 |
 |
Dear Peter, I think I am getting a much better sense of what you have suggested now. There is one final thing I hope you won't mind explaining a bit further: regarding the "trial" regressor, while you said it's for the whole trial, the delay isn't included in this, correct? Thus, the 'trial' regressor would include time points and duration for the stim_array period, cue, and probe. The delay, on the other hand, is modeled separately (and for each task condition - color, shape, and color+shape). Does it make sense to "chop up" the trial like this (i.e., model events before and after delay as "trial" while modeling the delay itself according to the task condition). I guess it makes sense because delay is the period with the task differences I am interested in. I was also considering treating this as a "block design". This way, for each trial I would model the three events including cue, delay, and probe as a single event with a boxcar function. This way, I would have color blocks, shape blocks, and color+shape block. I am unsure about the stim_array as this one is similar across all three conditions (the stim array always contain objects with shape and color). Perhaps the stim array doesn't have to be modeled or is modeled as "trial" regressor as you have proposed. Does it make any sense? This is a bit different from your recommendation, which I will definitely try, but I just thought maybe this approach is also worth a shot. I'm sorry for taking up your time and others' here so I am grateful for the patience and expertise that you are so willingly share. Best Miranda E. On Thu, Oct 13, 2016 at 5:30 AM, Zeidman, Peter <[log in to unmask]> wrote: > Dear Miranda > > Apologies for the delay in replying. > > > > Thank you for your quick and helpful response. I've checked my design > orthogonality. Except for a very small numbers of squares with the cos > values (I'm assuming these are the ones you referred to) exceeding 0.7, the > majority are < .2. > > > > Note that the important collinearities are the ones you wish to contrast. > If you are contrasting conditions A and B, and the regressors for these are > highly collinear, your results may be affected. If regressors that are not > of interest are collinear with each other – e.g. the motion regressors – > then that’s fine. > > > > Regarding your recommendation on creating the 4 regressors in GLM, just so > I know I understand you correctly, for the Trials (blocks covering the > while trial period), the vector would have all 1's, each 1 is for an event. > If I have the following 12 events: color stim_array, color cue, color > delay, color probe, shape stim_array, shape cue, shape delay, shape probe, > color+shape stim_array, color+shape cue, color+shape delay, color+shape > probe, then the vector would look like this: > > > > [1 1 1 1 1 1 1 1 1 1 1 1]. Would this be correct? Somehow I have a > suspicion that it isn't. > > > > I guess the Color Delay is more straightforward, I would just have '1' for > the color delay event so it would be something like this [0 0 0 0 0 0 1 0 0 > 0 0 0] > > > > Also, for the effects of interest, each event would also get the value of > 1 so the identity matrix would be: > > [1 1 1 1 0 0 0 0 0 0 0 0] > > [0 0 0 0 1 1 1 1 0 0 0 0] > > [0 0 0 0 0 0 0 0 1 1 1 1] > > > > First, note that you don’t need to model the stim_array, cue and probe > separately for color, shape and color+shape. You only need to model these > separately if you’re interested in the differences between conditions. > > > > With regards to the vectors you provide above, I think there may be some > confusion between conditions (regressors or columns in the GLM) and > contrasts. Each “condition” consists of the onset times and durations of > the trials, which the software convolves with the haemodynamic response > function to form a regressor in the GLM. My suggest is to have conditions: > trials, color delay, shape delay, color+shape delay. The “trial” regressor > models the average BOLD response of the whole trial (the stim_array, the > cue and the probe). The delay conditions model the additive BOLD response > of being in each delay period. > > > > Let me know if you need further clarification. > > > > Best > > Peter > > > > Could you please confirm? Again, thank you for your kind advice and > support! > > > > Best, > > > > Miranda E. > > > > On Wed, Oct 5, 2016 at 9:07 AM, Zeidman, Peter <[log in to unmask]> > wrote: > > Dear Miranda > > > > - Should all the event onsets (e.g., onset for stimulus arrays, cue, > delay, probe) be included as regressors? As this is an event-related > design, for other analyses, the natural answer would be yes. However, for > DCM do the events matter or only the conditions matter? > > > > The design considerations for DCM are identical to those for the standard > SPM analysis. If you model each of those parts of the trial, you may find > there is collinearity between the parameters (I’m assuming the onset, cue, > delay and probe are all presented on every trial with similar timing). You > could verify this by clicking Review in SPM, opening a single subject’s > SPM, click Design in the grey window, and click Design Orthogonality. With > highly collinear parameters (dark squares in that GUI, with values greater > than around 0.7), it will be hard to get significant results. > > > > That said, you said you’ve already found results that you’re happy with, > so it may be not be causing a big problem for you. One further caveat. The > basic deterministic DCM assumes all effects are driven by the stimuli – > there are no endogenous brain dynamics modelled. If you have a long delay > period, that could cause problems for you. > > > > Here’s one way to model this design, which is probably what I would try > first. Have 4 regressors in the GLM: Trials (blocks covering the whole > trial period), Color Delay (blocks or events for just the Color delay > periods), Shape Delay (blocks or events for just the Shape delay periods), > and Color+Shape Delay. Use this design to extract ROIs based on the Trials > regressor for the driving input region, and use contrasts comparing color > and shape delay periods to select the other ROIs. Then for the DCM, have > Trials as the driving input and the other regressors as modulatory inputs. > > > > (Another option would be to form a Color-Shape regressor that represents > the difference between color and shape, and use that as a modulatory input, > alongside Color+Shape.) > > > > I hope that helps! > > > > Best > > Peter > > >