specifying a full probabilistic model for the data-generating process, including complex nuisance parameters like propensity scores, making them vulnerable to modeling errors and difficult prior specification. This new framework bypasses explicit likelihood modeling entirely. Instead, it places priors directly on the causal estimands of interest—such as the Average Treatment Effect (ATE) or Conditional ATE (CATE)—and updates beliefs using a loss function derived from causal identification assumptions. This creates generalized posteriors that quantify uncertainty for causal effects.

The technical innovation lies in its flexibility and robustness. The framework can be layered on top of existing, high-performance causal ML estimators, including modern Neyman-orthogonal meta-learners, turning them into tools that provide full Bayesian uncertainty bands. The authors prove that for Neyman-orthogonal losses, the resulting generalized posteriors converge to their oracle counterparts and remain robust to estimation errors in the first-stage nuisance components (e.g., the propensity score model). With calibration, this provides valid frequentist uncertainty intervals even when these nuisance estimators converge at slower, non-parametric rates. Empirically, the method demonstrates calibrated uncertainty across various settings. This represents the first flexible framework for constructing generalized Bayesian posteriors specifically for causal machine learning, potentially making causal conclusions from complex models more trustworthy and actionable for decision-making in fields like medicine and policy.