2024-10-13-世界银行-后双重选择拉索在田间试验中的应用（英）_44页_1mb

Summary of PDS Lasso Performance in Field Experiments

Key Findings

1. Limited Impact of PDS Lasso

   • PDS Lasso rarely selects control variables (median of 2 variables), and standard errors are marginally smaller than those from ANCOVA, but improvements are minimal on average.
   • In over 75% of cases, treatment estimates and standard errors are similar to those obtained using ANCOVA (difference < 0.04 standard deviations).

2. Attrition Effects

   • PDS Lasso is more likely to select controls in cases of high attrition. However, even here, the impact on treatment estimates remains small.

3. Failure to Select Key Predictors

   • In some cases, PDS Lasso fails to select the lagged dependent variable—a common source of omitted variable bias, leading to larger standard errors.

4. Cross-Validation Risks

   • Cross-validation tends to select more variables, increasing the risk of overfitting and inflating standard errors, especially in small samples.
   • Recommended to use the default plug-in penalty for consistency.

Recommendations for Researchers

1. Amelioration Set

   • Include lagged variables and randomization strata in the amelioration set to prevent their exclusion.

2. Control Variables Selection

   • Avoid a “kitchen-sink” approach; limit inputted controls to reduce spurious associations and improve precision.

3. Handling Missing Data

   • Address missing values (8 of 18 papers had sample size reduction due to this). Dummy out missingness and include an indicator, or correct the amelioration set.

4. Interactions with Treatment

   • Include interacting variables in the amelioration set to avoid unintended selection as controls.

5. Implementation Checklist

   • Pre-specify controls and amelioration sets.
   • Avoid excessive controls and questionable missing-value handling.
   • Do not use cross-validation unless well-justified.

Conclusion

PDS Lasso offers a structured approach to covariate selection but fails to significantly reduce standard errors on average. Researchers should implement it judiciously and prioritize robust model specification over optimism from machine learning.