Working Paper: CEPR ID: DP10253
Authors: Laurent Gobillon; Thierry Magnac
Abstract: In this paper, we investigate the use of interactive effect or linear factor models in regional policy evaluation. We contrast treatment effect estimates obtained by Bai (2009)'s least squares method with the popular difference in differences estimates as well as with estimates obtained using synthetic control approaches as developed by Abadie and coauthors. We show that difference in differences are generically biased and we derive the support conditions that are required for the application of synthetic controls. We construct an extensive set of Monte Carlo experiments to compare the performance of these estimation methods in small samples. As an empirical illustration, we also apply them to the evaluation of the impact on local unemployment of an enterprise zone policy implemented in France in the 1990s.
Keywords: economic geography; enterprise zones; linear factor models; policy evaluation; synthetic controls
JEL Codes: C21; C23; H53; J64; R11
Edges that are evidenced by causal inference methods are in orange, and the rest are in light blue.
| Cause | Effect |
|---|---|
| their methods effectively control for cross-section dependence (C23) | provide valid estimates of the treatment effect on unemployment exits (J65) |
| Monte Carlo simulations demonstrate performance of different estimation methods (C15) | including interactive effects and synthetic controls (C32) |
| DiD estimates are biased when the data-generating process has interactive effects (C22) | unobserved heterogeneity leads to incorrect conclusions about treatment effects (C21) |
| synthetic control methods provide more reliable estimation of treatment effects (C90) | matching variables of treated units belong to the support of control units (C32) |