Predicting bee community responses to land-use changes: Effects of geographic and taxonomic biases

Abstract

Land-use change and intensification threaten bee populations worldwide, imperilling pollination services. Global models are needed to better characterise, project, and mitigate bees' responses to these human impacts. The available data are, however, geographically and taxonomically unrepresentative; most data are from North America and Western Europe, overrepresenting bumblebees and raising concerns that model results may not be generalizable to other regions and taxa. To assess whether the geographic and taxonomic biases of data could undermine effectiveness of models for conservation policy, we have collated from the published literature a global dataset of bee diversity at sites facing land-use change and intensification, and assess whether bee responses to these pressures vary across 11 regions (Western, Northern, Eastern and Southern Europe; North, Central and South America; Australia and New Zealand; South East Asia; Middle and Southern Africa) and between bumblebees and other bees. Our analyses highlight strong regionally-based responses of total abundance, species richness and Simpson's diversity to land use, caused by variation in the sensitivity of species and potentially in the nature of threats. These results suggest that global extrapolation of models based on geographically and taxonomically restricted data may underestimate the true uncertainty, increasing the risk of ecological surprises.

Figure 1. The predictive error and explanatory power of models that include only the intercept (NULL), LUI alone, subregion alone, additive effects, or interactive effects.

LUI = Land Use and Intensity. For explanatory power, solid bars show the marginal R²_glmm (the variance explained by fixed effects) and the hashed bars show the conditional R²_glmm (the variance explained by both random and fixed effects). Error bars show the standard error of the mean predictive error across 10 folds of cross validation. Note that the predictive error should only be compared among models assessing the same response variable, as absolute values depend on the measurement scale.

Figure 2. Predicted means of total (logged) abundance of bees for different land-use classes in each subregion, with 95% confidence intervals.

Also shown are significant results of multiple comparisons, testing differences between natural (Primary vegetation) and semi-natural land uses (Secondary vegetation) to human-dominated land uses, and differences between low, medium and high intensity cropland (*p < 0.05, **p < 0.01, *** p < 0.001).

Figure 3. The predictive error and explanatory power of models that include three way interactions between LUI, subregion and taxon (Bombus or not), and models with two way interactions between LUI and taxa, or LUI and Subregion.

LUI = Land Use and Intensity. For explanatory power, solid bars show the marginal R²_glmm (the variance explained by fixed effects) and the hashed bars show the conditional R²_glmm (the variance explained by both random and fixed effects). Error bars show the standard error of the mean predictive error across 10 folds of cross validation. Note that the predictive error should only be compared among models assessing the same response variable, as absolute values depend on the measurement scale.