Hi Jim, thanks for your reply, you are a 'reference' in this and other microarray related lists! However my doubts regard the idea of filtering out probes before statistical testing in order to reduce the potentially false positives. Usually after filtering genes on the basis of their A/P calls, I filter out the genes without relevant fold changes making ratios of means. At this point I run a statistical test with multiple hypothesis correction and then do data mining based on clustering, annotation and so on. The lists I produce for the users with 'candidates genes' has also an indication of fold changes that users try to biologically validate. If the data were always very reproducible, arithmetic means and geometric means should return a similar result, but this is not always the case. From all this my doubts. As you can see I use means at two step: 1) limiting step -> to filter out not changing genes. Could this be a wrong approach? But I noticed this could give better results when you apply statistical tests. 2) estimate of observations -> to give our user an 'indicative measure' of fold changes that they should validate. It would be great to chose the 'nearest to biology'. Thank you Regards Remo Sanges BioGeM On Mar 11, 2004, at 2:59 PM, James MacDonald wrote: > Hi Remo, > > If you have triplicates, I would recommend using a statistical test > (t-test, F-test, etc.) to determine which genes are differentially > expressed. You probably won't have much power to detect differences, > but > you can try to increase your power by using an empirical Bayes > adjustment to your variance estimate (using EBayes or limma packages). > > In general you would use log transformed data for most statisitical > tests, so if you want to filter your data further using fold change, > you > should use geometric means. > > HTH, > > Jim > > > > James W. MacDonald > Affymetrix and cDNA Microarray Core > University of Michigan Cancer Center > 1500 E. Medical Center Drive > 7410 CCGC > Ann Arbor MI 48109 > 734-647-5623 >>>> <sanges@biogem.it> 03/10/04 9:05 PM >>> > : > > I apologize if this topic is a little bit OT, > but I think this is the better list in which discuss about > mathematical and statistical issues relating Affymetrix chips. > > I work in a service but I have very poor feedback from users. > Furthermore I am a biologist so I ask your opinion based on your > experience and knowledge. > > Generally I use in the analysis gcrma background subtraction, > quantile normalization, pm-only and medianpolish. > Usually this analysis are conducted on an experiment in which > each point has a biological triplicate. > > Now I am thinking at the more robust way to infer fold changes. > For robust I means 'nearest to biology' and 'statistically acceptable'. > My problem is how to summarize replicate signals for each probe before > compute ratios. > > What do you think is the best way to have a mean signal from > replicates? > Arithmetic mean or geometric men? > Is one of this approach wrong? > Is the choose of the approach dependent from the homogeneity of > replicates? > > Thank you > > Best Regards > > Remo Sanges > BioGeM > > _______________________________________________ > Bioconductor mailing list > Bioconductor@stat.math.ethz.ch > https://www.stat.math.ethz.ch/mailman/listinfo/bioconductor > > _______________________________________________ > Bioconductor mailing list > Bioconductor@stat.math.ethz.ch > https://www.stat.math.ethz.ch/mailman/listinfo/bioconductor >