 |
 |
Dear Kunil > I would like to ask questions regarding how to implement the DCM/PEB design in the experimental setting and interpret the findings. My experimental design consists of three phases, 1) presentation of one of four contextual information (e.g. the smell of urine, fish stinks, water, and rose), 2) presentation of neutral facial information, and 3) the estimation of the pleasantness rating of the facial expression based on the contextual information. For simplicity, I will just use phases 1, 2, and 3 for further reference. The task design is to find neural correlates of the individual differences in contextual information usage when processing facial expressions. Ok that’s clear (I’ve attached the experimental design that you sent me). > At phase 2, GLM analysis suggested two different brain regions (sgACC and right IFG) associated with individual differences. For phase 3, we found the amygdala and the pgACC. We wanted to investigate the modulatory effects of the brain regions from phase 2 (sgACC and rIFG) to phase 3. However, my question is whether we can use the GLM findings from phase 2 to infer the connectivities at phase 3, vice versa. If it is okay to do it, then is it recommended to build a single full model containing all four brain regions, or can we build two DCM models containing three regions (I. e. sgACC, amygdala, and pgACC and IFG, amygdala, and pgACC). I know that the three models cannot be compared to each other for the selection of the best model because their brain regions (data) are different. Yes, you can use different contrasts to select different regions of interest. You will need to form a single DCM consisting of all regions (because you are modelling the entire timeseries with fMRI – unlike EEG/MEEG, where you could the timeseries into bits). > For another question, when we set the experimental input, we realized that we can either set the input as the onset regressor or the parametric modulation regressor. If we used the experimental input as the parametric modulation regressor, how should we interpret the "commonalities" in the PEB analysis? As the first regressor represents the mean modulatory changes in the PEB analysis, should we think of the mean modulatory changes in terms of the parametric modulation? I wonder if I can state in this manner: As the valence of the rating X increased, the average effective connectivity increased between regions A and B. Or is it recommended to make separate experimental input (e.g. neutral vs emotion) to investigate the modulatory changes? If we were to set two experimental inputs in phase 2, (e.g. onset regressor of phase 2 and the parametric modulator of phase 2, I wonder if I am investigating two independent connectivities at phase 2 with different inputs or one connectivity analysis with two different inputs. The first step is to work out which effects pertain to the within-subject level (DCM) and which pertain to the between-subject level (PEB). At the within-subject level, you mentioned you have two experimental factors, which are context and rating . Context has four levels and rating is continuous, so these will go in the DCM as modulatory parameters (matrix B), encoding the change in a connection strength due to context, or the change in connection strength per unit valence. You didn’t mention between-subject factors, so I’ll assume that you are just using PEB to model the average parameters across subjects (i.e. the commonalities that you mentioned). On the specifics of how to set up the DCM, you first need to establish what your hypotheses are. Are there particular regions or connections in the network, that you think may be modulated by context or valence? What do you think context or valence might be doing at the neural level? Ideally, you should be able to draw out each of your hypotheses as a network diagram, with modulatory effects on different connections. You then use the GUI to make a “full” DCM with all parameters of interest, and compare it to reduced models with particular parameters switched off. For example, you might hypothesise that context and valence could change the responsiveness of some of your regions (i.e., alter their sensitivity or synaptic gain). So include the separate contexts, and the parametric valence regressor, as modulators on the self-connections of particular regions. For the driving input, you could drive certain regions by phase 1 (2000ms blocks in one regressor, regardless of context), you could drive certain regions by phase 2 with a regressor having 4000ms blocks, and drive certain regions by phase 3 with variable-length blocks for the decision period. If you need to make any regressors you don’t have at the moment, you can create a new GLM with these in (for example, a new GLM with a phase 1 regressor). I hope that provides a useful starting point. Best Peter