2011). Once simplified, the selected environmental variables from each marginal model were then incorporated into a final conditional pCCA model for each habitat. Hence, we initially simplified models using a ‘bottom-up’ approach, before constructing the final ‘top-down’ model. Owing to differences across habitats and taxa in the set of explanatory environmental variables, we followed Okland (1999) and focussed on the explainable variability only of our models, decomposing total variance explained into proportions of variance explained by the main sets and subsets of environmental variables. Thus, the relative, as opposed to absolute, impact of sets of environmental variables Verteporfin manufacturer could be compared across habitats and taxa. For each taxon (Bombus and non-Bombus) and habitat, the environmental variables selected in the pCCA models were examined in more detail using randomisation tests; qualitative factors were assessed using the mean difference in species richness and abundance between sites with and without the factor, quantitative factors were tested using Spearman’s rank correlation. Pairwise similarity in community composition was quantified using the Chao-Sørenson abundance-based similarity index (Chao et al. 2005). Qualitative factors were tested using anova, comparing pairs of sites lacking the factor against pairs of sites where one site contained the factor; quantitative variables were assessed using Mantel tests of pairwise matrices of community similarity and Bray–Curtis similarity of the factor across sites. Two-tailed statistical significance was assessed using 999 Monte Carlo permutations as implemented in PopTools v3.2.5 (Hood 2010). To reduce the number of explanatory environmental variables and resultant multicollinearity within the pCCA models, variables were correlated using the phi coefficient and Pearson product-moment coefficient, respectively, and all but one of the significantly correlated variables removed. All P-values were adjusted for repeated testing using Storey & Tibshirani’s (2003) R package ‘Q-value’ v1.0. Where necessary, logarithmic or square-root transformation of environmental variables was used to achieve normality and proportional data were arcsin-square-root transformed. Counts of bees were pooled for the entire year per site, and the data transformed using log10(x + 1). To exclude casual occurrences and enhance the detection of relationships between community composition and environmental factors, species occurring in <5% of sites were excluded from pCCAs (McCune & Grace 2002). A total of 4608 individual bees (4327 from transects; 281 from pan traps) representing 54 species were recorded from the 40 sites. The most abundant taxon was the Bombini, representing 81% of our data, followed by the Colletidae (11%) and the Halictidae (5%).