Regression models are being increasingly used in rheumatology because of greater awareness of their application, and the ready availability of computerised statistical packages. Many researchers, therefore, now feel confident to undertake quite sophisticated multivariate analyses. One caution that we feel should be more widely debated and is pertinent to rheumatological datasets in particular, may limit the extent to which these analyses can be routinely undertaken without formal statistical assistance.1

The problem stems from the fact that in many aspects of medical research, data are collected on subjects and in most cases analysed using the individual as the basic ‘unit’ in the analysis. This is appropriate where the data are collected on single organ systems as the unit of the analysis is person specific. In the situation of having data on multiple joints or organs, this approach may not be wholly appropriate. For example, if one were examining risk factors for the knee, some variables, such as injury, will be joint specific rather than subject specific.

One example of this is the study by Doherty et al.2 In this study, factors that might predict progression of knee osteoarthritis were studied and included some which might be considered specific to the knee studied, and some which apply globally to the patient (table 1). As demonstrated in the paper, nearly all the knee specific features described showed a high degree of correlation between the right and left sides. The data could be analysed at the level of either patients or joints (sides). In the case of the former, data regarding knee specific factors have to be sacrificed and, as discussed below, choosing which knee to …