Abstract
Environmental changes in recent years have been linked to phenological shifts, which in turn affect the survival of various species. This research presents a novel Bayesian nonparametric approach for modeling the bivariate density of individual arrival and departure times of migratory birds across multiple sites and years. By combining the Hierarchical Dirichlet Process (HDP) and the Dependent Dirichlet Process (DDP), the proposed model—the Hierarchical Dependent Dirichlet Process (HDDP)—allows for site-specific weights and year-specific mixture locations. These locations are modeled as functions of environmental covariates using a multivariate extension of the Gaussian process. The framework is applied to capture-recapture data on blackcaps (Sylvia atricapilla) from the British Trust for Ornithology (BTO), demonstrating its utility in estimating migration patterns even when site-specific data is sparse.
Introduction
Phenology—the study of the timing of biological events—is a critical indicator of how species respond to climate change. For migratory birds, the timing of arrival and the duration of stay at breeding grounds are vital for reproductive success and survival. However, analyzing these patterns is often hindered by the limitations of "capture-recapture" (CR) data. In many monitoring schemes, such as the BTO’s Constant Effort Sites (CES), the number of birds caught at individual sites can be extremely low, making it difficult to derive statistically significant trends for a single location in isolation.
Furthermore, migration is a complex process involving distinct cohorts, such as "residents" who stay to breed and "transients" who are merely passing through. To address these challenges, this paper introduces a Hierarchical Dependent Dirichlet Process (HDDP) model that "borrows strength" across geographic sites while accounting for continuous environmental drivers.
Methodology: The HDDP Framework
The core of the proposed approach is a Bayesian nonparametric mixture model. Unlike traditional models that require a fixed number of clusters, Dirichlet Process Mixtures (DPM) allow the data to determine the number of migratory groups (clusters) dynamically.
Hierarchical and Dependent Components
The HDDP merges two powerful concepts:
- Hierarchical Dirichlet Process (HDP): This allows different geographic sites to share the same set of mixture components (migratory behaviors). While the behaviors are shared globally, each site can have different proportions of these behaviors. This structure is essential for handling data sparsity; a site with few captures can still be modeled accurately by leveraging the patterns observed at more data-rich sites.
- Dependent Dirichlet Process (DDP): This allows the characteristics of those behaviors—such as the mean arrival date—to vary as a function of external covariates.
Bivariate Density and Gaussian Processes
Rather than modeling arrival times alone, the model estimates the bivariate density of arrival time and length of stay. This joint modeling is crucial for distinguishing between cohorts (e.g., early-arriving breeders vs. late-arriving transients). To link these patterns to the environment, the authors utilize a Multivariate Gaussian Process (MGP). The MGP ensures that migration patterns in years with similar environmental conditions are more highly correlated, providing a robust way to integrate continuous climate data into the nonparametric framework.
Data and Application
The model was tested using both synthetic simulations and real-world data:
- Simulated Data: Synthetic datasets were used to verify that the model could recover true migration densities even when capture probabilities were as low as 26%.
- BTO Constant Effort Sites (CES): The primary application focused on 3,401 blackcaps (Sylvia atricapilla) across 10 UK sites over a 16-year period (1998–2013).
- Environmental Covariates: The North Atlantic Oscillation (NAO) index (averaged from January to April) was used as a proxy for large-scale climate trends affecting the UK.
Key Insights and Results
Robustness to Sparse Data
The hierarchical structure proved to be the model's greatest strength. By sharing information across the "franchise" of sites, the HDDP produced sensible density estimates where independent site models would have failed due to low sample sizes.
Identifying Migratory Cohorts
The bivariate approach successfully identified distinct population dynamics. In the blackcap dataset, the model revealed a significant peak of birds arriving before the sampling season officially began. These individuals likely represent the established breeding population, whereas later, shorter stays indicate transient birds moving through the site.
Climate Sensitivity
While the model is designed to capture environmental effects, the analysis of the 1998–2013 blackcap data showed that the 95% credible intervals for the NAO effect included zero. This suggests that during this specific period, there was no strong, uniform linear relationship between the NAO index and blackcap phenology across the UK, despite the variations observed at individual sites.
Limitations and Future Work
While the HDDP provides a flexible framework for density estimation, it is computationally intensive. The "Chinese Restaurant Franchise" sampling process scales linearly with the number of observations, which currently limits the volume of data that can be processed simultaneously.
Additionally, the current model assumes sites are "exchangeable," meaning it does not account for the physical distance between sites. Future iterations could incorporate spatial covariates, such as latitude and longitude, to better reflect the geographic reality that neighboring sites likely share more similar migration timings than distant ones. Finally, incorporating individual-level effects for birds that return to the same site across multiple years could further refine the model's accuracy.
Summary updated