Experimentation on heterogeneous experimental units

A frequent but often conveniently ignored problem in experimentation involving longitudinal and cross-sectional data is heterogeneity in variances across experimental units. Under such conditions conventional and popular OLS methods often perform poorly, producing biased results. Motivated by a practical problem from marketing consumer magazines at retail, we model the variance structure in this problem context in a manner that provides a simple solution to the estimation and inference problem. The proposed approach readily generalizes beyond the applications used in our illustrations.