Multilevel Modeling for Spatial Epidemiology of Cancer in Argentina.

GARCÍA F; DIAZ, MARIA DEL PILAR; ABALLAY LR; POU SA; STIMOLO MI

Barcelona

Conferencia; IBS XXIXth International Biometric Conference.; 2018

International Biometric Society

Introduction: Multilevel models are the most common strategy of statistical analysis for both spatially correlated data or geographic correlation studies. They also allow to analyze data collected at different level of spatial aggregation and provide a useful framework to model sources of variability. Cancer is the second leading cause of death in Argentina and several socio-economic and environmental exposure factors were identified associated with most incident cancers in this country. Argentina has around 90% of urban population and the standardized mortality ratios (SMR) distribution has been described as aggregate. For small areas, standardized mortality ratios are very instable and maps of this measure can be misleading. Objective: To assess heterogeneity within clusters, identifying which of the selected characteristics and factors are important for explain the spatial distribution of cancer mortality. Methods: Standardized mortality ratios (SMRi=(Yi/Ei)x100,000, Yi and Ei observed and expected number of deaths by cancer and cardiovascular diseases in Argentina , calculated over 1996-2015 period, for 511 counties nested into 24 provinces, within 5 regions, were analyzed spatially using statistical indexes (global and local measures) in order to confirm aggregated and different distributions between neighboring regions. Three main cut-points were chosen over time (1996, 2000 and 2010, two later ones coinciding with national census) to disease mapping. Multilevel Poisson model (two-levels) were performed by region to estimate factor effects using several covariates (socio-economic characteristics: percentage of households with unsatisfied basic need -UBN-, and others) and environmental exposure factors at the different levels (county and province). Prediction of random effects were obtained to explore small area phenomenon, disease mapping and model diagnostics. Results: The maps intuitively captures the existence of spatial patterns of the ASMRs. Significant autocorrelations were found for global cancer ASMR distributions in men and women populations, also for prostate cancer. Geographical differences in socioeconomic conditions and urbanization were associated with spatial variation in cancer mortality rates. There were additive effects of poverty (UBN) and urbanization (PUH), showing that low-UBN and high-PUH were associated with an increasing risk of death by cancer (in both sexes) compared to counties with the high-UBN and low-PUH conditions, respectively.Conclusion: Cancer mortality rates are spatially aggregated in Argentina and their distributions are associated with poverty and urbanization. The two-level weighted Poisson model reduces the spatial correlation in the residuals of predictions of mortality rates over counties