 |
 |
Hi John, Please, see below: On 6 March 2015 at 22:20, John Sheppard <[log in to unmask]> wrote: > Hi Anderson, > > Fortunately, I was able to get this model set up in FEAT alright -- I did > some of the editing directly to the .fsf file in gedit to get around the > lag issues. Thanks for the tips. > > If you don't mind, I have two follow-up questions: > > 1) I want to do another related analysis to the model you had sent > previously. For this analysis, I want to pool all the subjects together > into one big group (only one group, not two) and simply ask if, across all > categories of facial expression, the activation to faces is greater than > zero. This should be an F test, and I believe the model for this is a > simplified version of the one you already set up for me. I have prepared an > example of this model with seven subjects in the attached spreadsheet. The > output of interest is the "F1" test at the bottom, which should indicate > whether the activation is greater than zero across face categories. Can you > let me know if this model setup looks correct? > Looks fine. When running, keep using the "--permuteBlocks" option, and also include the option "-1", so that only sign-flippings are used. This is FMRI so the assumption of independent and symmetric errors is likely to be valid, so sign-flippings can be used. > > 2) Going back to the original model you supplied where we looked at group > differences, I noticed the group contrasts tested whether Group 1 > Group 2 > for each condition. Do I need to include a separate set of contrasts if I > want to test whether Group 2 > Group 1 for each condition? I know that when > doing two-group mean comparisons, one generally looks both ways to compare > groups (i.e., -1 1 and 1 -1). But I'm wondering about this in the context > of the F test: Would it have been equivalent to set up the contrasts as > Group 2 > Group 1, and then run an F test on these "reverse" contrasts? Or > would I need to separately analyze F tests of the two possible sets of > contrasts (one set where Group 1 > Group 2, the other set where Group 2 > > Group 1) to consider group effects arising from differences in either > direction (i.e., which group has higher activation)? Hopefully this > question made sense. > If you only look at the F-test, there's no need to add the reverse contrasts, as the F-test already looks into both directions. However, if you go on to investigate the each of the individual t-tests (e.g., if the F-test is significant), you'll probably want to add the other, reverse contrasts. There's no need to add another F-test with these other contrasts, neither modify the current F1. All the best, Anderson > > Many thanks, > John > > On Fri, Mar 6, 2015 at 2:59 AM, Anderson M. Winkler < > [log in to unmask]> wrote: > >> Hi John, >> >> Please, see below: >> >> >> On 5 March 2015 at 22:08, John Sheppard <[log in to unmask] >> > wrote: >> >>> Hi Anderson, >>> >>> Thanks for showing me how to set up this model! The only problem I am >>> having is that I have 61 subjects, and this results in a very large model >>> -- 305 COPE inputs x 69 EVs. Right now this is making it impractical to set >>> up the model in the FEAT GUI, since it becomes extremely slow to refresh >>> with so many variables in the "Full Model Setup." >>> >> >> There are workarounds: >> >> - Make a minimal model with just a few subjects, save. Then open the >> design.mat, design.con, design.fts and design.grp to see how do they look >> like internally. Use a spreadsheet software to make the actual, complete >> design, copy and paste into these files making sure to retain the same >> format, and making other small modifications where needed (e.g. fields >> /NumWaves, /NumPoints, etc). >> >> - Make a script that assembles the whole design. >> >> - Consider using Matlab or Octave, then save using the attached functions. >> >> >> >>> >>> Just out of curiosity, would it be possible to alter this model and use >>> *only* the first 8 EVs (i.e., the EVs that model the task conditions >>> for both groups) and then leave out the EVs that define each subject (9-15 >>> in the example)? I assume doing so would make the entire model incorrect, >>> since I believe those additional EVs (9-15) are needed to model the subject >>> means (each subject is a random effect?). Anyway, I was just trying to >>> think of any ways of slimming down the number of EVs to make it practical >>> to use the model in the FEAT GUI when I have 