Hierarchical regression for analyses of multiple outcomes

Area of Science:

• Epidemiology
• Biostatistics
• Toxicology

Background:

• Cohort mortality studies often examine associations between exposures and multiple causes of death.
• Standard methods using separate regression models can suffer from poor statistical precision due to sparse data.
• Accurate estimation of exposure-mortality associations is crucial for public health and risk assessment.

Purpose of the Study:

• To describe a novel hierarchical regression model for estimating outcome-specific relative rate functions and credible intervals.
• To address limitations of standard regression approaches in cohort mortality analyses with multiple outcomes.
• To enhance the precision and stability of association estimates between exposures and various causes of death.

Main Methods:

• Development of a hierarchical regression model incorporating background stratification for confounder control.
• Application of a hierarchical "shrinkage" approach to stabilize estimates of exposure associations with different causes of death.
• Illustration of the model using cancer mortality data from dioxin-exposed chemical workers and radiation-exposed atomic bomb survivors.

Main Results:

• Hierarchical regression yielded estimates with improved precision and less extreme values compared to standard regression.
• The proposed model demonstrated enhanced stability for exposure-mortality associations across multiple causes.
• The hierarchical approach successfully accommodated effect-measure modification in the analyzed cohorts.

Conclusions:

• Hierarchical regression offers a more precise and stable alternative to conventional methods for estimating associations with multiple mortality outcomes.
• The "shrinkage" technique effectively stabilizes estimates in the presence of sparse data.
• This approach is valuable for complex cohort mortality studies requiring robust estimation of cause-specific risks.