This article proposes using an expanded form of the Johnson S U family as a way to approximate nonnormal distributions in regression models. The distribution is one of the few that allows modeling heteroskedasticity and autocorrelation. The technique is evaluated with Monte Carlo simulation and illustrated through an empirical model of the West Texas cotton basis. Given nonnormality, this technique can substantially reduce the variance of slope parameter estimates relative to least squares procedures. Copyright 2003, Oxford University Press.