Statistical Analysis
We used mixed-effect models with appropriate error structures (see
below) to test the impact of drought on survival, growth and
photosynthesis. Preliminary models for growth showed that trends did not
vary between growth rate (Appendix Table A4) or relative growth rate
(Appendix Table A7) and we used the latter for interpretation since
relative growth rates were standardized by initial height.
For the first question concerning differences in species response to
drought in relation to their affinity for less vs. more seasonal areas,
we modeled species’ performance as a function of seasonality index,
drought treatment, and their interaction. For question 2, we tested the
role of each trait in governing drought response. We implemented
separate models per trait, where species’ performance was modeled as a
function of their trait values, drought treatment, and drought-trait
interaction. In all models, we used species-level trait means. Species
ID and block ID were included as random intercepts. Survival was modeled
using a Bernoulli error structure and Gaussian errors were assumed for
growth and photosynthesis. Because species differed in the extent of
variation in their response, we added a weighting term to separately
model the variance per species. Trait models for SRL were implemented
with and without the species Actinodaphne malabarica , which had
small seedlings with very light roots and high SRL values that were an
outlier compared to other species.
Model structure in Wilkinson-Rogers notation:
Performance variable ~ trait + treatment +
trait:treatment + (1|species) + (1|block)),
weights=varIdent(form=~1|species)
Finally, to understand the impact of the overall trait-based phenotype
on performance during drought (question 3), we used a two-step approach.
First, we performed a pairwise correlation among traits (Appendix Figure
A5) and a principal component analysis (PCA) on traits to obtain
composite phenotypes defined by trait combinations. Then, we modeled
individual survival, growth, and photosynthesis in control vs. drought
plants using GLMM with species-level intercepts and slopes. From these,
we extracted each species’ baseline performance (survival, growth,
photosynthesis) in well-watered conditions and the change in performance
with drought and conducted a PCA on these six variables to get species’
composite response to drought in relation to their baseline differences
in growth, survival and photosynthesis. For both PCAs, we performed
varimax rotation to simplify loadings for each factor. In the second
step, to explore the correspondence between traits and performance, we
performed Procrustean superimposition on the two PCA, using the axes
that explained >75% of the variation in each PCA.
Procrustes analysis tests the strength of association between two
ordinations for a common object (species in our case) with each other
(Peres-neto & Jackson 2001, Rüger et al. 2018), i.e., how well
species positions in trait space matched their positions in performance
space. Significance of the Procrustes correlation was assessed using
permutation tests with 10,000 random iterations.
All analyses were performed using R v 3.4. We used packages nlme
(Pinheiro et al. 2022) for mixed effects models with Gaussian
error structure lme4 (Bates et al. 2015) for models with
Bernoulli error, FactoMineR (Lê et al. 2008) for PCA and vegan
(Oksanen et al. 2020) for Procrustes analysis.