Hi Gordon, Tim, First of all, let me thank you for your answers. They have been very helpful. Using Gordon's recommendations (paired design, robust method and eBayes with a trend for the prior), I've been able to extract a subset of differentially methylated probes big enough for us to continue with our downstream analyses. That, at least, makes the biologists in our group a little bit happier. :-) Specially if we take into account that this is a very "hard" dataset, with which I have been struggling for months. I already knew Alicia Oshlack's work. In fact, the only normalization method we are currently applying to this upstream analysis is SWAN, which happens to correct for the different Infinium 450k probe designs. I am currently thinking of trying the normalization method of Touleimat and Tost implemented in the latest versions of the minfi package but, for now, we are happy with SWAN. Regarding the special variance structures in methylation data, I think this is something that is currently out of reach for me. I'll explain myself: I am inclined to think that methylation data has a special structure that deserves to be treated accordingly. Problem is, I think I do not have enough mathematical background to understand the subtleties of such a procedure. Fortunately, by writing in this list, I can learn a lot from answers like yours, while I try to continue studying statistics on my own. Tim works a lot with epigenetic information and has always been very helpful to me. As an example, I am aware of the correlation exhibited by nearby probes. But I wouldn't know how to include that into a model. Maybe, for example, as the bumphunter() function in minfi package does, could we include a dummy variable representing blocks of probes that vary accordingly in order to represent the clustered relationship of those regions? Thank you again for your help. Regards, Gustavo 2013/11/30 Gordon K Smyth <smyth@wehi.edu.au> > Hi Tim, > > I have never analysed data from any of the newer DNA methylation array > technologies, so you should ask someone who has how well the limma pipeline > seems to work on that type of data. > > For example, Alicia Oshlack's group develops methods for methylation > arrays and finds limma useful: > > http://genomebiology.com/content/13/6/R44 > > It seems very reasonable to me to suppose that methylation data has > special variance structures that could be advantageously taken into > account. I doubt though that this means abandoning the empirical Bayes > approach entirely, as it is not very dependent on normality or independence > and gives some benefit in a wide range of situations. You will notice that > I recommended to the original poster that they used eBayes() with > trend=TRUE. > > Best wishes > Gordon > > > On Fri, 29 Nov 2013, Tim Triche, Jr. wrote: > > Is eBayes generally regarded as appropriate for DNA methylation microarray >> probes, i.e. is the general consensus that departure from normality and >> shared variance structure is "normal enough" to shrink towards a global >> prior distribution whose parameters have been estimated from a potentially >> mixed population? e.g. concerns could involve the probe type (paired vs. >> unpaired), variance structure (which with m-values, one assumes these are >> at least approximately normal), and the degree of sharing (e.g. nearer >> probes tend to be more correlated than further probes, and probes in >> certain regions tend to be inherently more variable across datasets than >> probes in other regions). >> >> I've been meaning to ask for your thoughts on this for quite some time :-) >> >> Best, >> >> --t >> >> *He that would live in peace and at ease, * >> *Must not speak all he knows, nor judge all he sees.* >> >> Benjamin Franklin, Poor Richard's >> Almanack<http: archive.org="" details="" poorrichardsalma00franrich=""> >> > > ______________________________________________________________________ > The information in this email is confidential and inte...{{dropped:10}}