Title: Elastic-net for instrumental variables regression
Authors: Alena Skolkova - CERGE-EI (Czech Republic) [presenting]
Abstract: The purpose is to investigate the relative performance of the lasso, ridge and elastic-net estimators in obtaining first-stage predictions for IV estimation. Although the lasso estimator is currently established as the most popular regularization technique for prediction problems under the sparsity assumption, its performance under high correlation and grouping between instruments can be improved via the elastic-net. A Monte Carlo study demonstrates that in all analyzed scenarios the elastic-net estimator dominates other estimators in terms of the mean squared error and overall stability of the first-stage predictions. We also compare the relative performance of IV estimators that employ the lasso, ridge and elastic-net first-stage estimates. Finally, we confirm the superior performance of the elastic-net estimator in an empirical application with many IVs.