Working Paper: NBER ID: w30207
Authors: Dake Li; Mikkel Plagborg Møller; Christian K. Wolf
Abstract: We conduct a simulation study of Local Projection (LP) and Vector Autoregression (VAR) estimators of structural impulse responses across thousands of data generating processes, designed to mimic the properties of the universe of U.S. macroeconomic data. Our analysis considers various identification schemes and several variants of LP and VAR estimators. A clear bias-variance trade-off emerges: LP estimators have lower bias than VAR estimators but substantially higher variance at intermediate and long horizons. Consequently, unless researchers are overwhelmingly concerned with bias, shrinkage via Bayesian VARs or penalized LPs is attractive.
Keywords: Local Projections; Vector Autoregression; Impulse Response; Bias-Variance Tradeoff
JEL Codes: C32; C36
Edges that are evidenced by causal inference methods are in orange, and the rest are in light blue.
| Cause | Effect |
|---|---|
| LP estimators (C51) | lower bias than VAR estimators (C51) |
| VAR estimators (C51) | lower variance than LP estimators (C51) |
| Bias reduction (C46) | higher variance costs (J32) |
| Shrinkage methods (Bayesian VAR or penalized LP) (C51) | preferable unless bias is predominant concern (C90) |
| Bayesian VAR estimator (C51) | sensitive to prior selection (C52) |
| No single estimation method (C13) | dominates across all horizons (F01) |
| Penalized LP (C61) | optimal at short horizons (G19) |
| Bayesian VAR (C11) | favored at intermediate and long horizons (D15) |
| IV identification methods (C26) | can be biased but reduce dispersion significantly (C46) |
| External IV methods (C26) | may be justifiable despite limitations (H84) |