Hi I believe there are two approaches for using ANOVA with microarrays, 1) Calculate logratios, do normalisation and then fit the experimental conditions by an ANOVA model, or 2) Use the intensities of each channel, transformed with appropriate transformation (log, linlog.logshift,...), and use array, dye, spot effect and so in the ANOVA model in addition to the experimental conditions. Which means that the normalisation is done by factors in the ANOVA model limma is much used for the first approach, whereas I think Maanova is more used for the second approach. Does anybody have any experience on both approaches? WHat is recommended? Can the limma package be used for the second approach? Additional question: Can continuous covariates be fitted with limma? ---------------------------------------------------------------------- -------- Ingunn Berget Norwegian University of Life Sciences Department of Animal and Aquacultural Sciences Boks 5003 1432 ?s Norway

Limma can be used for the 2nd approach. See the section in the manual on Single Channel analysis. In my opinion, even if you use the MAANOVA approach, you need to normalize before analysis, because normalization takes care of intensity dependence of differential expression, which is not handled by the linear model. I prefer the Limma approach because of the clever modeling of the array random effect, which seems preferable to the gene-by-gene model of MAANOVA. (I have not used MAANOVA for 2 years, so perhaps I am out of date here.) The only possible problem with the Limma approach is that you cannot have other random effects, so e.g. if you have both technical and biological replicates, you need to fit the biological replicates as fixed blocks. In the experiments I have analyzed so far, this effect has not been large. I am not sure whether MAANOVA allows an unlimited number of random effects, currently. When I last used it, MAANOVA required balanced data, which meant that flagged spots were problematic. Limma handles flagged spots by weighting - but not in the Single Channel Analysis. --Naomi At 08:00 AM 12/8/2005, Ingunn Berget wrote: >Hi > >I believe there are two approaches for using ANOVA with microarrays, > >1) Calculate logratios, do normalisation and then fit the experimental >conditions by an ANOVA model, or >2) Use the intensities of each channel, transformed with appropriate >transformation (log, linlog.logshift,...), and use array, dye, spot effect >and so in the ANOVA model in addition to the experimental conditions. Which >means that the normalisation is done by factors in the ANOVA model > >limma is much used for the first approach, whereas I think Maanova is more >used for the second approach. > >Does anybody have any experience on both approaches? WHat is recommended? >Can the limma package be used for the second approach? > >Additional question: Can continuous covariates be fitted with limma? >--------------------------------------------------------------------- --------- >Ingunn Berget >Norwegian University of Life Sciences >Department of Animal and Aquacultural Sciences >Boks 5003 >1432 ?s >Norway > >_______________________________________________ >Bioconductor mailing list >Bioconductor at stat.math.ethz.ch >https://stat.ethz.ch/mailman/listinfo/bioconductor Naomi S. Altman 814-865-3791 (voice) Associate Professor Dept. of Statistics 814-863-7114 (fax) Penn State University 814-865-1348 (Statistics) University Park, PA 16802-2111

The difference between maanova and limma goes beyond what you mention. It is true that limma originally analized log-ratios and maanova fits intensity models. However, the main difference is that limma resorts to an empirical bayes procedure to assess significance in differential expression. Maanova, on the other hand, fits a gene by gene model on the intensities and allows to include fixed and random effects. For testing purposes, maanova provides parametric tests (assuming normality) or permutation based tests. The ratio versus intensity dicotomy is not the important thin in this case, for it can be shown that a gene-by-gene mixed model can be fit for ratios or intensities and still obtain the same result (the only restriction is that the array effect in the intensity model should be a fixed effect). Also, the limma package may fit intensity models in some cases (see Ch. 9 of limma user's guide). So the main question here is if we should use an EB procedure after the gene-by-gene linear model or not... I really don't have (a convincing) answer to that, but I'm partial to the idea of "borrowing information" across genes that EB procedures provide. The problem we have in practice is that the EB procedure implemented in limma only considers a single variance component. Anything else should be treated as a fixed effects (Even the biological subjects in some layouts!). And that may not be a good assumption for some experimental designs. JP Ingunn Berget wrote: >Hi > >I believe there are two approaches for using ANOVA with microarrays, > >1) Calculate logratios, do normalisation and then fit the experimental >conditions by an ANOVA model, or >2) Use the intensities of each channel, transformed with appropriate >transformation (log, linlog.logshift,...), and use array, dye, spot effect >and so in the ANOVA model in addition to the experimental conditions. Which >means that the normalisation is done by factors in the ANOVA model > >limma is much used for the first approach, whereas I think Maanova is more >used for the second approach. > >Does anybody have any experience on both approaches? WHat is recommended? >Can the limma package be used for the second approach? > >Additional question: Can continuous covariates be fitted with limma? >--------------------------------------------------------------------- --------- >Ingunn Berget >Norwegian University of Life Sciences >Department of Animal and Aquacultural Sciences >Boks 5003 >1432 ?s >Norway > >_______________________________________________ >Bioconductor mailing list >Bioconductor at stat.math.ethz.ch >https://stat.ethz.ch/mailman/listinfo/bioconductor > > > > -- ============================= Juan Pedro Steibel Graduate Student Department of Animal Science Michigan State University 1261 Anthony Hall East Lansing, MI 48823 USA Phone: 1-517-432-0671 E-mail: steibelj at msu.edu Web: http://www.msu.edu/~steibelj

