Hi,

Im currently using limma to interrogate methylation data using multiple regression models, typically in the format of methyaltion (beta) values ~ phenotypic variable + covariate 1 +...

However I was wondering, what is / is there a benefit or difference for an OLS regression using the lmFit, eBayes and toptable functions of limma to generate results vs. using multiple testing adjusted lm function for each cg?

We've conducted both limma-based and lm-based (parallelised) regressions using the same data, same regression model and seen ~10% difference in the top 100k hits (when arranged by P.Value).

Thanks, Matt

Yes, there's a big benefit from limma's empirical Bayes approach in terms of improved statistical power and better ROC curves, especially if your sample sizes are small.

There are plenty of published papers that show this. Just follow the references given in the limma documentation or do a Google Scholar search for published papers.