Fully parametric and semi-parametric regression models for common events with covariate measurement error in main...

Area of Science:

• Biostatistics
• Epidemiology
• Statistical Modeling

Background:

• Measurement error in covariates can bias exposure-disease relationship estimation in epidemiological studies.
• Traditional methods may rely on restrictive parametric models for measurement error, risking misspecification.
• Validation studies offer a way to correct for measurement error but are often small.

Purpose of the Study:

• To derive and evaluate likelihood functions for binary data in main/validation study designs with covariate measurement error.
• To propose empirical considerations for model selection over restrictive mathematical properties.
• To assess the performance of fully and semi-parametric methods in correcting for measurement error bias.

Main Methods:

• Derivation of the joint likelihood function for main and validation study data.
• Maximization of the likelihood function using standard statistical theory.
• Application of semi-parametric methods (Robins, Rotnitzky, Zhao, 1994; Robins, Hsieh, Newey, 1995) using nonparametric measurement error models.

Main Results:

• Empirical verification of a constant prevalence ratio model with gamma covariate measurement error.
• Reanalysis of occupational chemotherapeutic exposure data revealed a threefold increase in log relative risk.
• Semi-parametric estimates were consistent with parametric results, indicating robustness against misspecification bias.

Conclusions:

• Empirically driven model choices are crucial for handling measurement error in covariate data.
• Semi-parametric methods provide robust inference, free from bias due to measurement error model misspecification.
• Accounting for measurement error is essential for accurate estimation of exposure-disease relationships and can significantly alter risk estimates.