61 subjects (i.e. to prevent >>> the GUI from freezing due to memory issues). >>> >> >> Hmm, it won't work. It won't give any error message but the results will >> be incorrect. Maybe you can consider MANOVA, though, but it's a different >> story then. >> >> >>> >>> Assuming there's no alternative to the model you've already set up for >>> me, thanks very much for all your help with this. >>> >> >> Sure, no worries. >> >> All the best, >> >> Anderson >> >> >> >> >>> >>> Best, >>> John >>> >>> On Thu, Mar 5, 2015 at 3:53 AM, Anderson M. Winkler < >>> [log in to unmask]> wrote: >>> >>>> HI John, >>>> >>>> Please see at this link >>>> <https://dl.dropboxusercontent.com/u/2785709/outbox/mailinglist/design_johnsheppard.ods> >>>> how the design would look like, in this example considering 4 subjects in >>>> one group, 3 subjects in the other. Only the contrasts C1-C5 are needed; >>>> the others (marked with a dash, "-") aren't necessary, and are there just >>>> to show how the others were constructed. Define an F-test that has these 5 >>>> contrasts, and if this is significant, then we can say that there's a >>>> difference between the two groups in at least one of these conditions (here >>>> named A, B, C, D, and E). >>>> >>>> If you run this design in randomise, define one exchangeability block >>>> per subject, and use the options "-e design.grp" together with >>>> "--permuteBlocks". >>>> >>>> This is a design in which MANOVA could also be considered -- it's >>>> available in PALM (somewhat experimental, link here >>>> <https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/PALM>). >>>> >>>> All the best, >>>> >>>> Anderson >>>> >>>> >>>> On 4 March 2015 at 14:59, John Sheppard < >>>> [log in to unmask]> wrote: >>>> >>>>> Hi Anderson, >>>>> >>>>> Regarding my second question, I believe what I really want to do is >>>>> what you specified in your last email -- set up an F-test (ANOVA) to test >>>>> if there's any difference between the *means* (*not* the variance) of >>>>> the fMRI responses to the 5 categories between the two groups. In other >>>>> words, asking if there's a difference between groups A and B for angry, >>>>> fearful, happy, neutral, or sad faces. >>>>> >>>>> However, I'm a bit confused regarding how to set up this analysis in >>>>> FSL at the group level. Would it be necessary to include each subject as a >>>>> separate EV in the model (i.e., model each subject as a random effect)? >>>>> >>>>> Is this model on the FSL wiki the one I should follow when setting up >>>>> the ANOVA? >>>>> http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/GLM#ANOVA:_2-groups.2C_2-levels_per_subject_.282-way_Mixed_Effect_ANOVA.29 >>>>> >>>>> Any direction you can give me on how to set this model up in FSL would >>>>> be a huge help. >>>>> >>>>> Thank you! >>>>> John >>>>> >>>>> On Wed, Mar 4, 2015 at 3:15 AM, Anderson M. Winkler < >>>>> [log in to unmask]> wrote: >>>>> >>>>>> Hi John, >>>>>> >>>>>> Please, see below: >>>>>> >>>>>> >>>>>> On 3 March 2015 at 14:31, John Sheppard < >>>>>> [log in to unmask]> wrote: >>>>>> >>>>>>> Hi Anderson, >>>>>>> >>>>>>> Thanks very much for your reply. For my analysis, I have a block >>>>>>> design where subjects (belonging to two groups -- addicts and controls) >>>>>>> view 5 different categories of facial expressions (angry, fearful, happy, >>>>>>> neutral, and sad). The first thing I'd like to do is collapse across all >>>>>>> five categories of faces, merging all levels into a single average, and >>>>>>> then compare the global effects between the two groups. I assume for this I >>>>>>> should first do a second level analysis at the within-subject level where I >>>>>>> average the five conditions within-subject, and then setup a third level >>>>>>> analysis where I do a two group mean comparison to compare the means of the >>>>>>> across-condition average between the two groups. Does that sound right? >>>>>>> >>>>>> >>>>>> Yes, sounds right. >>>>>> >>>>>> >>>>>> >>>>>>> I do have a second question, though. I'm also hoping to setup an >>>>>>> F-map that tests the null hypothesis that there is no difference between >>>>>>> the two groups in the variance of the fMRI responses to the five categories >>>>>>> of facial expressions. Does that description make sense? I believe this >>>>>>> analysis would require a different model than simply collapsing across >>>>>>> levels and doing a mean comparison between groups. Do you have any idea how >>>>>>> I might go about