Master's Theses

Date of Award

2015

Document Type

Thesis

Department

Biology

First Advisor

Robert P. Anderson, Ph.D.

Second Advisor

Ana C. Carnaval

Keywords

Maxent, Madagascar, Niche Model

Abstract

Ecological niche models (ENMs) characterize the relationship between localities where a species is known to occur and the abiotic characteristics of these regions. While widely used, ENMs remain subject to several outstanding issues, including those related to model complexity and violation of modeling assumptions (e.g., representative sampling). Critical in resolving these issues is a better understanding of the effectiveness of model selection techniques. Here, I compare two strategies for optimizing ENMs: an information-criterion approach (AICc) and a sequential approach that assesses model performance on withheld data. I do so for a single species using two datasets, one with all available occurrence records, and the other with spatially filtered occurrence records (expected to reduce the effect of sampling bias).

I conduct these experiments making models with Maxent for a species with few occurrence records, the endemic Malagasy rodent Eliurus majori (subfamily Nesomyinae), using 19 bioclimatic variables. Candidate models were created across a wide range of complexities. For both datasets, both model-selection techniques chose simpler models than Maxent’s default settings. In the unfiltered dataset, the models selected as optimal by AICc had substantially fewer parameters than those selected by the sequential technique. In contrast, both techniques converged on similar settings when the spatially filtered dataset was used, possibly due to the relative lack of sampling bias present, which better fulfilled important niche modelling assumptions.

Nevertheless, the results of each respective selection technique, and default settings differed between unfiltered and filtered datasets. Qualitative examination of predictions in light of expert knowledge indicated that the selection techniques yielded more realistic models than did the default settings, and those models made with the filtered dataset more closely matched available distributional and natural history information for the species. To reach general conclusions regarding these issues, similar studies should be undertaken with a wide variety of simulated and real species datasets.

Available for download on Monday, January 01, 2018

Included in

Biology Commons

Share

COinS