Primary Submission Category: Machine Learning and Causal Inference
Anytime-Valid Inference for Double/Debiased Machine Learning of Causal Parameters
Authors: Abhinandan Dalal, Patrick Blöbaum, Shiva Kasiviswanathan, Aaditya Ramdas,
Presenting Author: Abhinandan Dalal*
Double (debiased) machine learning (DML) is widely used for learning causal/structural parameters due to its flexibility with high-dimensional nuisance functions and its ability to avoid bias from regularization or overfitting. However, the classic double-debiased framework is valid only for a fixed sample size, limiting its ability to either collect more data for sharper inference or stop early when stable estimates arise. This poses concerns in large-scale experiments with high costs or life-or-death decisions, and in observational studies where confidence intervals may fail to shrink even with more data because of partial identifiability.
We propose time-uniform counterparts to asymptotic DML, allowing valid inference and confidence intervals at any (possibly data-dependent) stopping time. Our conditions are only slightly stronger than standard DML requirements but guarantee anytime-valid inference. These results let any existing DML procedure become anytime-valid with minimal changes, making it highly adaptable. We demonstrate this with two examples: (a) local average treatment effect in online experiments with non-compliance, and (b) partial identification of average treatment effect in observational studies with unmeasured confounding.