On the nature and analysis of clustered data.

Studies in which data from multiple patients are collected per clinician or per practice are becoming common in primary care research, particularly with the increase of studies conducted in practicebased research networks. These studies generate data that are clustered. A special case of clustered data is an intervention study where clinicians or practices are randomized into an intervention or control group. In such cluster-randomized designs, all patients of a clinician or practice are assigned to the same treatment, and this design is often used when logistics of implementation or the need to avoid contamination of treatment arms is a priority. A major issue in the analysis of clustered data is that observations within a cluster are not independent, and the degree of similarity is typically measured by the intracluster correlation coeffi cient (ICC).1 Ignoring the intracluster correlation in the analysis could lead to incorrect P values, confi dence intervals that are too small, and biased estimates and effect sizes, all of which can lead to incorrect interpretation of associations between variables.2 Failure to take into account the clustered structure of the study design during the planning phase of the study also can lead to underpowered study designs in which the effective sample size and statistical power to detect differences are smaller than planned. In most situations, the numeric value of the intracluster correlation tends to be small and positive. Several authors have provided guidelines for interpreting the magnitude of the intraclass correlation3 with small, medium, and large values of the intraclass correlation coeffi cients reported as .05, .10, and .15. Small values of the intracluster correlation can be deceiving, however. Investigators need to be aware that the cluster effect is a combination of both the intracluster correlation and the cluster size. Small intracluster correlations Confl icts of interest: none reported