Statistical analysis
The Kolmogorov–Smirnov test was performed to determine whether all data were normal. Non-normal variables were log-transformed before the analyses. The significant differences of Rm among different biome zones and forest types were tested by the least significant difference (LSD) multiple-comparison post hoc test after the one-way analysis of variance. Spearman’s correlation analyses were performed to assess the relationships between important variables of soil properties and Rm. The above statistical analyses were performed using the SPSS software (version 19.0; SPSS Inc., Chicago, Illinois, USA), and significant differences were accepted at the p < 0.05 level of probability. Variation partitioning modeling was performed to assess the relative importance of three groups, namely, climate, soil physicochemical properties and microbial properties, in driving continental variations in soil Rm. We used the “forward.sel” function to avoid redundancy and multicollinearity in variation partitioning analysis that was conducted using the “Vegan” package.
First, we used “selected.forward” procedure to select most important predictors of Rm and then included these variables in structural equation modelling (SEM). SEM was performed to determine the pathways underlying the observed effects of environmental predictors on microbial respiration. SEM was conducted using “piecewiseSEM”, “nlme” and “lme4” packages. The piecewiseSEM could also account for random effects of sampling sites (to account for having more than one sample per site), by providing ”marginal” and ”conditional” contribution of environmental predictors in driving microbial diversity. We used the Fisher’s C test (when 0 ≤ Fisher’s C/df ≤ 2 and 0.05 <p ≤ 1.00) to confirm the goodness of the modelling results. We then modified our models according to the significance (p< 0.05) and the goodness of the model. We included the same predictors in SEM for Rm at different scales, which could disentangle the best predictors for Rm according to their differences in total standardized effects. We further conducted boosted regression tree analysis to quantify the relative importance of the moderator variables in regulating the spatial variation in soil Rm. A gaussian error structure was assumed during a 10-fold cross-validation to estimate the optimal number of trees. Tree complexity was set to 4, whilst the learning rate was kept at 0.01 and bagging fraction at 0.6. The BRT model was performed using the R package “gbm” (Ridgeway, 2013) combined with the “dismo” package in R 3.6.4.