Testing for non-normality in multivariate regression with nonspherical disturbances

Abstract

Statistical diagnostic testing is often associated with erratic conclusions due to the fact that a test against one certain specification may be highly sensitive to another specification. This paper concerns assessing normality of autocorrelated or heteroscedastic variables. It is shown why the type I error of skewnesslkurtosis test limits 100% if the data are not i.i.d. We propose a set of tests for non-normality, which are robust to autocorrelationiheteroscedasticity, covering a wide class of situations. The size and power of the tests are investigated by Monte Carlo techniques.