 |
 |
Hi Ben let me have a shot at this ... > Dear SPMers, > I have run an fMRI parametric working memory task. Now I have some > questions regarding the parametric analysis. > The task: > 4 conditions of working memory load (participants have to maintain > 1/2/3/4 items) > 1 baseline condition (during which participants view a fixation cross > My questions: > 1. Modeling the parametric response: > On the first level I have modeled the 4 load conditions as 1 > condition with a parametric modulation (entering a vector with > 1,2,3,4 according to the onset times of the 4 conditions) > As a result I get a design with: > First row: onset times > Second row: parametric modulation > (please correct me if I’m wrong) > Does the t-contrast 1 1 return the parametric contrast (the > load-dependent working memory network)? If not: what would the > appropriate contrast be? the regressors have a unique contribution to explain your data - this means that the 1st regressor will explain/fit data where the BOLD response is identical for the 4 loads whilst the 2nd regressor will explain/fit data where the BOLD response varies (above what can be explained by the 1st regressor) according to the load. So a F test [1 0; 0 1] will give you voxels where one of the regressor or their combination fit the data and T tests [1 0] [-1 0] [0 1] [0 -1] should be used to look their unique contributon (positive or negative fit). > 2. Would it be reasonable to include the baseline in this design? no there is no need for that - if your design is condition A and baseline, and you tell me which scans are condition A obviously I can deduce which ones are baseline, same with SPM Stats ... > 3. Alternatively I used a different first-level analysis: > > I entered 5 conditions: baseline / load1 / load2 / load3 / load4 and > their onsets > No parametric modulation. > Then entered the t-contrast: > -6 (baseline) 0 (load1) 1 (load2) 2 (load3) 3 (load4) > Would this approach display the load-dependent working memory network? > If it does: which approach (1 or 3) should be preferred? this is a almost good alternative - ie remove the baseline condition (note I assume that this is doing not much) - so now that you removed baseline you have load 1, 2, 3, 4. Each regressor will fit data for each condition separately - one disadvantage of this model is that you cannot ask if there are voxels that respond to the task but not vary with the load - one advantage is that now you can specify many different contrasts between conditions like [-1.5 -.5 .5 1.5] looking for a linear increase of the signal with the memory load or [-0.7945 -0.1014 0.3041 0.5918] for a log increase or [-18.4795 -13.8087 -1.1122 33.4004] for an exponential increase .. etc - note that you could also create several parametric regressors using exp or log but the difference will be that additional parametric regressors only explain above what is already modelled, ie the order you input data matters that is to say 1st linear fit, then the remaining non explained data are fitted by exp, etc .. the second kind of designs is more flexible if you don't have a priori on the shape of the response. Hope this helps Cyril -- The University of Edinburgh is a charitable body, registered in Scotland, with registration number SC005336.