Anthropogenic factors and the risk of highly pathogenic avian influenza H5N1: prospects from a spatial-based model

Paul, Mathilde; Tavornpanich, Saraya; Abrial, David; Gasqui, Patrick; Charras-Garrido, Myriam; Thanapongtharm, Weerapong; Xiao, Xiangming; Gilbert, Marius; Roger, Francois; Ducrot, Christian

Since we expected some non-linear relationships, continuous variables were transformed into categorical data before they were entered into the model. Four categories were chosen for each variable with the exception of free-grazing ducks (density of animals and density of farms) whose distribution allowed only 3 categories. We selected the thresholds that simultaneously fitted the non-linear relationships and had a sufficient number of statistical units per category. Finally, we added the level of relative risk in the duck population as a covariate in the chicken model and vice versa for the duck model. For each species, this thereby made it possible to adjust the analysis to account for the level of disease occurrence in the other species. The multivariate analysis was carried out using a stepwise backward elimination process; the significance of each variable in the full model was assessed in turn, with the least significant variable deleted and the process repeated until all of the remaining variables were significant at p value 0.05 (Fisher test) [6] . From the effect estimates computed in the final linear model, we deduced values of risk ratios (RR) and their confidence intervals (95%) for the different variables. For each variable, the reference category was defined as the category expected to be at the lowest level of risk based on our hypothesis and findings of previous studies in Thailand [8, 24] . Although spatial dependence already had been taken into account through the spatial contiguity in the Bayesian approach, we looked for any remaining spatial autocorrelation in the linear models. The semi-variograms of model residuals showed that autocorrelation may have played a role only in a very short-distance range (2 250 m). Computing the distances between each subdistrict to the next nearest subdistrict, we found that only a small portion of subdistricts (< 5%) actually had a chance of being influenced by their neighbours within that range. Thus the likelihood that spatial autocorrelation affected the results was assumed to be low. The statistical analysis was performed using R software v.2.9.2 5 . Figure 1 shows the geographical distribution of HPAI outbreaks in chicken and duck flocks during this time period. Figure 2 presents the spatially smoothed relative risk maps for chickens and ducks and shows that the two maps resemble each other fairly closely (Spearman rho = 0.91, p < 1eÀ16). The maps visually confirm the presence of a ''hot spot'' of HPAI risk in the central plain of Thailand where the relative risk was significantly higher than the national average (relative risk > 10 for both chickens and ducks). For ducks, however, the high-risk area tended to extend further across the western part of the central plain of Thailand. In contrast, the extreme south of Thailand appeared to be a high-risk area for chickens, with values of relative risk significantly > 10. On the contrary, some areas were especially low-risk for both chickens and ducks despite the occurrence of outbreaks (relative risk significantly < 0.5), as in northeastern Thailand and in the middle part of the peninsula. Northern Thailand had low values of relative risk (significantly < 0.5) only for chickens.