 |
 |
Dear Liang, On Thu, Aug 11, 2011 at 5:53 PM, liang wang <[log in to unmask]> wrote: > Dear Vladimir, > > Thanks very much for your kind explanations, which are very helpful. > However, several problems to me still exist. > > 1) When I setup a model (select 'IND' while specifying new data, select > 'LFP' in neural model step, and specify a couple of linear and nonlinear > connection), and run "invert DCM". Then Save the results as a specified > name. However, when I load that saved DCM file, I always get a error "please > specify # source locations" and the data type is automatically changed to > 'SSR'. I am not sure whether or not this is a bug for intracrani/local field > potential data analysis. > Yes, this is probably a bug. If you can change the DCM type back to IND and it works then there is no problem. I'll try to fix it when I have time. > 2) in last email, you said "you should look where you start getting activity > deviating from the baseline and put your input around that > time point. If your event is something like a button press then the input > will probably come before the button press". I completely understand this. > But I am not sure when the input happens. I think that the onset of the > input is the onset of the condition (or stimulus / event), right? this input > should be independent of the analysis window under "data and design", right? > Input timing is optimised so that it'll be where I said, close to the onset of the activity. You should just initialize it more or less to the right time. Everything in your analysis window that is before the input will be modelled as flat. But as I said the input timing will not necessarily stay as you defined it so it makes sense to include a baseline period. Also the code substract the first sample from the data so the window should start somewhere where there is no activity. > 3) do you think is that possible to investigate dynamic properties of a > causal model over time. If so, I would like to divide the data length into > different stage to do that. > I don't think that's a good idea because DCM models the actual physical process, in the case of DCM-IR - the system is at rest then comes the input, causes some dynamics and then it dissipates. You can't model the dynamics from the middle, you must always include the onset. What you can do is make the window longer or shorter at the end. This was done in Garrido et al. PNAS, 104(52):20961-20966, 2007 for DCM-ERP and could also apply to DCM-IR. But I don't know whether it'll help to answer your particular question. Best, Vladimir > Best wishes and thanks a lot. > Liang > > > 2011/8/10 Vladimir Litvak <[log in to unmask]> >> >> Dear Liang, >> >> On Wed, Aug 10, 2011 at 5:37 PM, liang wang <[log in to unmask]> wrote: >> > Thanks for your correction. I think I should use DCM for induced >> > responses, >> > but I am still confusing, though I have read several relevant papers. >> > >> > I do not clear know what's the main differences between DCM for evoked >> > response (ERP) and for induced responses (IR). From the definition for >> > two >> > terminology, evoked response is phase-locked to the stimulus, whereas >> > induced response is non-phase-locked to the stimulus. By reading SPM >> > manual, >> > I understand that DCM for IR models coupling within (i.e., linear) and >> > between (i.e., non-linear) frequencies across the sources. However, I >> > think >> > this different viewpoint by looking at coupling in temporal and >> > frequency >> > domain may not clear distinguish the two DCM variants. >> >> Lets make things simpler and define that evoked responses (a.k.a. >> ERP/ERF) for the purposes of DCM are waveforms computed by averaging >> EEG/MEG signals in the time domain around events of interest (which >> might be stimuli or responses). Induced responses are time-frequency >> matrices computed by averaging time-frequency images of power of >> EEG/MEG around events of interest. You get those images by using >> Fourier analysis or something similar. There is a different sense to >> evoked and induced that was recently discussed here >> (http://www.jiscmail.ac.uk/cgi-bin/wa.exe?A2=SPM;e02519b7.1105) , but >> lets ignore that kind of distinction for now. >> >> > >> > Also, SPM manual 16.11.1 says that "we will model the entire spectra, >> > including both the evoked and induced components", which means that DCM >> > for >> > IR covers both. >> >> That's exactly the distinction that is not very important for the >> present discussion. >> >> > I think this understanding is true if I epoch the data >> > around the stimulus onset. In the paper (Chen et al., J Neuro 2010), >> > they >> > epoched the data around the movement onset, whereby they could correctly >> > call induced response (non-phase-locked to the stimulus). >> >> Induced responses can be induced by stimulus or response. DCM-IR does >> not look at phase-locking. It models TF images that include both >> phase-locked and non phase-locked components. >> >> > In addition, I do >> > not know what's the onset (default 60ms) used for model specification? >> > which >> > region received the external input (that paper did not show this input)? >> > >> >> Yes , input is not mentioned in the paper but it's just an omission by >> the authors. There definitely were inputs in their model. I don't know >> exactly how they were set up but it couldn't be any other way. 60ms is >> something that is there for historical reasons and is not relevant for >> DCM-IR. As I mentioned before you should look where you start getting >> activity deviating from the baseline and put your input around that >> time point. If your event is something like a button press then the >> input will probably come before the button press. >> >> > To my understanding, DCM for IR likes bi-directional connections. If I >> > specify