Biases in GWAS – the dog that did not bark

Abstract Background Genome wide association studies (GWAS) of specific diseases are central to scientific discovery. Bias from inevitably recruiting only survivors of genetic make-up and disease specific competing risk has not been comprehensively considered. Methods We identified sources of bias using directed acyclic graphs, and tested for them in the UK Biobank GWAS by making comparisons across the survival distribution, proxied by age at recruitment. Results Associations of genetic variants with some diseases depended on their effect on survival. Variants associated with common harmful diseases had weaker or reversed associations with subsequent diseases that shared causes. Conclusion Genetic studies of diseases that involve surviving other common diseases are open to selection bias that can generate systematic type 2 error. GWAS ignoring such selection bias are most suitable for monogenetic diseases. Genetic effects on age at recruitment may indicate potential bias in disease-specific GWAS and relevance to population health.