Hi PLS experts,
I am running simple behavioral PLSC on structural GM and as behavioral I have some neurological soft signs parameters. I have only one group of subjects (24 subjects). I did multiple regression on data using sex, age and TIV as regressors and I have standardized the data. After I run the PLS I get first LV not significant with almost 60% covariance explained and then I have 2 LVs which are significant (one of them is not significant but trend could be seen). According to previous topic from evast187 (https://www.rotman-baycrest.on.ca/index.php?action=view_thread&id=1890&module=bbmodule&src=%40random45c35fcb17881), can I interpret those two LVs? I did some printscreens available here: https://dl.dropboxusercontent.com/u/70539709/PLS_results.docx
Thanks for your answers,
Matyas
HI Matayas - you can certainly interpret the significant LVs. The paradox of non-significant LV1 with behavior PLS is more frequent than you may imagine, partly b/c the behavior analysis includes a "grand mean" term whereas the task PLS does not because of mean-centering.
Thank you Randy, may I ask you then if I can subtract grand mean from structural data and then get more explained covariance by other LVs or that is only option for fMRI (task) data?
One more question quite out of topic, I cant decide whether to subtract effect of Total Incranial Volume from behavioral data too - I just subtract effect of age and sex from behavioral data- should I also use the effect of TIV? It depends on my decision or should be structural data and behavioral data been processed in the same way?
Thank you once more,
Matyas
subtracting the grand mean won't change the outcome statistically and remember that %cov is not the same as %var. If you have only 1 LV in your design, then that LV will by construction account for 100% of the covariance.
As for adjusting the behavioral data for "confounds" you can do that, though it changes the question you are asking. An alternative in some cases is to include factors such as age, sex, TIV in the behavior matrix. If the effects of interest are independent of these, then there should be an LV that separates those effects from ones related to age, etc. By regressing out those effects, you have no idea where and whether they have any relation to the brain data.
Hi Randy and all PLS experts,
I tried it the way you suggested - that is include the confound variables (age, sex, TIV and total GM) into the behavior matrix. I want just confirm the right interpretation - in LV 1 I got sigificant behavior of interest but also the confounds - so I should not interpret these I suppose. In the LV3 which is not significant but trend could be talked about I got 2nd and 5th behavior significant and not the confounds. So this I could use for interpretation? (Also if it has only about 13% of crossblock covariance explained? ). Here are the results I am talking about: https://dl.dropboxusercontent.com/u/70539709/PLS_result2.docx
Thanks once more,
Matyas
Hi Randy and all PLS experts,
I tried it the way you suggested - that is include the confound variables (age, sex, TIV and total GM) into the behavior matrix. I want just confirm the right interpretation - in LV 1 I got sigificant behavior of interest but also the confounds - so I should not interpret these I suppose. In the LV3 which is not significant but trend could be talked about I got 2nd and 5th behavior significant and not the confounds. So this I could use for interpretation? (Also if it has only about 13% of crossblock covariance explained? ). Here are the results I am talking about: https://dl.dropboxusercontent.com/u/70539709/PLS_result2.docx
Thanks once more,
Matyas
Hi there - sorry for the delay. The results look good and yes you can interpret the results the way you suggest. I would report LV1 as well to show how the covariates relate and show that is independent of the pattern on LV3
Hi Randy,
Thanks for your explanation of why this might happen with PLSC. I've been looking at some results from a structural mean-centered PLS. Leaving the default settings for mean-centering type (0), I also get the puzzling finding that the first LV, which explains the most variance is non-significant, while the second LV is significant. I looked more closely at the bootstrapped confidence intervals & for one of the groups, the 95% CI excludes the mean group value - I guess this means that there is too much variability in this group for PLS to estimate a stable contrast? How would you proceed?
Thanks,
John
Hi Randy,
Thanks for your explanation of why this might happen with PLSC. I've been looking at some results from a structural mean-centered PLS. Leaving the default settings for mean-centering type (0), I also get the puzzling finding that the first LV, which explains the most variance is non-significant, while the second LV is significant. I looked more closely at the bootstrapped confidence intervals & for one of the groups, the 95% CI excludes the mean group value - I guess this means that there is too much variability in this group for PLS to estimate a stable contrast? How would you proceed?
Thanks,
John
Good question - may be we can do a deeper dive into these data? Would you be willing to share the results file with me? The confidence interval issue is something we've seen before and usually is related heterogeneity in the sample or scaling differences.
Baycrest is an academic health sciences centre fully affiliated with the University of Toronto
Privacy Statement - Disclaimer - © 1989-2024 BAYCREST HEALTH SCIENCE. ALL RIGHTS RESERVED