Back to PLS Helpuni- vs. multi-variate results - common effects
Posted on 02/02/07 12:34:43
Number of posts: 100
Author: Agatha (---.student.Princeton.EDU)
Date: 01-14-07 13:54
hello again ~
in my analysis i have four conditions. in univariate GLM analysis i did not find differences between the conditions - but in a conjunction analysis of the individual Beta maps i've found that there are regions commonly activated across condition (i.e., a basic 'attention' network).
in the PLS task analysis i failed to find any significant effects - the LV's did not identify neither task-differences nor task-commonality (the behavioral PLS did find some reasonable effects suggesting the data are valid).
my questions is - under what circumstances could this occur? i would have expected the PLS to at least identify the common task-effect (i.e., it's akin to seeing V1 activation to checkerboards). the areas identified by the conjunction analysis share a time course... thus should be correlated... and should show up in some LV...
thanks very much for any advice!!
agatha
ps - are the error bars in the behavior PLS ('Correlation Plot Overview') uneven b/c they reflect a confidence interval rather than SE or STDev
Date: 01-14-07 13:54
hello again ~
in my analysis i have four conditions. in univariate GLM analysis i did not find differences between the conditions - but in a conjunction analysis of the individual Beta maps i've found that there are regions commonly activated across condition (i.e., a basic 'attention' network).
in the PLS task analysis i failed to find any significant effects - the LV's did not identify neither task-differences nor task-commonality (the behavioral PLS did find some reasonable effects suggesting the data are valid).
my questions is - under what circumstances could this occur? i would have expected the PLS to at least identify the common task-effect (i.e., it's akin to seeing V1 activation to checkerboards). the areas identified by the conjunction analysis share a time course... thus should be correlated... and should show up in some LV...
thanks very much for any advice!!
agatha
ps - are the error bars in the behavior PLS ('Correlation Plot Overview') uneven b/c they reflect a confidence interval rather than SE or STDev
Replies:
Re: uni- vs. multi-variate results - common effects
Posted on 02/02/07 12:35:01
Number of posts: 100
Author: Randy (---.rotman-baycrest.on.ca)
Date: 01-14-07 17:23
Agatha wrote:
> my questions is - under what circumstances could this occur? i
> would have expected the PLS to at least identify the common
> task-effect (i.e., it's akin to seeing V1 activation to
> checkerboards). the areas identified by the conjunction
> analysis share a time course... thus should be correlated...
> and should show up in some LV...
There are a number of possibilities. The first is that the common effect may be statistically reliable, but not strong enough to pass permutation tests. This scenario is possible, as illustrated in McIntosh & Lobaugh, Neuroimage 2004, Figure 1. A second is the difference in the level of analysis. Multivariate is most sensitive if you have a distributed response b/c every voxel contributes to the signal. If you have a handful of voxels that show an effect, PLS will weight these more heavily, but that does not mean other voxels are set to zero. The third could reflect the different approaches to the analysis. You said you did a conjunction analysis based on beta maps, which is not identical to what PLS does. If you did the univariate comparing the average of your tasks to the control condition, I suspect you would get the same answer as with PLS.
> ps - are the error bars in the behavior PLS ('Correlation Plot
> Overview') uneven b/c they reflect a confidence interval rather
> than SE or STDev
yes, and b/c these are obtained directly from the percentiles of the bootstrap distribution. If the distribution is skewed, the CI may not be centred around the correlation.
Date: 01-14-07 17:23
Agatha wrote:
> my questions is - under what circumstances could this occur? i
> would have expected the PLS to at least identify the common
> task-effect (i.e., it's akin to seeing V1 activation to
> checkerboards). the areas identified by the conjunction
> analysis share a time course... thus should be correlated...
> and should show up in some LV...
There are a number of possibilities. The first is that the common effect may be statistically reliable, but not strong enough to pass permutation tests. This scenario is possible, as illustrated in McIntosh & Lobaugh, Neuroimage 2004, Figure 1. A second is the difference in the level of analysis. Multivariate is most sensitive if you have a distributed response b/c every voxel contributes to the signal. If you have a handful of voxels that show an effect, PLS will weight these more heavily, but that does not mean other voxels are set to zero. The third could reflect the different approaches to the analysis. You said you did a conjunction analysis based on beta maps, which is not identical to what PLS does. If you did the univariate comparing the average of your tasks to the control condition, I suspect you would get the same answer as with PLS.
> ps - are the error bars in the behavior PLS ('Correlation Plot
> Overview') uneven b/c they reflect a confidence interval rather
> than SE or STDev
yes, and b/c these are obtained directly from the percentiles of the bootstrap distribution. If the distribution is skewed, the CI may not be centred around the correlation.
Thank You
Posted on 02/02/07 12:35:20
Number of posts: 100
Login to reply to this topic.
