A note on studentizing a test for heteroscedasticity

Breusch and Pagan (1979) have recently proposed a convenient test for heteroscedasticity in general linear models. This note derives the asymptotic distribution of their test under sequences of contiguous alternatives to the null hypothesis of homoscedasticity. The test is shown to possess asymptotically incorrect size (nominal significance level) except in the case of strictly Gaussian disturbances. A slight modification of the test is proposed which corrects this defect.