Alternative statistical techniques to evaluate linearity

With the increase in demand for linearity assessment, a wide variety of analysis techniques have been advocated. There is no consensus on optimal techniques. This article reviews different approaches that have been advocated by the College of American Pathologists, by the National Committee for Clinical Laboratory Standards, and by manufacturers of diagnostic methods and controls. This is not a review of all statistical techniques proposed for linearity assessment. I discuss four basic categories of methods, with the first and simplest being visual review. The second general approach includes the conventional statistical techniques based on least-squares regression; this includes an equation for the line and statistical tests for linearity and for curvature. The third general approach involves comparison of slopes for line segments, or "deltas." This approximates the visual assessment of linearity. The fourth approach involves comparing observed and expected values, with some allowance for differences. Comparisons can be made for observed single results, means of replicates, or components of variance. Allowances for error can be internal goals or goals that are recommended for clinical usefulness. Each approach has strengths in different aspects of linearity assessment, including intuitive appeal, statistical rigor, and objectivity. Similarly, each approach has limitations that are not always obvious.