世界银行-后双重选择拉索在田间试验中的应用（英）-2024.9-44页_1mb

Summary

PDS Lasso, a method for selecting control variables in randomized experiments, aims to improve precision and reduce bias. However, empirical analysis using 780 treatment effects from 18 published field experiments shows minimal practical benefits. On average, PDS Lasso reduces standard errors by less than 1% compared to the standard Ancova method. Variables selection is sparse, with few controls typically chosen, and in over half the cases, no variables are selected in the treatment regression. The method often fails to include the lagged dependent variable, potentially reducing precision. Cross-validation did not consistently outperform the default plug-in penalty. Implementing controls like lagged variables and strata effects in the amplify set is recommended. While PDS Lasso offers a structured approach, its limited gains in efficiency coupled with risks of underperformance suggest researchers should carefully weigh its use, especially given the prevalence of small samples and limited statistical power concerns in field experiments.