|
Abstract:
|
To date, the genetic architecture of most complex traits supports a highly polygenic model in which common variants have very small effect sizes, and rarer variants with potentially larger effect sizes. Sample size is the biggest barrier to identifying associations between genotype and phenotype, and in GWA studies, summary statistics shared between different studies can be combined in a meta-analysis to gain power in identifying associations. However, it is not uncommon for a highly significant meta-analysis test statistic to be dominated by the results of only one or a few studies. This could be the result of different underlying study populations with different allele frequencies, sequencing errors, imputation problems, or other artifacts of a particular study. This work focuses on methodology for classifying reproducible effects from potential false-positives, as well as estimating the power and sample size needed to replicate a significant effect in follow-up studies. This work examines several approaches to modeling reproducible effects, including likelihood based scoring and mixture models. We demonstrate the advantages of our approach compared to previous methodology.
|
