Study site and herbivory assessment
The study was conducted in the Gutianshan National Nature Reserve (29°14′N, 118°07′E) in south-east China. The reserve covers c. 80 km² of evergreen mixed broadleaved forest, with Castanopsis eyrei and Schima superba as dominant tree species. The subtropical monsoon climate is characterized by a mean annual temperature of 15.3°C and a mean annual precipitation of c. 2000 mm (Hu & Yu, 2008). Within the reserve, 27 study plots of 30 × 30 m2 were established in 2008. The plots were selected to represent the range of woody plant species richness (25–69 tree and shrub species per plot) and successional stages (< 20–> 80 yr) found in the reserve (Bruelheide et al., 2011).
Herbivory was assessed on saplings (20–100 cm in height) of 10 dominant tree and shrub species: Ardisia crenata Sims, Camellia fraterna Hance, Castanopsis eyrei (Champ. ex Benth.) Tutch., Cyclobalanopsis glauca (Thunb.) Oerst., Eurya muricata Dunn, Lithocarpus glaber (Thunb.) Nakai, Loropetalum chinense (R. Br.) Oliv., Machilus thunbergii Sieb. et Zucc., Neolitsea aurata (Hayata) Koidz. and Schima superba Gardn. et Champ. These 10 evergreen species accounted for c. 50% of the total biomass of the tree and shrub layers in the study plots (see Schuldt et al., 2010). A maximum of 10 saplings per species and plot were checked for herbivory. Herbivory was quantified as the overall leaf damage caused by chewing, mining, galling and (if visible) sucking insects on all leaves of the saplings (mean number of leaves per sapling = 45.4 ± 45.3 SD). Assessments were conducted at the end of the rainy season in June/July 2008, which also marks the end of a major activity period for arthropods in these forests (Schuldt et al., 2012). We used predefined percentage classes (estimated as 0%, < 1%, 1–5%, > 5–15%, > 15–35% and > 35%; see, for example, Scherber et al., 2010; Schuldt et al., 2010; Ness et al., 2011) to visually assess standing levels of leaf damage. The actual, mean amount of damage for each estimated percentage class was then checked in detail by analyzing samples of randomly collected leaves (20–30) for each class; these were digitally scanned to determine the exact amount of leaf damage as the ratio of removed to estimated total leaf area (Schuldt et al., 2010, 2012). For the statistical analyses, we then used the mean damage of the scanned leaves of each class to calculate mean damage levels for each sapling (i.e. to account for potential deviations in the visually estimated damage from the digitally verified mean damage levels; for details, see Schuldt et al., 2010).
Plant community data and general plot characteristics
For our analyses, we used a set of three morphological and four chemical leaf traits that are related to leaf quality and palatability, and that might thus particularly strongly affect herbivory (Coley & Barone, 1996; Perez-Harguindeguy et al., 2003; Poorter et al., 2004): leaf area (LA), specific leaf area (SLA) and leaf dry matter content (LDMC), as well as leaf C content, leaf C : N ratio, leaf C : P ratio and leaf polyphenolics content. The traits were measured for c. 80% of the 147 woody plant species recorded on the 27 study plots, and these species represented 95% of the total number of tree and shrub individuals at the study sites. As we used abundance-weighted indices to quantify functional community composition and diversity, these data should not be affected by the 5% of woody plant individuals for which trait values were missing. Data on leaf toughness, which has been shown in previous studies to potentially affect herbivory (Kitajima & Poorter, 2010), were only available for one-third of all species, and thus were not included in the analysis. However, Schuldt et al. (2012) showed that leaf toughness is probably not a limiting factor to herbivore damage in our study system. Details on trait measurements are provided in Kröber et al. (2012). In short, samples for trait measurements were taken from sun-exposed leaves of five to seven plant individuals in total, collected from up to seven plots per species in the summer of 2008. Trait measurements followed the standardized protocols of Cornelissen et al. (2003) and, for leaf polyphenolics, Hagermann (1987) (see Kröber et al., 2012). Our analysis focused on interspecific variation in trait values that determine community-level trait diversity, as intraspecific trait variability within species has been shown previously to have negligible effects on trait–environment relationships across our study plots (Kröber et al., 2012). Moreover, we show below that plot-level characteristics that can be expected to particularly strongly affect intraspecific trait variation (stand age, elevation and other abiotic conditions) were not retained in our final explanatory model, which further indicates that, unlike community-level trait diversity, intraspecific trait variation within species plays only a minor role in species-level variation in herbivory across the 27 study plots.
