Species distribution models (SDMs) are widely used as a correlative statistical approach to link species occurrences with their environment and predict spatio-temporal changes in species distributions. While the theory behind the connection between SDMs and population dynamics was established in the last two decades, this connection has not been extended to a more practical context. There is some evidence that the outputs of SDMs, probability of occurrence or habitat suitability, are positively correlated with population abundance but the strength of these reported correlations are highly variable across taxa. Also, there is only a single study to date that explored the relationship between intrinsic population growth rate and probability of occurrence across multiple species which found no generalizable pattern.
We applied a novel mark-recapture data analysis method called CJS-pop to estimate parameters related to population demography (for example, intrinsic growth rate) of 17 North American bird species with data from Mapping Avian Productivity and Survivorship (MAPS). CJS-pop is a Bayesian hierarchical model and requires only robust design mark-recapture data to estimate demographic parameters. We also used occurrence data from GBIF and built SDMs with a recently developed Bayesian SDM framework called Hierarchical Modeling of Species Communities (HMSC). We compare the probability of occurrence of MAPS sites of 17 bird species as estimated by HMSC with demographic parameters as estimated by CJS-pop to explore the link between SDMs and population dynamics.
Results/Conclusions
We found strong evidence that there is positive association between probability of occurrence and population size. Median probability of positive association was 0.9 across 17 species. This evidence was weaker for parameters related to intrinsic growth rate with positive association probability at 0.75. Probability of occurrence, however, explained very little of the spatial variation of demographic parameters across MAPS sites; for all parameters median R2 was lower than 0.1.
HMSC showed good discriminatory power between presence and absence points (Median AUC was 0.8) but it was not able to distinguish between MAPS sites with intrinsic growth rate larger than 0 (source populations) and lower than 0 (sink populations). This can be caused by high demographic noise among occurrences obtained from GBIF. Important portion of the presence points used in the models might be proxies for populations not in equilibrium.
Overall, we detected a general mismatch between SDMs and population dynamics. We argue that this mismatch has potential ramifications for practical applications of SDMs and for assumptions about what they estimate.