Variance partitioning methods for inferring the processes driving metacommunity structure: Experimental tests of effectiveness
Variance partitioning methods comprise a family of statistical approaches for separating variation in community structure among sites into components explained by environmental variables, spatial distance, both, or neither. It is widely believed that these approaches provide a straightforward, universally applicable, "off the shelf" way of inferring the processes generating metacommunity structure from observational data. For instance, if environmental variables explain most among-site variation in community structure, we infer a "species sorting" metacommunity in which local environmental conditions dictate local community structure, with processes like dispersal limitation, source-sink dynamics, and neutral drift being unimportant. However, variance partitioning typically fails to recover process from pattern when applied to simulated data generated by known processes (Smith & Lundholm 2010, Gilbert & Bennett 2010). As a complement to simulation-based tests of variance partitioning, I used protist microcosms as a model system to create metacommunities of real organisms, structured by known processes. By varying dispersal rate, microcosm size, the spatial arrangement of microcosms with different environmental conditions, and the initial distribution of competing ciliate protists among microcosms, I created metacommunities structured by different combinations of processes. I then asked whether these processes could be inferred via variance partitioning.
Species composition and abundances varied among treatments, among microcosms within treatments, among replicates, and over time. The absolute and relative magnitudes of the variance components varied widely among treatments, replicates, and over time, with the unexplained and joint spatial-environmental components generally being the largest (>30% of total variation in many cases). However, the relative magnitudes of the four variance components did not vary systematically with dispersal rate, microcosm size, and initial conditions. That is, variance partitioning methods failed to recover the processes that structured the metacommunities. This failure is striking because in many respects microcosm data comprise a best-case scenario for variance partitioning approaches. Sampling biases and errors are modest or absent, the relevant environmental variables are known, dispersal rates are experimentally-imposed and the same for all species, etc. These results may help explain why a recent meta-analysis of variance partitioning studies (Soininen 2014) found that results vary idiosyncratically among studies. The results highlight the importance of validating proposed methods for inferring process from pattern with simulated and real data generated by known processes.