Table 1.
Variable type summary. List and experiment names of each background environmental variable used in each modeling experiment. Predictor variables were grouped based on their abiotic, biotic, and anthropogenic characteristics. Variables with Pearson Correlation coefficient greater than 0.5 were not grouped together in their respective experimental subsets, and removed from the “All Uncorrelated Variables” group.
Table 2.
Final models summary. Summary of significant models (those which fit Omission Rate and AICc criteria) and final models selected by MaxEnt. Final models selected by MaxEnt are identified by the algorithm using pROC, omission rate, and AICc. RM denotes regularization multipliers of final selected models. FC denotes feature classes of final selected models.
Fig 1.
Model performance in terms of omission rate and AICc for each set of predictor variable experiments. Predictor variables for each experiment are shown as: all variables (yellow), climate only (green), landscape only (black), anthropogenic only (orange), grouped variables (purple), and grouped variables with bias surface correction (red and blue). Models calibrated using only climate variables had the overall lowest omission rates (green). Models calibrated with anthropogenic variables which used cattle density as a bias file; however, had the lowest amount of variance in both omission rates and AICc (red). This figure was created in BioRender. Van de Vuurst, P. (2025) https://BioRender.com/0b3muw.
Fig 2.
Bias file impact on model performance.
The overall reduction in model performance uncertainty between models without sampling bias correction (purple), and models which accounted for sampling bias (red) highlights the importance of accounting for sampling bias in ecological niche modeling. While models without bias correction were more well preforming (i.e., low omission rate and AICc of Purple points), model performance was more precise when bias was taken into account (red points). Sampling bias was corrected using cattle density data from Gridded Livestock of the World database [45]. Models which accounted for sampling bias using cattle density (red) had a lower amount of variance in both omission rates (sd = 0.006 vs 0.0110) and AICc (sd = 139 vs 395), with the differences in AICc being statistically significant per post hoc LSD analysis (p < 0.001).
Fig 3.
Anthropogenic predictor variable relationships to RABV.
A) Histogram of chicken density (number of individual chickens per square kilometer) at locations of RABV spillover used as training and testing data. Note the negative relationship between chicken density and RABV spillover locations. B) Histogram of poverty index at locations of RABV spillover used as training locations. Note the positive association between poverty index and RABV spillover locations. C) Histogram of human population density (number of individuals per km2) at locations of RABV spillover used as training locations. Note the negative association between human population density and RABV locations. These results indicate that RABV spillover risk is higher when chicken density is low, poverty index is high, and human population density is low. The majority of RABV spillover occurred in locations with fewer than 5000 chickens per km2, human density was less than 100 per km2, and where poverty index was greater than 70.
Fig 4.
Geographic projection of each final model per predictor variable model experiment.
Similarity of projected pixel to presence locations (i.e., suitability for RABV spillover) is shown from low (purple) to high (red). Maps correspond to predictor variable groups used including A) all uncorrelated variables, B) anthropogenic variables only, C) climate variables only, D) landscape variables only, E) all variables grouped by characteristics, F) all variables grouped by characteristic and with accessibility as a bias surface, and G) all variables grouped by characteristic and with cattle density as a bias surface. Predictor variable sets are broken down in Table 2. Note the differences between model outputs when cattle density data are used as a predictor variable vs as a sampling bias correction file (G). The total area predicted to have moderately or moderately-high suitability (yellow) was much larger (30.6% of the total area) when sampling bias was accounted for in panel G. High suitability for RABV spillover (red) was localized to northern portions of the country when only climate variables were used (C) than when anthropogenic (B) or landscape (D) variables were used. Maps created using ArcGIS Pro software version 2.5 with shape files from DIVA-GIS [51,84]. This figure was created using BioRender. Van de vuurst, P. https://BioRender.com/0b3muw1.