Controlling for continuous confounding factors: non- and semiparametric approaches

Area of Science:

• Epidemiology
• Biostatistics

Background:

• Confounding bias is a significant challenge in observational studies linking exposures to health events.
• Adjusting for confounders in regression models is a common method to mitigate this bias.

Purpose of the Study:

• To investigate the impact of coding continuous confounders in logistic regression.
• To explore advanced statistical methods for reducing residual confounding.

Main Methods:

• Simulations were used to assess residual confounding from different coding strategies (binary, categorical).
• Nonparametric and semiparametric approaches, including generalized partial linear models, spline regression, and fractional polynomials, were presented.
• Semiparametric models utilize both parametric and nonparametric functions for covariate coding.

Main Results:

• Coding continuous confounders as binary or broad categorical variables can lead to substantial residual confounding in logistic regression.
• Semiparametric models allow for flexible estimation of the relationship between continuous factors and health outcomes.
• These models can reveal dose-effect relationships, aiding in appropriate confounder coding.

Conclusions:

• Careful consideration of coding methods for continuous confounders is crucial in logistic regression.
• Semiparametric modeling offers a robust approach to handle continuous confounders and minimize residual bias.
• These methods facilitate accurate estimation of exposure-health event relationships in observational studies.