A Principal Components Approach to Cross-Section Dependence in Panels

The use of GLS to deal with cross-section dependence in panels is not feasible where N is large relative to T since the disturbance covariance matrix is rank deficient. Neither is it the appropriate response if the dependence results from omitted global variables or common shocks correlated with the included regressors. These can be proxied by the principal components of the residuals from a baseline regression. It is shown that the OLS estimates from a regression augmented by these principal components are unbiased and consistent using sequential limits for large T, large N. Simulations show that this leads to a substantial reduction in bias even for relatively small T and N panels. An empirical application indicates that the impact of cross section dependence seems to strengthen the case for long run PPP.