When should epidemiologic regressions use random coefficients?

Area of Science:

• Epidemiology
• Biostatistics
• Statistical Modeling

Background:

• Regression models with random coefficients are increasingly accessible in statistical software.
• They are known as generalized linear mixed models, hierarchical models, and multilevel models.
• Current epidemiologic analyses often rely on prevalent fixed-effects models.

Purpose of the Study:

• To advocate for regression models with random coefficients as a more scientifically defensible framework for epidemiologic analysis.
• To present these models as a rational compromise between model richness and estimability.
• To offer an alternative to standard variable-selection algorithms in epidemiologic research.

Main Methods:

• Utilized regression models with random coefficients.
• Applied principles of model richness (antiparsimony) and estimability.
• Illustrated methods with an analysis of diet, nutrition, and breast cancer data.

Main Results:

• Regression models with random coefficients offer a scientifically defensible framework for epidemiologic analysis.
• These models provide a rational compromise between model complexity and the ability to estimate parameters.
• They serve as an alternative to standard variable-selection methods, mitigating uncertainty distortion.

Conclusions:

• Regression models with random coefficients represent a more appropriate framework for epidemiologic analysis compared to fixed-effects models.
• These models balance the need for complexity with practical estimation.
• The approach offers improved uncertainty assessment over traditional methods.