Automatic Debiased Machine Learning of Causal and Structural Effects
Many causal and structural effects depend on regressions. Examples include average treatment effects, policy effects, average derivatives, regression decompositions, economic average equivalent variation, and parameters of economic structural models. The regressions may be high dimensional. Plugging...
| Main Authors: | , , |
|---|---|
| Format: | Report |
| Language: | unknown |
| Subjects: | |
| Online Access: | http://arxiv.org/pdf/1809.05224 |
| Summary: | Many causal and structural effects depend on regressions. Examples include average treatment effects, policy effects, average derivatives, regression decompositions, economic average equivalent variation, and parameters of economic structural models. The regressions may be high dimensional. Plugging machine learners into identifying equations can lead to poor inference due to bias and/or model selection. This paper gives automatic debiasing for estimating equations and valid asymptotic inference for the estimators of effects of interest. The debiasing is automatic in that its construction uses the identifying equations without the full form of the bias correction and is performed by machine learning. Novel results include convergence rates for Lasso and Dantzig learners of the bias correction, primitive conditions for asymptotic inference for important examples, and general conditions for GMM. A variety of regression learners and identifying equations are covered. Automatic debiased machine learning (Auto-DML) is applied to estimating the average treatment effect on the treated for the NSW job training data and to estimating demand elasticities from Nielsen scanner data while allowing preferences to be correlated with prices and income. |
|---|