Modeling Zero-inflated Count Data Using Generalized Poisson and Ordinal Logistic Regression Models in Medical...

Area of Science:

• Biostatistics
• Medical Research Methodology
• Data Analysis

Background:

• Statistical models are crucial for interpreting medical research findings.
• Count data in medicine often present challenges like over-dispersion and zero-inflation.
• Ordinal regression models are used for count data with a limited range, such as a maximum of 5.

Purpose of the Study:

• To evaluate the generalized Poisson (GP) model as an alternative to ordinal logistic regression (OLR) for medical count data.
• To compare the performance of GP and OLR models, particularly in scenarios with zero-inflated data.
• To assess model superiority using both simulated and real-world medical datasets.

Main Methods:

• Generated simulated count data with varying parameters (regression coefficients, sample sizes, zero proportions).
• Applied and compared the generalized Poisson (GP) and ordinal logistic regression (OLR) models.
• Utilized fit statistics and analyzed real-time medical datasets for comparative evaluation.

Main Results:

• The GP model consistently exhibited lower bias and mean squared error in simulations compared to the OLR model.
• Real-time data analysis showed the GP model had lower standard errors than the OLR model.
• The Bayesian information criterion generally favored the GP model, except under specific high-zero proportion conditions with large sample sizes.

Conclusions:

• The generalized Poisson (GP) model is a more advantageous statistical approach than the ordinal logistic regression (OLR) model for analyzing specific types of medical count data.
• The GP model offers improved ease of modeling and interpretation compared to the OLR model.
• This study supports the adoption of the GP model for handling over-dispersed and zero-inflated count data in medical research.