Statistical analysis
MAT and mean annual precipitation (MAP) of each plot were extracted from the relevant latitude and longitude of the global climate layers of WorldClim (1 km2 spatial resolution; http://www.worldclim.org/) using the extract function in the “raster” R package (v. 2.6.7), and the aspect of each plot was obtained by analyzing the digital elevation model data using the “raster” R package (Hijmans, 2020). We calculated the above-ground biomass (AGB) using a pantropical model (Chave et al. , 2014) in the “BIOMASS” R package (v. 2.1.1) (Réjou-Méchain et al. , 2017).
In this study, we totally collected 54 abiotic and biotic factors relative to the MR and Q 10. Regression analyses were used to examine the elevational trends of all response variables. Then, the 54 factors were clarified into seven groups: topography (elevation, aspect) and climate (MAT and MAP), soil environment (WHC, pH), soil texture (bulk density, sand content, silt content and clay content), plant community structure (AGB, species richness, Shannon diversity index (H’), Simpson index), plant carbon input (litter C, N, P concentration and C:N ratio (litter CNR), C:P ratio (litter CPR), N:P ratio (litter NPR), fine root biomass, fine root C, N, P concentration and C:N ratio (root CNR), C:P ratio (root CPR), N:P ratio (root NPR)), soil organic matter (soil C, N, P, C:N ratio (soil CNR), C:P ratio (soil CPR) and N:P ratio (soil NPR), soil available N and P), soil microbial biomass (soil microbial biomass C, N, P, C:N ratio (microbial CNR), C:P ratio (microbial CPR) and N:P ratio (microbial NPR)), soil microbial community structure (total phospholipid fatty acids, and its components of bacteria, gram-positive bacteria, gram-negative bacteria, actinomycetes, fungi, and gram-positive: negative bacteria ratio (GNR), actinomycetes: bacteria ratio (ABR), actinomycetes: fungi ratio (AFR), fungi: bacteria ratio (FBR)).
For each group of factors, we performed all subsets regression analysis to select the best model that had the lowest Bayesian information criterion (BIC) in predicting AccMR_MAT and Q 10, respectively (Table S1 and S2). If the difference of BIC was < 2 units (Burnham and Anderson, 2002), we obtained the model with the highest adjust R 2. Using this approach, we selected 14 and 12 variables from the best models for AccMR_MAT andQ 10, respectively. The unselected variables either had no significant influence on AccMR_MAT orQ 10, or were highly collinear with the selected variables. Then, the selected variables were used in structural equation modelling (SEM) to explain the variation of AccMR_MAT andQ 10 along the elevation. We dropped the non-significant path and variables in the SEM to simplify the initial model and improve the model fit. The indirect effect of each predictor was calculated by multiplying the standardized direct effects of a given predictor on AccMR_MAT or Q 10 via mediator in one route, and then we summed the multiple indirect effects and direct effect of the given predictor to quantify its total effect (Lefcheck, 2016). All the analyses were conducted in R 3.3.4. We used packages corrplot (Wei and Simko, 2013), leaps, piecewiseSEM (Lefcheck, 2016).