Batch-Adaptive Causal Annotations

A new paper from researchers Ezinne Nwankwo, Lauri Goldkind, and Angela Zhou presents Batch-Adaptive Causal Annotations, a method that tackles a critical bottleneck in causal inference: the high cost of labeling outcome data. In policy and decision-making, ground-truth outcomes are often missing or error-prone, requiring expensive annotation or follow-up. The team derived a closed-form solution for optimal batch sampling probability, minimizing the asymptotic variance of a doubly robust estimator (AIPW) for average treatment effect estimation with missing outcomes. This approach intelligently selects which data points to label, maximizing efficiency under budget constraints.

Tested on simulated and real-world datasets, including outreach interventions in homelessness services, the method achieves substantially lower mean-squared error than existing baselines. In a key result, it matches confidence intervals obtained with 361 random samples using only 90 optimized samples—a 75% savings in labeling budget. The framework extends to costly annotations of unstructured data like text or images, making it applicable in healthcare and social services. This work bridges machine learning and causal inference, offering a practical tool for resource-constrained settings.