Participation bias in the estimation of heritability and genetic correlation

Shuang Song, Stefania Benónísdóttir, Jun S. Liu, Augustine Kong

TL;DR

Participation bias—the fact that biobank volunteers aren't a random slice of the population—doesn't just skew demographic data; it fundamentally biases our estimates of heritability and genetic correlation. This paper introduces a novel statistical framework to correct these estimates by accounting for the genetic component of the "inclination to participate," revealing that we have likely been underestimating the genetic basis of many complex traits.

What Problem Does This Paper Address?

Large-scale genomic databases like the UK Biobank (UKBB) are the bedrock of modern genetics, but they suffer from a major flaw: participation bias. In the UKBB, only about 5.5% of invited individuals actually participated. These volunteers tend to be healthier, wealthier, and more educated than the general population.

When researchers estimate heritability (how much of a trait's variation is due to genetics) or genetic correlation (how much two traits share the same genetic roots), they usually assume random sampling. If participation itself has a genetic component—which it does—then "selecting" for participants creates a "collider bias." This distorts the relationship between genotypes and phenotypes in the sample, leading to misleading results that don't reflect the actual population.

Methods

The authors propose a correction method based on a liability-threshold model. They treat participation as a binary outcome of an underlying continuous "liability" score ($X$).

Key features of their approach include:

  • IBD-based Participation Statistics: They leverage a recent technique that uses "Identity By Descent" (IBD) segments shared between relatives to estimate the genetic component of participation without needing data from people who didn't join the study.
  • Mean Shift Analysis: They use the difference in average phenotype values (e.g., average BMI) between the biobank sample and a representative population survey to estimate the phenotypic correlation between that trait and participation.
  • Decoupling Correlations: Unlike previous methods (like Inverse Propensity Weighting), this model separates the genetic correlation from the non-genetic correlation between participation and other traits. It avoids the flawed assumption that the genetics of participation are entirely explained by other observed phenotypes like education.

Datasets

  • UK Biobank (UKBB): The primary dataset, specifically focusing on ~274,000 unrelated individuals of white British ancestry.
  • Health Survey for England (HSE): Used as the "representative population" baseline. Because the HSE uses a two-stage random probability sampling process and weights for non-response, it serves as a more accurate reflection of the general UK population.
  • GWAS Catalog: For participation summary statistics derived from IBD-based information.

Key Insights and Ideas

  • Participation is a Complex Trait: The authors argue that the "inclination to participate" should be treated as a complex behavioral trait in its own right, with its own unique genetic architecture, rather than just a byproduct of being educated or healthy.
  • The Underestimation Bias: Theoretically, participation bias can push estimates in either direction. However, for most traits studied, the authors found that ignoring participation bias leads to underestimation. For example, if a trait is genetically correlated with participation, the genetic variance in the sample shrinks more than the total variance, making the trait appear less "genetic" than it really is.
  • Collider Bias in Action: By conditioning on participation (an outcome influenced by both genetics and environment), an artificial negative correlation is introduced between those factors in the sample. This "washes out" some of the true genetic signals researchers are looking for.

Key Results

The authors applied their adjustment to 12 UKBB phenotypes (including BMI, height, education, and smoking):

  • Significant Correlations: 8 out of 12 phenotypes showed significant genetic correlations with participation.
  • Heritability Adjustments: For traits like Educational Attainment (EA), the heritability estimate increased from 0.203 to 0.217 after adjustment.
  • Genetic Correlation Shifts: The absolute values of genetic correlations between traits generally increased. For instance, the genetic correlation between Waist Circumference (WC) and EA was significantly higher after accounting for the fact that people with high EA and low WC are overrepresented in the biobank.
  • Smoking Status: "Smoke now" (current smoking status) showed a significant genetic correlation with participation only after the adjustment was applied ($p = 0.002$), highlighting how bias can hide real genetic relationships.

Limitations and Future Work

  • Direct vs. Indirect Effects: The current model uses participation statistics that capture direct genetic effects; it does not yet account for indirect "genetic nurture" effects from parents.
  • Sample Size: The IBD-based method for estimating participation genetics relies on sibling pairs, which are a relatively small subset of biobanks, limiting the precision of the adjustment.
  • Representative Baseline: The method relies on the HSE being a perfectly representative sample. If the baseline survey itself has hidden biases, the adjustments will be less effective.
  • Causality: Future work is needed to determine the direction of causality—does higher education cause people to participate more, or does the "inclination to participate" share a common biological pathway with cognitive traits?

Summary updated

Abstract

It is increasingly recognized that participation bias can pose problems for genetic studies. Recently, to overcome the challenge that genetic information of non-participants is unavailable, it is shown that by comparing the IBD (identity by descent) shared and not-shared segments among the participants, one can estimate the genetic component underlying participation. That, however, does not directly address how to adjust estimates of heritability and genetic correlation for phenotypes correlated with participation. Here, for phenotypes whose mean differences between population and sample are known, we demonstrate a way to do so by adopting a statistical framework that separates out the genetic and non-genetic correlations between participation and these phenotypes. Crucially, our method avoids making the assumption that the effect of the genetic component underlying participation is manifested entirely through these other phenotypes. Applying the method to 12 UK Biobank phenotypes, we found 8 have significant genetic correlations with participation, including body mass index, educational attainment, and smoking status. For most of these phenotypes, without adjustments, estimates of heritability and the absolute value of genetic correlation would have underestimation biases.
BibTeX
@inproceedings{song2025participation,
  author = {Shuang Song and Stefania Benónísdóttir and Jun S. Liu and Augustine Kong},
  title = {{Participation bias in the estimation of heritability and genetic correlation}},
  booktitle = {Proceedings of the National Academy of Sciences},
  year = {2025},
  doi = {10.1073/pnas.2425530122},
}
Figure