A0598
Title: Double/Debiased machine learning with gradient boosting for treatment effect
Authors: Jui-Chung Yang - National Tsing Hua University (Taiwan) [presenting]
Hui-Ching Chuang - Yuan Ze University (Taiwan)
Chung-Ming Kuan - National Taiwan University (Taiwan)
Abstract: A new double/debiased machine learning framework (DML) has been recently proposed for the estimation and inference of low-dimensional parameters in the presence of high-dimensional nuisance parameter. In DML, the low-dimensional parameters of interest are estimated using the Neyman orthogonal moment and the cross fitting technique, while the nuisance parameters are estimated by some machine learning algorithms with sufficient rates of convergence, namely, $o_p(n^{-1/4})$. We show that a) regression trees and random forests in general do not have the $o_p(n^{-1/4})$ rates and may result in serious bias and size distortion, and b) when unknown nuisance functions are additive, the gradient boosting with stumps provide consistent estimation and correct inference for the treatment effect. We apply our methods to recent debate of the treatment effect of the Big $N$ auditors to the audit quality. Consistent to previous result, which use 3,000 designs of the propensity score matching, the method supports the existing of the Big $N$ effect during 1988 to 2006 in U.S. DML-GB also identifies the non-linear associations of the firm characteristics and the audit quality.
