The Tangled Web of Community Ecology: Making Sense of Complex Data

Ecological communities are governed by complicated processes that give rise to observable patterns. Making sense of these patterns, much less inferring the underlying processes, has proved challenging for several reasons. Manipulative experiments in natural communities may not be feasible due to large numbers of variables, lack of adequate replication, or the risk of undesirable consequences (e.g., introducing an invasive species). The multivariate nature of ecological datasets presents analytical problems as well; many statistical techniques familiar to ecologists have difficulty handling large numbers of potentially collinear variables. I present results from three studies of spider communities in which I employ a combination of familiar and less familiar statistical approaches to elucidate the factors influencing community structure in spiders. These approaches include null model analyses, nonmetric multidimensional scaling (NMS) for variable reduction of predictor and response data matrices, multiple regression, and observed variable structural equation modeling (SEM). While NMS has been employed as a multivariate descriptive analysis, examples of its use in further analyses are rare. SEM is a technique widely applied in other fields, but has only recently been used in ecological studies. General results from analyses of these three studies suggest that: 1) significant patterns of spider species co-occurrence based on null model analyses are consistent with a hypothesis of shared habitat preferences rather than one of species interactions, 2) in multiple regressions using NMS axes as predictor and response variables to compare the roles of plant species composition and habitat architecture in influencing spider species composition, the plants explained as much or more variation as the architecture, and 3) based on SEM analyses using NMS axes for spider species, plant species, arthropod orders and habitat architecture as variables, plant species composition acts both indirectly (through its effect on arthropods and architecture) and directly. The combination in these analyses of a traditionally descriptive multivariate approach (NMS) with null models, a classic regression approach, and SEM permits the analysis of otherwise statistically intractable datasets (the original data matrices). This suite of approaches provides new insights into spider community structure, and can be applied by ecologists working in other systems as well.