Of course, tested SNPs are not chosen randomly, but the relative lack of success of candidate–gene association studies, apparently well powered for nominal significance, suggests that the prior odds have generally been exaggerated [Ioannidis, 2005]. However, a genomewide scan, for which SNPs are provided by a commercial supplier of genotyping chips, is in a sense little different from a random selection, except that the prior odds are substantially shortened. Yet, unless the chip has complete coverage of the genome, there remains a discrepancy between the number of SNPs tested and the actual prior odds of any SNP being associated. Our view is that a responsible use of P–value thresholds must reflect the multiplicity of the whole genome, rather than of the set of SNPs tested. This is in line with the previous approaches [International HapMap Consortium, 2005; WTCCC, 2007] and facilitates the comparison of results between different marker panels as well as those obtained by imputation methods that estimate the genotypes of markers not directly typed [Marchini et al., 2007]. However, permutation methods for multiple test correction only take account
