Abstract A nested error linear regression model with unequal error variances is considered. This model is appropriate for combining results from different groups, experiments or studies. New estimators of the between experiment variance, σ v 2 , and the error variances, σ i 2 , are obtained by ignoring lower-order terms in the minimum norm quadratic unbiased estimation (MINQUE) equations. The resulting two-step estimator, β w , of regression parameters, β , is well defined. Asymptotic normality of β w is established, and a consistent estimator of Cov ( β w ) is also obtained, as the number of experiments, k, increases. A “delete-group” jackknife method is employed to obtain consistent estimators of var ( σ v 2 ) and Cov ( β w ) which in turn lead to asymptotically valid tests and confidence intervals on σ v 2 and the elements of β . Results of a simulation study on the finite-sample properties of the proposed methods are also reported.