Machine Labor

• 成果类型: Article
• 署名作者: Angrist, Joshua D.; Frandsen, Brigham
• 署名单位: Massachusetts Institute of Technology (MIT); National Bureau of Economic Research; Brigham Young University
• 刊物名称: JOURNAL OF LABOR ECONOMICS
• ISSN/ISSBN: 0734-306X
• DOI: 10.1086/717933
• 发表日期: 2022
• 页码: S97-S140
• 关键词: instrumental variable estimation; weak instruments; propensity score; learning-methods; School quality; linear-models; estimators; regression; inference; selection

摘要

The utility of machine learning (ML) for regression-based causal inference is illustrated by using lasso to select control variables for estimates of college characteristics' wage effects. Post-double-selection lasso offers a path to data-driven sensitivity analysis. ML also seems useful for an instrumental variables (IV) first stage, since two-stage least squares (2SLS) bias reflects overfitting. While ML-based instrument selection can improve on 2SLS, split-sample IV and limited information maximum likelihood do better. Finally, we use ML to choose IV controls. Here, ML creates artificial exclusion restrictions, generating spurious findings. On balance, ML seems ill-suited to IV applications in labor economics.