On the Robustness of Dirichlet-multinomial Regression in the Context of Modeling Pollination Networks
Recent studies have suggested that the structure of plant-pollinator networks is driven by two opposing theories: neutrality and linkage rules. However, relatively few studies have tried to exploit both of these theories in building pollination webs. This thesis proposes Dirichlet-Multinomial (DM) regression to model plant-pollinator interactions as a function of plant-pollinator characteristics (e.g. complementary phenotypic traits), for evaluating the contribution of each process to network structure. DM regression models first arose in econometrics for modeling consumers' choice behaviour. Further, this thesis (i) evaluates the robustness of DM regression to misspecification of dispersion structure, and (ii) compares the performance of DM regression to grouped conditional logit (GCL) regression through simulation studies. Results of these studies suggest that DM regression is a robust statistical method for modeling qualitative plant-pollinator interaction networks and outperforms the GCL regression when data are indeed over-dispersed. Finally, using DM regression seems to significantly improve model fit.