setting up this model? >>>>>>> >>>>>>> Please let me know if this description is too vague. >>>>>>> >>>>>> >>>>>> You can definitely do an F-test (ANOVA) to test if there's any >>>>>> difference between the *means* of the responses to the 5 categories >>>>>> between the two groups. That is, asking whether if there's any difference >>>>>> between groups A and B in angry, fearful, happy, neutral or sad. >>>>>> >>>>>> However, to see if there's any difference between the *variances* of >>>>>> the two groups across the 5 categories (i.e., not an ANOVA), we can devise >>>>>> a test, but it will require a good deal of coding, though. >>>>>> >>>>>> All the best, >>>>>> >>>>>> Anderson >>>>>> >>>>>> >>>>>> >>>>>>> >>>>>>> Thanks again! >>>>>>> John >>>>>>> >>>>>>> On Tue, Mar 3, 2015 at 2:03 AM, Anderson M. Winkler < >>>>>>> [log in to unmask]> wrote: >>>>>>> >>>>>>>> Hi John, >>>>>>>> >>>>>>>> Different hypotheses can be considered with repeated measurements, >>>>>>>> and the design changes accordingly: these can consider averages of the >>>>>>>> measurements (i.e., to investigate different global effects between groups, >>>>>>>> collapsing all levels), or differences between levels (i.e., to see if the >>>>>>>> levels differ among themselves and then among groups), or to see if the >>>>>>>> groups differ at each level, without checking for a global effect or if the >>>>>>>> levels differ). >>>>>>>> >>>>>>>> Each case requires a slightly different model. Nonetheless, it's >>>>>>>> also possible to put all into a single, large model, but the contrasts >>>>>>>> would then be different, and the assumptions for each also be different. >>>>>>>> >>>>>>>> From your description it sounds as if you wanted to merge all >>>>>>>> levels into a single average, and then compare the groups. This is the case >>>>>>>> when the multiple measurements are obtained to, e.g., improve SNR. Is this >>>>>>>> really what you'd like to do? If you could give more details it may be >>>>>>>> easier to try to help. >>>>>>>> >>>>>>>> All the best, >>>>>>>> >>>>>>>> Anderson >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> On 2 March 2015 at 19:19, John Sheppard < >>>>>>>> [log in to unmask]> wrote: >>>>>>>> >>>>>>>>> Hi everyone, >>>>>>>>> >>>>>>>>> I have a question related to setting up a 2x5 repeated-measured >>>>>>>>> ANOVA model in FSL (2 groups x 5 factors), and testing the main effect of >>>>>>>>> group across all levels. I have done my best to understand the design >>>>>>>>> examples on the FSL GLM Wiki page ( >>>>>>>>> http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/GLM#Experimental_Designs_-_Repeated_measures >>>>>>>>> ). >>>>>>>>> >>>>>>>>> In particular, my data have 5 experimental conditions per each >>>>>>>>> subject, and I want to test for an effect of group across all the 5 >>>>>>>>> conditions pooled together. If I understand the Wiki correctly (looking >>>>>>>>> under ANOVA: 2-groups, 2-levels per subject (2-way mixed effect ANOVA)), >>>>>>>>> the only way to test for group differences in this case is to do a >>>>>>>>> within-subject average across all conditions/levels, and then run a 2 group >>>>>>>>> mean comparison in a 3rd level analysis. I just want to confirm whether the >>>>>>>>> approach I outline below is the correct way to go about this. >>>>>>>>> >>>>>>>>> After setting up the GLM for each experimental condition with >>>>>>>>> stimfiles in the first level analysis (within-subject), I then took an >>>>>>>>> average across all five conditions (again, within-subject) as a >>>>>>>>> second-level analysis. Next, in a third level analysis, I loaded the cope >>>>>>>>> images of the "average" for each individual subject, indicating each >>>>>>>>> subject's group with two EVs. I then performed a 2 group mean comparison on >>>>>>>>> the second-level averages, which gave me results for the group main effect >>>>>>>>> contrasts in both directions (1 -1 for control group - addict group, -1 1 >>>>>>>>> for addict group - control group). >>>>>>>>> >>>>>>>>> Is this approach the proper way to test for a main effect of group >>>>>>>>> across all conditions under this setup (2x5 ANOVA with two groups and five >>>>>>>>> repeated measures)? Any advice would be great. Please let me know if I have >>>>>>>>> not described the analysis clearly enough. >>>>>>>>> >>>>>>>>> Thank you very much! >>>>>>>>> >>>>>>>>> John Sheppard >>>>>>>>> >>>>>>>>> [log in to unmask] >>>>>>>>> >>>>>>>> >>>>>>>> >>>>>>> >>>>>> >>>>> >>>> >>> >> >