Partitioning biochemical reference data into subgroups: comparison of existing methods.

Four existing methods for partitioning biochemical reference data into subgroups are compared. Two of these, the method of Sinton et al. and that of Ichihara and Kawai, are based on a quotient of a difference between the subgroups and the reference interval for the combined distribution. The criterion of Sinton et al. appears rather stringent and could lead to recommendations to apply a common reference interval in many cases where establishment of group-specific reference intervals would be more useful. The method of Ichihara and Kawai is similar to that of Sinton et al., but their criterion, based on a quantity derived from between-group and within-group variances, seems to lead to inconsistent results when applied to some model cases. These two methods have the common weakness of using gross differences between subgroup distributions as an indicator of differences between their reference limits, while distributions with different means can actually have equal reference limits and those with equal means can have different reference limits. The idea of Harris and Boyd to require that the proportions of the subgroup distributions outside the common reference limits be kept reasonably close to the ideal value of 2.5% as a prerequisite for using common reference limits seems to have been a major improvement. The other two methods considered, that of Harris and Boyd and the "new method" follow this idea. The partitioning criteria of Harris and Boyd have previously been shown to provide a poor correlation to those proportions, however, and the weaknesses of their method are summarized in a list of five drawbacks. Different versions of the new method offer improvements to these drawbacks.