Predictors of free-roaming domestic dogs' contact network centrality and their relevance for rabies control

Free roaming domestic dogs (FRDD) are the main vectors for rabies transmission to humans worldwide. To eradicate rabies from a dog population, current recommendations focus on random vaccination with at least 70% coverage. Studies suggest that targeting high-risk subpopulations could reduce the required vaccination coverage, and increase the likelihood of success of elimination campaigns. The centrality of a dog in a contact network can be used as a measure of its potential contribution to disease transmission. Our objectives were to investigate social networks of FRDD in eleven study sites in Chad, Guatemala, Indonesia and Uganda, and to identify characteristics of dogs, and their owners, associated with their centrality in the networks. In all study sites, networks had small-world properties and right-skewed degree distributions, suggesting that vaccinating highly connected dogs would be more effective than random vaccination. Dogs were more connected in rural than urban settings, and the likelihood of contacts was negatively correlated with the distance between dogs’ households. While heterogeneity in dog's connectedness was observed in all networks, factors predicting centrality and likelihood of contacts varied across networks and countries. We therefore hypothesize that the investigated dog and owner characteristics resulted in different contact patterns depending on the social, cultural and economic context. We suggest to invest into understanding of the sociocultural structures impacting dog ownership and thus driving dog ecology, a requirement to assess the potential of targeted vaccination in dog populations.

questionnaire forms are available as supplementary material. The MFA was performed at the country level, except for Chad where the analysis at the country level fully separated the owners by study site. Therefore, to categorize the wealth within our two study areas in Chad, a MFA was performed separately for each study site (Rural 1 and rural 2). The number of components included in the analysis was selected based on a visual basis of the eigenvalues histogram to explain the variability of the data with as few dimensions as possible (Supplementary Figure S3). The number of components was set to two for all the countries, explaining in total 56.1%, 71.4%, 22.8% and 31% of the variance for Chad-rural 1, Chad-rural 2, Indonesia and Uganda, respectively.

Hierarchical clustering analysis
The number of clusters selected through hierarchical clustering analysis (HCA) is the one with the higher relative loss of inertia, with the inertia being the sum of within-and between-class inertia. The selected partition is the one for which an additional merge between clusters leads to a high loss in inertia between clusters 2 . A cluster can be constituted of a single household or groups of households. The HC resulted in the creation of three clusters in each study area (Supplementary Tables S6-S9). The MFA and the hierarchical clustering was performed using the FactoMineR package in R software 32 . Supplementary Fig. S4. Permutation-based multivariate linear regression model results with betweenness as output. A. Significance and coefficient sign of the dog-level network. B: Proportion of deviance explained by each variable of the dog-level network. C. Significance and coefficient sign of the householdlevel network. D. Proportion of deviance explained by each variable of the household-level network. Empty field denote variables that were not explored or where not selected for the best models by the PBLM.
Dog-level factors: dog's sex (Sex -0: male (baseline), 1: female); body conditioning score (BCS) of 2 and lower (Low BCS), BCS of 4 and higher (High BCS) with the baseline of BCS = 3; being a guardian dog (Guardian =1, dummy variable), hunting dog (Hunting, dummy variable), shepherd dog (Shepherd, dummy variable) or raised for meat (Source of meat, dummy variable); free-roaming time (FRT -range from 0 to 10); number of dogs collared per household (NDC, contiguous variable) and distance per 100 meters from dogs home to the centroid of the study site (Distance, continuous variable). Household-level factors: wealth category based on the Multiple Factor Analysis or the income when available (Wealth, with the lowest level (i.e. poorest) being the baseline); owner finalizing primary school (Primary school), finalizing secondary school (Secondary school), finalizing professional training or university (Higher Education), with absence of formal education being the baseline; belonging to the main local ethnicity (Ethnicity, dummy variable), being catholic (Catholic, dummy variable), being evangelic (Evangelic, dummy variable); number of dogs collared per household (NDC, contiguous variable) and distance per 100 meters from household to the centroid of the study area (Distance, continuous variable). Supplementary Fig. S5. Dog-level permutation-based linear model results with degree as output, excluding the proximity events with a RSSI below -75dBm. A. Significance and coefficient sign. B: Proportion of deviance explained by each variable. Empty field denote variables that were not explored or where not selected for the best models by the MRQAP.
Dog-level factors: dog's sex (Sex -0: male (baseline), 1: female); body conditioning score (BCS) of 2 and lower (Low BCS), BCS of 4 and higher (High BCS) with the baseline of BCS = 3; being a guardian dog (Guardian =1, dummy variable), hunting dog (Hunting, dummy variable), shepherd dog (Shepherd, dummy variable) or raised for meat (Source of meat, dummy variable); free-roaming time (FRT -range from 0 to 10); number of dogs collared (NDC, contiguous variable) and distance per 100 meters from dogs home to the centroid of the study site (Distance, continuous variable). Fig. S6. Dog-level permutation-based linear model results with betweenness as output, excluding the proximity events with a RSSI below -75dBm. A. Significance and coefficient sign. B: Proportion of deviance explained by each variable. Empty field denote variables that were not explored or where not selected for the best models by the PBLM.

Supplementary
Dog-level factors: dog's sex (Sex -0: male (baseline), 1: female); body conditioning score (BCS) of 2 and lower (Low BCS), BCS of 4 and higher (High BCS) with the baseline of BCS = 3; being a guardian dog (Guardian =1, dummy variable), hunting dog (Hunting, dummy variable), shepherd dog (Shepherd, dummy variable) or raised for meat (Source of meat, dummy variable); free-roaming time (FRT -range from 0 to 10); number of dogs collared (NDC, contiguous variable) and distance per 100 meters from dogs home to the centroid of the study site (Distance, continuous variable).
Supplementary Fig. S7. Odds ratios derived from the multiple regression quadratic assignment procedure for having a contact between dogs, using the modified netlogit function (A) or excluding the proximity events with a RSSI below -75dBm (B) having the same level per variable investigated. Positive: OR > 1, negative OR < 1. Empty field denote variables that were not explored or where not selected for the best models by the MRQAP.
Dog-level factors: same sex (male versus female), same age category (more or less than two years old), same BCS category (more or less than 2), same reason for keeping the dog (guardian, hunting, shepherd, source of meat), similar free-roaming time (FRT, always free-roaming, free-roaming by day, by night, a few hours per day or never) and distance per 100 meters between the households (Distance, continuous variable). Household-level factors: same wealth category (Wealth, cluster 1 to 4), same education level (Education: no formal education, primary education, secondary education, higher education), same ethnicity (Ethnicity, various levels depending on the study site), same religion (Religion, various levels depending on the study site) and distance per 100 meters between the households (Distance, continuous variable). Number of adjacent nodes to a given node Betweenness Number of times a node sits on the shortest path between two other nodes Relative degree Degree divided by the network size Normalized betweenness Betweenness divided by the maximum possible betweenness in the network. The normalized betweenness corresponds to the fraction of all possible shortest paths in the network on which a dog lie. * We choose not to consider isolated nodes because it is impossible to differentiate truly isolated dogs (i.e. dogs that did not have contact with any other dog during the data collection period) and dogs of which their collar was defective. Table S2. Average shortest path length and clustering coefficient of the observed and simulated networks and small world index. For the average path length, the p-value corresponds to the proportion of simulated random networks having an average path length shorter than the observed average path length. For the clustering coefficient, the p-value corresponds to the proportion of simulated networks having a clustering coefficient larger than the observed clustering coefficient.

Chad
Guatemala Indonesia Uganda