Understanding community structure: a data-driven multivariate approach.

Habitat is known to influence community structure yet, because these effects are complex, elucidating these relationships has proven difficult. Multiple aspects of vegetation architecture or plant species composition, for example, may simultaneously affect animal communities and their constituent species. Many traditional statistical approaches (e.g., regression) have difficulty in handling large numbers of collinear variables. On the other hand, multivariate methods, such as ordination, are well suited to handle these large datasets, but they have primarily been used in ecology as descriptive techniques, and less frequently as a data reduction tool for predictor variables in regression. Here, I employ a multivariate approach for variable reduction of both the predictor and response variables to investigate the influences of vegetation architecture and plant species on community composition in spiders using multiple regression. This allows retention of the information in the original dataset while producing statistically tractable variables for use in further analyses. I used nonmetric multidimensional scaling to reduce the number of variables for predictor (habitat architecture and plant species) and response (spider species) data matrices, and used these new variables in multiple regression analyses. These axes can be interpreted based on their correlations with the original variables, allowing for recovery of biologically meaningful information from regressions. Consequently, the important variables are determined by the data themselves, rather than by a priori assumptions of the researcher. Contrary to expectations based on previous work in spiders and other animals, plant species composition explained more variation in spider communities than did habitat architecture, and was also a stronger predictor of other community structure variables (overall abundance, species richness, and species diversity). I discuss possible ecological explanations for these results, and the advantages of the proposed method.