【Ｍetrics seminar】Empirical Bayes with Optimal Shrinkage Trees
研討會日期 : 2022-12-13
時間 : 14:30
主講人 : Professor Yu-Chang Chen (陳由常)
地點 : Conference room B110
主持人 : Professor Le-Yu Chen
演講者簡介 : Professor Yu-Chang Chen (陳由常) received his Ph.D. in Economics from the UC San Diego in 2022. He is currently an Assistant Professor at the National Taiwan University. His research interests are Econometrics, Public Economics and Labor Economics.
演講摘要 : This paper studies the estimation of treatment effects in settings where there are many treatment arms while each treatment arm only has a moderate sample size. Examples in economics include neighborhood effects, hospital effects, and teacher value-added measures. We propose an Empirical Bayes (EB) method, which we refer to as “EB with Optimal Shrinkage Trees”, that overcomes the high-dimensionality by leveraging the auxiliary information contained in treatment characteristics. Our method starts with using a decision tree to group treatment arms with similar treatment characteristics. Then, we apply the Empirical Bayes methodology and shrink the individual effect estimate from each arm toward the average among the group it belongs to. We show that our procedure can reduce the total mean squared errors (MSE) of the treatment effect estimators compared to estimators with no shrinkage and conventional EB estimators that do not consider treatment characteristics. Also, we propose a consistent model selection procedure to approximate the optimal tree that minimizes the MSE. Simulations calibrated to real-world settings confirm that the method can greatly reduce the estimation errors, especially when the treatment characteristics are highly correlated to treatment effects.