Flexible Item Response Models for Count Data: The Count Thresholds Model.

Area of Science:

• Psychometrics
• Statistical Modeling
• Data Analysis

Background:

• Traditional count data models often assume fixed response distributions (e.g., Poisson).
• Existing item response theory (IRT) models may not adequately capture the nuances of count data with varying distributions.
• A need exists for flexible IRT models tailored for count data.

Purpose of the Study:

• Introduce a new item response theory (IRT) model specifically designed for count data.
• Develop a flexible model that does not assume a fixed response distribution.
• Demonstrate the model's performance and flexibility compared to existing methods.

Main Methods:

• Developed a novel IRT model for count data where response distributions are determined by difficulty functions.
• Utilized sparse parameterizations with fixed parametric or basis function approximations for difficulty.
• The model generalizes binary IRT models (e.g., Rasch, normal-ogive) when responses are dichotomized.

Main Results:

• The proposed IRT model for count data demonstrates competitive performance against advanced count data models.
• Simulations confirm accurate recovery of model parameters and response distributions.
• An application highlights the model's flexibility in handling strongly varying response distributions.

Conclusions:

• The new IRT model provides a flexible and effective alternative for analyzing count data.
• Its ability to adapt to different response distributions makes it suitable for diverse applications.
• The model advances the application of IRT to a broader range of data types.