Estimating imprecision profiles in biochemical analysis.

We describe a computer program IMPROFIL which determines an imprecision profile of an analytical method from replicated measurements of samples. It calculates the variance function, the coefficient of variation, the power of definition, the critical limit, the limit of detection and the lower limit of quantification. The primary property, the variance function, is determined by two alternative methods: the conventional maximum approximate conditional likelihood method and the newly developed weighted absolute deviation method. For all quantities, confidence intervals are obtained using the bootstrap procedure. The program combines the use of robust numerical techniques, user-friendliness and integration into a spreadsheet program for data pre- and post-processing. The algorithms used are described in detail. Tests with synthetic data sets are used to validate the method and to establish its powers and limitations. Finally, its application to a practical analytical task (tumor marker CA 15-3 in human sera) is reported. For the method to yield a reliable estimate of the variance function and the derived properties, certain minimum requirements on the raw data must be met: They have to be spread throughout the concentration range of interest, there should not be less than three replicates per specimen, and there must be at least of the order of 25 (better at 50) specimens.