To the Editor.—The article in the January 2004 issue, “Evaluation of Linearity in the Clinical Laboratory,” 1 contained a fine review of the issue of linearity testing, as well as a demonstration of its benefits. In describing the conditions that can lead to a nonlinear response, the authors might have noted an article by Lum et al,2 which demonstrates a direct relationship between nonlinear performance on the College of American Pathologists (CAP) (ungraded) Linearity (LN) Surveys and unacceptable scores in the corresponding (graded) Surveys. This is strong evidence to support the conclusions by Jhang et al,1 that linearity assessments can detect important lingering problems that escape quality control.
Unfortunately, I must point out a significant misstatement concerning the NCCLS EP6-A3 procedure. The article correctly points out that the CAP LN Surveys and the NCCLS protocol use the same polynomial procedure to determine linearity, as originally developed by Kroll and Emancipator.4 However, the authors then describe the statistical tests used in the CAP LN Surveys and state “This statistical approach has gained acceptance as the best statistical method to evaluate linearity of quantitative tests and has been adopted as an approved guideline (NCCLS EP6-A).” This is not correct; the statistical tests in the CAP LN Surveys are very different than those in the EP6-A protocol in the following aspects: (1) the CAP procedure tests the average nonlinear error across the tested range, whereas the NCCLS protocol tests each point individually against predetermined goals; (2) the CAP procedure determines the statistical significance of nonlinearity using complex statistics based on parametric assumptions about nonlinear error, whereas the NCCLS protocol uses simple arithmetic comparisons with no assumptions about the underlying distribution of differences; and (3) to test for adequate precision, the NCCLS procedure includes a visual review of the graphical presentation of the data, while the CAP procedure relies on another parametric test (but uses the same graphical techniques).
The CAP statistical tests were reviewed by NCCLS, but they did not survive the rigorous consensus process. It was felt that (1) the use of average nonlinear error might not be sensitive to the most frequent nature of nonlinearity, which occurs at one of the endpoints, while the remainder of response can be linear; (2) the parametric assumptions are not supported by evidence; and (3) the complex statistical procedures cannot be easily reproduced by the intended users of EP6-A.
A more complete description of the EP6-A procedure and a description of the long process of developing a consensus linearity standard are available in a recent article in Clinical Chemistry News.5