Model evaluation
We built the phylogeny-only, environment-only, and combined (phylogeny and environment) models using leaf trait data from NEON and evaluated their explanatory power using the Bayesian R2 of the predicted values for 88 out-of-sample test trees. For all 8 leaf traits, the combined model explained the largest amount of variance in the held-out test data (average R2 across the 8 traits = 0.64), substantially outperforming both the environment-only (average R2 = 0.35) and phylogeny-only (average R2 = 0.52) models (Figure 1, Figure S.2). A hierarchical clustering of model residuals supported two major trait classes (Figure S.3): traits mainly involved in photosynthesis (Croft et al., 2017) and traits involved in leaf structure. The combined model had the highest performance for predicting LMA (R2 = 0.81) and the lowest performance for predicting ChlA (R2 = 0.51). Uncertainty in predictions was accurately estimated across all traits and for all models (Figure S.4), with the combined model showing mean 95% coverage values ranging from 94.3% to 98.8%. The importance of different environmental drivers varied among traits, supporting an important role of climate in driving leaf economics in local communities (Ordoñez et al. 2009). Precipitation and temperature were mainly important for traits involved in photosynthesis (N%, ChlA, ChlB and Carotenoids), generally having a positive effect on their concentration (except for ChlB, Figure S.5). Net radiation showed a negative effect on N%, while vapor pressure generally had a negative effect on pigments but a positive effect on traits associated with leaf toughness and durability (cellulose, lignin and C%). Elevation was the most important topographic predictor and the strongest environmental driver of LMA (Figure S.5), consistent with previous studies (Reich & Oleksyn, 2004, Hedin 2004, Poorter et al., 2009, Kitajima et al., 2016). The joint model structure also captured the strong correlation among LMA and N% characteristic of the leaf economics spectrum (Reich et al, 1997, Wright et al., 2004).
Role of species and environment in predicting traits at the continental scale
We used variance partitioning on the Bayesian R2 from out-of-sample test data to explore the relative contributions of inter- and intraspecific trait variation at continental scale (Supplement 4). On average, interspecific variation (pure phylogenetic and species effect) accounted for 25% of the total explained variation across the 8 traits, intraspecific variation (pure environment effect) accounted for 13%, and joint phylogenetic-environment effects accounted for 23%. The relative importance of inter- and intraspecific effects varied widely among traits. Species and phylogeny explained most of the variation in structural traits (e.g. LMA, C% and lignin%). For these traits, often used in large scale distribution studies, species distributions may contain more information about traits than direct predictions from the environment, as previously suggested by CWM models (Clark 2016, Yang & Swenson, 2018). These results are aligned with previous studies suggesting that the relative extent of intra-species variation among communities is negatively related to spatial extent (Siefert, et al. 2015) and that at continental and global scales, patterns of leaf traits are mainly driven by leaf economic strategies at the species level (Wright et al., 2005). In contrast, the intraspecific component of the model accounted for as much or more of the variance than the interspecific one for pigments.
This difference between structural and photosynthesis traits may be explained by being driven by different kinds of tradeoffs. Structural traits are more affected by leaf lifespan and toughness, which varies widely across species (Wright et al. 2004, Kitajima et al. 2012, Osnas et al. 2018, Lichstein et al. 2021), whereas photosynthetic traits are less variable among species due to the fundamental need for all species to maximize carbon gain (subject to ecological conditions and tradeoffs; Wright et al. 2004; Maire et al, 2015). Our analysis only quantifies intraspecific variation from the upper (sun-lit) canopy. Although this is a common practice (Pérez-Harguindeguy et al. 2013), it likely results in underrepresenting intraspecific variation by ignoring leaf variability across the light gradient, a major source of intraspecific variation (Osnas et al., 2018). Furthermore, 62% of species in our analysis were only sampled within a single NEON site, thus representing only a small fraction of species’ true environmental ranges. Thus, our analysis likely underestimates the true level of intraspecific trait variation.