Back to PLS HelpI'm Online
Posted on 03/01/17 11:50:16
Number of posts: 4
Hello all,
I've conducted a non-rotated task PLS analysis (2 groups, 3 conditions) to explore group and condition effects. The contrast for a group main effect is significant (p<0.011). However, the confidence intervals in the brain scores plot are very very large and all cross zero. https://www.dropbox.com/s/6ce1cwoj6jqr9e8/nonrotated_MEage_BSPlot.JPG?dl=0
In terms of the brain scores themselves, I think I am correct in thinking that I would not necessarily expect the brain score plot (all positive values) and design score plot (positives and negatives) to look the same - from browsing of previous Message Board posts I think it’s confirmed to me that: e.g. if a contrast is entered so one group has 1 1 1 and other group -1 -1 -1, then the brain score plot may not necessarily follow the valence of the design score plot and contrast, but should follow the general pattern so that one group has higher brain scores than the other group, but all could be positive, which is the case in these results.
In terms of the confidence intervals though, could they be so large, with none that do not cross zero?
(I have used the boot_results ll and ulusc to double-check plotting – and I downloaded the latest version from the PLS webpage (which says Version 6.1311050 But I saw on an older post that this should still be the latest version.))
I tried a mean-centered task PLS analysis to see what comes out of a more data-driven approach… There is one significant LV (p<0.018) which appears to be a similar group effect: brain scores for all 3 conditions for one group are positive, and all 3 conditions for the other group are negative, and here all error bars do not cross zero. https://www.dropbox.com/s/704cgkux4b9ykxs/meancentered_LV1_BSPlot.JPG?dl=0
The spatial map pattern associated is similar to that associated with that from the non-rotated PLS for the MEgroup contrast.
I am not sure where I have gone wrong here…any help and advice would be appreciated!
Also, the brain scores from the mean-centering PLS LV are exactly flipped values for the two groups (i.e. groupAcond1 orig_usc -40, llusc 36 and groupBcond1 orig_usc 40, ulusc -36)… is this to be expected?
Thank you very much,
Carolyn
Replies:
Posted on 03/01/17 11:57:16
Number of posts: 394
that is odd. Can you post the results file so I can take a look at the bootstrap distribution for the non-rotated case?
I'm Online
Posted on 03/01/17 12:31:40
Number of posts: 4
Hello Randy,
Here is the link: https://www.dropbox.com/s/n9s8842ao01reta/gp_cond_gpxcond_BfMRIresult.mat?dl=0
Thank you!
Carolyn
Posted on 03/02/17 22:05:05
Number of posts: 394
Hi Carolyn,
I've taken a look at your results and I've identified a red flag but not the source. If you look at the bootstrap distributions, there are negative values in it that pull the confidence intervals down below zero
try: hist(squeeze(result.boot_result.distrib(1,1,:)))
and you'll see this
I've id'd the resampling that show this so the question is whether there is something odd with the persons in that sample?
find(squeeze(result.boot_result.distrib(1,1,:))<0)
What would be instructive is to save the datamat in the results file (there is a tickbox for this) and then we can explore further. I suspect it has to do with mean centering but am not 100% sure.
I'm Online
Posted on 03/03/17 04:45:22
Number of posts: 4
Thank you for taking a look Randy,
Hmm, I see what you mean about the bootstrap distributions.
I've rerun it and here's the results with the stacked datamats...2 groups (N of 24 and 35), 3 conditions, variability data.
https://www.dropbox.com/s/d4pxy64c78uaq2q/gp_cond_gpxcond_MC0datamats_BfMRIresult.mat?dl=0
Kind regards,
Carolyn
Posted on 03/05/17 15:48:22
Number of posts: 394
Thank you for taking a look Randy,
Hmm, I see what you mean about the bootstrap distributions.
I've rerun it and here's the results with the stacked datamats...2 groups (N of 24 and 35), 3 conditions, variability data.
https://www.dropbox.com/s/d4pxy64c78uaq2q/gp_cond_gpxcond_MC0datamats_BfMRIresult.mat?dl=0
Kind regards,
Carolyn
So, I think I know what's going on but I don't have a solution yet. It seems that for certain bootstrap samples there are collections of voxels that are much higher in group two than group one, hence have a negative weight in the singular vector and drag the scores down for both groups so that they drop below zero on those iterations. These voxels also have high SE estimates in the bootstrap so end up not being reliable.
What I am not clear on is why this doesn't happen when you do the mean-centred version. Can you please post the results file from the mean-centred analysis so I can compare?
I'm Online
Posted on 03/06/17 09:22:37
Number of posts: 4
Hello again Randy,
Here is the mean-centered version: https://www.dropbox.com/s/8rmu4jk1wpdvmx0/Group_3Cond_MC1datamats_BfMRIresult.mat?dl=0
Originally I used the mean-centering type 1 to focus on group differenes, as the non-rotated PLS came out with a group difference…
I just ran the mean-centering type 0 though, and this is quite a different outcome: https://www.dropbox.com/s/ji64qb0xgdee3z0/Group_3Cond_MC0datamats_BfMRIresult.mat?dl=0
Kind regards,
Carolyn
Login to reply to this topic.
