A Modified Debiased Inverse-Variance Weighted Estimator in Two-Sample Summary-Data Mendelian Randomization

Area of Science:

• Genetic Epidemiology
• Statistical Genetics
• Biostatistics

Background:

• Mendelian randomization (MR) uses genetic variants as instrumental variables to infer causal effects from observational data.
• A key challenge in MR is the presence of many weak instruments, where genetic variants have modest associations with the exposure.
• Conventional methods like inverse-variance weighted (IVW) can be biased with weak instruments.

Purpose of the Study:

• To address limitations of existing MR estimators, particularly the debiased IVW (dIVW) and penalized IVW (pIVW) estimators.
• To propose a novel modified debiased IVW (mdIVW) estimator with improved statistical properties.
• To extend the mdIVW method to handle instrumental variable selection and pleiotropy.

Main Methods:

• Development of the modified debiased IVW (mdIVW) estimator by applying a shrinkage factor to the dIVW estimator.
• Theoretical analysis to prove the second-order bias, variance, and mean squared error properties of mdIVW.
• Extension of mdIVW to account for instrumental variable selection and balanced horizontal pleiotropy.
• Extensive simulation studies and real data analysis to compare mdIVW with existing methods.

Main Results:

• The dIVW estimator can exaggerate causal effect estimates, especially with small sample sizes.
• The pIVW estimator offers better statistical properties but is more complex.
• The proposed mdIVW estimator demonstrates second-order bias properties and achieves smaller variance and mean squared error compared to dIVW and pIVW.
• The extended mdIVW method effectively handles instrumental variable selection and pleiotropy.

Conclusions:

• The mdIVW estimator provides a robust and statistically superior approach for Mendelian randomization with many weak instruments.
• The method offers improved accuracy and efficiency over existing techniques.
• The findings have significant implications for causal inference in genetic epidemiology and observational studies.