Cody Schank (PhD expected 2018) is the lead author of a collaborative paper that compares different statistical methods to model the potential distribution of Baird’s Tapir. The abstract is below and the article can be found here:
Aim
We test a new species distribution modelling (SDM) framework, while comparing results to more common distribution modelling techniques. This framework allows for the combination of presence-only (PO) and presence-absence (PA) data and accounts for imperfect detection and spatial bias in presence data. The new framework tested here is based on a Poisson point process model, which allows for predictions of population size. We compared these estimates to those provided by experts on the species.
Species and Location
Presence data on Baird’s tapir (Tapirus bairdii) throughout its range from southern México to northern Colombia were used in this research, primarily from the years 2000 to 2016.
Methods
Four SDM frameworks are compared as follows: (1) Maxent, (2) a presence-only (PO) SDM based on a Poisson point process model (PPM), (3) a presence-absence (PA) SDM also based on a PPM and (4) an Integrated framework which combines the previous two models. Model averaging was used to produce a single set of coefficient estimates and predictive maps for each model framework. A hotspot analysis (Gi*) was used to identify habitat cores from the predicted intensity of the Integrated model framework.
Results
Important variables to model the distribution of Baird’s tapir included land cover, human pressure and topography. Accounting for spatial bias in the presence data affected which variables were important in the model. Maxent and the Integrated model produced predictive maps with similar patterns and were considered to be more in agreement with expert knowledge compared to the PO and PA models.
Main conclusions
Total abundance as predicted by the model was higher than expert opinion on the species, but local density estimates from our model were similar to available independent assessments. We suggest that these results warrant further validation and testing through collection of independent test data, development of more precise predictor layers and improvements to the model framework.