Selecting Invalid Instruments to Improve Mendelian Randomization with Two-Sample Summary Data

Area of Science:

• Genetics
• Epidemiology
• Statistical genetics

Background:

• Mendelian randomization (MR) is crucial for estimating causal relationships between risk factors and diseases.
• Instrument selection is fundamental in MR, with potential trade-offs between instrument validity and statistical power.
• Large genome-wide association studies (GWAS) provide numerous genetic variants, complicating optimal instrument selection.

Purpose of the Study:

• To develop a "focused" instrument selection method for Mendelian randomization (MR) that minimizes estimated asymptotic mean squared error.
• To propose a novel strategy for constructing confidence intervals for post-selection estimators in MR, addressing potential coverage loss.
• To evaluate the optimal instrument selection strategy in empirical applications, including lipid drug target validation and vitamin D effect investigation.

Main Methods:

• Developed a "focused" instrument selection approach to minimize the mean squared error of causal effect estimates in MR.
• Proposed a new method for constructing confidence intervals for post-selection causal effect estimators to maintain asymptotic coverage.
• Applied the methods to real-world data for validating lipid drug targets and assessing vitamin D's effects on various outcomes.

Main Results:

• The "focused" instrument selection method effectively minimizes the mean squared error of causal effect estimates.
• The proposed confidence interval strategy provides robust coverage in settings with many weak and potentially invalid instruments.
• Empirical applications demonstrated that optimal instrument selection includes numerous potentially invalid instruments, not just a few biologically justified ones.

Conclusions:

• Optimal instrument selection in MR, particularly with many weak and potentially invalid instruments, involves a balance between bias and variance.
• The "focused" instrument selection method and associated confidence interval strategy offer improved causal inference in complex genetic association settings.
• Findings challenge the exclusive reliance on a small set of "valid" instruments, advocating for the inclusion of more instruments to enhance precision.