Correlation robust inference when economic distance is latent

• Department of Economics

This seminar features a paper presentation by Mark W. Watson (Princeton University).

Valid standard errors for estimated parameters need to account for potential correlation between observations. Autocorrelation robust standard errors are routinely used in time series applications, and analogous methods are available for spatially correlated data. However, these methods require knowledge of the distance between the observations, but in many applications the relevant economic distance is unknown. We show that this economic distance can be recovered from covariance patterns of strongly correlated variables observed at the same latent locations as the variables of interest. We provide conditions under which sample values of these auxiliary variables can be used to construct weights for projection-based standard errors and critical values that guarantee large-sample validity of standard inference methods. Empirical examples show the usefulness of the proposed method.

Jointly with Anna Bykhovskaya and Ulrich Müller