Avoiding regression analysis pitfalls when calibrating white blood cell differential counters.

Abstract

Method comparison and calibration are important for clinical hematologists and for manufacturers of new methods and instruments. Automation of the white blood cell (WBC) differential count poses unique problems in this regard because the reference method is a microscopic manual cell count of 200 WBCs. A highly variable reference method can obscure the true relation between the reference and the test methods and reduce sensitivity to miscalibration. A computer model based on published cell distributions and on the statistical properties of proportions was created to enable testing of alternative approaches to calibration. Ordinary linear regression (OLR) and standard Deming regression (SDR) were applied to a comparison of the results obtained with a virtual cell counter (counting 2,000 WBCs) and that obtained with a virtual 200-cell reference method for monocytes (mean population percentage of 7.1%) and lymphocytes (mean population percentage of 29%). Both regression approaches can misrepresent the known mathematical relation between the methods. This problem almost always occurs when OLR is used and the proportion of the cell of interest is low, as in the case of monocytes (7.1% of the total count). OLR comparisons, however, introduce significant error even for cell counts of 29%. SDR performed much better for both cell types.