Predictive model for severe food insecurity for the Brazilian municipalities

In 2004 a national survey employed the Brazilian Food Insecurity Scale; however the results did not allow reliable disaggregation of the estimates on a local level. The aim of this study was to build a statistical model for prediction of severe food insecurity in Brazil, thus allowing to obtain reliable estimates of the food insecurity magnitude for the 5,564 Brazilian municipalities.  The statistical modeling included four phases. Stage 1: Selection of variables within the food security / severe food insecurity dimensions, through linear regression, using the criterion of p-value <0.05. The selected variables should be present in both the nationwide 2004 survey and the 2000 Census, subsequently used to calculate municipal estimates for severe food insecurity. Stage 2: Obtaining a reduced model, using the expanded sample, with Odds ratios adjusted by multiple logistic regressions using the criteria of p<0.05 and odds ratios statistically different from 1.0, Stage 3: Model Adjustment; at this stage other variables which did not enter the initial model were tested and the model was refined by removing non significant variables. The Wald test was applied to verify the significance of coefficients which formed the logistic equation. Stage 4: Testing the proposed model. The final Food Insecurity model included the variables: income per capita, education, sex and color of the family head; presence of piped water and bathroom in the household; total of children in the household, number of residents per bedroom and state where the home is located. A goodness-of-fit model test was performed on the observed results using Nagelkerke test (0.84). The measure of the model´s prediction efficiency was also evaluated by the construction of the ROC curve, with an area of 0.83. The proposed model showed an excellent fit and may be used for the prediction of food insecurity in Brazilian municipalities.