Back to PLS HelpPosted on 01/21/15 15:16:22
Number of posts: 36
Greetings PLS experts,
I conducted a PLS analysis with 30 subjects split into 2 groups (2 conditions), and found 1 significant LV. However, I then ran another analysis with a slightly different pool of subjects, which provided a significant LV1 very similar to that from the previous analysis.
Just to get a better idea, I decided to do a conjunction (as in just to get a visual summary, and a not statistical analysis) with the 2 different LVs. However, the LVs have a different number of "elements" as found in found in their respective result.boot_result.compare_u. I then just padded the one LV with the lower element value with zeros to match the other "higher" LV, and then just continued with the conjunction.
Although the difference is small (at least compared to the total) between the two LV's (less than 500 elements), I was wondering if with the 2 different element values, could potential overlapping regions between the 2 LVs be missed in this type of conjunction (as in would extending the smaller LV with zeros be incorrect)? Or would the elements between the 2 LVs up to where they had the same row value be consistent as to where they are represented. I'm curious since both my LVs show 3 major clusters in what appear to be similar reigons, but the conjunction I've been doing thus far has only shown overlap for 1 of the 3. Sorry if question is confusing...Any input would be great. Thank you.
Paul
Replies:
Posted on 01/24/15 09:31:11
Number of posts: 394
Greetings PLS experts,
I conducted a PLS analysis with 30 subjects split into 2 groups (2 conditions), and found 1 significant LV. However, I then ran another analysis with a slightly different pool of subjects, which provided a significant LV1 very similar to that from the previous analysis.
Just to get a better idea, I decided to do a conjunction (as in just to get a visual summary, and a not statistical analysis) with the 2 different LVs. However, the LVs have a different number of "elements" as found in found in their respective result.boot_result.compare_u. I then just padded the one LV with the lower element value with zeros to match the other "higher" LV, and then just continued with the conjunction.
Although the difference is small (at least compared to the total) between the two LV's (less than 500 elements), I was wondering if with the 2 different element values, could potential overlapping regions between the 2 LVs be missed in this type of conjunction (as in would extending the smaller LV with zeros be incorrect)? Or would the elements between the 2 LVs up to where they had the same row value be consistent as to where they are represented. I'm curious since both my LVs show 3 major clusters in what appear to be similar reigons, but the conjunction I've been doing thus far has only shown overlap for 1 of the 3. Sorry if question is confusing...Any input would be great. Thank you.
Paul
Hi Paul,
I am not 100% sure what you are asking, but let me see if my answer suffices. First what are the data? fMRI, EEG, structural MRI? For fMRI analysis, PLSGUI will create a voxel mask of the union of all subjects to ensure that there are data for all voxels across subjects. The common mask can change when you add or subtract subjects. An extreme example would be if you have a group with focal lesions - the common mask for controls and patients together would exclude data from the lesioned areas.
If you can provide a bit more info on what you are trying to assess I should be able to help further.
Randy
Posted on 01/24/15 13:06:11
Number of posts: 36
Hi Randy,
thanks for the response. Sorry for making it confusing. Essentially, I want to compare the LV1 from two seperate PLS fMRI analyses, with close to the same group of subjects, but slight differences (subject removal), just to get a better idea of how similar the LV1's are that are produced. Just to get a visual summary, I thought I would look at the overlap between the 2 LV's, just to get a better idea if the major clusters were similar. So I decided to do a conjunction summary, by doing what I assumed was an element * element multiplication of each LV (the result.boot_result.compare_u). However, when comparing the two different LV's, their total element count was different. I thus added zeroes to the "smaller" length LV to make the size of the LVs the same. I then proceeded with the element by element multiplication.
I just wasn't sure if this was an appropriate way to go about trying to getting a visual summary of overlapping regions, due to the different element length (and if adding zeros to an LV would be in appropriate), or in general, if trying to get a summary of overlapping regions from two seperate PLS analyses is "wrong" in any sense. I became concerned because although it would look like 3 major clusters would show overlap, the "conjunction" I did only showed 1 of the 3, so I wasn't sure if what I did was appropriate.
I hope that makes it a bit less confusing (and sorry if it still is). Thank you.
Posted on 01/24/15 14:51:59
Number of posts: 394
I think I get it now. One thing to be cautious is that you cannot simply add zeros to the smaller vector to make it the same length unless you match the zeros to the spatial location that is not common across the two data sets. It is likely that the missing voxels are in discontinuous locations. Are you accounting for that? If not, there is an "st_coords" vector that indicates the matrix locations for each of the analysed voxels that you can use to identify where you want add zeros to the smaller dataset.
an easy way to do this is (assume LV1 is dataset1 and lv1 and the smaller dataset and lv1_st_coords is the matrix location vector for lv1)
%create a zero vector the same size as LV1 to put lv1 data into
lv1_bigger=zeros(size(LV1));
lv1_bigger(lv1_st_coords)=lv1; %this will put data in locations lv1_st_coords and the rest will be zero
You can then look at the overlap of LV1 and lv1 in the same space.
Posted on 01/24/15 15:16:41
Number of posts: 36
Great, thank you so much Randy.
Indeed, I was worried I went in with the wrong assumption. I thought that perhaps the overlapping element count between the two different sized LVs would account for the same spatial coordinates, and whatever additonal elements were counted for in the bigger LV, would not overlap with any elements with the smaller LV, and thus I thought multiplying with zero would be of no concern, since there would be no overlap for the elements (under my wrong assumption). Thanks for clearing that up!
Login to reply to this topic.
