Print

Print


On Wed, Nov 20, 2013 at 6:34 PM, Michael Cohen <[log in to unmask]> wrote:
Hi,
I have been having a few issues in trying to do a PPI analysis, so I was just wondering if somebody might be able to help clarify things for me.

To briefly introduce my paradigm, we present some to-be-learned items that are high-value, and some that are low-value.  We're trying to do a brain-behavior correlation between PPI beta values representing connectivity at encoding, and the degree to which value influences memory at recall.  We're using the McLaren gPPI approach, where the high-value PPI regressor = 1 and the low-value PPI regressor = -1 in the model.  (Actually, there are 2 separable parts of each trial, so I have two pairs of PPI regressors in the full model, but hopefully that's not a problem.) We're using Matlab scripts developed at UCLA that allow us to compute PPI regressors in SPM while doing the rest of the analysis in FSL, to account for the issue raised in the Gitelman et al. (2003) paper, which, as I understand it, is particularly salient for an event-related design, as we have.

So here are the issues that I'm having with the analysis.  (These are independent issues, so if you can help with just 1 or 2 of them, that would still be very helpful.)

1. In trying to break down the PPI effects across the different conditions, I was given some advice (originating from Jeanette Mumford) that the physiological regressor should be included together with the relevant PPI regressor, instead of just looking at effects of the PPI regressor by itself.  I understand that this puts back effects of non-task-related connectivity that would otherwise be removed in the model, but I'm not sure under what circumstances it's necessary to do that.  If I'm just trying to see what's driving the brain-behavior correlation in the overall PPI effect (e.g., whether it's individual differences in connectivity during encoding of high-value items, or individual differences in connectivity during encoding of low-value items), my intuition is that it's better not to include the physiological regressor, but I'm not entirely sure.  Any advice on this?

>> The physiological regressor needs to be in the model. When you form your contrast, the weight of physiological regressor will be 0. If you have a contrast of 1 over the high-value PPI column and 0s for all other columns, then the contrast indicates the change from baseline.

 

2. It seems like most of our PPI effects only show up going in one direction, even though the PPI analysis should theoretically be nondirectional.  So for example, we see a brain-behavior correlation in hippocampus-VLPFC connectivity when using a hippocampal seed, but not when using a VLPFC seed, using the same ROIs.  Would this cause concern about whether the effect is real, or is it typical for PPI?

This is typical. Usually what you will see is the PPI difference being positive between A and B and negative between B and A. The difference is that you've flipped the X and Y axis. A to B might have slopes of 3 and 2 for the two conditions, so 3-2 would be 1. If you flip the axis, then it would be 1/3 and 1/2, and the difference would be -1/6. Additionally, the deconvolution step makes each direction slightly different as well.

 

3. Although we're getting a correlation between the magnitude of the PPI effect and the magnitude of the behavioral effect between individuals, we're not getting a main effect for the PPI, and it's not clear why that is.  I did find some other papers in the PPI literature that reported a within-subjects correlation but no main effect (e.g., Ofen et al., 2012, in J. Neurosci., Cremers et al., 2010, in Neuroimage, and Passamonti et al., 2009, in J. Neurosci.), but it seems like this could still be a major issue with our data.  I'll attach a scatterplot from one of our analyses to help illustrate this. Is there anything non-intuitive about PPI that might lead to this pattern of results?

I am assuming that you have a design matrix of 2 columns, the first column being a constant and the second column being your selectivity index. From the scatterplot, it  look like you did not remove the mean of the covariate. In this example, the coefficient for the constant column should be negative, and I would expect it might be significant. There is nothing that says the connectivity must be significantly different than 0 in all subjects, especially when its related to behavior. If you scanned people with very high selectivity index, then the PPI would be positive, if you scanned people with very low selectivity index, then the PPI would be negative -- at least in your example plot.
 

4. In setting up the PPI analysis, things seem to get more complicated because we're trying to mix FSL and SPM, but there's one particular piece that I'm uncertain about, which I tried implementing in two different ways.  The first approach was to do a nuisance analysis that incorporated all preprocessing steps, and then use the residuals of that analysis as the input for subsequent analyses, including to extract the seed and to run the first-level PPI analysis.  The second approach was to do everything on filtered_func_data from the univariate analysis, as Jill O'Reilly has suggested, while not running a separate nuisance analysis.  My understanding is that when using the gPPI approach, it's not necessary to do the nuisance analysis.  However, when I tried comparing the two methods, all of my correlations seemed a bit stronger when using the nuisance analysis.  I can't think of any way that this approach is "cheating" (we do also add in 6 empty EVs to the first-level PPI model to account for the extra degrees of freedom lost by putting the motion parameters in the nuisance analysis), so I think the nuisance model is just a bit better because it's getting rid of more noise...but again, I'm hoping that somebody who understands this stuff more deeply than I do can tell me if this is reasonable.

Did you remove the effect of motion from the timeseries before doing the deconvolution step? If not, then this could definitely be the source of the difference. In SPM, you have the option to adjust the timeseries and remove the nuisance effects before deconvolution.

If you look at your two models, Are the design matrices for the task regressors, PPI regressors, and phsyiological regressors the same? I suspect that they are different and the improved effects could be due to the differences in the deconvolution step.

 Hope this helps.

You might want to see if you can integrate the gPPI toolbox into the analysis stream. 


Anyway, any guidance that any of you might be able to give would be very helpful...

Many thanks,
Michael





--
Michael S. Cohen, M.S., C. Phil
Ph.D. Candidate
Department of Psychology
University of California, Los Angeles
Los Angeles, CA 90095