Old trends and alternative views on spatial analyses in ecology

Background/Question/Methods

Organisms, along with their environments, present remarkable patterns of variation across space.  Unravelling the multiple mechanisms behind the myriad, yet complex patterns that dominate in nature is one of the most important intellectual challenges in the science of ecology.  Much as astronomers cannot perform planetary experiments, ecologists cannot easily manipulate most biodiversity patterns because these patterns occur at many different scales.  Therefore, ecologists often need to use spatial models to understand the processes that structure biodiversity patterns at large scales such as species richness, species distributions and species co-occurrence across space.  Although spatial models can provide great insights into these processes, ecologists tend to see spatial autocorrelation as a nuisance that needs to be filtered out of data rather than an interesting property to be studied.  Beyond the standard nuisance viewpoint, an alternative and less acknowledged perspective is that the spatial legacy of ecological data can help us challenge models and improve our understanding of ecological phenomena.   While these views can be reconciled, we need to understand what are the implications of spatial structure on different model components, namely residual and predictors.  To reconcile these two views, I use different simulation scenarios in which the spatial autocorrelation levels in these two model components are manipulated.  

Results/Conclusions

The simulation results reveal a number of interesting outcomes.  First, the analyses of model residuals as a way to determine whether a spatial model should be considered or not is often misleading.  Second, autocorrelation in both predictors and residuals has a far greater effect on model accuracy than when only residuals are autocorrelated.  Third, the idea that we loose statistical power by using a spatial model is not always true when in fact we may even gain power by using one.  Finally, models based on autocorrelated data are more accurate than non-autocorrelated data.  As ecologists commonly, if not always, deal with autocorrelated data, we should see spatial complexity as an opportunity rather than the view in which modeling of spatial data provide difficult challenges.  However, we need to understand the nature of our data in order to select the best approach to pursue.  Based on these simulation results, I will present a number of guidelines that should be helpful to ecologists while making these decisions.