 |
 |
Thanks very much for your feedback. Setting a dummy regressor at the last volume does not really work, as Helmut suggests. I have figured out a way to set-up session-specific contrasts, but am unsure about the scaling issues that arise. Assume my task has 3 sessions, and each session has 4 conditions of interest, coded with unique regressors, A, B, C and D. Lets say I am interested in the B > C contrast. Lets also say that there are no events for condition C in session 2. At the 1st level, I don’t replicate the contrast over sessions and I am able to specify the B>C contrast in sessions 1 and 3 but not for session 2. E.g., my contrast vector looks like this: [0 .5 -.5 0 0 0 0 0 0 .5 -.5 0] In comparison to someone who has events for all conditions, where the contrast vector would look like this: [0 .33 -.33 0 0 .33 -.33 0 0 .33 -.33 0] I’m a little concerned that these contrasts may be scaled differently and thus not appropriate for comparison at the 2nd level because of the different number of sessions involved in each contrast. I have read somewhere that this can be addressed by ensuring the contrast weights add to [1 and -1] as I have done. I was wondering whether this is correct or whether I need to perform some additional scaling? SPM8 only allows scaling of contrasts that replicate over sessions. Thanks again, Ari Pinar *PhD Candidate* *Monash Clinical & Imaging Neuroscience* *School of Psychology and Psychiatry &* *Monash Biomedical Imaging* *Monash University* *A: 770 Blackburn Rd, Clayton, 3168, **Vic, Australia* *E: [log in to unmask] * On 19 October 2013 03:37, H. Nebl <[log in to unmask]> wrote: > A few comments. > > > Concerning contrasts: If there are no trials of a particular type then > don't use this session for any related contrast. There's simply nothing to > look at. Some dummy regressor might reflect anything but the assumed brain > activation for this condition. When averaging across sessions you will also > contaminate the estimates for these sessions. Leaving this aside, I would > look at differences between error types only if you have at least a few > trials for each of the categories. Otherwise it's probably comparing noise > with noise. It might also be problematic because some subjects are prone to > mistakes (many trials, better to estimate) and others aren't (few trials, > noisy, possibly bad estimate). > > > Concerning the proposed dummy regressor, does this work? I just tried with > a data set consisting of a single run with 234 volumes, TR = 2 s. For the > single-subject model I used the default microtime resolution of 16 and a > microtime onset of 8. > > When the onset of the dummy regressor is identical to the end of the > experiment = 468 s with a duration of 0 s, then the regressor can't be > properly estimated (beta not uniquely specified) and one cannot define > corresponding contrasts. However, the voxels of the beta image are not 0 or > NaN but have very small values +/- 10^-13. > > The earliest onset that results in a unique beta for parameter > estimability is 466.8124, but 466.8125 does not work - why? With 466.8124 > the values in the design matrix range between ~ +/- 10^-9 after high-pass > filtering (default 128 Hz, seems like it introduces some riples), the last > volume has a value of 10^-7 (BOLD response starts). The beta image is full > of extreme values with a range of approx. +/- 10^6. So definitely nothing > meaningful. In comparison the beta zero has values of 100 - 250. The beta > estimates of the other regressors are somewhat different compared to the > model without the dummy regressor, but it's the same pattern. > > > Now this is a single observation of course. In this particular case it > seems the dummy regressor is a nonsense regressor. Probably it's much > better to have one df less/more instead of anything like that. But I would > be interested in Karl's message quoted above. > > > > Best, > > Helmut >