Species distribution modelling, where the goal is to predict the distribution of a species as a function of environmental variables, is a highly-researched topic of interest to ecologists, biologists, and climate change scientists. Often, the best available data is “presence-only data”, which consists of reported opportunistic species sightings (presences) with no corresponding absence information. Many different methods for species distribution modelling have been developed for presence-only data across different disciplines, including the popular MAXENT.
In this seminar, I will establish a link between MAXENT and Poisson point process models, a method which analyses presence-only data as a point process. This link ties MAXENT to other methods that have been connected to Poisson point process models, such as pseudo-absence logistic regression.
As a consequence of these equivalences, it is possible to “borrow strength” across the different modelling frameworks. Two particular extensions I have proposed are PPM-LASSO (combining strengths of point process models and MAXENT) and downweighted Poisson regression (combining strengths of point process models and logistic regression). Particular advantages of these extensions I will highlight include a clear framework for choice of pseudo-absences, the ability to diagnose model adequacy and, where appropriate, fit models with point interactions, model-based control of observer bias, and estimation-consistent choice of the LASSO penalty available through a new method, which will be compared to other penalty choices.