Aligning Incentives to Balance Covariates in Experiments with Selection Bias

Jiachun Li, Yang Meng, David Simchi-Levi, Chonghuan Wang

Keywords: Causal Inference, Game Theory, Machine Learning, Decision Making, Incentives

Abstract: Estimating the average treatment effect (ATE) when participants can self-select into treatment or control groups based on their preferences can lead to significant selection bias and large variance of the estimation. We propose an incentivization framework that realigns participant preferences to balance covariates, thereby reducing bias and variance in treatment effect estimation. Our approach leverages incentive mechanisms solved under budget constraints to redistribute participants towards underrepresented groups. We provide theoretical guarantees for variance reduction using the Augmented Inverse Probability Weighting (AIPW) estimator and analyze the impact of unobserved confounders, showing that aligning incentives mitigates bias in treatment effect estimation. To achieve these goals, we introduce a low-switching learning-to-incentivize algorithm that dynamically adjusts incentives while adhering to resource constraints, achieving consistent and asymptotically efficient ATE estimation.

Submission Number: 18