Abstract:
|
Selection of best results from genetic association studies introduces a winner's curse bias. Top hits of a study selected based on largest estimated effect sizes or smallest P-values tend to have more modest estimates upon independent re-evaluation. Bayesian methods including those concerned with evaluating the proportion of false discoveries among top hits are immune to winner's curse, provided that priors, such as the effect size distribution, and the frequency of false positives accurately reflect the reality. Common methods assume that false positives represent a large portion of findings and have exactly zero effect size, while effect sizes for true signals follow a specific distribution, such as normal. In reality, there is a large portion of variants carrying tiny, practically undetectable effects, while variants with larger effect sizes are progressively less abundant. Thus, we assume a model without a special null class of variants and rather define the effect size distribution that includes a class of variants with negligible effect size magnitude. This model is realistic, flexible and gives a way for combining association signals across several studies.
|
ASA Meetings Department
732 North Washington Street, Alexandria, VA 22314
(703) 684-1221 • meetings@amstat.org
Copyright © American Statistical Association.