Needles in a haystack: a machine learning approach to instrumental variables selection

Sammanfattning: This paper explores the comparative merits of two different machine learning algorithms for variable selection in an instrumental variables (IV) setting with many weak instruments. We apply a new method, post-L2boosting, to Angrist and Krueger's (1991) classical paper about the effect of schooling on earnings. We compare the performance of the post-L2boosting with another recently suggested method, post-LASSO estimation, on the same data. Among the methods used in this paper, post-LASSO is superior for increasing the first stage F-statistic of the IV estimation, implying that it more effectively reduces finite sample bias. However, our findings are not conclusive as further research is needed regarding the effects of different tuning techniques for the hyper parameters.