Low spatial autocorrelation in mountain biodiversity data and model residuals

Abstract

Spatial autocorrelation (SAC) is a common feature of ecological data where observations tend to be more similar at some geographic distance(s) than expected by chance. Despite the implications of SAC for data dependencies, its impact on the performance of species distribution models (SDMs) remains controversial, with reports of both strong and negligible impacts on inference. Yet, no study has comprehensively assessed the prevalence and the strength of SAC in the residuals of SDMs over entire geographic areas. Here, we used a large-scale spatial inventory in the western Swiss Alps to provide a thorough assessment of the importance of SAC for (1) 850 species belonging to nine taxonomic groups, (2) six predictors commonly used for modeling species distributions, and (3) residuals obtained from SDMs fitted with two algorithms with the six predictors included as covariates. We used various statistical tools to evaluate (1) the global level of SAC, (2) the spatial pattern and spatial extent of SAC, and (3) whether local clusters of SAC can be detected. We further investigated the effect of the sampling design on SAC levels. Overall, while environmental predictors expectedly displayed high SAC levels, SAC in biodiversity data was rather low overall and vanished rapidly at a distance of~5-10 km. We found low evidence for the existence of local clusters of SAC. Most importantly, model residuals were not spatially autocorrelated, suggesting that inferences derived from SDMs are unlikely to be affected by SAC. Further, our results suggest that the influence of SAC can be reduced by a careful sampling design. Overall, our results suggest that SAC is not a major concern for rugged mountain landscapes.

Snow cover persistence as a useful predictor of alpine plant distributions

“…Plant data come from a field campaign carried out from 2002 to 2009, 912 plots of 4 m 2 were sampled (Figure 1; [Buri et al, 2017; Dubuis et al, 2011]), following a random‐stratified equal sampling design (Hirzel & Guisan, 2002), with elevation (10 classes, from 375 m to 3201 m a.s.l. ), slope (three classes, 0–5°, 5–25° and >25°) and aspect (five classes, North, East, South, West and no aspect if slope <5°) as stratifying factors, and a minimal distance of 200 m between plots to minimize spatial auto‐correlation (Chevalier et al, 2021; Pottier et al, 2013). The sampling was limited to open, non‐forest vegetation.…”

Finding what you don’t know: Testing SDM methods for poorly known species

“…To avoid this problem, we used environmental data at a high resolution (80 m × 80 m). Smaller grid sizes may increase the influence of spatial autocorrelation of environmental variables (Qi & Wu, 1996), though the topographic complexity of mountains justifies the use of higher resolutions, as montane regions exhibit lower spatial autocorrelation at high resolutions (Chevalier et al, 2021). The variables included in the model were elevation, slope, aspect, distance to the nearest stream, and per cent canopy cover (http://www.nconemap.gov.datasets; https://www.mrlc.gov/data?f%255B0%255D=category%253ATree%2520Canopy).…”

Food and habitats requirements of the Eurasian Scops Owl (Otus scops) in Switzerland revealed by very high‐resolution multi‐scale models

“…To avoid bias, we checked for SAC with Mantel tests, with 10 000 data permutations per predictor (Lichstein 2007), though SAC in predictors is usually not seen as a problem in models (Chevalier et al . 2021). We discarded two topo‐climatic variables due to their overall low AUC at their best scale, ‘GDD’ with 0.58 and ‘SRad’ with 0.59, therefore only keeping ‘Prec’ in further analysis.…”

“…Within this new pool of variables, we tested the correlations among all variables; when a pair of variables showed a correlation > 0.7, we removed the variable showing the lowest AUC in order to limit collinearity and avoid overfitting the models (Lomba et al 2010. To avoid bias, we checked for SAC with Mantel tests, with 10 000 data permutations per predictor (Lichstein 2007), though SAC in predictors is usually not seen as a problem in models (Chevalier et al 2021). We discarded two topo-climatic variables due to their overall low AUC at their best scale, 'GDD' with 0.58 and 'SRad' with 0.59, therefore only keeping 'Prec' in further analysis.…”