Is it Correlation or Causation? Uncertainty Quantification in Estimating Causal Effects with Unobserved Confounding
Miruna Oprescu, Cornell University
Estimating the heterogeneous effects of interventions from observational data is crucial for data-driven decision-making across many domains. Recent advances have yielded robust machine learning methods for estimating these treatment effects, but these methods often overlook the risk of unobserved confounders — latent variables that can arbitrarily and unknowingly bias causal estimates.
When we allow for hidden confounding, we cannot obtain unbiased point estimates of the conditional average treatment effect (CATE); however, we can instead bound the true, unbiased CATE. In this talk, I introduce the B-Learner, a flexible meta-learner that efficiently learns precise bounds on the CATE function, accounting for hidden confounders. The B-Learner can leverage various function estimators, such as random forests and deep neural networks, to produce these bounds.
The B-Learner has desirable theoretical properties: validity (correct bounds with high probability), sharpness (tightest possible bounds), and efficiency (less data required). Experimental comparisons using synthetic and real-world data showcase its practical utility, consistently providing tighter and more accurate bounds on the CATE than existing approaches.
Join me to explore how the B-Learner transforms causal inference, offering reliable insights despite hidden confounding, and enhancing the robustness of data-driven decision-making processes. This work is based on Oprescu, Miruna, et al. "B-learner: Quasi-oracle bounds on heterogeneous causal effects under hidden confounding." International Conference on Machine Learning. PMLR, 2023.