Phylogenetic data were obtained from an ultrametric phylogenetic tree of all angiosperm woody species recorded in the 27 study plots (Michalski & Durka, 2013). Woody plant species richness was recorded at the time of plot establishment in 2008 and was based on a complete inventory of all tree and shrub individuals of a height > 1 m (Bruelheide et al., 2011).
We also accounted for general plot characteristics, such as stand age, tree density, canopy cover, herb cover, elevation and aspect (see Bruelheide et al., 2011), as they might potentially confound diversity-functioning relationships in observational studies. Many of these characteristics were strongly correlated with each other, and we used principal components analysis (PCA) on these variables to obtain orthogonal predictor axes (for details of this analysis, see Schuldt et al., 2010). Only the first principal component axis (PC1abio), which represented stand age and age-dependent aspects of stand structure and biomass, was related to herbivore damage (Schuldt et al., 2010), and therefore was included in our analyses to account for diversity-independent plot effects. Other plot characteristics, as well as sapling height and the total number of saplings sampled, were shown by Schuldt et al. (2010) to have no effect on herbivory patterns of the study species.
Diversity metrics and statistical analysis
In many cases, it remains unclear whether ecological functions are more strongly affected by CWM trait values, the variability within single traits or the diversity of multiple traits (Butterfield & Suding, 2013; Dias et al., 2013), and to what extent phylogenetic diversity provides additional information (Cadotte et al., 2009). To quantify the functional and phylogenetic aspects of the woody plant communities, we thus used a three-fold approach calculating: (1) Rao's quadratic entropy Q (Rao, 1982) to assess plot-level trait and phylogenetic diversity; (2) CWM trait values to identify mass ratio effects of single traits; and (3) functional and phylogenetic relatedness between each of our focal species and all other species in the study plots to measure species-specific diversity effects.
Rao's Q is calculated as the variance in pairwise dissimilarities among all individuals in a community. It can easily be applied to both functional and phylogenetic data, calculated for single as well as multiple traits, and weighted by abundance data (Schleuter et al., 2010; Pavoine & Bonsall, 2011). It thus enables a comparison between different facets of diversity using a consistent statistical framework (Pavoine & Bonsall, 2011). Moreover, as a measure of trait dispersion, Rao's Q complements measures of CWM trait values (Ricotta & Moretti, 2011). Whereas CWM quantifies a community's average functional trait value, weighted by the relative abundances of all individuals in this community, Rao's Q provides a measure of trait variation around this mean. We calculated both CWM values and Rao's Q for single traits (CWMsingle.trait, Qsingle.trait), as well as two multivariate versions of Rao's Q that assessed the overall diversity of morphological (Qmorph) and chemical (Qchem) leaf traits. We also tested for the effects of an overall Rao's Q measure that integrates both the leaf morphological and chemical traits, but, as this measure was less strongly related to herbivory than was Qchem, we kept the distinction between morphological and chemical leaf trait diversity to allow for a better mechanistic interpretation of potential effects (although traits such as LDMC and C content might be related to some extent by both influencing leaf palatability (Poorter et al., 2009), the former also includes a strong morphological component (Kitajima & Poorter, 2010), and distinguishing between these effects via morphological and chemical trait diversity yielded straightforward results). Calculations of Rao's Q were based on standardized trait values (mean = 0, SD = 1) and a Euclidean species distance matrix. For the multivariate measures of Rao's Q based on the three morphological and four chemical traits, we used all axes of a PCA (as these axes are orthogonal to each other) on the standardized traits for the distance matrix to avoid collinearity effects (Böhnke et al., 2013; Purschke et al., 2013). For the phylogenetic data, we correspondingly calculated Rao's Q from a phylogenetic cophenetic distance matrix (Qphylo). All measures of functional and phylogenetic diversity were weighted by plot-level abundance data to account for the relative impact of dominant vs rare species on community-level metrics.
