Abstract
Almost every recent Alzheimer's disease (AD) genome-wide association study (GWAS) has performed meta-analysis to combine studies with clinical diagnosis of AD with studies that use proxy phenotypes based on parental disease history. Here, we report major limitations in current GWAS-by-proxy (GWAX) practices due to uncorrected survival bias and non-random participation of parental illness survey, which cause substantial discrepancies between AD GWAS and GWAX results. We demonstrate that current AD GWAX provide highly misleading genetic correlations between AD risk and higher education which subsequently affects a variety of genetic epidemiologic applications involving AD and cognition. Our study sheds important light on the design and analysis of mid-aged biobank cohorts and underscores the need for caution when interpreting genetic association results based on proxy-reported parental disease history.
Introduction: The Education Paradox
Genome-wide association studies (GWAS) require massive sample sizes to detect the subtle genetic signals associated with complex diseases. For late-onset conditions like Alzheimer’s Disease (AD), researchers face a significant hurdle: large population biobanks, such as the UK Biobank, consist primarily of middle-aged individuals who are often too young to have received a clinical AD diagnosis.
To circumvent this, researchers have increasingly adopted GWAS-by-proxy (GWAX). In this approach, participants report whether their parents had AD, effectively turning the participant into a genetic proxy for their parents. While GWAX has successfully identified major risk genes like APOE, this study reveals that the method fails at the "infinitesimal" level—the thousands of small-effect variants that determine genetic correlations and polygenic risk scores. Most notably, GWAX produces the counter-intuitive finding that higher education increases AD risk, whereas clinical data consistently shows it is protective.
Methodology
The researchers employed several genomic tools to dissect these systematic biases:
- GSUB (GWAS-by-subtraction): A new, computationally efficient method to "regress out" real AD signals from GWAX data to isolate the bias component.
- Bias Correction Benchmarking: Testing existing methods like the Marioni approach (filtering by parental age) and the Jansen approach (creating a continuous "disease load" phenotype).
- Datasets: The study utilized the UK Biobank for GWAX data, the Health and Retirement Study (HRS) for validation of cognitive decline prediction, and the All of Us Research Program to investigate participation bias.
The Survival and Participation Trap
The study identifies two primary drivers of bias in GWAX:
- Survival Bias: To be a "proxy case," a parent must live long enough to develop AD. This inadvertently selects for longevity genes. Consequently, GWAX shows a spurious negative correlation between AD and heart failure—not because of biological protection, but because those with high genetic risk for heart failure die before they can develop AD.
- Participation/Reporting Bias: Individuals who report their parents' medical history tend to have higher educational attainment and health awareness. Because the tendency to answer surveys is itself heritable, this trait bleeds into GWAX results, creating a false positive genetic correlation between education and AD.
Key Results
Flipped Correlations and the Education Paradox
The most striking finding is the reversal of genetic correlations. Clinical GWAS shows a protective genetic correlation between AD and education (approx. -0.13). In contrast, standard GWAX shows a positive correlation (approx. +0.17), suggesting education is a risk factor.
Attenuation of Effect Sizes
While the effect sizes in GWAX correlate strongly with clinical GWAS (Pearson correlation = 0.97), the actual magnitude of these effects is significantly attenuated. The OLS slope of 0.63 indicates that GWAX consistently underestimates the strength of genetic associations compared to clinical case-control studies.
Polygenic Risk Score (PRS) Performance
Polygenic Risk Scores built from GWAX data performed significantly worse at predicting actual cognitive decline in the independent HRS cohort compared to PRS built from traditional clinical GWAS. This suggests that the biases in GWAX severely limit its utility for precision medicine and risk prediction.
Discussion and Future Directions
While the Jansen approach (using age-based disease load) successfully corrected the direction of the education correlation, it does not yet produce SNP effect sizes on a scale comparable to traditional studies. The authors emphasize that clinical case-control GWAS remains the "gold standard," though it may have its own unique survival biases.
This research serves as a critical call for transparency. The authors urge the scientific community to stop releasing only meta-analyzed (GWAS + GWAX) results. Instead, separate summary statistics should be provided to allow researchers to account for these pervasive biases in downstream genetic epidemiologic applications.
Summary updated