In modern statistical genetics, researchers rarely work with raw, individual-level sequence data for every study. Instead, they rely on summary statistics generated by Genome-Wide Association Studies (GWAS). These summary statistics allow teams to perform powerful downstream analyses, such as estimating genetic correlations between different diseases, running large-scale meta-analyses, or conducting Mendelian Randomization (MR) to infer causal relationships between exposures and clinical outcomes.
Almost all of these downstream frameworks share a foundational mathematical assumption: sample independence. They assume that the cohort used to study Trait A consists of entirely different human beings than the cohort used to study Trait B.
However, the rise of massive, shared biobanks (such as the UK Biobank) has broken this assumption. Because these centralized repositories are incredibly rich, independent research teams frequently use the exact same participant populations to study different traits. This creates sample overlap—a hidden redundancy where the same individuals are counted multiple times across different datasets. When statistical models assume independence but are fed overlapping data, the mathematical machinery breaks down, generating systematic biases and false discoveries.
Why Overlap Mimics Biology
To understand why sample overlap is so destructive, consider what happens when we try to estimate the genetic correlation between two distinct traits. We want to know if the same biological pathways influence both phenotypes (pleiotropy).
If two cohorts are completely independent, the only reason their summary statistics should consistently correlate across the genome is a shared biological mechanism. But if the cohorts contain the same individuals, a second source of correlation appears: environmental and technical noise. If a participant has a unique environmental factor, a specific lifestyle habit, or even a shared measurement error in their data, that non-genetic correlation will manifest across both datasets.
As Zhang et al. demonstrate in Comparison of methods for estimating genetic correlation between complex traits using GWAS summary statistics, these shared participants introduce technical correlations that are mathematically indistinguishable from true biological genetic correlation. If a statistical method cannot differentiate between the two, it will mistake redundant sampling for genuine genetic architecture.
In a single-author study, Zhao notes in On Genetic Correlation Estimation With Summary Statistics From Genome-Wide Association Studies that this overlap introduces complex biases that can either artificially inflate or deflate genetic correlation estimates. The direction and severity of the bias depend entirely on the nature of the underlying non-genetic correlations, making it an elusive target to correct without specialized statistical adjustments.
The Breakdown of Causal Inference
The stakes are even higher in Mendelian Randomization (MR). MR uses genetic variants as instrumental variables to mimic a randomized controlled trial, allowing researchers to determine if an exposure (like vitamin levels) causally changes an outcome (like cancer risk).
"Two-sample MR" is designed specifically to avoid the classic "weak instrument bias" that plagues single-sample studies. By measuring the genetic effects on the exposure in Sample 1 and the genetic effects on the outcome in Sample 2, any random error in the first sample should be completely uncorrelated with the error in the second sample.
When samples overlap, this firewall vanishes. As Cheng et al. explain in MR-LDP: a two-sample Mendelian randomization for GWAS summary statistics accounting for linkage disequilibrium and horizontal pleiotropy, participant overlap causes the causal estimates to become highly unreliable, often biasing the results back toward the observational, confounded association.
This is not a theoretical problem; it routinely limits real-world epidemiological discoveries. For instance:
- In a study on cardiovascular risk factors, de Ruiter et al. noted in Smoking as a Risk Factor for Cardiovascular Disease in Females and Males: Observational and Mendelian Randomisation Analyses in the UK Biobank that overlap between the discovery GWAS and the outcome data threatens the validity of the causal inference.
- Similarly, Gao et al., writing in Genetically Determined Plasma Docosahexaenoic Acid Showed a Causal Association with Female Reproductive Longevity-Related Phenotype: A Mendelian Randomization Study, acknowledged that overlapping samples between major databases like the UK Biobank and ReproGen introduce severe risks of Type I errors (false positives).
To protect against this, researchers must either strictly isolate their cohorts—as emphasized by Xu et al. in Identifying the mediating role of socioeconomic status on the relationship between schizophrenia and major depressive disorder: a Mendelian randomisation analysis, who enforced independent cohorts for exposure, mediator, and outcome variables—or apply explicit mathematical corrections to adjust for the shared variance, as discussed by Qi et al. in Genetically predicted vitamins supplementation and risk of skin cancers: a Mendelian randomization study.
Overfitting and False Replications
Beyond correlation and causation, sample overlap degrades our ability to predict disease risk and validate discoveries.
When building polygenic risk scores (PRS) that combine thousands of genetic variants into a single predictive index, models must be trained on one dataset and tested on another. If the same individuals exist in both datasets, the model essentially "cheats." It memorizes the specific noise of those individuals rather than learning generalizable biological signals. Maier et al. demonstrate in Improving genetic prediction by leveraging genetic correlations among human diseases and traits that this participant redundancy leads to severe statistical overfitting, creating artificially inflated accuracy estimates that vanish when the model is applied to truly independent populations.
A parallel failure occurs in large-scale meta-analyses. When combining multiple genetic studies to boost statistical power, the standard assumption is that each study brings new, independent observations to the table. If studies secretly share cohorts, the sample size is effectively lower than the model believes. As Weissenkampen et al. show in Model-based assessment of replicability for genome-wide association meta-analysis, this correlation inflates statistical significance, generating highly confident "discoveries" that fail to replicate because they were artifacts of redundant sampling.
Algorithmic Solutions and Trade-offs
Because we cannot stop researchers from utilizing the same high-quality biobanks, the solution must be algorithmic. The field has evolved toward frameworks that either explicitly model participant redundancy or are mathematically insulated from it.
One approach is to build joint probabilistic models that account for shared sample structures from the ground up. For example, Yang et al. introduced LPG: A four-group probabilistic approach to leveraging pleiotropy in genome-wide association studies, which models pleiotropy while adjusting for the fact that individuals may appear across multiple datasets. Similarly, to prevent genetic confounding in observational studies, Zhao et al. developed the PENGUIN framework in Controlling for polygenic genetic confounding in epidemiologic association studies, utilizing variance component estimation to maintain unbiased results even when working with highly overlapping biobank resources.
However, accounting for overlap often requires a fundamental trade-off between precision and robustness. In their comprehensive benchmark, Zhang et al. revealed that highly precise, likelihood-based methods (such as HDL) offer superior statistical power but are acutely sensitive to sample overlap, breaking down when independence is violated. Conversely, more traditional summary-statistic methods (like LDSC and GNOVA) trade away a degree of precision to achieve crucial robustness, successfully distinguishing genuine biological overlap from the technical artifacts of shared participant populations.
Key Foundational Work
Improving genetic prediction by leveraging genetic correlations among human diseases and traits (2018) — contributed 2 concept elements to this direction.
LPG: A four-group probabilistic approach to leveraging pleiotropy in genome-wide association studies (2018) — contributed 1 concept elements to this direction.
MR-LDP: a two-sample Mendelian randomization for GWAS summary statistics accounting for linkage disequilibrium and horizontal pleiotropy (2019) — contributed 1 concept elements to this direction.