BackgroundData analytic approaches to Affymetrix®microarray data include: (a) a covariate model, in which the observed signal is some estimated linear function of perfect match (PM) and mismatch (MM) signals; (b) a difference model [PM-MM]; and (c) a PM-only model, in which MM data is not utilized.MethodsBy decomposing the correlations among the variables in the statistical model and making certain assumptions, we theoretically derive the statistical model that reflects the actual gene expression level under a variety of conditions expected in microarray data.Results and conclusionWhen modeling non-systematic variation, the covariate model provides maximum flexibility and often reflects the actual gene expression levels better than the difference model. However, the PM-only model demonstrates superior power in an overwhelming majority of realistic situations, which provides theoretical support for the current trend to employ PM-only models in microarray data analyzes.