Back to PLS HelpPosted on 09/30/10 18:51:44
Number of posts: 3
In my experiment, I would like to try PLS regression between number of sources (as independent variable) and number of clinical parameters.
As my data has only 1 condition (difference between pre and post treatment) in 16 sources and 6 bands (EEG) and corresponding clinical changes in symptoms measured on 7 different scales, the option of permutation seems not to be applicable.
The bootstrapping in this PLS soft can only confirm the reliability of the model (a model with certain LVs number already determined, if I understand it correctly).
What would be in your opinion the best way (in my case) to select the number of LV?
I had earlier some experience with PCA and thought of may be doing PCA on my sources data, selecting the number of LVs in PCA using the Kaiser criterion or some other cross validation test (V-fold) and then just manually taking this number of LVs in to the PLS model (not sure if this might be at all statistically justifiable).
Another crazy idea would be to extrapolate the Kaiser criterion on to the Singular value, and take only the LVs that are over 1 (if Singular value is a square of eigenvalue, 1 as a threshold would not change).
I would be thankful for all your comments!
Cheers,
Ilya
Replies:
I'm Online
Posted on 10/01/10 13:01:02
Number of posts: 291
First, the number of LV that generated by our PLS software is determined by which option that you choose to run PLS:
For Mean-Centering PLS: nLVs = nConditions x nGroups;
For Non-Rotated Task PLS: nLVs = nContrasts;
For Regular Behavior PLS: nLVs = nConditions x nGroups x nBehaviorMeasure;
For Non-Rotated Behavior PLS: nLVs = nContrasts;
For Multiblock PLS: nLVs = nConditions x nGroups + nConditions x nGroups x nBehaviorMeasure;
Second, there are also several restrictions:
1. Since nConditions only provide nConditions-1 degrees of freedom, the maximum number of contrasts is limited by nConditions;
2. In order to run Behavior PLS, you need to have at least 3 subjects;
3. In order to run Mean-Centering PLS, you need to have at least 2 conditions. However, if you have only 1 condition and you have more than 1 groups, our PLS will reconstruct datamat so that it will become single group with multiple conditions.
Therefore, in most case we have more than 1 conditions and more than 3 subjects.
Posted on 10/01/10 14:04:37
Number of posts: 394
Posted on 10/02/10 07:27:38
Number of posts: 3
Hi
thank you for your replies.
Randy, the general answer that I'm trying to find out with your software is
what source in which frequency (I use precalculated spectra instead
of ERPs in ERP module) contributes most to the changes in separate
(each measure at a time) behavioral measures (for 1st behavior, 2nd
and so on) and if possible to what extent (regression like plot for example).
So far I came up with the following general plan: to see which LV (at this
point I should not forget about significance of the component: shell I take
only significant ones?) is contributing the most to the selected
behavioral measure (for example behavioral 2) I can look at the plot of
scalp scores for behavioral analysis and pick the highest correlation (can I
find a significance of this correlation some where?).
Then I can see which frequency bins in which "electrodes" have "diamonds"
(at the salience plot) to find out that these bins do significantly
contribute to that LV.
Can I see the loading of each of these bins somewhere to see also which from
these contribute to this LV
the most (to weigh them)? So that I can assess and pick only those frequency
bins from certain electrodes that contribute the most to the change
in the behavioral condition.
I hope I’m not completely out of this world there and got the concept
correctly)
Cheers
Ilya
P.S. unfortunatelly i only get the permutation value of 0.01 for the first component and the rest of them have permutation values around (0.6-0.9).
So I might need to be stuck with only one LV
Login to reply to this topic.
