PUATE: Semiparametric Efficient Average Treatment Effect Estimation from Treated (Positive) and Unlabeled Units
Advances in Neural Information Processing Systems (NeurIPS), 2025.
- causal inference
- PU learning
- missing data
- semiparametric efficiency
In one sentence. Efficiency bounds and efficient ATE estimators for the case where the data contain a treated group and an unlabeled group whose treatment status is unknown.
A common data problem
Textbook treatment effect estimation assumes each unit is known to be treated or untreated. In practice one often has a record of who was treated and a much larger pool of units whose treatment status was simply never recorded. The unlabeled pool is a mixture of treated and untreated units, in unknown proportion.
This is a variant of learning from positive and unlabeled data, and equivalently a special case of treatment effect estimation with missing data — but the missingness is in the treatment indicator rather than in the outcome, which changes the identification and efficiency analysis.
Results
- Derivation of the semiparametric efficiency bounds for this observation scheme, which characterize the lowest asymptotic variance achievable by any regular estimator.
- Construction of semiparametric efficient ATE estimators that attain those bounds.
The results connect two literatures that rarely talk to each other: causal inference with missing data, and weakly supervised learning.