Evaluating and implementing block jackknife resampling Mendelian randomization to mitigate bias induced by...

Area of Science:

• Genetics and Bioinformatics
• Statistical Genetics
• Epidemiology

Background:

• Participant overlap in genome-wide association studies (GWAS) can introduce overfitting bias into Mendelian randomization (MR) and polygenic risk score (PRS) analyses.
• This bias can affect the accuracy of causal inference and genetic risk prediction.

Purpose of the Study:

• To evaluate a block jackknife resampling framework for GWAS and PRS construction to mitigate overfitting bias in MR.
• To implement and assess this framework in a causal inference setting using UK Biobank data.

Main Methods:

• Simulated PRS and MR under three scenarios: external GWAS weights, overlapping GWAS sample weights, and block jackknife resampling.
• Applied the block jackknife resampling MR framework to examine body mass index effects on circulating biomarkers.
• Extended the framework to sex-stratified, multivariate, and bidirectional analyses.

Main Results:

• Block-jackknifing PRS showed no overfitting bias (mean R2=0.034) compared to externally weighted PRS (mean R2=0.040) at a stringent P-value threshold (P < 5 × 10-8).
• Overlapping sample PRS exhibited higher variance explained (mean R2=0.048).
• Overfitting worsened with liberal P-value thresholds (P < 0.05), but jackknife estimates remained robust (mean R2=0.084) versus overlapping (mean R2=0.103) and external (mean R2=0.086).
• Applied MR analysis yielded comparable estimates to external instruments, while overfitted scores produced narrower confidence intervals.

Conclusions:

• The block jackknife resampling framework effectively mitigates overfitting bias in PRS construction and MR analyses.
• This method provides robust and reliable estimates, particularly in the presence of sample overlap or when using more liberal P-value thresholds.
• The framework's extension to complex settings enhances its utility for causal inference research.