Predict first–assemble later versus assemble first–predict later: Revisiting the dilemma for functional biogeography

Abstract

Predicting community‐level trait indices, such as community‐weighted mean or functional dispersion, in space or time, is crucial, as they are markers of ecosystem functions, proxies for ecosystem services, and are now integral part of conservation planning. On one hand, since members of a species share similar traits, a natural way is to model species distributions as a function of the environment first (i.e. with species distribution models), and then to reconstruct trait indices for any locality of interest (predict‐first). On the other hand, community‐level trait indices can be seen as the direct result of environmental filtering, and thus their distributions may directly be modelled in response to the environment (assemble‐first). Although relatively different, these two approaches have been used interchangeably in trait‐based ecology with unknown consequences on their usability. Here, using plant community (4463 plots) and trait data (LNC, leaf nitrogen content; PLH, plant height; SLA, specific leaf area, for >800 species) covering the French Alps, we compared the two approaches to predict community mean and functional dispersion indices, accounting or not for abundance (CM, community mean; CWM, community weighted mean; FDis, functional dispersion; uFDis, unweighted functional dispersion). We tested both interpolation versus extrapolation capabilities of the two approaches. To support the empirical findings, we also run the same comparative analysis on simulated community data. While both approaches produced similar and skillful CM/CWM predictions (R2 > 0.63 for CWM PLH) for interpolation, the assemble‐first outperformed the predict‐first approach for extrapolation to unobserved environmental conditions (R2 = 0.60 against 0.54 for CWM PLH). Functional dispersion was generally less well predicted, although the assemble‐first approach again provided better predictions for both interpolation (R2 = 0.31 against 0.27 for FDis PLH) and extrapolation (R2 = 0.28 against 0.18 for FDis PLH). The predict‐first approach systematically overestimated FDis. Results were confirmed by the simulation experiments. We conclude that the direct modelling of community‐level indices provides more robust spatial predictions over space and time and should thus be preferred. This approach also does not suffer from having generally too many species predicted by the predict‐first approach.