Abstract Background Validating a candidate method is commonly done by methods comparison (MC) using an accepted comparator to measure samples and to assess their agreement. The most powerful way of doing so is by parametric modeling: an “errors-in-variables,” or Deming, regression. The measurement variance is rarely constant across the measuring interval, and so Deming regression requires suitable weights—the inverse of each measurement's variance—to achieve its theoretical good performance. Sometimes, independently of the data set, we have an explicit mathematical model, the precision profile, connecting each assay's variability to the analyte concentration. But many studies have no such external information, and the analysis must rely on the data set alone, with some assumptions about the form of the precision profile. Methods The mathematical theory melding mathematical precision profile models with errors-in-variables (Deming) regression is sketched. Results Weighted Deming regression is outlined, and R codes for implementing it in the known and unknown precision profile settings are discussed. The implementation includes a jackknife approach for standard errors, confidence intervals for the regression parameters, residuals for diagnostics on normality and linearity, and a methodology for identifying and testing outliers. Conclusions Existing weighted Deming regression publications assume either constant coefficient of variation or constant variance. Precision profile models fill this gap and allow for more general settings.
Hawkins et al. (Tue,) studied this question.