A demonstration of modeling count data with an application to physical activity

Area of Science:

• Biostatistics
• Public Health
• Epidemiology

Background:

• Counting outcomes with many zeros, like physical activity days, present statistical challenges.
• Standard Poisson regression often fails to adequately model such skewed data distributions.
• Accurate statistical modeling is crucial for understanding health behaviors and their predictors.

Purpose of the Study:

• To evaluate and compare several modified Poisson regression models for analyzing skewed count data.
• To identify the most suitable statistical model for assessing predictors of vigorous physical activity (VPA) in Latina women.
• To demonstrate the utility of underutilized regression models for health-related counting outcomes.

Main Methods:

• Fitted five regression models: Poisson, overdispersed Poisson, negative binomial, zero-inflated Poisson (ZIP), and zero-inflated negative binomial (ZINB).
• Utilized data on vigorous physical activity (VPA) among Latina women.
• Employed analytical and graphical methods for model selection.

Main Results:

• Standard Poisson regression showed a poor fit due to 82% of subjects reporting zero VPA days.
• Negative binomial and ZIP models provided significantly improved fits.
• The zero-inflated Poisson (ZIP) model demonstrated the best overall fit for the VPA data.

Conclusions:

• Zero-inflated Poisson (ZIP) regression is a valuable tool for analyzing skewed count data common in health research.
• No VPA days were associated with poorer self-reported health, lower cultural assimilation, and higher BMI.
• Higher VPA frequency correlated with increased education and, marginally, with older age.