The repeatability of DNA methylation measures may also affect the power of epigenome-wide association studies: Table 1.

We read with great interest the article of Tsai and Bell recently published in this journal.1 The authors conducted an interesting simulation study to derive power and sample size estimates to be used in epigenome-wide association studies (EWAS). Their study, and others,2,3 did not explicitly consider that DNA methylation values may be subject to measurement error. Whereas measures (genotypes) used in genome-wide association studies are highly reliable,4 this does not appear to be the case for EWAS. Recent studies have assessed the repeatability of the Illumina450K assay by examining intraclass correlation coefficients (ICC) between technical replicates of leukocyte DNA samples, and reported median ICCs for methylation β-values of 0.305 and 0.36.6 Using DNA extracted from Guthrie cards (dried blood spots) and whole-blood cell lymphocytes from our case-control studies of various cancers,7 we computed ICCs based on 129 technical replicate pairs. We observed similar median ICCs for β-values of 0.36 [interquartile range (IQR): 0.13 to 0.67] for Guthrie cards (based on 67 pairs) and 0.50 (IQR: 0.26 to 0.76) for lymphocytes (based on 62 pairs), indicating that for a substantial proportion of probes in the Illumina450K assay, the repeatability of methylation level is moderate to low. Applying a batch-effect removal method (ComBat8) did not meaningfully change the ICCs. Although heterogeneity in cell composition may explain somewhat lower ICCs for DNA obtained from Guthrie cards, this is less likely for lymphocyte DNA samples. In tissues other than blood, the repeatability of methylation measures might be higher, but to our knowledge this has not been reported in the literature. To complement the findings of Tsai and Bell, we estimated the statistical power of EWAS under several scenarios of ICC values. We assumed a normal distribution for methylation values and that measurement error was non-differential.10,Table 1 shows that the ICC also strongly influences power. This is consistent with the findings presented by Tsai and Bell in Figure 5,1 showing the loss of power associated with increasing pooled standard deviation. EWAS power under various scenarios for the repeatability of methylation measuresa aICC: intraclass correlation coefficient (1 represents perfect repeatability). We assumed a sample size of 50 case-control pairs and a standard deviation (without measurement error) of 0.15, as found in ref. 1, although we observed smaller standard deviations in our samples (typically ranging from 0.01 to 0.10). The power is expressed as a percentage, and was calculated assuming a normal distribution for methylation values and non-differential measurement error.10 bMean difference in methylation β-values between cases and controls. Methylation β-values range from 0 to 1. EWAS power under various scenarios for the repeatability of methylation measuresa aICC: intraclass correlation coefficient (1 represents perfect repeatability). We assumed a sample size of 50 case-control pairs and a standard deviation (without measurement error) of 0.15, as found in ref. 1, although we observed smaller standard deviations in our samples (typically ranging from 0.01 to 0.10). The power is expressed as a percentage, and was calculated assuming a normal distribution for methylation values and non-differential measurement error.10 bMean difference in methylation β-values between cases and controls. Methylation β-values range from 0 to 1. Interestingly, the strongest and most consistent signals of differentially methylated markers of smoking in peripheral blood have been discovered at CpG sites with high repeatability.5,9 Since CpG sites with low repeatability (low ICCs) tend to either be measured with substantial error or exhibit very little variation in the population (or both), the ICC for a given CpG site might provide valuable information about whether differential methylation between cases and controls is likely to be observed. For a given significance threshold, probes with lower ICCs have a relatively low power, and measurement error biases effect estimates towards the null. Probes for which power is insufficient to detect meaningful methylation differences should be interpreted with caution, as statistically significant results are more likely to represent false-positive signals.11