Two-Sample Mendelian Randomization

The defining feature of 2SMR is the use of two independent study populations.

• Exposure Data: Genetic variant-exposure associations (e.g., BMI) are estimated from a dedicated GWAS (e.g., GIANT consortium).
• Outcome Data: Genetic variant-outcome associations (e.g., coronary artery disease) are estimated from a completely separate GWAS (e.g., CARDIoGRAMplusC4D).
• Bias Mitigation: This separation eliminates the winner's curse and confounding that can inflate effect estimates in single-sample designs.

2SMR is typically performed using publicly available GWAS summary statistics rather than individual-level participant data.

• Data Structure: Input consists of SNP identifiers, effect alleles, beta coefficients, standard errors, and p-values for millions of variants.
• Accessibility: Enables researchers to conduct causal analyses without requiring access to sensitive, individual-level genotype-phenotype data.
• Harmonization: A critical preprocessing step aligns the effect alleles between exposure and outcome datasets to ensure consistent directionality.

The reliability of a 2SMR analysis hinges on the strength of the genetic instruments.

• F-Statistic: A standard metric for instrument strength; a value greater than 10 is the conventional threshold to avoid weak instrument bias.
• Variance Explained: Instruments must collectively explain a meaningful proportion of the variance in the exposure phenotype.
• Weak Instrument Bias: In 2SMR, weak instruments bias the causal estimate toward the confounded observational association in the outcome dataset, making robust instrument selection paramount.

A robust 2SMR workflow mandates a suite of sensitivity analyses to assess and correct for horizontal pleiotropy.

• MR-Egger Regression: Allows for directional pleiotropy by fitting an unconstrained intercept, though at the cost of statistical power.
• Weighted Median Estimator: Provides consistent causal estimates even when up to 50% of the weight comes from invalid instruments.
• MR-PRESSO: Detects and removes outlier SNPs that contribute disproportionately to pleiotropic bias, providing a corrected estimate.
• Convergence: A causal inference is considered robust only when multiple sensitivity methods with different assumptions yield consistent results.

The 2SMR framework is extensible beyond a single exposure-outcome pair.

• Bidirectional MR: The roles of exposure and outcome are reversed to test for reverse causation (e.g., does BMI cause depression, or does depression cause BMI?).
• Multivariable MR (MVMR): Estimates the direct causal effect of multiple correlated exposures (e.g., LDL, HDL, triglycerides) on an outcome simultaneously, accounting for shared genetic architecture.
• Mediation Analysis: 2SMR can be used in a two-step framework to identify mediating traits on the causal pathway from exposure to outcome.

Valid causal inference requires that the two samples are drawn from comparable underlying populations.

• Ancestry Matching: Exposure and outcome GWAS should be performed in populations with similar genetic ancestry (e.g., both European) to avoid bias from differing linkage disequilibrium structures.
• Cohort Overlap: Any sample overlap between the exposure and outcome datasets can reintroduce confounding and bias similar to single-sample MR.
• Trans-Ethnic MR: Emerging methods adapt 2SMR for cross-population analyses, leveraging distinct LD patterns to refine causal signals.