An evaluation of typing procedures for quantitative data: results of a Monte Carlo study

The accuracy of clustering as a method for typing is considered. Clustering is compared to a method which uses maximum likelihood estimates (MLEs), the criteria for comparison being the specificity and the sensitivity of the methods. In general, the clustering procedure is more sensitive (fewer false negatives) and less specific (more false positives) than the method using MLEs. Both procedures work equally well if the positives make up between 20 to 80% of the population and there is a considerable difference between positive and negative responses. Both methods work poorly when there is a considerable overlap in response. For attributes having a prevalence of less than 0.1 and a spacing of 3.0 or more standard deviations, it is worthwhile to compute maximum likelihood estimates and use a Bayes method for typing. Also considered are the implications of these results with respect to MHC typing. In particular we present methods for determining whether clustering is in order and a discussion of the possibility of using the MLEs along with Bayes method for considering other models for the MHC data and for typing with selected panels.