Quantifying epidemiologic risk factors using non-parametric regression: model selection remains the greatest...

Area of Science:

• Epidemiology
• Biostatistics
• Statistical Modeling

Background:

• Standard logistic regression struggles with continuous exposure variables in case-control studies.
• Accurate characterization of exposure-response curves is crucial for risk estimation.
• Semi-parametric models offer a flexible alternative to traditional parametric approaches.

Purpose of the Study:

• To apply semi-parametric generalized linear models to case-control data.
• To estimate the dose-response relationship between alcohol consumption and oral cancer risk.
• To evaluate the utility of semi-parametric models with cross-validation for objective curve characterization.

Main Methods:

• Utilized semi-parametric generalized linear models with regression or smoothing splines for continuous exposure.
• Employed a cross-validation score minimization procedure for model selection.
• Applied the methodology to case-control data on alcohol consumption and oral cancer risk in African Americans.

Main Results:

• Consistent findings across linear, cubic, and smoothing splines indicated no risk-free threshold for alcohol consumption.
• The semi-parametric approach revealed a continuous increase in oral cancer risk with alcohol intake.
• A standard step-function model failed to detect this dose-response relationship.

Conclusions:

• Semi-parametric modeling provides a robust and reproducible method for characterizing continuous exposure-response curves.
• The study suggests a lack of a safe threshold for alcohol consumption in relation to oral cancer.
• Despite potential challenges like multiple local minima, the semi-parametric approach is a practical advancement.