Self-Report Alzheimer's Disease Statuses in UK Biobank Distort Downstream Analyses

Abstract

Genetic studies have identified Alzheimer's disease (AD)-linked variants through genome-wide association studies (GWAS) and proxy-based GWAS (GWAX), yet inconsistent causal inferences still persist. Here, we systematically evaluated how self-reported AD diagnoses in the UK Biobank (UKB) distort Mendelian randomization (MR) analyses. Using seven AD datasets (four GWAS and three GWAX, including IGAP, N = 63 926; FinnGen R10, N = 191 061; UKB G30, N = 420 531; UKB2024, N = 434 286, and GWAX2017, N = 74 366; GWAX2018, N = 548 955; GWAX2021, N = 408 691) and six education subtypes (including years of schooling, N = 1 131 881; hardest math class completed, N = 430 445; self-reported math ability, N = 564 698; college completion, N = 280 007; cognition test performance, N = 257 841; and non-cognitive skills, N = 257 841). We also assessed the heterogeneity and pleiotropy across these datasets. We found opposing causal directions between GWAS and GWAX cohorts. In GWAS datasets, genetic variations related to education were causally linked to a lower risk of AD (OR < 1, p < 0.05), with years of schooling showing the strongest protective effects (OR = 0.71 in IGAP, p < 0.05). Conversely, UKB-based GWAX analyses paradoxically linked education-related traits to increased AD risk (OR > 1, p < 0.05), directly conflicting with the protective associations in clinical AD GWAS results. Genetic heterogeneity was observed in both AD GWAS and GWAX datasets. Pleiotropy was noted in AD outcomes, but MR estimates remained stable after outlier adjustments. Our findings reveal that self-reported AD statuses in UKB distorted genetic effect estimates, particularly for education subtypes requiring validation. The research urges caution in interpreting MR results from GWAX studies that use self-reported endpoints and highlights the need for rigorous phenotyping in biobank studies.