Statistical models for longitudinal zero-inflated count data with applications to the substance abuse field

Area of Science:

• Biostatistics
• Longitudinal Data Analysis
• Count Data Modeling

Background:

• Longitudinal studies often involve count data with excess zeros.
• Statistical models like hurdle and zero-inflated Poisson are used but require careful selection.
• Existing knowledge gaps exist in comparing these models for complex longitudinal data.

Purpose of the Study:

• To comprehensively review and compare hurdle and zero-inflated Poisson models for longitudinal zero-inflated count data.
• To evaluate model performance under various real-data scenarios, simulating features of substance abuse studies.
• To provide guidance on model selection based on conceptual fit, computational efficiency, and accuracy.

Main Methods:

• Comparative analysis of hurdle and zero-inflated Poisson models.
• Simulation studies designed to mimic features of longitudinal substance abuse data.
• Evaluation of models based on conceptual framework, computational advantages, and estimation accuracy.

Main Results:

• The hurdle model is often conceptually superior for zero-inflated count data.
• The zero-inflated Poisson model can yield inaccurate estimates, particularly in specific data configurations.
• Model performance is enhanced by larger sample sizes, reduced missing data, and lower covariate correlations.
• Including random effects diminishes the computational advantage of the hurdle model.

Conclusions:

• Model selection for longitudinal zero-inflated count data depends on data characteristics and study objectives.
• The hurdle model is generally recommended when conceptually appropriate, but caution is advised with the zero-inflated Poisson model.
• Simulation results offer practical insights for researchers in substance abuse and related fields analyzing complex count data.