Estimation and Inference in the Presence of Neighborhood Unobservables


Recent literature emphasizes the importance of accounting for local unobserved effects, which may stem from either omitted neighborhood characteristics or shocks. As in a panel data setting, we exploit neighborhood data transformations to eliminate these effects. We consider a class of estimators indexed by a proximity threshold, allowing the researcher to select the neighborhood. Gaining insights from the recent dyadic data literature, we study their asymptotic properties allowing for overlapping neighborhoods. We develop a Hausman-like testing strategy to detect the presence of correlated local unobserved effects. Our test does not require assumptions on the neighborhood unobserved heterogeneity distribution and allows practitioners to select the best neighborhood specification to transform the data. We illustrate our approach using data from Miguel and Kremer (2004). Our method allows us to recover the same direct average treatment effect by considering treatment spillovers as local unobserved effects.