An extended random-effects approach to modeling repeated, overdispersed count data.

Area of Science:

• Statistics
• Biostatistics
• Statistical Modeling

Background:

• Non-Gaussian count data are often modeled using the exponential family, particularly the Poisson model.
• Existing models are often restrictive, necessitating extensions for overdispersion and data hierarchies (e.g., repeated measures).
• While models exist for overdispersion (e.g., negative-binomial) and hierarchies (e.g., random effects), combined models are less common.

Purpose of the Study:

• To propose a generalized linear model that simultaneously accommodates overdispersion and clustering in count data.
• To extend existing count data models, including classical overdispersion models and generalized linear mixed models.
• To provide a flexible framework for analyzing complex count data structures.

Main Methods:

• Development of a generalized linear model incorporating two distinct sets of random effects: gamma for overdispersion and normal for clustering.
• Comparison and extension of classical overdispersion models and generalized linear mixed models.
• Exploration of estimation options, settling on maximum likelihood estimation with analytic and hybrid integration methods, implemented in SAS NLMIXED.

Main Results:

• The proposed model effectively extends existing statistical frameworks for count data analysis.
• The methodology is demonstrated through application to a real-world study involving epileptic seizures.
• The chosen estimation techniques provide viable options for parameter estimation in the proposed model.

Conclusions:

• The developed generalized linear model offers a robust approach for analyzing count data with both overdispersion and hierarchical structures.
• This model provides a valuable tool for researchers dealing with complex data in fields like biostatistics.
• The study highlights the utility of combining gamma and normal random effects for comprehensive count data modeling.