Recorded presentation

Basile Chaix, PhD, Research team on social determinants of health and utilization of care, National Institute of Health and Medical Research, INSERM U444, Faculté Saint-Antoine, 27 rue Chaligny, Paris Cedex 12, 75571, France, 33144738443, chaix@u444.jussieu.fr and Juan Merlo, MD PhD Asso Prof, Department of Community Medicine (section of Preventive Medicine), Malmö University Hospital, Faculty of Medicine, Lund University, S-205 02, Malmö, Sweden.

Measuring the extent to which phenomena occur in cluster is highly informative for public health policymakers. For continuous outcome variables, multilevel models provide a convenient measure of clustering in the form of the intraclass correlation coefficient (ICC). However, for binary outcome variables, only approximate definitions exist for the ICC. Klaus Larsen and colleagues propose to express clustering in the well-known odds ratio scale with an index termed the median odds ratio (MOR) that can be easily computed from multilevel logistic models. On the other hand, alternating logistic regression (ALR) models are now recognized as an interesting alternative to multilevel logistic models. They quantify clustering with a pairwise odds ratio that considers similarity / dissimilarity between individuals residing in the same area rather than heterogeneity between areas. Using Swedish data from the HSS-2000 survey, we considered area-level variations of the propensity to use private rather than public healthcare providers to illustrate the different measures of clustering. The area-level variance was highly significant, the ICC was equal to 0.08, the MOR to 1.81 and the POR to 1.37 (95% CI: 1.19, 1.57). Both the multilevel and ALR models indicated that the amount of clustering was higher among older individuals. This complex pattern of variations was in part explained by the percent of inhabitants with a higher education in the area of residence. Since the ICC, the MOR and the POR provide information on clustering under different forms, we finally discuss their statistical consistency and interpretability in a public health user-oriented perspective.

Learning Objectives: