A More Robust Approach to Multivariable Mendelian Randomization.

Area of Science:

• Statistical Genetics
• Epidemiology
• Bioinformatics

Background:

• Multivariable Mendelian randomization (MVMR) infers causal effects of multiple exposures on an outcome using genetic variants.
• MVMR faces challenges with weak instruments, leading to bias and inflated Type I errors, unlike univariable Mendelian randomization.

Purpose of the Study:

• To develop a more accurate theoretical framework for MVMR estimators under varying instrument strengths.
• To propose a novel estimator that mitigates bias and improves finite-sample performance in the presence of many weak instruments.

Main Methods:

• Introduction of a new asymptotic regime accommodating varying instrument strengths.
• Analysis of the multivariable inverse-variance weighted method under the new regime.
• Development of a closed-form modification and a spectral regularization technique for MVMR estimators.

Main Results:

• The standard multivariable inverse-variance weighted method shows bias and narrow confidence intervals with many weak instruments.
• The proposed spectral-regularized estimator is consistent and asymptotically normal under many weak instruments.
• Simulations and real data applications confirm enhanced robustness of the proposed methods.

Conclusions:

• The novel asymptotic framework provides a more accurate theoretical basis for MVMR.
• The proposed spectral-regularized estimator significantly improves the reliability of MVMR analyses with weak instruments.
• This work enhances the robustness and accuracy of causal inference using multivariable Mendelian randomization.