Back to PLS HelpPosted on 12/11/13 09:09:12
Number of posts: 13
Dear PLS experts,
I have a block-design fMRI task with three different conditions and two external variables reflecting behavioral performance. I am interested to look at the networks of the brain that are not only modulated (similarly and/or differentially) across different conditions, but also at the same time, are correlated with two behavioral measures. To be able to do such analysis, I have used multi-block PLS.
I got two significant patterns: LV1 reflected condition C (baseline) vs. Condition A&B(task). Since I am interested in negative saliencies of LV1 (shown as blue in bootstrap plot) which reflect higher activation during task, I assume the sign of correlation between the brain scores and behavioral data (the one you get from Brain scores vs behavioral data plot) should be reversed for interpretation. Am I right?
Assuming that I am correct, I interpreted LV1 as a network which reflected higher activation during condition A&B than C, and this network is positively correlated with my two behavioral measures (Again, be aware that I reversed the sign). However, when I look at the datamat correlation response for the regions strongly contributed to this LV, it seems that the sign of the correlation for most of the regions is opposite to what was seen from the brain scores plot for behavioral analysis. It is indeed very strage! Am I missed something.
Can anybody explain why this is the case?
Many Thanks
Kambiz
Replies:
Posted on 12/11/13 09:22:20
Number of posts: 394
Hi Kambiz,
What are the signs for the correlations between brain scores and your behavior in the results? That may help explain the confusion.
Posted on 12/11/13 10:07:24
Number of posts: 13
Hi Kambiz,
What are the signs for the correlations between brain scores and your behavior in the results? That may help explain the confusion.
Dear Randy,
Thanks a lot for your reply!
Here come the full details of the LV that induces confusion.
LV2: Results of task PLs (Brain Scores): conA conB conC: -50 -10 +45
This is very clear that the LV reflects those brain areas with greater activation during condition A&B than condition C.
The strongest negative saliencies are: (-24 -74 46): BSR=-22.93 and (-48 -2 46): BSR=-17.54
Correlation between brain score of this LV and two behavioral measures are as follow:
Correlation between Brain scores of condition A and behavioral measure 1: r=0.38 (I assume I have to reverse the sign for this correlation because condition A reflected negative saliences on task PLS). Note that for this correlation CI does not cross zero.
Correlation between Brain scores of condition A and behavioral measure 2:
r = - 0.19 (again, the sign should be reversed) and CI does not cross zero.
Other correlations (condition B/C and behavioral measures) turn out to be unreliable given the zero-crossed CI.
Up to here, everything is clear. In sum, we have a network that reflect greater activation during task (condition A and B) than baseline (condition C). In addition greater activation of this network is negatively correlated with behavioral measure 1 and positively with behavioral measure 2 only for Condition A. Am I right?
Now, I want to check whether the strongest regions of LV2 show the same correlation with behavioral measure 1 and 2. So, I use datamat correlation response and enter the two aforementioned clusters:
Correlation between (-24 -74 46) and behavioral measure 1: r=- 0.10
Correlation between (-24 -74 46) and behavioral measure 2: r= 0.10
This is completely opposite to the sign of aforementioned correlation between brain score and behavioral measures. The same correlation has been observed for most of the clusters reliably contributed to this LV. I am well aware of the fact that the network is more important than any single cluster, but the fact that most of the clusters exhibit an opposite correlation with behavioral measures as compared to BS and behavioral measures sounds intriguing.
Do you have any idea what is going wrong?
Many thanks,
Kami
Posted on 12/11/13 12:58:08
Number of posts: 394
Dear Randy,
Thanks a lot for your reply!
Here come the full details of the LV that induces confusion.
LV2: Results of task PLs (Brain Scores): conA conB conC: -50 -10 +45
This is very clear that the LV reflects those brain areas with greater activation during condition A&B than condition C.
The strongest negative saliencies are: (-24 -74 46): BSR=-22.93 and (-48 -2 46): BSR=-17.54
Correlation between brain score of this LV and two behavioral measures are as follow:
Correlation between Brain scores of condition A and behavioral measure 1: r=0.38 (I assume I have to reverse the sign for this correlation because condition A reflected negative saliences on task PLS). Note that for this correlation CI does not cross zero.