Thanks to both Naomi and and JP for informative answers. IB ----- Original Message ----- From: "Juan Pedro Steibel" <steibelj@msu.edu> To: "Ingunn Berget" <ingunn.berget at="" umb.no=""> Cc: <bioconductor at="" stat.math.ethz.ch=""> Sent: Thursday, December 08, 2005 3:51 PM Subject: Re: [BioC] Limma vs Maanova, and use of covariates > The difference between maanova and limma goes beyond what you mention. > It is true that limma originally analized log-ratios and maanova fits > intensity models. However, the main difference is that limma resorts to an > empirical bayes procedure to assess significance in differential > expression. Maanova, on the other hand, fits a gene by gene model on the > intensities and allows to include fixed and random effects. For testing > purposes, maanova provides parametric tests (assuming normality) or > permutation based tests. > > The ratio versus intensity dicotomy is not the important thin in this > case, for it can be shown that a gene-by-gene mixed model can be fit for > ratios or intensities and still obtain the same result (the only > restriction is that the array effect in the intensity model should be a > fixed effect). Also, the limma package may fit intensity models in some > cases (see Ch. 9 of limma user's guide). > > So the main question here is if we should use an EB procedure after the > gene-by-gene linear model or not... > > I really don't have (a convincing) answer to that, but I'm partial to the > idea of "borrowing information" across genes that EB procedures provide. > The problem we have in practice is that the EB procedure implemented in > limma only considers a single variance component. Anything else should be > treated as a fixed effects (Even the biological subjects in some > layouts!). And that may not be a good assumption for some experimental > designs. > > JP > > Ingunn Berget wrote: > >>Hi >> >>I believe there are two approaches for using ANOVA with microarrays, >> >>1) Calculate logratios, do normalisation and then fit the experimental >>conditions by an ANOVA model, or >>2) Use the intensities of each channel, transformed with appropriate >>transformation (log, linlog.logshift,...), and use array, dye, spot effect >>and so in the ANOVA model in addition to the experimental conditions. >>Which means that the normalisation is done by factors in the ANOVA model >> >>limma is much used for the first approach, whereas I think Maanova is more >>used for the second approach. >> >>Does anybody have any experience on both approaches? WHat is recommended? >>Can the limma package be used for the second approach? >> >>Additional question: Can continuous covariates be fitted with limma? >>-------------------------------------------------------------------- ---------- >>Ingunn Berget >>Norwegian University of Life Sciences >>Department of Animal and Aquacultural Sciences >>Boks 5003 >>1432 ?s >>Norway >> >>_______________________________________________ >>Bioconductor mailing list >>Bioconductor at stat.math.ethz.ch >>https://stat.ethz.ch/mailman/listinfo/bioconductor >> >> >> > > -- > ============================= > Juan Pedro Steibel > Graduate Student > Department of Animal Science Michigan State University > 1261 Anthony Hall > East Lansing, MI > 48823 USA Phone: 1-517-432-0671 > E-mail: steibelj at msu.edu > Web: http://www.msu.edu/~steibelj > ============================= > >