Program Evaluation and Causal Inference With High-Dimensional Data

• DOI: 10.3982/ecta12723
• Publication Date: 2017-01-11T11:39:17Z
• Authors: A. Belloni; V. Chernozhukov; I. Fern�ndez-Val; C. Hansen
• Keywords: FOS: Computer and information sciences; Endogeneity; Propensity score; 330; Economics; Econometrics (econ.EM); Business & economics; Mathematics - Statistics Theory; Machine Learning (stat.ML); Statistics Theory (math.ST); Heterogenous treatment effects; 01 natural sciences; Social sciences; Methodology (stat.ME); FOS: Economics and business; interdisciplinary applications; Economic theory; Local average and quantile treatment effects; Statistics - Machine Learning; Machine learning; FOS: Mathematics; Local effects of treatment on the treated; Econometrics; 0101 mathematics; Inference after model selection; Statistics - Methodology; Economics - Econometrics; Marketing; Moment-condition models; ddc:330; mathematical methods; Randomized control trials; Physical sciences; Mathematical methods in social sciences; Causality; Statistics theory; Lasso; Science & technology; Instruments; Neyman orthogonality; Mathematics; Statistics & probability

In this paper, we provide efficient estimators and honest confidence bands for a variety of treatment effects including local average (LATE) quantile (LQTE) in data-rich environments.We can handle very many control variables, endogenous receipt treatment, heterogeneous effects, function-valued outcomes.Our framework covers the special case exogenous either conditional on controls or unconditionally as randomized trials.In latter case, our approach produces (functional) (ATE) (QTE).To make informative inference possible, assume that key reduced form predictive relationships are approximately sparse.This assumption allows use regularization selection methods to estimate those relations, post-regularization post-selection uniformly valid (honest) across wide-range models.We show ingredient enabling is orthogonal doubly robust moment conditions estimating certain functional parameters.We illustrate proposed with an application effect 401(k) eligibility participation accumulated assets.The results program evaluation obtained consequence more general condition framework, which arises from structural equation models econometrics.Here too crucial conditions, be constructed initial conditions.We (function-valued) parameters within where any high-quality, modern machine learning (e.g., boosted trees, deep neural networks, random forests, their aggregated hybrid versions) used learn nonparametric/high-dimensional components model.These include number supporting auxilliary major independent interest: namely, (1) prove uniform validity multiplier bootstrap, (2) offer delta method, (3) sparsity-based estimation regression functions outcomes.