TY - JOUR T1 - Beta regression improves the detection of differential DNA methylation for epigenetic epidemiology JF - bioRxiv DO - 10.1101/054643 SP - 054643 AU - Timothy J. Triche, Jr. AU - Peter W. Laird AU - Kimberly D. Siegmund Y1 - 2016/01/01 UR - http://biorxiv.org/content/early/2016/05/23/054643.abstract N2 - Background DNA methylation is the most readily assayed epigenetic mark, possessing confirmed relationships with gene expression, imprinting, and chromatin accessibility.Given the increasingly widespread use of DNA methylation microarrays in population-scale epidemiological applications, we sought to determine which methods provided the greatest statistical power to reproducibly detect differences in DNA methylation across various conditions,using publicly available data sets on tissue type and aging.Results Beta regression, as proposed originally by Ferrari and Cribari-Neto, yielded more validated hits in each of our comparisons than any other method under consideration, both in a regression setting and in comparisons to two-group tests such as the Wilcoxon-Mann-Whitney, Student t, and Welch t tests.In large cohorts of whole blood samples, we corrected for compositional differences and batch effects, and found that marginal likelihood ratio tests from beta regression models uniformly dominate popular alternatives based on linear models.The superior sensitivity and specificity exhibited by beta regression in epidemiologically relevant cohort sizes corresponded to approximately a 2% increase in sensitivity at the same specificity when compared to linear models fitted on raw beta values (proportion of signal intensity due to the methylated allele), M-values, or rankquantile normalized values.Conclusions Investigators should consider beta regression to maximize statistical power in studies of DNA methylation using microarrays.At epidemiologically relevant sample sizes, with typical quality control procedures (compositional and batch effect correction), cross-cohort agreement uniformly favors beta regression over popular alternatives. ER -