Challenges in Modelling Spatio-Temporal Climatic Correlates of Local Losses of Wild Bees Using Dynamic Occupancy Models

Aim Impacts of climate change on biodiversity are well documented. Much of the evidence linking climate change to species distribution changes derives from studies using long-term climate averages in a correlative spatial framework. While useful, these static species distribution models (SDMs) give little insight into the process behind the correlations. Here, we model changes in wild bee occupancy dynamics as a function of temperature covariates that vary in space and time. We aim to detect fine-scale, climate-associated distribution changes in wild bees beyond those captured by traditional SDMs and aim to assess the challenges of applying dynamic occupancy models to large-scale opportunistic datasets.

Location Great Britain.

Methods We use dynamic occupancy models to examine the relationship between temperature and local losses for 106 wild bee species. We focus on one spatial metric (mean long-term temperature average, akin to SDM approaches) and one spatio-temporal metric (the annual temperature anomaly). We use a risk-of-bias assessment to evaluate how data limitations may affect inference in our dynamic occupancy models.

Results Mean long-term temperature for a site was associated with the probability of local loss for > 60% of species. In general, mean temperature was negatively associated with the probability of local loss for southerly distributed species (meaning a lower probability of loss at warmer sites), while the relationship was reversed for northerly species. The annual temperature anomaly was only influential for one species.

Main Conclusions Our results mirror the large-scale spatial pattern of SDMs, and although ecologically plausible, we find little signal of fine-scale, climate-associated wild bee distribution changes. We attribute this lack of signal to data-model mismatches as revealed by the detailed risk-of-bias assessment. We conclude that while dynamic occupancy models offer promise for integrating spatio-temporal covariates, their use may be limited when applied to large-scale opportunistic datasets that lack systematic sampling effort.