Master Thesis

In method comparison experiments with replications, two measurement procedures (test method and comparative method) are compared using regression analysis. The ordinary linear regression approach often yields biased estimates of regression coefficients because the predictor (measurements of comparative method) contains random error. Instead, a linear error-in-variables regression model should be used. This work focuses on the approaches based on linear structural error-invariables model, called averaged Deming regression, and Linnet’s weighted Deming regression as they are recommended by CLSI (Clinical and Laboratory Standards Institute) for the analysis of method comparison data. Theoretical foundations of these approaches are analyzed and compared with ML estimations for linear structural error-in-variables model and the performance of both methods is investigated using Monte-Carlo simulations. The performance of different methods of calculation of confidence intervals for the estimates of regression coefficients is also examined using Monte-Carlo simulations. Simulation studies show that the averaged Deming regression and Linnet’s weighted Deming regression approaches combined with jackknife or bootstrap quantile confidence intervals yield reasonable estimations of regression coefficients, even if the model assumptions are considerably violated. The case of matrix effect in the test measurement procedure is also considered. Simulation studies suggest biased estimates of regression parameters calculated via averaged Deming regression or Linnet’s weighted Deming regression, if there is a strong matrix effect in the test method. In that case, another approach based on the linear structural ME model with error in equation may be used for the estimation of regression coefficients. The simulation studies show proper results of this method if there is a strong matrix effect in the test method. Chapter