A general variance model in methods comparison

Valid statistical tests of paired data require correct models of how measurement variance depends on analyte concentration. One often-used assumption is that the variance is constant across the range; another is that the coefficient of variation is constant. But in many data sets, neither of these holds. A variance model containing both a constant variance and a constant coefficient of variation term is recommended as an often-useful additional analysis tool for methods comparison. The more general variance model is fitted to a simulated data set, and one from a clinical chemistry methods comparison. It is used to provide more reliable average versus difference plots, to fit weighted Deming regressions, and to provide valid paired data analyses. The calculations are implemented in r software.