Correlation between Brain scores of condition A and behavioral measure 2:
r = - 0.19 (again, the sign should be reversed) and CI does not cross zero.
Other correlations (condition B/C and behavioral measures) turn out to be unreliable given the zero-crossed CI.
Up to here, everything is clear. In sum, we have a network that reflect greater activation during task (condition A and B) than baseline (condition C). In addition greater activation of this network is negatively correlated with behavioral measure 1 and positively with behavioral measure 2 only for Condition A. Am I right?
Now, I want to check whether the strongest regions of LV2 show the same correlation with behavioral measure 1 and 2. So, I use datamat correlation response and enter the two aforementioned clusters:
Correlation between (-24 -74 46) and behavioral measure 1: r=- 0.10
Correlation between (-24 -74 46) and behavioral measure 2: r= 0.10
This is completely opposite to the sign of aforementioned correlation between brain score and behavioral measures. The same correlation has been observed for most of the clusters reliably contributed to this LV. I am well aware of the fact that the network is more important than any single cluster, but the fact that most of the clusters exhibit an opposite correlation with behavioral measures as compared to BS and behavioral measures sounds intriguing.
Do you have any idea what is going wrong?
Many thanks,
Kami
Ah, I see. You should not reverse the signs of the correlations between brain scores and behavior. Remember that the brain score is the weighted sum across the entire brain, where postiive and negative weights both contribute, so leave the sign of the correlation as is.
Posted on 12/12/13 05:34:04
Number of posts: 13
Ah, I see. You should not reverse the signs of the correlations between brain scores and behavior. Remember that the brain score is the weighted sum across the entire brain, where postiive and negative weights both contribute, so leave the sign of the correlation as is.
Dear Randy,
Thanks a lot for your reply! I still feel that there is something fishy going on. I think this is a very critical issue to be solved since I know couple of papers in which the authors reversed the sign. I would appreciate your patience in order to get the problem solved.
If I don’t reverse the sign, as you suggested, then everything goes wrong for all LVs (excepect an small part of LV2). Let me just give you an example from the first LV:
LV1: Task PLS (Brain score): ConA ConB ConC: -38 -5 43
It is clear that this LV reflects those brain areas with greater activation during condition A&B than condition C. Note that the sign for brain scores of the task conditions (A&B) are negative
Correlation between brain score of this LV and two behavioral measures are as follow:
Correlation between Brain scores of condition A and behavioral measure 1: r = - 0.38 (without reversing the sign).
Correlation between Brain scores of condition A and behavioral measure 2: r = 0.25 (again, the sign hasn’t been reversed)
The correlation between brain scores of condition B/C and behavioral measure 1/2 also mimics the similar correlations reported above (i.e. negative with behavioral measure 1 and positive with behavioral measure 2).
In sum, we have a network that reflects greater activation during task (condition A and B) than baseline (condition C). In addition greater activation of this network is negatively correlated with behavioral measure 1 and positively correlated with behavioral measure 2 for all 3 conditions.
Now, I want to check whether the strongest regions of LV1 show the same correlation with behavioral measure 1 and 2. So, I use datamat correlation response and check for the response of regions contributed to LV1. Here are results for one the most reliable cluster:
Correlation between the most reliable cluster of LV1 and behavioral measure 1: r= 0.25
Correlation between the most reliable cluster of LV1and behavioral measure 2: r= - 0.20
Apparently, this is exactly opposite to aforementioned correlations between brain score and behavioral measures. This is also true for majority of regions, if not all, contributed to this LV. Thereby, if we don’t reverse the sign, the direction of correlation between brain score of LV1 and behavioral measures do not concur with the direction of correlation between regions contributed to LV1 and behavioral measures.
I am really confused now. What could be the reason?
I can off course send you *-bfMRIResult.mat if needed!
Thanks in advance for your help!
/Kami
Login to reply to this topic.
