TL;DR
This paper introduces BAMBI, an R package designed to fit complex mixture models to angular (directional) data, a type of data common in fields like geology and bioinformatics. BAMBI provides a unified, user-friendly platform for analyzing this data, addressing a significant gap in available statistical software.
What Problem Does This Paper Address?
Directional data, where values wrap around (like angles on a circle), are prevalent in many scientific disciplines. While methods for analyzing single directional values (univariate) are well-established, analyzing pairs of directional values (bivariate) or mixtures of these distributions has been more challenging due to a lack of specialized statistical models and, crucially, a lack of accessible software to implement them. This gap has hindered the practical application of advanced statistical techniques to bivariate angular data.
Methods
BAMBI implements Bayesian inference for fitting finite mixture models to both univariate and bivariate angular data. The core of the package revolves around fitting models based on three key bivariate distributions: the wrapped normal, the von Mises sine, and the von Mises cosine. For fitting these models, BAMBI leverages advanced Markov Chain Monte Carlo (MCMC) techniques, including Hamiltonian Monte Carlo (HMC) and a blocked Gibbs sampler, to efficiently explore the complex posterior distributions of the model parameters. The package also includes functionalities for density evaluation, random data generation, and various diagnostic tools for assessing model fit and convergence.
Datasets
The paper demonstrates BAMBI's capabilities using two real-world datasets:
- wind: Univariate wind direction data from Saturna Island, British Columbia.
- tim8: Bivariate backbone dihedral angle data (φ, ψ) for the protein Triose Phosphate Isomerase (8TIM).
Key Insights and Ideas
- Unified Platform for Angular Mixtures: BAMBI provides a single, coherent R package for a range of bivariate angular mixture models, which were previously scattered or unavailable.
- Bayesian Approach for Complex Models: The use of Bayesian inference and advanced MCMC methods (like HMC) allows for robust analysis of complex, potentially multimodal, angular mixture models.
- Addressing Software Gaps: The primary contribution is filling a critical void in statistical software for bivariate angular data analysis, making these advanced methods accessible to a wider audience.
- Flexibility in Model Choice: The package supports multiple distributions (wrapped normal, von Mises sine, von Mises cosine) and allows for fitting both single component models and finite mixtures, offering flexibility in modeling diverse data patterns.
- Diagnostic Tools are Crucial: The paper emphasizes the importance of diagnostic tools (trace plots, convergence diagnostics) for ensuring the reliability of MCMC-based results, especially with multimodal posteriors.
Key Results
The paper demonstrates BAMBI's functionality through detailed examples of fitting mixture models to the tim8 (bivariate) and wind (univariate) datasets. For the tim8 dataset, the analysis shows that optimal models were found to have 4 components for vmsin, vmcos, and wnorm2 distributions. Model comparison using LOOIC indicated no significant difference in fit among these three optimal bivariate models. For the wind dataset, optimal models were found to have 2 components for both vm and wnorm distributions, with no significant difference in their goodness of fit. The paper also showcases BAMBI's ability to provide parameter estimates, credible intervals, and visualizations (contour and density plots) for the fitted models.
Limitations and Future Work
The authors acknowledge several areas for future development:
- Additional Distributions: Implementing more angular distributions, such as the wrapped Cauchy.
- Enhanced Simulation: Developing more methods for density evaluation and random simulation from fitted models.
- Advanced Visualizations: Incorporating toroidal plots for visualizing bivariate angles.
- MCMC Efficiency: Exploring parallel tempering or related methods to improve MCMC exploration.
- Handling Overfitting: Investigating better methods for handling overfitting heterogeneity in high-dimensional mixture models, potentially using sparse priors.
Summary updated