Abstract
Two tests for heteroscedasticity in linear regression and a heteroscedasticity-consistent estimator of convariance (White, 1980) are studied, through simulation, for their impact on inference on the regression coefficients in moderate sized samples. One of the tests, the direct test, and the estimator of convariance are both incorporated in recent regression software. The evidence of this study suggests that can give rise to misleading inference. The other test for heteroscedasticity, the simple test, performs far better.