Ith more than 5,000 persons per square kilometer were considered as “urban”. Other communes were classified as “rural”. Human behaviors were documented through a dedicated questionnaire. For each of the 1,578 communes considered, the percentage of surface covered by each landscape class (vegetation and water bodies), as well as the values of climatic, NDVI and cattle density covariates were computed with the Quantum GIS INK1117MedChemExpress INK1117 software [37]. Malagasy commune administrative boundaries and data come from the layers data merged by the Office for the Coordination of Humanitarian Affairs (OCHA) and based on data obtained from the Malagasy National Disaster Management Office in 2011.Multiple Factor AnalysisSynthetic variables characterizing the environment of communes were computed using a MFA combining the previously mentioned climatic and landscape variables [38,39]. By performing aPLOS Neglected Tropical Diseases | DOI:10.1371/journal.pntd.July 14,5 /Rift Valley Fever Risk Factors in Madagascarfactor analysis inside each variable category and then between categories, MFA produces a quantitative summary of the initial set of variables taking the form of a set of linear combination of variables, referred to as factors [39]. The climatic category included the annual means of day and night LST, the annual mean and seasonality of precipitation. The landscape category included the percentage of the surface of the commune covered by each landscape category and the annual mean and seasonality of NDVI. The value of each factor was computed for each of the 1,578 Malagasy communes. Correlation between MFA factor values and cattle density distribution was assessed using Pearson product-moment correlation coefficient test.Statistical analysisAs a first step univariate analyses of association between suspected risk factors and cattle or human RVFV serological status were undertaken using Chi square tests for categorical factors and generalized linear models for quantitative factors. Risk factors with significance level 0.20 were then included as explanatory variables in GLMMs, with cattle or human individual serological status as the binomial response. In these models, it was assumed that the relationships between serological prevalence and quantitative factors were linear on the logit scale. To account for interdependency of serological status of individuals sampled in the same locality, the smallest administrative unit–the commune for the cattle model and the city/village for human model- were included in the models as a random effect. Multicollinearity among explanatory variables was assessed using Variance Inflation Factors (VIF) and correlation tests. Collinear factors were not included in a same model. The selection of the best models was based on the Akaike Information Criterion (AIC). When needed, a multi-model inference approach was used to Linaprazan web estimate model-averaged fixed effects (mafe) and the relative importance (RI) of each explanatory variable [40]. Within the set of models tested, only those with an AIC within 2 units difference from the best model were considered [40]. Internal validity of sets of models was evaluated using the Receiver Operating Characteristic (ROC) curve method [41]. In addition, we calculated the 10-fold cross-validation prediction. Because, it is not possible to perform 10-fold cross-validation on GLMM, this procedure was applied to Generalized Linear Models that were similar to the selected GLMM except that did not include t.Ith more than 5,000 persons per square kilometer were considered as “urban”. Other communes were classified as “rural”. Human behaviors were documented through a dedicated questionnaire. For each of the 1,578 communes considered, the percentage of surface covered by each landscape class (vegetation and water bodies), as well as the values of climatic, NDVI and cattle density covariates were computed with the Quantum GIS software [37]. Malagasy commune administrative boundaries and data come from the layers data merged by the Office for the Coordination of Humanitarian Affairs (OCHA) and based on data obtained from the Malagasy National Disaster Management Office in 2011.Multiple Factor AnalysisSynthetic variables characterizing the environment of communes were computed using a MFA combining the previously mentioned climatic and landscape variables [38,39]. By performing aPLOS Neglected Tropical Diseases | DOI:10.1371/journal.pntd.July 14,5 /Rift Valley Fever Risk Factors in Madagascarfactor analysis inside each variable category and then between categories, MFA produces a quantitative summary of the initial set of variables taking the form of a set of linear combination of variables, referred to as factors [39]. The climatic category included the annual means of day and night LST, the annual mean and seasonality of precipitation. The landscape category included the percentage of the surface of the commune covered by each landscape category and the annual mean and seasonality of NDVI. The value of each factor was computed for each of the 1,578 Malagasy communes. Correlation between MFA factor values and cattle density distribution was assessed using Pearson product-moment correlation coefficient test.Statistical analysisAs a first step univariate analyses of association between suspected risk factors and cattle or human RVFV serological status were undertaken using Chi square tests for categorical factors and generalized linear models for quantitative factors. Risk factors with significance level 0.20 were then included as explanatory variables in GLMMs, with cattle or human individual serological status as the binomial response. In these models, it was assumed that the relationships between serological prevalence and quantitative factors were linear on the logit scale. To account for interdependency of serological status of individuals sampled in the same locality, the smallest administrative unit–the commune for the cattle model and the city/village for human model- were included in the models as a random effect. Multicollinearity among explanatory variables was assessed using Variance Inflation Factors (VIF) and correlation tests. Collinear factors were not included in a same model. The selection of the best models was based on the Akaike Information Criterion (AIC). When needed, a multi-model inference approach was used to estimate model-averaged fixed effects (mafe) and the relative importance (RI) of each explanatory variable [40]. Within the set of models tested, only those with an AIC within 2 units difference from the best model were considered [40]. Internal validity of sets of models was evaluated using the Receiver Operating Characteristic (ROC) curve method [41]. In addition, we calculated the 10-fold cross-validation prediction. Because, it is not possible to perform 10-fold cross-validation on GLMM, this procedure was applied to Generalized Linear Models that were similar to the selected GLMM except that did not include t.
