hi Claudia, On Wed, Mar 26, 2014 at 7:41 AM, Claudia Pommerenke < claudia.pommerenke@med.uni-goettingen.de> wrote: > Dear Michael, > > I have been switching from edgeR/DESeq to DESeq2 and I'm happy to include > more complex experimental design to the analysis. Thanks for developing and > providing this R-package! :o) > > I hope I can bother you with some questions about your package: > 1. when analysing for let's say two variables (e.g. ~ type + condition), > will this be a difference to change the order of the variables (~ condition > + type)? Probably this will do a blocking and hence order is important here > similar to edgeR? > ​it makes no difference except that the results() function with no arguments will automatically extract results for the last variable in the design formula, and for the last level of this variable over the first level if this variable is a factor.​ But otherwise, no it doesn't make a difference if you are using the arguments of results() to specify which results tables to construct. > 2. For more than 2 variables (e.g. ~type + condition + time) would this be > meaning blocking will also be done for the further variables (e.g. type for > time)? If I intend no blocking I probably need to do ~ 0 + type + condition > + time? > ​Yes, adding variables "accounts for" or "controls for" other effects, whether there are 2 variables, 3, 4 etc..​ We do not recommend using the "0" in design formula for DESeq2, because by default we shrink non-intercept effects (see the description of betaPrior in ?DESeq), but we want to have an unshrunken intercept. Can you say more what you mean by "intend no blocking", in my mind, blocking is a property of the experiment. > 3. Is it also possible to study nested effects and other interactions > (e.g. condition:time)? > > ​Yes, you can include interaction terms in the design formula. If you are using DESeq2 v1.2 you can pull out the effects by 'name' argument to results(), using a name in resultsNames(dds). In DESeq2 >= 1.3​, I have added instructions and examples to the ?results manual page, but still need to add a section to the vignette on interaction terms. ​Mike​ > I would be happy getting any answer of you. > Thank you, > Claudia > > > [[alternative HTML version deleted]]

Dear Michael, I have been switching from edgeR/DESeq to DESeq2 and I'm happy to include more complex experimental design to the analysis. Thanks for developing and providing this R-package! :o) I hope I can bother you with some questions about your package: 1. when analysing for let's say two variables (e.g. ~ type + condition), will this be a difference to change the order of the variables (~ condition + type)? Probably this will do a blocking and hence order is important here similar to edgeR? 2. For more than 2 variables (e.g. ~type + condition + time) would this be meaning blocking will also be done for the further variables (e.g. type for time)? If I intend no blocking I probably need to do ~ 0 + type + condition + time? 3. Is it also possible to study nested effects and other interactions (e.g. condition:time)? I would be happy getting any answer of you. Thank you, Claudia

hi Claudia, On Thu, Mar 27, 2014 at 5:17 AM, Claudia Pommerenke < claudia.pommerenke@med.uni-goettingen.de> wrote: > Dear Michael, > > > My understanding from working with the edgeR package is, that you can > consider paired samples or batch effects when comparing two or more > treatments. Accounting for paired samples seems to be a specific case of > "blocking" in experimental design. In edgeR doing ~type+condition I can > retrieve the 'condition' comparison including the correction for the paired > effect 'type'. Is this also true for DESeq2? And if so, does the order > matter in the formula e.g.: ~type+condition the same as ~condition+type? > > > ​Yes, using a design formula '~ type + condition' allows you to retrieve the condition comparison correcting for the type effect. This is generally true of linear models and generalized linear models (we call this "multi-factor design" in the vignette, and it was also implemented in the previous DESeq as nbinomGLMTest). The order only matters in so far as I mentioned in my previous email. Mike​ [[alternative HTML version deleted]]

Dear Michael, thanks for your instantaneous reply! > I hope I can bother you with some questions about your package: > 1. when analysing for let's say two variables (e.g. ~ type + > condition), will this be a difference to change the order of the > variables (~ condition + type)? Probably this will do a blocking > and hence order is important here similar to edgeR? > > > ​it makes no difference except that the results() function with no > arguments will automatically extract results for the last variable in > the design formula, and for the last level of this variable over the > first level if this variable is a factor.​ But otherwise, no it > doesn't make a difference if you are using the arguments of results() > to specify which results tables to construct. > > 2. For more than 2 variables (e.g. ~type + condition + time) would > this be meaning blocking will also be done for the further > variables (e.g. type for time)? If I intend no blocking I probably > need to do ~ 0 + type + condition + time? > > > ​Yes, adding variables "accounts for" or "controls for" other effects, > whether there are 2 variables, 3, 4 etc..​ We do not recommend using > the "0" in design formula for DESeq2, because by default we shrink > non-intercept effects (see the description of betaPrior in ?DESeq), > but we want to have an unshrunken intercept. Can you say more what you > mean by "intend no blocking", in my mind, blocking is a property of > the experiment. > My understanding from working with the edgeR package is, that you can consider paired samples or batch effects when comparing two or more treatments. Accounting for paired samples seems to be a specific case of "blocking" in experimental design. In edgeR doing ~type+condition I can retrieve the 'condition' comparison including the correction for the paired effect 'type'. Is this also true for DESeq2? And if so, does the order matter in the formula e.g.: ~type+condition the same as ~condition+type? > > 3. Is it also possible to study nested effects and other > interactions (e.g. condition:time)? > > > ​Yes, you can include interaction terms in the design formula. If you > are using DESeq2 v1.2 you can pull out the effects by 'name' argument > to results(), using a name in resultsNames(dds). In DESeq2 >= 1.3​, I > have added instructions and examples to the ?results manual page, but > still need to add a section to the vignette on interaction terms. > Cool! Thanks a lot, Claudia [[alternative HTML version deleted]]