Use of cluster analysis to define COPD phenotypes: Table 1–

The current classification of airways disorders is imprecise, with an overlap of phenotypes ( e.g. asthma, chronic bronchitis and emphysema), resulting in difficulties in differentiating the disorders from each other. This has led to considerable diagnostic, management and prognostic uncertainty. The traditional approach has been to present this phenotypic overlap in the Venn diagram format 1; however, this results in ≥15 phenotypes, whose pathogenesis or response to treatment have not been clearly defined 2, 3. More recent work 4–8, including that of Burgel et al. 8 published in the current issue of the European Respiratory Journal , has used cluster analysis to characterise different types of airways disorders. But what is cluster analysis, is it a reasonable approach to take, and how valid are the conclusions? Cluster analysis is a collection of methods for defining groups of individuals based on measured characteristics, so that they are grouped based on their differences (or similarities), into clusters 9–11. The groupings are constructed such that the degree of association is strong between members of the same cluster and weak between members of different clusters 4. Cluster analysis is distinct from other ways of trying to understand multivariate data, which include principal component and factor analysis, discriminant analysis and multivariate regression. Principal component (as used by Burgel et al. 8) and factor analysis produce linear combinations of measured variables, in the sense that new derived variables are produced by multiplying each of the original variables by a scaling parameter and adding the resulting numbers. Discriminant analysis (as also used by Moore et al. 5) starts with known groups and finds scaled combinations of the measured variables that best distinguish those known groups. Multivariate regression can have a set of response variables predicted by a set …