A typical GWAS for one disease includes one logistic regression per SNP, or 500,000 or more statistical tests. These tests are not all independent as SNPs that are located close to one another can be correlated due to linkage disequilibrium. Even so, with 105–106 statistical tests, very small p-values by conventional standards are expected by chance. As noted above, p-values < 5×10−8 (akin to a Bonferroni correction of the traditional 0.05 Type 1 error level for 1,000,000 statistical tests) (Pe’er et al., 2008) are generally required for significance. Experience suggests that findings more significant than this threshold tend to replicate well across studies. However, unless power is exceptional, it is generally incorrect to exclude a SNP from consideration if does not exceed this threshold. Indeed, some SNPs that are unimpressive in an initial study (e.g., p=0.001) can eventually replicate well and exceed the critical threshold. As emphasized above, replication is essential.
