Lung cancer and COPD rates in Apulia: a multilevel multimember model for smoothing disease mapping

Background: If spatial representations of hospitalization rates are used, a problem of instability arises when they are calculated on small areas, owing to the small number of expected and observed cases. Aim of this study is to assess the effect of smoothing, based on the assumption that hospitalization rates, when calculated at the municipal level, may be influenced by both the neighboring municipalities and the health service organization, as well as by environmental risk factors associated with the disease under study. Methods: To smooth rates we hypothesize that each municipality belongs to two independent hierarchical levels; at one of these levels subjects may belong to a plurality of superior hierarchical objects. Two different models, so-called Multilevel Multimembership Models, are fitted. In the first the structure of random effects was: the municipality heterogeneity, the spatial dependence of the municipalities and the local health service organization. In the second we replaced the local health service organization effect with the environmental risk effect for each municipality area. The models were applied to spatially represent the rates of hospitalization for lung cancer and chronic obstructive pulmonary disease, determined through the hospital discharge forms recorded in Apulia for the year 2006. Results: The effect of smoothing was greater in smaller municipalities and in those with a more unstable Risk Adjusted Rate (RAR) due to the small number of cases and of population at risk. When a hierarchical level representing the ASL is inserted, the model fits the data better. Conclusion: Maps of hospitalization rates for lung cancer and chronic obstructive pulmonary disease, shaded with the rates obtained at the end of the smoothing procedure, change the visual picture of the disease distribution over the whole territory, and if detected by the model, seem to express a geographical distribution pattern in specific areas of the region. In the case of lung cancer, the models show a clear difference between RAR and smoothed RAR. The inclusion of a random effect indicating the ASL contributed to improve the graphic representation of the results, whereas the environmental risk was not found to be a better hierarchical level than the municipality for fitting of the model.