Dear List,
In order to figure out how to do SEM with brain data, we've taken to the attention to motion dataset which is from the original Buchel & Friston SEM paper (Modulation of connectivity in visual pathways by attention: cortical interactions evaluated with structural equation modelling and fMRI.; 1997). As best as possible we've attempted to replicate the order of steps in the paper in SPM8 and have had miserable luck.
I was wondering if anyone else had attempted this before? Or if the authors are monitoring the list maybe they can shed some light.
To give an overview of what we've done. We modeled the design using the included regressors in the attention to motion dataset. This was pretty straightforward and as best as we can tell the GLM is identical. We preformed what we believe is the appropriate test in the paper. The paper described a contrast of attention vs. no attention and no attention vs. stationary. As the reported results were F-values we assume this amounts to (1 -1 0; 0 1 -1) assuming col1 is attention, col2 no-attention and col3 is stationary. And here's where our problems begin:
1-In SPM8 we get coordinates that are almost identical to those for subject 1 in the paper (although usually off by 1 voxel in the x,y,z), but flipped on the X (probably due to our defaults). Which is why I'm assuming the attention to motion dataset is subject 1 from the Buchel & Friston paper. Thus, I think we got the contrast correct but the F-values are off by large amounts. For instance, instead of V1 having an f-value of 255 we get 55.
2-Nevertheless, we managed to get ROIs in almost identical regions as the original paper. We adjusted with effects of interest and grabbed the first eigenvariate. Next we attempted to generate the covariance matrix for SEM. The paper says 96 timepoints were used for the SEM model, however the experiment has only 8 blocks x 10 TPs per condition. I'm assuming for each block of attention or no-attention an extra 2 TPs were added to account for hrf delay? Ergo 12 timepoints per condition instead of 10?
3-We normalized the timeseries in each region for attention and no-attention, generated the covariance matrix and used Mplus to set up the first model reported in the paper (the simple model of V1-V5-PP). However, our results are radically different from what is reported in table 2. The biggest discrepancy is that the paper reports very weak V5-->PP relations in the no-attention condition, whereas the relations are fairly strong in our output.
So. We've either got the wrong subject, wrong regions or all the changes since spm96 to spm8 make this sort of replication impossible. I don't suppose someone has the actual timeseries used as input to the SEM sitting around in a text file? Even just the covariance matrix would be a huge help!
Alternatively, if someone has thoughts on how to extract and parse the timeseries for SEM that would be helpful. My understanding is there's a lot of manual work on the front-end to take a timeseries and chop it up into separate conditions in order to generate the covariance matrix. What is less clear to me is how own accounts for HRF delay. Cut the timeseries 6 seconds following the first stimulus onset of the block and add a further 6 seconds at the end seems reasonable, if a little arbitrary.
Apologies for the long message. Thanks in advance for any help or insight.
|