 |
 |
Dear Lynn > > Most of your questions come from fMRI users these days. However, we have > FDG PET data on 30 subjects (one condition). They are assigned to one of > three groups (0,1,2) and it is the metabolic differences between groups that > are of interest. There is reason to expect global differences between the > groups. We would like to use a confounding covariate. > > Is SPM appropriate to use with these data? > SPM could certainly be used for these data. > If so: > 1) What is the correct way to set up the design? When you launch the "Statistical analysis" module you will be prompted to "Select design type". You should select "Compare groups: 1 scan per subject". Specify number of groups etc. Select one when asked how many confounding covariates you want, and specify their values in the same order as you specified your scans (you will want to check this in the output to make sure that the right values are entered for each scan). Since you expect that there may be global differences between the groups (and I assume these differences are of interest) you should do no global normalisation. I also assume that some form of model-based quantification has been performed on your images (otherwise you can't really assess global differences) in which case you will not want to perform any scaling of the grand mean either. After that it should be pretty straightforward. > 2) Group assignment is not random but is based upon DSM diagnosis, ranked > from no symptoms (0, normal) to few symptoms (1, indeterminate) to many > symptoms (2, definite). Would this be considered a "fixed effects" model? > The "random" in the "random effects" model does not imply that subjects are randomly assigned to groups, but rather that the specific sample at hand is a random sample from the entire population of subjects with this specific condition. e.g. in your case you have picked ten subjects with symptom severity 2, and have obtained a certain realisation (your specific ten scans). Had you picked ten other subjects with symptom severity 2, you would have obtained ten other scans that were different from your first, hence the randomness. Therefore, one cannot based on a "non-random" group assignment say that one has a fixed effects model. However in this specific case, where you have only one condition, there is no difference between the random and the fixed effects models and you will effectively be performing a one-way ANCOVA. The recent interest in the random effects model stems from the realisation that the between subject variance is different from the within subject variance. It is typically larger and may well have a different spatial distribution. Hence, one needs to consider both the within and between subject variances. Since there is no within subject variance in your case, this doesn't apply. You should realise that since the between subject variance, which is what you will be dealing with, is typically quite large and you will be left with quite few degrees of freedom your power will not be tremendeously large. You may want to consider using the non-parametric methods developed by Andrew Holmes since they will not depend on high degrees of freedom for their validity and they will allow you to smooth the variance map quite severly, resulting in higher power. > > Thanks for your help. > > Good luck Jesper %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%