Background: A crucial question when developing reference intervals is whether different subpopulations need their own reference interval or if a single joint reference interval can be used. It is reasonable to use partitioned reference intervals in situations where a single interval results in considerable variation in sensitivity between subpopulations. The aim of partitioning is to harmonize the sensitivity of the reference intervals, i.e., to make the sensitivity similar for all patients, regardless of patient characteristics. Statistical criteria to identify when partitioning is adequate have been developed over the last two decades. These criteria are applicable when considering two subpopulations, but recently a procedure for considering several subpopulations has been developed. When several subpopulations are considered, there is a possibility that some subpopulations could form a group or cluster that could share a common reference interval. However, there is no formal systematic approach to indicate how to divide these subpopulations into clusters. The aim of this study was to suggest such a systematic approach for clustering.
Methods: A clustering technique was applied to data including several subpopulations. The technique is based on measuring the distance between separated reference limits and successively pooling subpopulations divided by short distances. A cluster is defined by a group of subpopulations that are close to each other and that differ from subpopulations in another cluster. A cluster recruits new subpopulations as long as the subpopulations can be pooled without violating a partitioning criterion.
Conclusions: We have suggested a procedure for partitioning a number of Gaussian (or Gaussian-transformable) subpopulations into clusters. This is the only formalized procedure indicating how to analyze several subpopulations and identify a suitable number of groups and reference intervals. Using a computer program developed for partitioning issues, the approach was easy to adopt.