2.3.2 Distribution of ant assemblages along vertical and horizontal gradients
In order to understand how ant assemblages changed vertically and horizontally, we first applied linear regression models using the abundance (ln( x+1) transformed) and richness of ants in each sampling point as response variables, x , with vertical height as continuous explanatory variable and horizontal position of vertical transects as categorical explanatory variable. We then used both abundance-based (ln (x+1) transformed) assemblage data and species presence/absence data for the following assemblage composition analyses. Two sampling points (comprising one ant individual each) from one emergent layer were excluded from the analyses due to the lack of other replicates at this height level, resulting in total of 61 ant assemblages as our sample size. We generated non-metric multi-dimensional scaling (NMDS) plots using Bray-Curtis distance index for the abundance-based community and Jaccard distance index for the presence/absence data of ant assemblages at each sampling point grouped by vertical level and transect. To keep sample size same for analyses in vertical and horizontal directions, we combined the 12 vertical levels into six vertical groups for NMDS analyses. We tested for statistical significance of these groupings by analyses of between-point assemblage presence/absence variances using the adonis function in the vegan package (999 permutations) (Oksanen et.al. 2013, R Core Team, 2013). We then calculated the pairwise assemblage dissimilarity across sampling points and further partitioned it into assemblage turnover (balanced variation in abundance) and assemblage nestedness (abundance gradient) between each pair of samples to understand which component drives beta diversity patterns at different horizontal and vertical distances (Si et al. 2017). This will allow us to distinguish the potential effects of resource limitation, which is expected to lead simply to richness gradients but no turnover pattern, as opposed to environmental filtering, which is predicted to lead to species turnover as low canopy species are replaced by high canopy ones. The analyses were calculated using the beta.pair.abund function (for abundance-based dissimilarity) and beta.pair function (For presence-absence data) in the betapart package (Baselga and Orme 2012). For ant assemblage composition analysis, as results from measures using abundance-based pairwise Bray-Curtis dissimilarity and presence-absense based Jaccard species dissimilarity were similar, we only present results using abundance-based pairwise Bray-Curtis dissimilarity in the main text (see online supplementary materials for presence-absence analyses).
We used four different methods to quantify the distance between sampling points: 1) straight line distance between points, 2) maximum surface travel distance, 3) horizontal distance between points, and 4) vertical distance between points. To simplify distance calculations, we assumed that all sampling points were on a straight line from above when calculating the distance between points, although there were some small deviations from this. Straight line distance modelled the case where there was vegetation structure to directly connect the two sampling points for ant movement (Fig. S1). Maximum surface travel distance was calculated as the horizontal distance between the sampling points added to the vertical height of each sampling point above the ground, modelling the scenario where the ground provides the only horizontal connectivity between two points (Fig. S1). We conducted Multiple Regression on Distance Matrices (MRM) using the MRM function in the ecodist package (Goslee and Urban 2007) to test for relationships between different distance measures and total pairwise dissimilarity, assemblage turnover and nestedness. The tests of significance of MRM coefficients were performed using 1000 permutations.
We then tested the effects of distance on assemblage dissimilarity between and within vertical transects and vertical strata. We first grouped the assemblages within the same vertical transect/vertical stratum, and then tested effects of horizontal/vertical distance on pairwise assemblage dissimilarity between transects/strata by MRM. To measure within vertical transect/vertical stratum patterns, we conducted post-hoc MRM tests on vertical/horizontal distance effects of pairwise assemblage dissimilarity across sampling points within each transect and stratum. We then conducted linear regression to compare the mean assemblage dissimilarity value and linear effects of distance on pairwise dissimilarity across two dimensions by comparing the intercept and the slope from the MRM analyses of each transect/stratum. To test for the different horizontal turnover patterns between vertical strata expected if ant mosaics effect is present, we conducted linear regression with horizontal pairwise dissimilarity within each stratum as the response variable, horizontal distance and vertical height as well as their interaction as explanatory variables.
Correlates of ant assemblage distribution patternsWe conducted constrained ordinations to further understand how microclimate (air temperature relative humidity, and PPFD) and microhabitat structure (total leaf area) affect ant assemblage composition. We first conducted Detrended Correspondence Analysis (DCA) using decorana function in vegan package and found maximum axis length greater than 4 (range: 3.91-6.96) indicating that Canonical Correspondence Analysis (CCA) constrained ordinations that assume unimodal responses of species to environmental gradients were appropriate. We then first checked the collinearity among all explanatory variables using vif.cca function to reduce redundancy in the model. Among all microclimate and microhabitat factors, no strong collinearity (VIF<10) was been detected and so all predictors were included in the Canonical Correspondence Analysis (CCA). We then conducted CCA ordinations using ccafunction in the vegan package with all explanatory variables (air temperature, relative humidity, and PPFD). We conducted backward model selection using ordistep function in the vegan package to identify the most significant variables in affecting the assemblage composition based on permutation tests using 1000 permutations (Blanchet et al. 2008). As the model selection process for CCA analysis requires samples with all environmental factors available, only 48 out of total 61 ant assemblages that had all environmental information available were included for this analysis.
Results