Data analysis
We compiled a dataset with species-mean trait values, as
well as an individual-level dataset, to examine the robustness
of our analysis to the inclusion of intraspecific variation.
In the species-level dataset, we standardized the data to
correct for the effects of local environment and ontogenetic stage on trait phenotypes. To do so, we used two measures
of individual tree stature: diameter at breast height and
overall height (measured with a laser rangefinder), and two
measures of crown light exposure (CE) (Poorter & Arets
2003). We estimated CE indices, which provide an ordinal
estimate of the local light environment, separately for the
entire individual and for the collected twig and leaf sample.
Thus, we had four intercorrelated measures of individual
stature, which we collapsed into a single measure using the
Non-linear Iterative Partial Least Squares algorithm, as
implemented in the ade4 package of R (Dray & Dufour
2007). Only two of the 16 traits, d13C and bark thickness,
varied significantly with this multivariate factor, and so we
corrected for these correlations by substituting the residuals
from linear regressions of these variables against individual
stature.
Although leaf traits were measured on every individual,
wood traits and chemistry were not. We therefore estimated
unobserved trait values using Multiple Imputation with
Chained Equations (MICE), as implemented in the mice
package of R (van Buuren & Groothuis-Oudshoorn,
Unpublished). Missing values constituted 28.6 and 13.5%
of the individual- and species-level datasets, respectively.
Unobserved values were estimated through predictive mean
matching using all other data as predictors, rather than
assigning column mean values as is done under other
imputation procedures (e.g., Wright et al. 2004). The robustness
of the data imputation procedure was evaluated by
assessing the convergence of the Gibbs sampler at the heart
of MICE by plotting the means and standard deviations of
five imputations of data. No trends were observed in the
mean or variance of the imputed data over the course of
1000 iterations. We are therefore confident in the robustness
of the data resulting from the imputation procedure.
To test the hypothesis that the spectrum of stem traits
is orthogonal to the spectrum of leaf traits, we used
multiple factor analysis (MFA), a multivariate ordination
method that permits examination of common structures
in datasets with many variables that can be separated into
different groups of variables (Escofier & Page`s 1990).
MFA involves two steps. First, a principal component
analysis (PCA) is performed on each group of variables
which is then !normalized" by dividing all its elements by
the square root of the first eigenvalue obtained from the
PCA. In our dataset, the groups were defined as in
Table 1. Second, the normalized datasets are merged to
form a unique matrix and a global PCA is performed on
this matrix. The individual datasets are then projected
onto the global analysis. In this way, variables in each
group are permitted to maintain free covariances amongst
themselves, and the relationships between groups of
variables can be examined without the influence of
within-group covariance. We use as a test statistic the
between group correlation coefficient, RV, which is scaled
from 0 if every variable in one group is completely
uncorrelated with every variable in the other group(s), to
1 if the two groups are completely homothetic. Under the
hypothesis of orthogonality of leaf and stem traits
economics spectra, the RV coefficient of a MFA
performed on groups as defined in Table 1 should be
smaller than the RV of a MFA performed on randomly
generated groupings of the same data. We created a
sampling distribution for our test statistic using 1000
permutations of variable assignments to two groups, and used a one-tailed test with a = 0.05 to test for
orthogonality between leaf and stem trait groups. To
examine the placement of the leaf area variables on the
global ordination, we projected them afterwards on the
global analysis. We conducted the same MFA analyses
and permutation tests for individual and species-level
datasets, both with and without data imputations, to
verify the robustness of our results to within-species
variation.
To test for any sampling bias in our cross-species
comparisons related to their evolutionary history, we also
performed species-level analysis using phylogenetically
independent contrasts (PICs). We recovered a phylogenetic
hypothesis for our 668 species using the PhyloMatic utility
(Webb & Donoghue 2005), based on the Davies et al. (2004)
phylogenetic hypothesis for relationships amongst angiosperm
families, with polytomies applied within most families
and genera. PICs were calculated in the ape module of R
(Paradis et al. 2004), as the difference in mean trait values for
pairs of sister species and nodes (n = 667). For this analysis,
branch lengths were scaled to 1.
To determine the nature of relationships both within and
amongst functional traits defining leaf and stem trait groups,
we examined pairwise correlations amongst all variables.
We conducted this analysis for the species-level dataset both
with and without PICs.
For all analyses, leaf toughness, surface area, leaf tissue
density and SLA were log-transformed to more closely meet
the assumption of normality. Other traits were approximately
normally distributed without transformation.
All analyses were conducted in the R 2.10 statistical
platform (R Development Core Team 2010).