Correction to: Is population structure in the genetic biobank era irrelevant, a challenge, or an opportunity?

Daniel J. Lawson, Neil M Davies, Simon Haworth, Bilal Ashraf, Laurence J Howe, Andrew Crawford, Gibran Hemani, George Davey Smith, Nicholas J. TimpsonHuman 2019 · DOI · OpenAlex

Introduction

In the era of massive genetic biobanks, "population structure"—the systematic difference in allele frequencies between sub-populations due to shared ancestry—remains a significant challenge. Without proper adjustment, genetic variants (SNPs) associated with specific groups may be incorrectly linked to traits prevalent in those groups, leading to false-positive associations. This paper provides a critical correction to the causal models used to address this issue, emphasizing that standard ancestry corrections are not universally effective.

The Causal Framework

The authors refine the original publication's framework by employing Directed Acyclic Graphs (DAGs) to explicitly distinguish between true ancestry ($A$), which is unobserved, and observed ancestry ($A_0$), typically represented by proxies like Principal Components (PCs). By mapping the flow of influence from SNPs ($G$) to Traits ($T$) and Outcomes ($O$), the paper identifies four distinct scenarios of statistical adjustment.

Scenarios of Statistical Adjustment

1. Accurate Correction and Undercorrection

Accurate correction occurs when the observed ancestry ($A_0$) successfully blocks the confounding path between true ancestry ($A$) and the trait ($T$). However, "undercorrection" is common when ancestry serves as a proxy for shared history or geography ($H$), which in turn influences environmental factors ($E$). If these environmental paths are not captured by genetic proxies, bias persists.

Causal diagrams illustrating accurate correction (left) and undercorrection due to environmental links (right).

2. Overcorrection and Unbiased Causal Inference

Overcorrection happens when the SNPs under study are themselves used to define the observed ancestry, potentially "correcting away" the true biological signal. This is particularly risky when ancestry is associated with the causal pathway of the trait. Conversely, the authors demonstrate that unbiased causal inference (e.g., in Mendelian Randomization) is possible even with complex population structure, provided there is no direct effect of ancestry on the final outcome ($O$) that bypasses the intermediate trait ($T$).

Causal diagrams illustrating overcorrection (left) and the conditions for unbiased causal inference (right).

Key Insights

  • The Proxy Gap: Observed ancestry ($A_0$) is rarely a perfect substitute for true ancestry ($A$).
  • Environmental Confounding: Genetic corrections often fail to account for non-genetic paths mediated through environment and history.
  • Methodological Robustness: Causal inference methods like Mendelian Randomization can be robust to population structure under specific structural assumptions.

Conclusion

While these revised models are simplifications, they highlight the necessity of formal causal thinking in genetic epidemiology. Researchers must look beyond simple ancestry adjustment and consider the complex interplay of geography, history, and environment to truly disentangle genetic effects from population structure.

Summary updated

BibTeX
@inproceedings{lawson2019correction,
  author = {Daniel J. Lawson and Neil M Davies and Simon Haworth and Bilal Ashraf and Laurence J Howe and Andrew Crawford and Gibran Hemani and George Davey Smith and Nicholas J. Timpson and Daniel John Lawson},
  title = {{Correction to: Is population structure in the genetic biobank era irrelevant, a challenge, or an opportunity?}},
  booktitle = {Human},
  year = {2019},
  doi = {10.1007/s00439-019-02051-3},
}
Figure