Modelo hierarquizado : uma proposta de modelagem aplicada à investigação de fatores de risco para diarréia grave

Abstract

Os estudos epidemiológicos de doenças que acometem as crianças geralmente envolvem grande número de variáveis. As associações entre potenciais fatores de risco e doença freqüentemente são avaliadas através de modelagem estatística sem descrição das estratégias empregadas. No estudo realizado apresenta-se uma abordagem hierarquizada aplicada à avaliação de fatores de risco para diarréia grave. As variáveis foram hierarquicamente agrupadas em características socioeconômicas, ambientais, reprodutivas maternas, nutricionais e demográficas. Na análise univariada todas as variáveis associaram-se com diarréia grave. Em cada bloco selecionaram-se fatores de confusão através de um algoritmo, utilizando-se o processo retrógado de seleção, através do módulo em passos, segundo um p=0,10. Os fatores de risco foram avaliados através de regressão logística após o ajuste para fatores de confusão de cada conjunto e para aqueles hierarquicamente superiores. As variáveis incluídas no modelo permitiram identificar corretamente uma proporção elevada de casos (gamma=0,74) e todos os blocos contribuíram significativamente para a modelagem.

In epidemiological investigations of infant diseases, data are usually collected on a large set of variables. The associations between presumptive risk factors and the outcome is commonly evaluated through statistical modeling, but the model-building strategies are seldom described. In the project, data collected in a case-control study of risk factors for dehydrating diarrhea in infants have been used to present a hierarchized approach to the assessment of risk factors. The variables were grouped into a hierarchy of categories, ranging from distal determinants to proximate ones. These categories included, in this order, socioeconomic, environmental, reproductive maternal, nutritional and demographic sets. According to the univariate analyses all variables were associated with the outcome. As the purpose was to identify a parsimonious model to explain the data, in each set the confounders were selected through backward elimination, according to an alpha level of 0.10. The risk factors were evaluated through logistic regression after adjustment for confounders in the same set or in hierarchically superior sets. This approach allows researchers to quantify the contribution of each level of adjustment, to understand the modelbuilding strategy as well as interpret the independent associations. The goodness-of-fit assessed at each set showed significant improvements in the model. The gamma coefficient of association was employed to quantify the proportion of cases and controls correctly identified by comparing the observed value with that predicted by the variables in the model. The final model resulted in a gamma of 0.74. The children’s ages did not improve the prediction of cases and controls, but they have been kept in the model as they affect some exposures such as breastfeeding.