A Modification to Two-Stage Least Squares With Genetic Applications.

Area of Science:

• Genetics and Bioinformatics
• Statistical Genetics
• Computational Biology

Background:

• Two-stage least squares (2SLS) is standard for inferring causal links between exposures (genes/proteins) and outcomes (diseases/traits) in TWAS/PWAS.
• A common challenge in two-sample TWAS/PWAS is the smaller sample size in stage 1 compared to stage 2, leading to attenuation bias and reduced statistical power.

Purpose of the Study:

• To introduce reverse two-stage least squares (r2SLS), a new method designed to mitigate bias and enhance statistical power in TWAS/PWAS.
• To theoretically establish the asymptotic properties of the r2SLS estimator and compare its efficiency with conventional 2SLS.

Main Methods:

• Developed r2SLS, which predicts the outcome using genetic variants as instrumental variables (IVs) in stage 1 and tests the association with observed gene expression in stage 2.
• Provided theoretical analysis of r2SLS estimator's asymptotic unbiasedness and normal distribution.
• Investigated conditions for asymptotic equivalence and superiority of r2SLS over 2SLS, including strategies for invalid IV selection.

Main Results:

• Demonstrated through simulations and real data analyses (GTEx, UKB-PPP, GWAS) that r2SLS can offer improved type I error control.
• Showcased higher statistical power and greater robustness to weak IVs compared to the conventional 2SLS method.
• Confirmed theoretical advantages of r2SLS in reducing attenuation bias and estimation uncertainty.

Conclusions:

• r2SLS presents a statistically advantageous alternative to 2SLS for causal inference in large-scale genetic association studies.
• The method shows promise for improving the reliability and power of TWAS/PWAS, especially under typical two-sample study designs.
• r2SLS offers a robust framework for exploring gene-trait associations with better performance characteristics than traditional approaches.