Modeling Multivariate Count Time Series Data with a Vector Poisson Log-Normal Additive Model: Applications to Testing...

Area of Science:

• Psychology
• Education
• Statistics

Background:

• Single-case designs (SCDs) analyze time series data to detect intervention effects.
• Multiple-baseline designs in SCDs often involve multiple count outcomes.
• Standard Poisson models struggle with overdispersion and correlations common in SCD data.

Purpose of the Study:

• To present a novel statistical model, the vector Poisson log-normal additive (V-PLN-A) model.
• To address challenges of overdispersion, correlation, and change processes in multivariate count time series data from SCDs.
• To accurately estimate treatment effects in educational intervention studies.

Main Methods:

• Developed a vector Poisson log-normal additive (V-PLN-A) model.
• Incorporated a multivariate normal distribution to handle correlations and overdispersion.
• Applied the model to an educational intervention study and conducted simulation analyses using Bayesian methods.

Main Results:

• The V-PLN-A model demonstrated satisfactory parameter recovery with sufficient timepoints in Bayesian analysis.
• Ignoring change processes and overdispersion resulted in biased treatment effect estimates.
• The model successfully handled correlated and overdispersed multivariate count time series data.

Conclusions:

• The V-PLN-A model provides a robust framework for analyzing complex data in single-case designs.
• Accurate estimation of treatment effects in SCDs requires accounting for data characteristics like overdispersion and autocorrelation.
• This approach enhances the reliability of findings in educational and psychological intervention research.