A sensitivity test for species distribution models used for gap analysis in New Mexico



Journal Title

Journal ISSN

Volume Title


Texas Tech University


Gap Analysis Program is a landscape-level evaluation of plant communities and animal richness and is useful in wildlife conservation. Models to predict species distributions are fundamental for Gap analysis to assess animal richness patterns. Basic models combine species' locality, together with their vegetation associations (base variables). Additional environmental variables (filter variables) are used for further adjustment and to represent particular habitat associations. However, there is a need to know the contribution that filter variables provide to the models. This research evaluated model sensitivity to simplification of habitat associations in the model, by systematically removing selected filter variables. This was done to quantify indirectly the value added by these variables and to examine if the response pattern was consistent with expected model performance.

Distribution predictions from the New Mexico Gap Analysis Project (NMGAP) were used as baseline data for this study. A representative sample of species was evaluated by using subset combinations of filter variables as model input. Altered species distribution estimates were examined for differences in area relative to the full model.

Model sensitivity to different single filter variables and combinations was found to be highly variable among different types of models (number of filter variables used) and the species they represented. Between extreme high and low values, results indicated that in general all filter variables presented some level of influence on the models (assumed adjustment). However, for those models formed with two or three filter variables most of their least influential variables did not differ (p>0.05) from a threshold value set as minimum acceptable change (5% change in area). This indicated that more risks of variables with no effect may occur when adding more than two filter variables (potential correlation). Consistent with model performance expectations, general response to perturbation suggested that each additional filter variable used in the model produced a cumulative effect for most models, and that this effect was greater (p<0.05) for models representing species with restricted distributions than for those with widespread species distributions. Sensitivity analysis is recommended at the stage of review of preliminary wildlife species distribution maps produced for Gap analysis to detect model weaknesses.