The methods of direct gradient analysis and variation partitioning are the most widely used frameworks to evaluate the contributions of species sorting to metacommunity structure. In many cases, however, species are also driven by spatial processes that are independent of environmental heterogeneity (e.g., neutral dynamics). As such, spatial autocorrelation can occur independently in both species (due to limited dispersal) and the environmental data, leading to spurious correlations between species distributions and the spatialized (i.e., spatially autocorrelated) environment. In these cases, the method of variation partitioning may present high Type I error rates (i.e., reject the null hypothesis more often than the pre-established critical level) and inflated estimates regarding the environmental component that is used to estimate the importance of species sorting. In this paper, we (1) demonstrate that metacommunities driven by neutral dynamics (via limited dispersal) alone or in combination with species sorting leads to inflated estimates and Type I error rates when testing for the importance of species sorting; and (2) propose a general and flexible new variation partitioning procedure to adjust for spurious contributions due to spatial autocorrelation from the environmental fraction. We used simulated metacommunity data driven by pure neutral, pure species sorting, and mixed (i.e., neutral + species sorting dynamics) processes to evaluate the performances of our new methodological framework. We also demonstrate the utility of the proposed framework with an empirical plant dataset in which we show that half of the variation initially due to the environment by the standard variation partitioning framework was due to spurious correlations.

Mantel tests are widely used in ecology to assess the significance of the relationship between two distance matrices computed between pairs of samples. However, recent studies demonstrated that the presence of spatial autocorrelation in both distance matrices induced inflations of parameter estimates and type I error rates. These results also hold for partial Mantel test which is supposed to control for the spatial structures. To address the issue of spatial autocorrelation in testing the Mantel statistic, we developed a new procedure based on spatially constrained randomizations using Moran spectral randomization. A simulation study was conducted to assess the performance of this new procedure. Different scenarios were considered by manipulating the number of variables, the number of samples, the regularity of the sampling design and the level of spatial autocorrelation. As identified by previous studies, we found that Mantel statistic and its associated type I error rate are inflated in simple and partial Mantel tests when both distances matrices are spatially structured. We showed that these biases increased with the number of variables, decreased with the number of samples and were slightly lower for regular than irregular sampling. The new procedure succeeded in correcting the spurious inflations of the parameter estimates and type I error rates in any of the presented scenarios. Our results suggest that studies from several fields (e.g. genetic or community ecology) could have been overestimating the relationship between two distances matrices when both presented spatial autocorrelation. We proposed an alternative solution applicable in every field to correctly compute Mantel statistic with a fair type I error rate.