Response and sensitivity of urban plants with different seed dispersal modes

To check whether the responses of spontaneous plant richness to driving factors differ among dispersal modes, generalized linear mixed effect models (GLMMs) were used to analyze the driving factors of the richness of species with the four seed dispersal modes: anemochory, autochory, hydrochory and zoochory. We fitted a model with species richness as the response variable, using a Poisson distribution and “glmer” function within R package lme4 (version 1.1-35.1) with patch ID nested within site ID, and site ID nested within city ID as random effect (1|patch ID/site ID/city ID). As fixed effects, we used all natural factors, dispersal limitation, and habitat quality described above, as well as the interaction between these factors and seed dispersal modes (hereafter called ‘full model’). To start, the independent variables were standardized and centered to ensure uniform dimensionality. To avoid collinearity, if the correlation coefficient between independent variables was greater than 0.7, only the variable of the highest correlation with the dependent variables was chosen for further modelling. After model construction, the variance inflation factor (VIF) of the model was checked, and variables with VIF greater than 5 were removed from the model sequentially until all variables had a VIF ≤ 5. Residual diagnostics of models was then checked using the “simulateResiduals” function in R package DHARMa (version 0.4.6). The “Anova” function in the R package car (version 3.1-2) was employed to assess both the main effects and interaction effects in the full model. To assess differences in the response to predictor variables between dispersal modes, we used the “emtrends” function in the R package emmeans (version 1.10.1) to compare trend estimates of for the different dispersal modes. To further explore the driving factors within a dispersal mode, we separately ran the same GLMM analysis described above for each group. The “dredge” function in the R package MuMIn (version 1.47.5) was used on each separate model, which created a suite of models with all possible combinations of the initial variables and sorted them according to the Akaike Information Criterion (AIC). We then conducted model averaging on all models with ΔAIC &lt; 2 using the function “model.avg”. We used the “r.squaredGLMM” function in the R package MuMIn to calculate R2, then manually performed variance decomposition to assess the relative contribution of each driving factor and the random effects. We ran all analyses using the software R 4.3.2. <br>