Primary Submission Category: Weighting
Off-policy evaluation using debiased calibration weighting
Authors: Jae-kwang Kim, Yumou Qiu, Yonghyun Kwon,
Presenting Author: Yuyang Li*
Calibration weighting is an important tool for improving efficiency of the design-based estimators in survey sampling. We propose a unified framework for debiased calibration in reducing selection bias and improving the efficiency of the estimator in the context of causal inference. Our approach is based on the generalized entropy as the objective function for optimization. The constraint incorporating the design weights is used to correct the selection bias while the benchmarking constraints are used to reduce the variance. The resulting calibration estimator is asymptotically equivalent to the design-optimal regression estimator of Deville and Sarndal (1992). We also propose the cross entropy as the optimal entropy function for calibration. To test our theory, we perform a limited simulation study.