Conditional-cumulant-of-exposure method in logistic missing covariate regression

Area of Science:

• Biostatistics
• Statistical Modeling
• Epidemiology

Background:

• Missing covariate data poses challenges in logistic regression analysis.
• Existing methods, like those by Satten and Kupper, focus on exposure probability for odds ratio estimation.
• Semiparametric estimators face limitations when validation data is restricted to controls.

Purpose of the Study:

• To extend existing methods for estimating parameters in logistic regression with missing covariates.
• To propose a novel approach utilizing the cumulant-generating function of missing covariates.
• To address limitations of inverse probability weighting when validation is control-only.

Main Methods:

• Approximation of a partial likelihood.
• Estimation of lower-order cumulants of the conditional distribution of unobserved data.
• Solving estimating equations for logistic regression parameters, with a simplified version using conditional mean and variance imputation.

Main Results:

• The proposed method effectively estimates logistic regression parameters even with missing covariates.
• It provides a workable solution when control-only validation prevents standard semiparametric estimators.
• The estimator performs well generally, though it shows inconsistency with large relative risk parameters.

Conclusions:

• The developed method offers a robust alternative for logistic regression with missing covariates, especially in control-only validation scenarios.
• The technique demonstrates practical utility through simulations and real data analysis.
• Further research may be needed to address the inconsistency observed with large relative risks.