I apologize if this is a duplicate question,I read limmas paper (filtering section) it referred that after normalizing microarray data, for downstream analysis, it is usually worthwhile to remove probes that appear not be expressed but on the other hand in this post Non-specific filtering methodogies for ExpressionSet in R/Bioconductor , it referred that, filtering is not much needed if we use limma for microarray analysis , now if you think it is better to apply filtering before the linear modelling and empirical Bayes steps, please let me know what kind of filtering I can use without conflict with distribution of variance ?
I think it is reasonable to filter out probe sets that can be confidently identified as non-expressed before the eBayes step. The purpose of eBayes is to squeeze the variance of each probe set toward the average variance of all probe sets, under the assumption that this average is a reasonable representation of the overall behavior of the data. Ideally, that average should only reflect the biology of your system, not the technical noise in probes that aren't measuring anything. So removal of non-expressed probes should help ensure that. However, it might make little difference in practice.
One thing you definitely want to avoid is to filter probe sets based on their variance. Doing so would significantly bias the eBayes estimate of the average variance of the data set and compromise your entire analysis.
actually the approach of "non-specific intensity filtering", is beneficial prior using limma, in order to remove non-interesting probes(i.e. probes that not expressed in the majority of your samples, & also to reduce the number of multiple hypothesis tests). On the other hand, is also data-dependent and it's up to you if and how you could filter your data prior statistical inference.