the null hypothesis. Figure 2a demonstrates this effect on the 438 whole exomes. The PLINK/SEQ suite computes from data the so-called i-stat, which is an estimate of the minimal achievable p-value for a gene. The i-stat can be used by setting a threshold (e.g., 10−3) and only correcting for the number of genes that have the i-stat below the threshold following the idea that for the genes with i-stat above the threshold there is no power to find an association. Another way to correct for multiple testing is to compute an experiment-wide significance threshold by permutations of phenotype labels, create the empirical distribution of minimal p-values for all genes across permutations, and compare the minimal p-value from the real data to that distribution (Figure 2b). This approach efficiently controls Type-I error and is less conservative than the Bonferroni correction. Importantly, the p-value threshold computed by permutations is dependent on both the study and on the statistical test. However, the experiment-wide correction via permutation is not robust to confounding and it is essential to assess the quality of the distribution of test statistics, for those genes that have i-stats less than the threshold, to ensure appropriate calibration of the distribution. Nevertheless,
