Evaluation of regression procedures for methods comparison studies.

Using simulation, I evaluated five regression procedures that are used for analyzing methods comparison data: ordinary least-squares regression analysis, weighted least-squares regression analysis, the Deming method, a weighted modification of the Deming method, and a rank procedure. I recorded the following performance measures: plus or minus bias of the slope estimate, the root mean squared error of the slope estimate, and correctness of hypothesis testing. I evaluated the unweighted regression procedures by using a simulated comparison of two electrolyte methods; only the Deming method gave unbiased slope estimates. Using the jackknife method, a computer program that estimates the standard error, I showed that hypothesis testing was correct for the Deming method. I used all regression procedures on data from a simulated comparison of two measurement methods with proportional analytical errors dispersed over one decade. Bias of the slope estimates was not a problem for these cases. The weighted least-squares regressionanalysis and the weighted Deming method were most efficient (lowest root mean squared error); the other procedures required 1.6 to 2.2 times as many observations to attain the same precision for the slope estimate. Hypothesis testing was correct by the weighted Deming method with the jackknife principle for standard error computation; the other methods rejected the null hypothesis 1.4 to 4.4 times too frequently. In conclusion, it is preferable to use an alternative to ordinary least-squares regression analysis for methods comparison studies.