Bayesian clustering of multiple zero-inflated outcomes.

Area of Science:

• Statistics
• Computational Statistics
• Bayesian Inference

Background:

• Many real-world applications involve count data with an excess proportion of zeros.
• Traditional models may not adequately capture the complexity of multiple, related zero-inflated count processes.
• Clustering subjects based on these complex count patterns is a significant challenge.

Purpose of the Study:

• To develop a novel Bayesian approach for clustering subjects from multiple zero-inflated count processes.
• To propose a flexible joint model that handles excess zeros and underlying count distributions.
• To enable a two-level clustering based on zero-inflation patterns and sampling distributions.

Main Methods:

• A joint hurdle model is specified for multiple zero-inflated count processes, each using a shifted Negative Binomial distribution.
• An enriched finite mixture model with a random number of components is used for flexible modeling of subject-specific parameters.
• Posterior inference is conducted using specialized Markov chain Monte Carlo (MCMC) algorithms.

Main Results:

• The proposed Bayesian method effectively clusters subjects based on both zero/non-zero patterns and the underlying count distributions.
• The model allows for conditional independence between processes, significantly reducing parameter complexity compared to multivariate methods.
• Demonstrated application on WhatsApp usage data showcases the practical utility of the approach.

Conclusions:

• The novel Bayesian framework provides a powerful tool for analyzing and clustering complex zero-inflated count data from multiple sources.
• The two-level clustering mechanism offers deeper insights into subject heterogeneity.
• This approach advances the analysis of multivariate count data in various scientific domains.