Bayesian Correction for Misclassification in Multilevel Count Data Models.

Area of Science:

• Statistics
• Biostatistics
• Epidemiology

Background:

• Covariate misclassification is a known issue in regression, causing biased estimates.
• Its impact on hierarchical count models is less understood.
• Hierarchical generalized linear models (HGLMs) are widely used in various scientific fields.

Purpose of the Study:

• To propose a fully Bayesian approach for modeling covariate misclassification within HGLMs.
• To investigate the performance of this approach using simulations and real data.
• To extend misclassification modeling to the domain of HGLMs.

Main Methods:

• Developed a Bayesian hierarchical model to simultaneously account for a misclassified covariate and the hierarchical response.
• Considered models incorporating single and multiple diagnostic tests for misclassification.
• Employed simulation studies to assess bias reduction and interval estimator performance.

Main Results:

• The proposed Bayesian model effectively accounts for covariate misclassification, reducing bias.
• Simulation studies demonstrated improved performance of interval estimators when misclassification is addressed.
• A real data example showed a significant shift in the relationship's direction and significance when misclassification was handled.

Conclusions:

• Ignoring covariate misclassification in HGLMs can lead to erroneous conclusions.
• The proposed Bayesian method provides a robust approach to handle misclassification in HGLMs.
• This methodology enhances the reliability of statistical inference in complex hierarchical data structures.