New heteroskedasticity-robust standard errors for the linear regression model

Linear regressions fitted to cross-sectional data oftentimes display heteroskedasticity, that is, error variances that are not constant. A common modeling strategy consists of estimating the regression parameters by ordinary least squares and then performing hypothesis testing inference using standard errors that are robust to heteroskedasticity. These tests have the correct size asymptotically regardless of whether the error variances are constant. In finite samples, however, they can be quite size-distorted. In this paper, we propose new heteroskedasticity-consistent covariance matrix estimators that deliver more reliable testing inferences in samples of small sizes.