 |
 |
Please see my two/three cents below. There is no easy answer, but the field needs to have more discussions like these as more and more researchers are using fMRI. On Mon, Aug 6, 2012 at 9:49 AM, Gabor Oederland <[log in to unmask]> wrote: > Dear SPM users, > > > I would like to rise a question about the validity of group analyses in SPM when using within-subject designs with more than one within-subject factor. Although the topic has been discussed various times (see threads about "Flexible factorial", "x by x - ANOVA" and so on) , I somehow have the feeling that there is no definite answer so far. This is somewhat confusing especially for beginners as there are hundreds of fMRI papers by now. >>> You are correct that there is no definite answer yet. I haven't been able to create simulations for more complex ANOVAs yet, but its on the todo list. There are two issues that need to be dealt with: (1) purely within-subject designs; and (2) within- and between-subject designs. (1) If you are only looking at the within-subject effects (comparing 2 or more levels of a factor) then the only issue is whether or not SPM can correct for violations of sphericity in these larger models. To the degree that the assumption that there are no violations of sphericity holds, then the results are valid. As far as I know, this has not been tested in larger models. In simulations of 2x2 designs (from Rik Henson), if you correct for violations for sphericity, then you can get accurate and slightly more powerful/robust results. In larger designs, this has not been demonstrated and in the lab I work with (the developers of GLM Flex) we have opted for a combined approach to reduce the probability of violations of sphericity: (a) we use partioned variance for each factor; (b) we use pooled variance to correct for violations of sphericity within each factor that has multiple levels. (2) If you have a within- and between-subject designs (or are comparing a single level of a factor - which is a between-subject effect), then you need to use GLM Flex or use multiple models (one for the within-subject effects and one for the between-subject effects) because SPM only uses one error term. > > When looking at published papers dealing with two and more within-subject factors, some explicitly or implicitly use Flexible factorial, others seem to use Full factorial, and others don't state which model they use. Some seem to add "subject" as a between-subject factor, other papers seem not. I have been talking to people who are into fMRI successfully for many years. They are no statistical experts, which means that they might run incorrect analyses in a pure statistical sense, but they probably represent the behavior of "normal" fMRI researchers. >>> I'd suggest you take a look at my poster comparing the models (http://www.martinos.org/~mclaren/ftp/presentations/OHBM2011_v3.pdf). Also, the flexible factorial design is the same as standard statistical packages. You are definitely right that the field needs to be more transparent about their group models. > > > In summary, there seem to be two "popular" options: > Option 1 is to use "Full factorial" setting "independence" to "no" and "variance" to "equal" (without "subject" as a factor) > Option 2 is to use "Flexible factorial" (in general with "subject" as a factor) > > > I well understand the different statistical meaning of these models. But nonetheless both options are published regularly (at least I got this impression when looking at various papers). Is this just due to lack of knowledge/disinterest of reviewers? >>> I can't speak for others, but I always pay close attention to the methods as different methods could significantly change the results. Using the incorrect model could increase the number of significant effects as can be seen on my poster above. I think the field needs to have more statisticians and be more careful about the methods that are employed and how the methods are described. In case one would want to publish his/her own analysis, which of the two options mentioned above (or any other) would you prefer? Could one justify his/her own "incorrect" analysis by refering to other "incorrect" studies (so keeping with a tradition ;-) ? >>> I don't think that saying because something was published and was done "incorrectly" that others should also do their statistics "incorrectly". Keep in mind that SPM continually updates their software alogorithms, so we already commonly change what is being done today compared to 10 or more years ago. > > > I'm aware of the scripts known as GLM flex (http://nmr.mgh.harvard.edu/harvardagingbrain/People/AaronSchultz/Blog/Blog.html), which work with different error terms and help to overcome the problems with traditional SPM analyses. Donald McLaren has written a lot about it in previous threads. Still, I somehow was a little surprised that there was no real discussion about the meaning for studies already published. >>> There will be some discussion about this in an upcoming paper. I also wonder whether SPM12 is going to adopt the options of SPM8? >>> I can't speak for what will be going into SPM12. I'm sure Aaron wouldn't mind having a version of GLM Flex integrated into SPM12. One small caveat, GLM Flex does not use the SPM.mat structure; however, in its own output structure are a number of fields that allow OrthoView to be used seamlessly to view and plot the data. > > > Any comment is greatly appreciated! > > > Gabor