Primary Submission Category: Heterogeneous Treatment Effects
Debiasing Treatment Effect Estimation for Privacy-Protected Data: A Model Audition and Calibration Approach
Authors: Ta-Wei Huang, Eva Ascarza,
Presenting Author: Ta-Wei Huang*
The growing concern for data privacy and recent regulatory changes have led organizations to implement privacy-preserving measures to protect sensitive customer information. However, there is a concern about whether these measures may hinder their ability to personalize their interventions. In this research, we examine the impact of two commonly used privacy protection methods on heterogeneous treatment effects (HTE) estimation: adding substantial noise to the data in a differentially-private way and excluding protected customer characteristics (such as gender or race) from tracking. We find that these mechanisms significantly impact the prediction accuracy of current HTE estimation methods, resulting in suboptimal targeting policies.
To overcome this problem, we propose a generic post-processing approach that combines recent advances in multi-group fairness and HTE estimation. This bias correction mechanism divides the experimental data into three folds: the first is to construct an initial HTE model, the second is to identify subgroups with large prediction errors and calibrate the model by a boosting procedure using the information on those groups, and the third is to stop the calibration procedure. Using a set of simulation analyses and real-world applications, we show that the proposed method significantly improves the accuracy of HTE estimation and provides more effective targeting policies when the data is collected under the above privacy-preserving measures.