unidirectional connection, the model will converge very fast and >> > predicted data for that region involved will be empty. Is that >> > reasonable? >> >> It's not necessarily a matter of unidirectional vs. bidirectional but >> the model needs to be complex enough to generate interesting dynamics. >> If all your connections are linear, the activity will dissipate very >> fast and won't be able to fit anything. Under these circumstances the >> Bayesian scheme might prefer to shrink the parameters to zero. So try >> for instance making the connections nonlinear. >> >> > I can not specify linear connection within a source (disable option), is >> > this >> > connection specified as default? >> >> Yes. There must be at least intrinsic linear connection. >> >> > >> > Last question, could you guide me how to check how much variance is >> > explained by the specified frequency modes, because Chen's paper >> > mentioned >> > this information. Otherwise, I need compute them manually. >> > >> >> You can find in DCM.xY.S the singular values corresponding to the >> modes. You can use them to compute variance explained by something >> like: >> >> S = DCM.xY.S; >> >> VE = S.^2./sum(S.^2); >> >> Vladimir >> >> > Sorry about a lot of questions from a new DCM explorer. >> > >> > Best, >> > Liang >> > >> > 2011/8/9 Vladimir Litvak <[log in to unmask]> >> >> >> >> Dear Liang, >> >> >> >> It sounds like you are somewhat confused by the different DCM >> >> variants. DCM for induced responses does not distinguish between >> >> forward and backward connections, just between linear and nonlinear. >> >> You should usually have some well-defined model (or a series of >> >> models) to specify which are which. See Chen et al. NeuroImage, >> >> 45(2):453-462, 2009 and Chen et al. J Neurosci. 23;30(25):8393-9 for >> >> examples. >> >> >> >> SPM and DCM cannot handle variable length trials. Also since DCM for >> >> induced responses models the average response your must take a set of >> >> trials that when averaged will yield something meaningful. So I >> >> suggest that you select a set of trials with comparable durations and >> >> cut them to the same length. Usually for induced responses your trials >> >> will be longer than 200-300ms especially for areas high in the >> >> hierarchy. But you can just look at your averaged TF data and select >> >> the trial boundaries so that the interesting features will be >> >> included. The input should be more or less at the time when the >> >> activity starts departing from baseline. >> >> >> >> Best, >> >> >> >> Vladimir >> >> >> >> On Mon, Aug 8, 2011 at 10:18 PM, liang wang <[log in to unmask]> wrote: >> >> > Hi DCM experts, >> >> > >> >> > I would like to use DCM to investigate information flow direction >> >> > across >> >> > several attention regions (mainly focusing on frontal parietal >> >> > attention >> >> > network). The regions selected are ventral medial frontal gyrus >> >> > (vMFG), >> >> > middle frontal gyrus (MFG), dorsal premotor motor cortex (dPMC), >> >> > intraparietal sulcus (IPS) and inferior parietal lobe (IPL). I have a >> >> > couple >> >> > of questions for specifying the connections. >> >> > >> >> > 1) to my understanding, all the regions are high-level regions. Is >> >> > that >> >> > possible that I just select those regions if I have a hypothesis >> >> > about >> >> > the >> >> > connection direction. I am not sure whether or not it is necessary to >> >> > specify forward (bottom-up) and backward (top-down) connections in >> >> > this >> >> > situation (or leave them alone). The reason for no visual region is >> >> > that >> >> > the >> >> > electrodes did not cover that area. >> >> > >> >> > 2) I want to focus on induced responses (not evoked response). Could >> >> > I >> >> > set >> >> > the start window, like starting at 200ms? and chane 60ms onset >> >> > (default >> >> > setup for neural response to my understanding) to 0? >> >> > >> >> > 3) because of variable stimuli duration for my experiment, is it >> >> > possible to >> >> > make each trail have different data length? what is the minimum >> >> > (restricted) >> >> > data length for DCM analysis? (because sometime one of trails has >> >> > very >> >> > short >> >> > duration, like 300ms). >> >> > >> >> > 4) (the last question may be not related to this topic) could you >> >> > give >> >> > me >> >> > some advice how to specify forward or backward connection. Let me say >> >> > if >> >> > I >> >> > have several regions including thalamus and few cortical regions. For >> >> > an >> >> > attention task, thalamus is an important area, partly because this >> >> > area >> >> > projects a lot of anatomical connections to posterior part of the >> >> > brain. >> >> > However, it is hard to simply specify it as low- or high-level >> >> > region. >> >> > To >> >> > me, a simple solution is trying each possible model and compare those >> >> > model >> >> > based on BMC tool (too much time will be taken). Maybe there is >> >> > alternative >> >> > ways to easy achieve that. thanks. >> >> > >> >> > Please let me know if you think some steps are important for this >> >> > type >> >> > of >> >> > analysis that I need pay more attention on. >> >> > >> >> > Best, >> >> > Liang >> >> > >> >> > -- >> >> > Liang Wang, PhD >> >> > Neuroscience of Attention and Perception Laboratory >> >> > Princeton Neuroscience Institute >> >> > Princeton University >> >> > Princeton, NJ, 08540 >> >> > >> >> > >> >> > >> > >> > >> > >> > -- >> > Liang Wang, PhD >> > Neuroscience of Attention and Perception Laboratory >> > Princeton Neuroscience Institute >> > Princeton University >> > Princeton, NJ, 08540 >> > >> > >> > > > > > -- > Liang Wang, PhD > Neuroscience of Attention and Perception Laboratory > Princeton Neuroscience Institute > Princeton University > Princeton, NJ, 08540 > > >