Incorporating location uncertainty improves inference with stop-level North American Breeding Bird Survey data
Ecological models should account for uncertainty to be most effective and useful. Yet, uncertainty from model covariates—unlike that from other sources, such as sampling error or process variability—is seldom explicitly incorporated. This can cause underestimates of uncertainty to cascade through model parameter estimates, predictions, and downstream uses. Burner et al. proposed a method for quantifying uncertainty in covariates and incorporating it into models using informative Bayesian priors. This method was applied to stop-level Breeding Bird Survey (BBS) analyses, where land cover uncertainty at each stop arises from substantial stop location uncertainty. A limited validation of model-estimated land cover, using stops with known locations, indicated the method’s potential effectiveness, but it was not rigorously evaluated. We conduct a robust simulation-based test, generating stop locations, extracting land cover, and simulating bird communities across 210 BBS routes in the upper Midwest. We compare 3 models: a “known” model with true land cover, a “naive” model assuming consistent 800-m stop spacing, and a “full” model using informative priors to estimate land cover. Species parameter estimates and predicted prevalence patterns across gradients in land cover from the full model approached those of the known model and were substantially closer to the true values used in simulations relative to those from the naive model. Naive model parameters were more biased relative to the other models, and credible intervals of predicted species prevalence rarely included the true simulated values. The full model also produced land cover covariate estimates closer to true simulation values relative to the mean informative priors. Our results show that, for the BBS, informative priors enable more accurate stop-level analyses despite location uncertainty. In contrast, naive models that ignore this uncertainty yield poor inferences. More broadly, we demonstrate empirically the utility of informative priors to account for covariate uncertainty in ecological models.