Testing concordance of instrumental variable effects in generalized linear models with application to Mendelian...

Area of Science:

• Epidemiology
• Biostatistics
• Genetic Epidemiology

Background:

• Instrumental variable (IV) regression addresses unmeasured confounding in observational studies.
• Existing structural mean models for IV analysis can face identification issues with weak instruments, limiting Mendelian randomization.
• Generalized linear models are used to define causal effects with multiple genetic variants as IVs.

Purpose of the Study:

• To propose a new method for testing causal effects in the presence of unmeasured confounders.
• To evaluate the concordance between IV effects on an intermediate exposure and disease outcome as a causality test.
• To address limitations of existing IV methods in genetic epidemiology.

Main Methods:

• Developed generalized least squares (GLS) estimators to test causality by assessing concordance of IV effects.
• Analyzed the performance of GLS estimators in generalized linear models with multiple IVs.
• Investigated asymptotic properties for continuous exposure and dichotomous outcomes in logistic models.

Main Results:

• GLS estimators provide valid and consistent tests of causality.
• Estimators are asymptotically conservative for continuous exposure/dichotomous outcome in logistic models.
• Consistency is achieved for rare outcomes via log-linear approximation.

Conclusions:

• The proposed concordance test offers a robust approach to inferring causality with instrumental variables.
• This method enhances the applicability of Mendelian randomization in genetic epidemiology.
• GLS estimators demonstrate validity and consistency, with specific performance characteristics noted for logistic models.