 |
 |
Dear SPM masters, I have some questions on the statistical analysis of ERPs (60 electrodes) with multiple conditions, 20 subjects. General procedure: The procedure that I inferred from the literature is: first test the main effects in sensor space, identify the time intervals where the effects arise, do source reconstruction on each time interval and test the effects in source space for each time interval. Is this a correct procedure or is there a better one? This sounds like double-dipping, is this incorrect but tolerated or is it fully correct? Design: I have a factorial design with three factors, one of which has four parametrically varying levels (from 1 to 4). - I would like to test parametric modulation within the factor that has four levels. I tried to do it by choosing "multiple regression" in the Factorial Design Specification module (open by clicking on "Specify 2-nd level"), but I don't understand how I can insert the different subjects, (while it is straightforward for the paired t-test). - I would like to test the effects of the three factors all together. What design should I use? Do I need to build my own design matrix? Can you indicate me examples on how this is done? Contrast when testing the results: I have performed a statistical analysis to compare two conditions across subjects (design: paired t-test), and I'm confronted with two different results: - I assume to expect a specific polarity for the effect (which is meaningful, I'm looking at a N400), so I use a t-contrast. The N400 pops up with p(FWE)=0.000 at the cluster level (see figure T24x24x251). Fine. - Now I search for potentially novel effects assuming no a-priori polarity, so I use an F-contrast (as suggested in the SPM manual). Even though the N400 pops up, its effect is not FWE-significant at the cluster level (see figure F24x24x251). This is surprising to me because the effect is really strong, and it comes out highly significant also with the Fieldtrip non-parametric cluster analysis (which is pretty conservative). Is it normal to have a so big difference between the t-test and the F-test? Are there any indications on how to test for effects with unknown polarity (which you would like to be able to do all the time with EEG) without the risk of seen huge effects non-significant? Or should I take it as a hint that the effect in my data is not that robust? Source reconstruction: - Should I crop the ERPs to the time interval of the effect (let's say 20 ms long), or should I do source reconstruction on all the epoch and then choose the temporal window when seeing the results? Apologies if these basic questions have been already covered. Thanks a lot for your help, Best, Marco -- Marco Buiatti, PhD CEA/DSV/I2BM / NeuroSpin INSERM U992 - Cognitive Neuroimaging Unit Bât 145 - Point Courrier 156 Gif sur Yvette F-91191 FRANCE Ph: +33(0)169.08.65.21 Fax: +33(0)169.08.79.73 E-mail: [log in to unmask] http://www.unicog.org/pm/pmwiki.php/Main/MarcoBuiatti ***********************************************