Abstract
A Monte Carlo study is used to examine the size and power of t tests formed using a variety of estimation procedures appropriate in the context of heteroskedasticity when there are no replicated observations. There are three main results: (1) the ordinary least squares estimator is quite robust with respect to inference; (2) an estimated generalized least squares estimator, formed using a possibly-erroneous assumption that the functional form of the heteroskedasticity is multiplicative, has highest power among the estimators considered, but has a too-large size; and (3) the advantages of the jackknife do not appear until the degree of heteroskedasticity is unrealistically large