Spatial aggregation confounds null model-based inference of species association networks

Species association networks inferred from spatially replicated abundance data are typically constructed by comparing observed pairwise similarities to a null distribution generated by permuting the site-by-species matrix while holding row and column marginals fixed. I show that this widely used approach has inflated Type I error whenever individuals within a species are spatially aggregated. High Type I error is driven by the fact that permutational null models do not preserve the shape of the intra-specific spatial species abundance distribution (SSAD). I demonstrate this mechanism with a mathematical argument, simulations of purely neutral communities with no true species associations, and reanalysis of a large, taxonomically and geographically diverse compilation of spatially replicated abundance data. I focus throughout on the specific and widely used class of null-model permutation methods; whether the same artifact affects other approaches to inferring species associations, such as joint species distribution models or the compositional correlation methods common in microbial ecology, remains an open and, I argue, important question for future work.