We assessed how the aboveground, belowground, and overall biomass
differed between treatments, splitting the dataset into observations
from each phytometer species. We ran a mixed-effects model (GLMM)
relating biomass (transformed to the log10 scale) as a
function of the conditioning species, the sterilization status of the
soil, and the interaction between the two. The pot ID number of the
conditioned training soil was used as a random intercept with a fixed
mean. Conditioned soils came from individual pots in the training stage
that may differ in their abiotic and biotic features, so we chose to use
mixed-effects models to account for the variance in the strength of
feedback due to these differences. If the random effect was not
significant (i.e. individual pots from the training stage did not differ
in their effect on the feedback), we ran the same formula as a
generalized linear model (GLM). For GLMMs or GLMs of aboveground,
belowground, and overall biomass data, the most parsimonious model was
selected through comparison of AIC between full and reduced models. The
type of model, whether an interaction term was used, and the
R2 value for each model is indicated (Table 1). To
determine if any of the simple main effects were significant, we ran the
same formula as an ANOVA using the linear model to calculate degrees of
freedom and sum of squares error. We were particularly interested in
comparing the effects of live and sterilized eastern redcedar soil to
live and sterile home soils for each phytometer species. To elucidate
this relationship for each phytometer species, we performed post
hoc pairwise comparisons to obtain the estimated marginal means (also
called least-squares means) using the emmeans package (Russell
2021).
We visualized differences in phytometer biomass between live and sterile
home and redcedar soils using effects plots that were derived from the
linear model fit for each set of contrasts (Ho et al. 2019;
Wilschut and van Kleunen 2021). These plots illustrate simple mean
differences between contrasts of interest with 95% confidence intervals
using the sample data. The second part of these plots shows the modeled
means and 95% confidence intervals paired with raw data points (Figure
2).