Ramos SR, Delgado RA, Martínez CH, Santana PS

Language: Spanish
References: 6
Page: 223-232
PDF size: 203.35 Kb.

ABSTRACT

We discuss in this article the performance of 4 solutions to the methods comparison problem: ordinary least squares regression (P1), Passing-Bablok regression (P2), Deming regression with the λ coefficient estimated from the variances of replicate measurements (P3) and the Deming regression with the λ coefficient estimated from the variances of each serie of observations (P4), when the corresponding analytical errors can be assumed to be constant throughout the analytical range of interest, but different. The case of the constant and equal analytical errors was presented in an accompanying article [Delgado Ramos A, Ramos Salazar R, Martínez Canalejo H, Santana Porbén S. Procederes de regresión lineal como soluciones al problema de la comparación de métodos. I. Errores analíticos constantes e iguales. LABORAT-Acta (Archivos Mexicanos de Laboratorio Clínico). 2000]. The performance of the Deming regression with preset values of the λ coefficient: 2.25 (the case of sodium), 4 (the case of Albumine), and 2.25 (the case of glucose) was also discussed (P5). The theoretical model, the simulated analytical scenarios, the statistical-mathematical simulation algorithm and the quality specifications were previously published. The ordinary least squares regression and the Passing-Bablok regression returned biased slope estimates and resulted in a high rejection frequency rate of the null hypothesis H_o: Β = 1, and therefore, were not satisfactory solutions to the methods comparison problem. Regarding the ordinary least squares regression, the slope estimation error was distorted by a systematic component. The poor performance of the Passing-Bablok could not be explained on the basis of the gathered evidences in this study, though it can be hypothesized that this procedure is rather sensible to differences between the analytical errors of the methods under comparison. The Deming regression with the λ coefficient estimated from the variances of each serie of observations (P4) was neither a satisfactory solution to the methods comparison problem, when also returning biased slope estimates. The remaining versions of the Deming regression (preset values of the λ coefficient, and λ coefficient estimated from the variances of replicate measurements, respectively) returned unbiased slope estimates, but at the cost of an increased rejection rate of the null hypothesis H_o: Β = 1. It might be possible that the real level of accuracy of these versions of the Deming Regression lies between 90 - 95%. It is concluded that when the analytical errors of the methods under comparison are constant and different, the Deming regression with a properly estimated λ coefficient is the only solution to the methods comparison problem.

REFERENCES

1. Hollis S. Analysis of method comparison studies. Ann Clin Biochem 1996; 33: 1-4.

2. Delgado RA, Ramos SR, Martínez CH, Santana PS. Procederes de regresión lineal como soluciones al problema de la comparación de métodos. I. Errores analíticos constantes e iguales. Laborat Acta. 2000.

3. Martínez CH, Santana PS. Manual de Procedimientos Bioestadísticos. La Habana: Editorial de Ciencias Médicas, 1990.

4. Linnet K. Estimation of the linear relationship between the measurement of two methods with proportional error. Stat Med 1990; 9: 1463-1473.

5. Linnet K. Performance of Deming regression analysis in case of misspecified analytical error ratio in method comparison studies. Clin Chem 1998; 44: 1024-1031.

6. Stöckl D, DeWitte K, Thienpont LM. Validity of linear regression in method comparison studies: limited by the statistical method or the quality of the analytical input data? Clin Chem 1998; 44: 2340-2346.