Model selection for the North American Breeding Bird Survey.

Area of Science:

• Ecology and Evolutionary Biology
• Quantitative Ecology
• Ornithology

Background:

• The North American Breeding Bird Survey (BBS) is a crucial dataset for analyzing bird population dynamics.
• Traditional model selection methods are often unsuitable for complex hierarchical models used with BBS data.
• Leave-one-out cross-validation (LOOCV) is effective but computationally intensive for large datasets like the BBS.

Purpose of the Study:

• To conduct the first large-scale formal model selection for BBS data using LOOCV on sampled observations.
• To compare four alternative hierarchical models differing in year effect structures and overdispersion descriptions.
• To evaluate model performance for 548 North American bird species.

Main Methods:

• Applied LOOCV to stratified random samples of BBS observations.
• Compared four hierarchical models: two year effect structures (D vs. S models) and two overdispersion structures (H vs. normal distribution).
• Utilized a hierarchical model across species to determine posterior probabilities for best-fit models.

Main Results:

• Models with conditionally independent year effects (D models) were generally favored over slope-based year effects (S models).
• Models with t-distributed overdispersion (H models) were favored over normally distributed overdispersion.
• The favored D and H models were recommended for 544 out of 548 species, with minimal impact on overall population trend estimates.

Conclusions:

• Recommend adopting D and H models for BBS data analysis, moving away from the conventional S model for most species.
• The favored models provide a more robust analysis of bird population trends, especially for species with significant population changes.
• WAIC, a simpler alternative to LOOCV, was found to be unreliable for this large-scale model selection task.