In each plot, and for each of the 10 focal species, we further calculated a species-specific phylogenetic distance measure (Qspecphylo), based on the mean phylogenetic distance between an individual of a given focal species and all other woody plant individuals in a given study plot (Webb et al., 2002, 2006) – for consistency, we again expressed this measure as Rao's Q, which, in the abundance-weighted case, is analogous to the MPD (mean phylogenetic distance) used in other studies (Vellend et al., 2011). Recent studies have shown that not only the overall phylogenetic diversity, but, in particular, the phylogenetic distance of a focal individual to all other individuals in a community, can determine herbivore effects (Webb et al., 2006; Paine et al., 2012; Parker et al., 2012). The species-specific measure of Rao's Q was also calculated for trait data, and we included both multivariate relatedness measures for our focal species based on morphological (Qspecmorph) and chemical (Qspecchem) leaf traits and measures for each individual trait (QspecT, where T is the respective trait) in our analysis. Species-specific indices were calculated from the same distance matrices as used for the calculation of plot-level Rao's Q, but by contrasting individuals of the respective focal species to all other individuals in each of the communities. Again, all measures were weighted by plot-level abundance data.
We used generalized linear mixed models with a binomial error structure (as a recommended way to analyze proportion data; Zuur et al., 2009), fitted by Laplace approximation (Bolker et al., 2009), to analyze the effects of functional and phylogenetic diversity metrics on the degree of herbivore damage of the 10 study species across the 27 study plots, whilst accounting for the effects of woody plant species richness and general plot characteristics. To determine which functional and phylogenetic characteristics particularly affect herbivory, and to assess whether their effects were complementary to simple species richness effects and independent of plot characteristics, we constructed five sets of models. These contained: (1) all predictors; (2) PC1abio and all functional metrics (functional diversity sensu Diaz et al., 2007); (3) PC1abio and phylogenetic metrics; (4) PC1abio and woody plant species richness; and (5) only PC1abio. PC1abio was included in all model sets to account for potentially confounding plot characteristics. Species identity, with individuals nested within species, and plot identity were considered as crossed random effects. The use of species identity as a random factor accounts for all interspecific differences in the levels of herbivory, leaving individual-level differences as the only source of variation. We also included a random factor with the total number of observations as factor levels to account for potential overdispersion in the data (Bates et al., 2013). Before the analysis, predictors were checked for collinearity and, where there was strong correlation (> 0.7) among predictors, we excluded those that were less strongly related to herbivory (e.g. CWMC : N and CWMC : P, which were strongly correlated with CWMPhenol, but less strongly correlated with herbivory than CWMPhenol, and several correlated species-specific Qspec measures; see Supporting Information Table S1 for a correlation matrix and a list of excluded variables). The final set of predictors included the general plot characteristics PC1abio, woody plant species richness, the phylogenetic diversity measure Qphylo, the multivariate chemical trait diversity Qchem, the single-trait dispersion variables QLDMC, QC, QC : N, QPhenol, the CWM values CWMLA, CWMLDMC, CWMC, CWMPhenol, and the species-specific measures Qspecphylo, QspecLA, QspecLDMC, QspecC, QspecC : N, QspecC : P and QspecPhenol. We also included the interaction between woody plant species richness and overall phylogenetic diversity Qphylo, as this has been shown recently to influence species richness effects in grasslands (Dinnage, 2013). All predictors were standardized to a mean of zero and a standard deviation of unity before the analysis. Each model set was simplified by sequential deletion of predictors based on the reduction in the corrected Akaike information criterion (AICc) values to obtain the most parsimonious, minimal adequate model (which may potentially also contain variables that are not statistically significant at P < 0.05 if deletion of these variables would have markedly decreased the AICc fit; see Burnham & Anderson, 2004). The five resulting minimal adequate models were compared on the basis of their AICc values (ΔAICc) and AICc weights, with particularly low AICc values and high AICc weights indicating the best model fit (Burnham & Anderson, 2004). Model residuals were checked to comply with modeling assumptions. All analyses were performed with R 3.0.0 (http://www.R-project.org) and the package lme4 (Bates et al., 2013).