Causal EpiNets: Precision-corrected Bounds on Individual Treatment Effects using Epistemic Neural Networks

Researchers introduce Causal EpiNets, a neural network framework that improves estimation of individual treatment effects using Probability of Necessity and Sufficiency bounds. The method resolves critical limitations in finite-sample estimation by guaranteeing structural constraint satisfaction and correcting extremum bias, achieving better coverage and validity than standard plug-in estimators.

This research addresses a fundamental problem in causal inference: estimating personalized treatment effects from limited data. Traditional approaches to individual treatment effect estimation suffer from identifiability issues and systematic estimation failures in realistic sample sizes. The Probability of Necessity and Sufficiency framework provides intersection bounds that characterize individual-level causality by combining experimental and observational data, but previous implementations produced invalid results by violating probability constraints and generating spuriously tight intervals.

The Causal EpiNets approach introduces two key innovations. An anchored neural architecture enforces structural probability constraints by design, eliminating the constraint violations plaguing standard estimators. Precision-corrected intersection-bound inference then addresses extremum bias—a statistical pathology where max-min operations distort estimates—by leveraging Epistemic Neural Networks for uncertainty quantification at scale. This combination enables valid inference in high-dimensional settings where classical methods break down.

The practical implications extend across fields relying on personalized decision-making: precision medicine, targeted policy interventions, and adaptive treatments. Clinicians and policymakers currently lack reliable tools for understanding whether specific treatments work for specific individuals. This framework provides theoretically grounded bounds that maintain statistical validity even with finite data and complex covariates.

The empirical validation demonstrates that Causal EpiNets achieve nominal coverage rates and exact constraint satisfaction in high-dimensional regimes—conditions where standard plug-in estimators systematically fail to provide valid inference. The scalability provided by neural networks makes this approach viable for realistic applications with many confounding variables and large feature spaces.

• →Causal EpiNets solve structural constraint violations and extremum bias in individual treatment effect estimation from finite samples.
• →Anchored neural architecture guarantees probability constraint satisfaction by construction rather than through post-hoc corrections.
• →Epistemic Neural Networks enable scalable uncertainty quantification for precision-corrected intersection bound inference in high dimensions.
• →Method achieves nominal coverage and maintains exact constraint validity where standard estimators systematically fail.
• →Applicable to precision medicine, policy evaluation, and any domain requiring valid personalized treatment effect estimates.