Back to PLS HelpPosted on 12/06/16 05:57:56
Number of posts: 2
Dear PLS experts,
We are doing PLS correlation analysis – both task and behavioral PLS.
When running behavioral analysis on our original data we got significant results (5 significant LVs with p=0.000 (LV5=0.002), and singular values from 280.89 (LV1). However, when we scrambled the behavioral data we found equally significant results (5 significant LVs, with p=0.000, and singular values from 213.58).
When running task PLS in the original data (patients vs. controls), we found significant group differences (1 LV, with p=0.000, and singular value 66.28). However, when scrambling the data (randomly mixed patients and controls in the two groups) we also found significant results (1 LV, with p=0.000, and singular value 70.41).
Could you please help us understand why this is happening?
Best
Katja
Replies:
Posted on 12/06/16 08:35:31
Number of posts: 394
Hi Katja,
Keep in mind that the permutation test creates a distribution where in your original data are assessed against random reordings. It is often the case that some reorderings will produce singular values equal to or higher than the original sample, which I suspect is what you are seeing This is the nature of probability and again reinforces why complementary statistics are necessary (see McIntosh & Lobaugh, 2004,Neuroimage: https://www.dropbox.com/s/pwszup8m0f92us3/pls_review.pdf?dl=0).
Its difficult to give a specific answer without knowing a bit more about your data and the permutation you did. I will note that the permutation test alone is not sufficient, which is why we doing bootstrap to complement to give an estimate of reliability. Also, the split-half resampling routine is a good complement.
Can you give more details on the data you are analysing?
Posted on 12/08/16 04:34:01
Number of posts: 2
Posted on 12/08/16 07:50:21
Number of posts: 394
Thanks Katja,
How do the bootstrap resamplings look for the original and your permuted data?
As I noted in my previous response, what you see with your permuted data is not unexpected given that we are dealing with probabilities. This is why the bootstrap is helpful. The split-half is still a work-in-progress but does give a good indication of the stabiliity of the relationship.
Login to reply to this topic.
