Testing for error cross section independence with an application to US health expenditure

This paper considers the problem of testing for error cross section independence in a panel where statistical units may be subject to unobserved common effects, spatial spill overs, or both. We review a number of diagnostics that are used for testing for error cross section independence in panels, including tests based on spacings and spatial statistics. We then argue that commonly used spatial statistics might give misleading results when cross section correlation arising from common effects is not taken into account. Hence, we study the properties of spatial statistics applied to residuals obtained from an augmented regression, where common factors have been approximated by principal components (Bai, 2009). Small sample properties of our testing strategy are investigated in a Monte Carlo study. Results show that spatial tests applied to de-factored residuals detect well the presence of spatial correlation in the data. The paper concludes with a small empirical exercise on US health expenditure.