Speaker
Description
Quantitative evidence synthesis aims for general insights into the direction, magnitude and variability of ecological effects. Two common forms of quantitative synthesis are meta-analysis (i.e., analyses of effect sizes collated or calculated from existing studies), and analyses of data compiled to address specific questions. Both approaches frequently quantify the heterogeneity of effect sizes across studies to describe how consistent ecological effects are across studies. However, both these types of quantitative synthesis typically assume constant between-study heterogeneity (meta-analysis) or residual variation (i.e., homoscedasticity for analyses of raw data). This assumption is equivalent to assuming that included studies sample effect sizes from a single population. Despite the rapid rise of meta-analysis in ecology and conservation biology, neither the consequences of this assumption, nor the opportunities available when it is relaxed, have received much attention. Here, I demonstrate the utility of multilevel location-scale models and cross-validation for model selection in quantitative evidence synthesis. First, I revisit a meta-analysis of spatial scale-dependence in plant native-exotic species richness relationships. I quantify relationships among unexplained variation, grain size, and spatial extent, and examine how relaxing the assumption of constant heterogeneity impacts predictive performance. The second case study uses a data compilation of habitat fragment diversity studies. I focus on the relationship between fragment size and local (i.e., patch-scale) species richness, examine whether residual variation is related to fragment size and other study-level predictors, and how heteroscedastic residual variation impacts predictive performance. For both case studies, the assumption of constant between-study heterogeneity limits model predictive performance, particularly the ability of models to make predictions to new studies.
| Status Group | Postdoctoral Researcher |
|---|
