Weighted Dirichlet Process Mixture Models to Accommodate Complex Sample Designs for Linear and Quantile Regression.

Area of Science:

• Biostatistics
• Statistical Modeling
• Computational Statistics

Background:

• Standard randomization-based inference can yield unstable estimators in certain populations.
• Bayesian model-based approaches offer a potentially better balance of bias correction and efficiency.
• Existing Bayesian methods are limited in estimating general distributional functions like quantiles.

Purpose of the Study:

• To adapt an extended Dirichlet Process Mixture model for flexible covariate-dependent inference.
• To enable general estimation of distributional functions, including quantile regression parameters.
• To apply the model for analyzing serum dioxin levels and age relationships in the US population.

Main Methods:

• Utilized an extended Dirichlet Process Mixture model with covariate-dependent DP priors.
• The model allows for adaptive complexity, using many mixture components or few as data permit.
• Applied linear and quantile regression to National Health and Nutrition Examination Survey (NHANES) data.

Main Results:

• The proposed model successfully accommodates complex sample designs.
• Demonstrated efficient estimation for both mean-level (linear regression) and distributional (quantile regression) analyses.
• Provided insights into the relationship between serum dioxin levels and age across the US population.

Conclusions:

• The extended Dirichlet Process Mixture model offers a powerful and flexible Bayesian approach for complex survey data.
• This method overcomes limitations of previous approaches by allowing general distributional function estimation.
• The findings highlight the utility of advanced statistical modeling for public health research using large-scale surveys.