Local influence diagnostics for hierarchical count data models with overdispersion and excess zeros.

Area of Science:

• Biostatistics
• Statistical Modeling
• Longitudinal Data Analysis

Background:

• Hierarchical count data often exhibit overdispersion and excess zeros, complicating standard statistical modeling.
• Existing models, such as Poisson-normal generalized linear-mixed models with gamma random effects, address these issues but require rigorous assessment.
• Model assessment is crucial due to the parametric complexity of these advanced statistical methods.

Purpose of the Study:

• To develop and apply local influence measures for assessing hierarchical count data models.
• To identify subjects with undue influence on model fit and parameter estimates.
• To provide interpretable influence components for statistical and clinical insights.

Main Methods:

• Derivation of local influence measures tailored for models with overdispersion and excess zeros.
• Application of these measures to detect influential subjects in longitudinal clinical trial data.
• Analysis of influence on fixed effects, variance components, and overdispersion parameters.

Main Results:

• The proposed local influence method effectively detects influential subjects in complex count data models.
• Identified a potentially small but clinically significant subgroup of patients in an epilepsy clinical trial.
• Provided statistically and clinically relevant insights beyond previous analyses of the trial data.

Conclusions:

• Local influence diagnostics are essential for validating complex statistical models for hierarchical count data.
• The method enhances understanding of subject-specific effects in longitudinal studies.
• This approach can uncover important patient subgroups, improving clinical trial interpretation.