On Sat, 2 May 2009 13:45:25 +0100, Jesper Andersson <[log in to unmask]> wrote: Dear Jesper, thank you so much for taking the time to provide an informative and detailed explanation for my post. I had managed to confuse myself somewhat! I appreciate the help Rebecca >Dear Rebecca, > >> I am trying to run a paired group analysis using an additional >> covariate and >> would like some advice. I apologise if it seems basic, but I am >> still trying to >> get a grip on the analysis. >> >> I have a group who were scanned twice and I want to compare their >> performance pre and post, as well as adding a RT covariate. >> However, I am using contrasts e.g. happy-neutral (pre) > happy-neutral >> (post). >> For my additional RT covariate do I have to calculate my mean RT as >> happy- >> neutral for each group? > >There is really no "have to" here. What you need to do is to think >about what question you want to ask of your data. Let as say e.g. you >were to put in the pre RT values as a covariate (mean corrected with >the mean pre RT value). You would then ask a question like > >"Where in the brain is the "change" (post vs pre) in processing of >happy faces (controlled for faces) correlated to pre (pre training?) >reaction time"? > >Only you can know if this is a reasonable question, that might >potentially have an interesting answer. > >Similarly, if you were to put in the delta RT values you would be >asking a question like > >"Where in the brain is the "change" (post vs pre) in processing of >happy faces (controlled for faces) correlated to changes (pre vs post) >in reaction time"? > >Again, is this a reasonable question? Only you can say. > >As for mean correction, the point of the mean correction is to make >sure that the estimates you get for your correlation are not >contaminated by an overall mean effect in the data (such as the >average activation across all subjects). Therefore the mean-correction >should give you a regressor with zero mean. So whatever you put in >there, it is the mean of those values that should be subtracted. > >> Also I am a bit confused about how to set up the contrasts to look at >> pre>post with covariate. I have read a couple of previous posts and >> remain >> confused. >> EV1 =group >> EV2-3 = partciapnts (obviously I have more this is merely for example) >> EV4= RT covariate >> >> EV1 EV2 EV3 EV4 >> con1 1 0 0 0 >> con2 -1 0 0 0 >> >> The above should give Pre>post and the post>pre activation. I >> initailly though >> that this will show activation which already includes the additional >> contrast as >> it is defined as EV4, but after reading previous posts I am not >> sure. Is this not >> the case? > >This would only look at pre>post (or vice versa) since the contrast >does not include the RT covariate. What the covariate will do here is >to remove any variance that could be explained by reaction time, prior >to looking at the pre-post effect. So, let us e,g, say there is an >effect pre>post, but let us also say that there is a consistent >difference in RT (they are all faster post). Then by including the RT >covariate you will effectively explain the difference with that, so >that nothing is left for the group to explain and you will not find >anything e.g. in your pre>post contrast. This may sound paradoxical >(and maybe even unwanted), but it actually makes sense (sometimes). If >an effect that can be explained by pre vs post can EQUALLY WELL be >explained by a difference in reaction time, then you cannot really say >which of these things that caused the change in response. It may be >the pre vs post (whatever that was) but it may also be that if the >subjects had somehow managed to improve their reaction times in some >other fashion you would have obtained the same effect. And if there is >any ambiguity, GLM will always be conservative (i.e. leave the effects >out of the contrast). > > >> If not how would I define the contrast? I thought it would be con3 >> and con4 >> EV1 EV2 EV3 EV4 >> con3 1 0 0 1 >> con4 -1 0 0 1 >> con5 0 0 0 1 >> >> But reading previous posts contrast 5 was recommeded. I am unclear >> how this >> tells me where my differences are between my groups adding in the >> additonal >> covariate. >> I would really appreciate it if someone could explain this to me. > >Again, I think you need to be much more clear about what question you >are actually trying to ask of your data. I cannot see how con3 or con4 >could ever be valid questions (basically adding a group effect and a >correlation with some continous variable together). Contrast 5 MAY be >the question you want to ask. It asks > >"Where in the brain is the response correlated to the reaction time, >after I have explained away anything that could be explained by group >(which I can only assume corresponds to pre vs post in this case). > >Designs can get rather complicated when there are several levels of >subtraction, especially when adding in also continous covariates. But >at the end of the day it is all a matter of common sense (no maths >training needed) to make sure that for each level of subtraction you >put into simple words what that subtraction means, and then at the >next level of subtraction you simply add another layer of simple words. > >For example (happy_faces vs neutral_faces) gives you >"where in the brain are happy faces processed when controlling for >faces?". > >If you then add another layer (pre_happy_faces - pre_neutral_faces) - >(post_happy_faces - post_neutral_faces) it turns into > >"Where in the brain does training change the processing of happy faces >when controlled for faces?" > >Here I have pretended that pre vs post pertains to training, but just >replace it by whatever is appropriate. > >And in this way you build up your contrasts and questions gradually >until you have the question you really want to ask. > >I hope this was helpful. > >Good luck Jesper