The use of two-sample methods for Mendelian randomization analyses on single large datasets.

Area of Science:

• Genetics
• Epidemiology
• Biostatistics

Background:

• Genome-wide association studies (GWAS) from biobanks enable one-sample Mendelian randomization (MR) for causal inference.
• Standard MR methods assume independence of gene-exposure and gene-outcome associations, which is violated by confounding in one-sample MR.
• Confounding introduces correlation between estimates from the same individuals, challenging standard MR assumptions.

Purpose of the Study:

• To evaluate the performance of various one-sample MR methods under confounding.
• To compare the utility of two-sample MR methods when applied to one-sample data.
• To identify conditions under which MR-Egger performs acceptably in one-sample MR.

Main Methods:

• Simulations mimicking UK Biobank studies were used to assess bias and precision.
• Methods evaluated include fixed-effect/random-effects meta-analysis, weighted median, weighted mode, and MR-Egger regression.
• Scenarios varied in causal effect presence, confounding levels, and pleiotropy type (none, balanced, directional).

Main Results:

• Two-sample MR methods performed similarly in one-sample MR as in two-sample MR, even with confounding.
• MR-Egger showed bias proportional to confounding, but this was reduced with high instrument strength variability (IGX2 of 97%).
• All tested methods demonstrated robustness to confounding and pleiotropy to varying degrees.

Conclusions:

• Two-sample MR methods are generally applicable to one-sample MR in large biobanks.
• MR-Egger is not recommended for one-sample MR due to confounding bias.
• MR-Egger may be used cautiously if confounding is minimal or instrument strength variability is very high.