In a GWAS, logistic regressions are done for every SNP (i.e., a total of ~1 million regression models). Given the number of statistical tests, p-values that are very small by traditional standards are to be expected merely by the play of chance (e.g., 10 p-values < 0.00001 and 100 p-values < 0.0001). Thus, the standard of evidence that has emerged for a compelling GWAS finding is rigorous: (a) a strong association in an initial sample, (b) precise replication in one or more independent samples (i.e., the same SNP, allele, and direction of association), and (c) a cumulative p-value < 5×10−8 (Chanock et al., 2007). The 5×10−8 threshold is akin to a Bonferroni correction of the traditional 0.05 Type 1 error level for 1,000,000 statistical tests (although the full argument is more complex as some of these tests are not independent due to linkage disequilibrium) (Pe’er et al., 2008). P-values that are smaller than expected by chance and which replicate well in other samples highlight a genomic region associated with a disorder (and potentially causal).
