Estimating Heteroscedastic Variances in Linear Models

Abstract We describe an estimator of heteroscedastic variances in the Gauss-Markov linear model where E(ε) = 0 and with σ i 2 and unknown. It may be thought of as an approximation to the MINQUE method which results in computational economy, positive estimates, and decreased mean square error. Properties of this almost unbiased estimator are stated. It is compared with other estimators, and extensions to more general models are discussed.