Noncollapsibility and its role in quantifying confounding bias in logistic regression.

Area of Science:

• Epidemiology
• Biostatistics
• Statistical Modeling

Background:

• Confounding bias is a significant concern in epidemiological research.
• The 10% change-in-estimate criterion is commonly used to identify confounding bias by comparing univariable and multivariable regression models.
• This criterion can lead to erroneous conclusions in logistic regression due to the noncollapsibility phenomenon.

Purpose of the Study:

• To clarify the impact of noncollapsibility on confounding bias assessment in logistic regression.
• To provide guidance for researchers on accurately determining confounding bias in the presence of noncollapsibility.
• To illustrate the limitations of the change-in-estimate criterion.

Main Methods:

• A Monte Carlo simulation study was conducted to investigate confounding bias and noncollapsibility effects in logistic regression.
• An empirical data example was utilized to demonstrate the issues with the change-in-estimate criterion.
• Comparison of effect estimates from univariable, multivariable regression, and inverse probability weighting.

Main Results:

• The change-in-estimate criterion can underestimate or overestimate confounding bias due to noncollapsibility.
• Noncollapsibility effects can lead to different, yet valid, adjusted exposure effect estimates between multivariable regression and inverse probability weighting.
• In the data example, noncollapsibility caused an underestimation of confounding bias by the change-in-estimate method.

Conclusions:

• The difference between univariable and multivariable effect estimates in logistic regression may reflect both confounding bias and noncollapsibility.
• Confounders should ideally be identified during the study design phase based on subject matter expertise.
• Comparing unadjusted and inverse probability weighted estimates is a recommended approach to quantify confounding bias.