Figure 4 illustrates the differences in correlation structure for two chips. For each HapMap SNP we found it's best “tag” (the SNP on the chip with which it has the highest r 2) and generated a histogram of these maximized r 2 values. To recover coverage we simply count the proportion of SNPs for which the best tag r 2 is ≥0.8, coloured red in the bottom row of figure 4. In this sense, informally, it is useful to think of coverage as assuming that there is power one for every “tagged” SNP and no power for every other SNP. This is of course false, in ways which help to explain why coverage differences do not translate into power differences. When a SNP is common and the effect size is moderate or large, there will still be good power to detect it even if the best SNP on the chip only has r 2 = 0.5 or less. At the other extreme, for rare SNPs, unless the effect size is very large, power would be low even if the SNP
