At 03:12 AM 6/08/2004, Richard Friedman wrote: >Fellow Expressionists, > > Does Limma automatically perform a multiple >comparison adjustment for non-orthogonal contrasts? Yes. classifyTestsF() is a classification which takes full account of non-orthogonality. Unpublished method though. Also "holm" and some other options are valid even across non-orthogonal contrats. > If not, can you recommend another program that can >be used in conjunction with Limma to do this? > > Also, I find the Bayesian theory in the paper >by Gordon Smyth ("Linear Models and Empirical >Bayes Methods..." tough going) Can anyone >please recommend a book or books on Bayesian methods >that can bridge the gap between basic statistics texts >(e.g. Hoel, "An Introduction to Mathematical Statistics" and >Zar "Biostatistical Analysis" and this article)? I don't know of any accessible refs, even for stat majors I'm afraid. And unfortunately Bayes refs may not be much help. The empirical Bayes arithmetic requires different quantities to be computed compared to full Bayes. I hope that the final formula in the paper make intuitive sense even though the math derivation might be hard. Gordon >Thanks and best wishes, >Rich

Gordon, Thank you for answering my questions. The last equation in your paper makes intuitive sense to me. I'm wondering if you can take the time to answer two more questions: 1. Say I have the following case: Level A (3 replicates) Level B (2 replicates) Level C(1 replicate) Level D(1 replicate) Can I legitimately calculate a P value for the contrast Level A to level C in the linear model even though I have only one replicate on Level C. I am not talking about just Limma here. I am talking about the linear model in general. Also, I realize that one replicate is poor experimental design. This is what I was given to analyze. 2. If I wished to apply a multiple test correction to the pvalues from non-orthogonal contrasts, would the following procedure be legitimate:: 1. Generate a pvalue for each contrast in the set of nonothogonal contrasts for each gene using classifyTestsF(). 2. Correct the pvalues using a multiple test correction such as FDR. I realize that no multiple-test correction is entirely satisfactory, I just want to get an approximate estimate of the p-values for each contrast as a guide to further experimentation and literature searching. Best wishes, Rich On Aug 5, 2004, at 9:26 PM, Gordon Smyth wrote: > At 03:12 AM 6/08/2004, Richard Friedman wrote: >> Fellow Expressionists, >> >> Does Limma automatically perform a multiple >> comparison adjustment for non-orthogonal contrasts? > > Yes. classifyTestsF() is a classification which takes full account of > non-orthogonality. Unpublished method though. Also "holm" and some > other options are valid even across non-orthogonal contrats. > >> If not, can you recommend another program that can >> be used in conjunction with Limma to do this? >> >> Also, I find the Bayesian theory in the paper >> by Gordon Smyth ("Linear Models and Empirical >> Bayes Methods..." tough going) Can anyone >> please recommend a book or books on Bayesian methods >> that can bridge the gap between basic statistics texts >> (e.g. Hoel, "An Introduction to Mathematical Statistics" and >> Zar "Biostatistical Analysis" and this article)? > > I don't know of any accessible refs, even for stat majors I'm afraid. > And unfortunately Bayes refs may not be much help. The empirical Bayes > arithmetic requires different quantities to be computed compared to > full Bayes. > > I hope that the final formula in the paper make intuitive sense even > though the math derivation might be hard. > > Gordon > >> Thanks and best wishes, >> Rich > > ------------------------------------------------------------ Richard A. Friedman, PhD Associate Research Scientist Herbert Irving Comprehensive Cancer Center Oncoinformatics Core Lecturer Department of Biomedical Informatics Box 95, Room 130BB or P&S 1-420C Columbia University Medical Center 630 W. 168th St. New York, NY 10032 (212)305-6901 (5-6901) (voice) friedman@cancercenter.columbia.edu http://cancercenter.columbia.edu/~friedman/ In Memoriam, Francis Crick

At 11:36 PM 6/08/2004, Richard Friedman wrote: >Gordon, > > Thank you for answering my questions. The last equation in your > paper makes intuitive sense to me. > > I'm wondering if you can take the time to answer two more questions: > >1. Say I have the following case: > >Level A (3 replicates) >Level B (2 replicates) >Level C(1 replicate) >Level D(1 replicate) > >Can I legitimately calculate a P value for the contrast Level A to level C >in the linear model even though I have only one replicate on Level C. I am >not talking about just Limma here. I am talking about the linear model in >general. Given assumption of common variance across levels, yes. > Also, >I realize that one replicate is poor experimental design. This is what I >was given to analyze. > >2. If I wished to apply a multiple test correction to the pvalues from >non-orthogonal contrasts, would the following procedure be legitimate:: > > 1. Generate a pvalue for each contrast in the set of nonothogonal > contrasts for each gene using classifyTestsF(). > 2. Correct the pvalues using a multiple test correction such as FDR. Nothing special about this design. All usual things, e.g. in limma, apply. Gordon >I realize that no multiple-test correction is entirely satisfactory, I >just want to get an approximate estimate of the p-values for each contrast >as a guide to further experimentation and literature searching. > >Best wishes, >Rich