Causal Inference and Off-Policy Evaluation

• causal inference
• off-policy evaluation
• external validity
• synthetic control

The question

Identification is only the first obstacle. The harder practical problems come from the gap between the data one has and the population one cares about: treatment status is unrecorded for most units, the logged population differs from the deployment population, or there is exactly one treated unit and a pool of imperfect comparisons.

Each case has a semiparametric structure worth working out precisely — what is identified, what the efficiency bound is, and which estimator attains it.

Three settings

• Unlabeled treatment status. PUATE (NeurIPS 2025) derives efficiency bounds and efficient ATE estimators when the data consist of a treated group and an unlabeled group of unknown composition — causal inference with missing data meeting weakly supervised learning.
• Distribution shift between study and target. Off-policy evaluation and learning under a covariate shift (NeurIPS 2020, spotlight) gives the efficiency bound and doubly robust efficient estimators when the evaluation population is not the logging population — the external validity problem, stated in bandit terms.
• One treated unit. Asymptotically unbiased synthetic control by density matching (Journal of Causal Inference) fixes the endogeneity bias in synthetic control weights and returns the full distribution of the treatment effect rather than a point estimate.

Structural estimation

Importance weighting for conditional moment restrictions (ICLR 2022, spotlight) turns a continuum of conditional restrictions into an unconditional risk minimization problem, which allows nonparametric instrumental variable models to be estimated with neural networks and other flexible learners.

Bias correction for these estimators is developed under debiased machine learning and Riesz regression; the design of the experiments that generate the data is developed under adaptive experimental design.