Bias in two-sample Mendelian randomization when using heritable covariable-adjusted summary associations

GWAS
bias
causal inference
covariable adjustment
mendelian randomization
statistical methods
  • Core Problem: This study demonstrates that using GWAS summary statistics that have been adjusted for heritable covariables (e.g., adjusting BMI for height) can introduce bias into Two-Sample Mendelian Randomization (2SMR) estimates.
  • Mechanism: Adjusting the exposure GWAS for a heritable covariable makes the resulting MR estimate prone to bias toward the null (zero), as the genetic variant is no longer a valid instrument for the total effect of the unadjusted exposure.
  • Recommendation: Researchers should prioritize using unadjusted GWAS summary statistics to estimate the total causal effect of the exposure and must be transparent about any adjustments made to the summary data used.
Published

23 January 2026

PubMed: 33619569 DOI: 10.1093/ije/dyaa266 Overview generated by: Gemini 2.5 Flash, 28/11/2025

Key Finding: Bias from Adjusted GWAS Summary Data

This paper investigates the potential bias introduced into Two-Sample Mendelian Randomization (2SMR) estimates when the summary association results from Genome-Wide Association Studies (GWASs)—used for either the exposure or the outcome—have been adjusted for heritable covariables (e.g., adjusting a Body Mass Index GWAS for height).

The key finding is that performing 2SMR using summary statistics adjusted for a heritable covariable can introduce bias, and the direction and magnitude of this bias depend crucially on whether the adjustment was applied to the exposure GWAS, the outcome GWAS, or both, and the underlying causal structure.

Mechanism of Bias

Standard MR relies on the assumption that the genetic variant’s effect on the exposure is unconfounded. When a GWAS adjusts for a heritable covariable, the resulting genetic effect estimate (\(\hat{\Gamma}\)) is no longer the total effect of the variant on the phenotype, but rather the direct effect net of the covariable.

The authors show via simulations and theoretical exploration that:

  • Adjustment in Exposure GWAS Only: If the genetic association with the exposure (\(\hat{\Gamma}_{XZ}\)) is adjusted for a heritable covariable, the MR estimate is generally biased toward the null (zero). This occurs because the IV estimate now targets a highly conditional and often complex causal effect parameter, which may not be the total effect intended by the MR analysis.
  • Adjustment in Outcome GWAS Only: If the genetic association with the outcome (\(\hat{\Gamma}_{YZ}\)) is adjusted, the resulting 2SMR estimate will be estimating the effect of the exposure on the residual outcome (the part of the outcome not explained by the covariable). This estimate may also be biased relative to the total causal effect of the exposure on the unadjusted outcome.

Recommendations and Implications for MR Practice

The paper advises extreme caution when using GWAS summary data that has been adjusted for heritable traits:

  • Prioritize Unadjusted GWAS: The most straightforward approach is to prioritize unadjusted GWAS summary statistics for both the exposure and the outcome when aiming to estimate the total causal effect of the exposure.
  • Targeting Direct Effects: If the researcher specifically wants to estimate a direct causal effect (e.g., the effect of BMI on T2D, independent of cholesterol), then adjustments in both the exposure and outcome GWAS (or using multivariable MR) may be appropriate, but the interpretation becomes highly specific and complex.
  • Transparency: Researchers must be fully transparent about whether the GWAS data they used was adjusted for heritable covariables, as this deeply affects the interpretation and validity of the final 2SMR causal estimate.

Conclusion

The study underscores a critical source of potential bias in the widely used 2SMR framework, demonstrating that genetic variants associated with conditional phenotypes are not necessarily valid instruments for estimating the total causal effect of the unadjusted phenotype. Researchers must carefully select GWAS data based on the adjustment status of the summary statistics.