MacKinnon and White (1983) examine modified versions of the heteroskedasticity-consistent covariance matrix estimator (HC) and compare their finite sample properties. They find that the jackknife-based estimator, HC3, performs better than other estimators in small samples. They also compare the power of heteroskedasticity tests and find that HC3 is more reliable than the conventional OLS estimator even in the absence of detected heteroskedasticity. The paper discusses the use of modified critical values based on Edgeworth approximations and finds that they improve the performance of tests in large samples. They also evaluate alternative heteroskedasticity tests, finding that White's "portmanteau" test performs well. The study concludes that HC3 is the preferred estimator for small samples and that using HC3 instead of OLS is advisable even when there is little evidence of heteroskedasticity. The paper is based on extensive sampling experiments and provides detailed results on the performance of various estimators and tests under different conditions.MacKinnon and White (1983) examine modified versions of the heteroskedasticity-consistent covariance matrix estimator (HC) and compare their finite sample properties. They find that the jackknife-based estimator, HC3, performs better than other estimators in small samples. They also compare the power of heteroskedasticity tests and find that HC3 is more reliable than the conventional OLS estimator even in the absence of detected heteroskedasticity. The paper discusses the use of modified critical values based on Edgeworth approximations and finds that they improve the performance of tests in large samples. They also evaluate alternative heteroskedasticity tests, finding that White's "portmanteau" test performs well. The study concludes that HC3 is the preferred estimator for small samples and that using HC3 instead of OLS is advisable even when there is little evidence of heteroskedasticity. The paper is based on extensive sampling experiments and provides detailed results on the performance of various estimators and tests under different conditions.