Statistical power in the detection of matrix effects.

Matrix-induced bias can adversely affect the performance evaluation a clinical laboratory receives on proficiency testing results. Therefore, it is vital that matrix effects from a matrix-biased system are detected and that laboratories using these systems are not falsely penalized on their proficiency testing results. The College of American Pathologists has developed an experimental protocol to test whether an observed bias is, in fact, due to the proficiency testing sample matrix rather than true performance or calibration problems. The probability of detecting a matrix effect using this protocol given a matrix-biased system is defined as statistical power. Five parameters are known to affect the probability of detection: (1) size of the bias in the proficiency testing material, (2) lack of fit coefficient of variation-natural variation in patient samples used to define the relationship between the test and reference methods, (3) pure error coefficient of variation-random variation in the test method, (4) the number of fresh patient samples, and (5) the number of replicates for each sample. The level of significance of the statistical test will also affect the probability of detection. Power curves are generated to show the effect these five parameters have on the determination of power. With the exception of bias, power is most influenced by the components of variance, lack of fit (nonlinear), and pure (random) error. Of these two components, the lack of fit error, which represents an uncontrollable source of error, is usually more influential than pure error, which can be reduced by a larger number of replicates. A large increase in power will result from an increase of fresh patient samples from 10 to 20, and a moderate increase in power will result from an increase of fresh patient samples from 20 to 40; no noticeable increase in power is seen with greater than 40 fresh patient samples. Large increases in power were observed for increases in the number of replicates per sample from one to two, two to three, and three to five. To markedly increase power further, 10 replicates would have to be assayed.