Statistical analyses
Contrast between urban and rural soil, plant traits and AMF
variables: To test for differences between urban and rural environments
in soil and plant traits (question 1), and AMF variables
(AMF-colonization rates, spore density, and diversity; question 2), we
ran independent one-way ANOVAs to test for the effect of the environment
factor (three levels: RS, DUS, OUS). Variables such as soil properties
(e.g. N concentration, pH), spore density, richness, and diversity were
pooled at the site level; spore diversity was calculated using the
Shannon-Wiener diversity index. On the other hand, because replication
for variables such as plant morphometrics and AMF-colonization rates
were at the plant level, we implemented the lmer function from
package lmer4 in R (Bates et al. , 2015) to perform linear
mixed models considering environment and site as fixed and
random factors, respectively. The significance of the fixed effect was
evaluated using a Type II Wald’s test running the Anova function
from the car package in R (Fox & Weisberg, 2019), assessing the
significance of the random effect with a likelihood-ratio test (LRT). To
discard potential spatial autocorrelation, Mantel’s tests were used to
assess the correlation between geographic distance and Euclidean
distance for each variable recorded using the package ade4 in R
(Chessel et al. , 2004); however, in none of the considered
dependent variables spatial autocorrelation was detected. To test for
changes in community composition based on spore morphospecies between
environments, we conducted a permutational analysis of variance
(PERMANOVA) with 999 permutations using the adonis function from
the vegan package in R (Oksanen et al. , 2020). Finally, a
Principal Component Analysis (PCA; prcomp , R), based on a
correlation matrix, was used to test for differences in the
multidimensional variation of soil properties and plant attributes
between each environment (question 1). A one-way ANOVA tested theenvironment (RS, DUS, OUS) effect on PC1. This PC1 summarized the
variation on soil nutrient concentration (N, P, K, Na, and Ca) and pH.
Before PCA, we equalized P and Ca variances by dividing raw values by
100.
Soil and AMF associations: To explore the soil properties that
may predict AMF-colonization rates on roots of R. nudiflora(question 3), we performed pairwise correlations between
AMF-colonization rate and soil attributes at the global level (i.e.
using data from all environments) as well as for each environment
separately. As a next step, we used the Z-scores of PC1 (see previous
section) to explore the association between PC1 and AMF-colonization
rate (using pooled data). However, because PCA visual inspection and
previous statistical analyses on the first principal component (PC1; see
previous section) indicated strong differences in Z-scores between urban
and rural environments we ran an independent linear model (i.e.
AMF-colonization rate ~ PC1) within each environment.
Structural equation modeling: Because analyses based on principal
components obscure the relationship between soil attributes mediating
the levels of AMF-colonization rate, we implemented Structural Equation
Modelling (SEM) using lavaan package in R (Rossel, 2012), and
tested for contrasts between urban (DUS, OUS) and rural (RS)
environments using a multigroup test. We used SEM approach to
disentangle the direct and indirect effects that soil nutrients and pH
have on AMF-colonization rate (question 3). We designed an initial
causal model (Fig. S2) based on previous knowledge of soil properties
and nutrient interactions (Fig. S2). We considered Ca as one of the main
drivers of pH, which in turn can affect K, P, and N availability (Osman,
2013). However, the two later macronutrients (P and N) have been
reported to negatively affect K availability (Osman, 2013; but see Dibb
& Thompson, 1985). In turn, it is expected that N and P will have
important and negative effects on the AMF-colonization rate (Salvioli di
Fossalunga & Novero, 2019; Chen et al. , 2020). Finally, Na can
also increase the benefit of the association with AMF in the context of
salinity stress (Evelin et al. , 2009), and, in soils with low K
availability, AMF can play an important role to improve plant K
nutrition (Garcia & Zimmermann, 2014). The goodness-of-fit of the
initial causal and alternative models (Fig. S2) was assessed with an LRT
to test the null hypothesis that the predicted covariance matrix of the
model is not different from the observed (Iriondo et al. , 2003).
A significant χ2 indicates that the model does
not fit the observed covariance matrix. Because a good-fit model may
result from an inadequate statistical power (Mitchell, 1992), we
calculated additional indices of goodness-of-fit, such as the
Comparative Fit Index (CFI; cut-off good fit ≥ 0.95), the Root Square
Mean Error of Approximation (RMSEA; cut-off good fit ≤ 0.06), and the
Standardized Root Mean Square Residual (SRMR; cut-off good fit ≤ 0.08),
which are insensitive to sample size (Hooper et al., 2008). All
these indices were estimated using the fitMeasures function in
the lavaan package in R (Rossel, 2012), and were used in the
process of model selection.
Once we obtained a fitted and general causal model, we ran a multigroup
test to assess whether path coefficients contrast between DUS, OUS, and
RS environments. The multigroup test imposes cross-group constraints on
the path model regression coefficients, and then compared with an LRT,
the constrained and unconstrained models using the lavTestLRTfunction (lavaan package in R; Rossel, 2012). In particular, we
used the function lavTestScore, a Lagrange Multiplier Test (LMT),
as a guide to identify a set of constrained path coefficients that, if
released, would result in a significantly better model (i.e. a lower
goodness-of-fit χ2 ; Bentler, 1989).
All statistical analyses were performed using R version 4.0.3 (R Core
Team, 2020). Furthermore, all analyses, were evaluated with transformed
variables when needed to meet the normality assumption.