Back to PLS HelpPosted on 07/16/15 19:41:54
Number of posts: 9
Hi everyone,
I am doing a non-rotated structural PLS analysis to see whether grey matter volume is correlated with four different behavioural variables. I specified four contrasts, weighing each of the behavioural variables against the sum of the other three to check whether there are any unique contributions of one variable. The first latent variable came out significant (p = 0.30). This is the first contrast: https://www.dropbox.com/s/hz7bfm2v1s1nozc/contrast.png?dl=0
However, when I look at the correlation plot, instead of contrasting one variable with the other three, the correlations go in the same direction for all variables and their confidence intervals overlap. https://www.dropbox.com/s/wjzemz0rub6jf61/corrPlot.png?dl=0 If I interpret this correctly, this means that behavioural variables 1, 2, and 4 (the CI of variable 3 crosses zero) contribute to the observed pattern in the same direction and that there is no significant difference between variable 1, 2, and 4, as their CIs overlap. Is that interpretation correct? If so, would someone be able to explain why the LV is significant, even though the correlation plot does not reflect the contrast I specified very well?
Thank you so much,
Kristina
Replies:
Posted on 07/17/15 09:42:15
Number of posts: 394
HI Kristina,
Another way of framing the contrast you entered is "where are the correlations for the first variable higher than the average of the other three?", which is what your correlation plot reflects even if the correlations are all positive.
However, the p-value you wrote down is 0.30, which by convention is not significant. Is that number correct?
Also, did you try to regular behavior PLS to see what a data driven approach provided?
cheers
Randy
Posted on 07/20/15 21:24:21
Number of posts: 9
Hi Randy,
thanks for your reply. The p-value of 0.30 was just a typo; it was supposed to read .030, sorry.
Ok, I slightly misunderstood what PLS does for that contrast, but your answer clarified that. So the fact that the CIs of the first and the second behavioural variables overlap doesn't matter, because for my contrast it is only about the averages of the other three correlations and their CIs?
I did run a regular PLS on the whole brain. The non-rotated PLS was a follow up that I did within the pattern of regions identified by the regular PLS analysis, because I wanted to see whether any volumetric changes are driven by the first variable, i.e. whether the correlations are stronger for the first variable than for the other three. The pattern of the correlations with the behavioural variables looks pretty similar for both the regular and the non-rotated analyses. [This is the correlation plot for the regular PLS https://www.dropbox.com/s/nr2myt0kqipmgi6/corr_regularPLS.png?dl=0]
I was really trying to get at the question whether there are any regions for which the correlation with the first behavioural variable is greater than for the second, the third, AND the fourth variable. Do you know if there is any way to do that? I can't think of a contrast that could do it. Would it be valid to run 3 separate analyses for the first variable > second variable, first variable > third variable, and first variable > fourth variable and then see which regions come out in all analyses? Or is that too univariate a way of thinking?
Thanks,
Kristina
Posted on 07/21/15 16:01:04
Number of posts: 394
Hi Randy,
thanks for your reply. The p-value of 0.30 was just a typo; it was supposed to read .030, sorry.
Ok, I slightly misunderstood what PLS does for that contrast, but your answer clarified that. So the fact that the CIs of the first and the second behavioural variables overlap doesn't matter, because for my contrast it is only about the averages of the other three correlations and their CIs?
I did run a regular PLS on the whole brain. The non-rotated PLS was a follow up that I did within the pattern of regions identified by the regular PLS analysis, because I wanted to see whether any volumetric changes are driven by the first variable, i.e. whether the correlations are stronger for the first variable than for the other three. The pattern of the correlations with the behavioural variables looks pretty similar for both the regular and the non-rotated analyses. [This is the correlation plot for the regular PLS https://www.dropbox.com/s/nr2myt0kqipmgi6/corr_regularPLS.png?dl=0]
I was really trying to get at the question whether there are any regions for which the correlation with the first behavioural variable is greater than for the second, the third, AND the fourth variable. Do you know if there is any way to do that? I can't think of a contrast that could do it. Would it be valid to run 3 separate analyses for the first variable > second variable, first variable > third variable, and first variable > fourth variable and then see which regions come out in all analyses? Or is that too univariate a way of thinking?
Thanks,
Kristina
HI Kristina - to answer the question of whether there are regions for which the correlation with the 1st variable is higher than the 2nd, 3rd and 4th, you would use the same contrast you already have. What you may wish to run are the other other 2 contrasts to complete the parameterization of your design
e,g, (3 -1 -1 -1) {the one you did}; (0 2 -1 -1); (0 0 1 -1)
this may help clarify the uniquesness of the pattern you see.
cheers
Randy
Posted on 07/25/15 21:57:47
Number of posts: 3
HI Kristina - to answer the question of whether there are regions for which the correlation with the 1st variable is higher than the 2nd, 3rd and 4th, you would use the same contrast you already have. What you may wish to run are the other other 2 contrasts to complete the parameterization of your design
e,g, (3 -1 -1 -1) {the one you did}; (0 2 -1 -1); (0 0 1 -1)
this may help clarify the uniquesness of the pattern you see.
cheers
Randy
Hi Randy,
I am also running a similar analysis to the one Kristina is doing but am having a bit of trouble understanding your last reply.
So from previous posts, I get that running a contrast of 0 2 -1 -1 will give me the regions where the correlations are higher for the second variable compared to the average of the 3rd and 4th variables. And this contrast forms the X matrix which is fitted to the Y matrix (the first variable in this contrast, due to its 0 weighting). However, I'm a bit confused as to what this result is telling me, especially in the context of my original contrast of 3 -1 -1 -1. If the original analysis is significant, but the 0 2 -1 -1 and the 0 0 1 -1 contrasts are not, could this be interpreted as the first variable having unique contributions to the pattern? Or have I completely misunderstood the point of running the other 2 contrasts?
Any help would be much appreciated, thank you!
Haeme
Posted on 07/26/15 11:35:13
Number of posts: 394
Hi Randy,
I am also running a similar analysis to the one Kristina is doing but am having a bit of trouble understanding your last reply.
So from previous posts, I get that running a contrast of 0 2 -1 -1 will give me the regions where the correlations are higher for the second variable compared to the average of the 3rd and 4th variables. And this contrast forms the X matrix which is fitted to the Y matrix (the first variable in this contrast, due to its 0 weighting). However, I'm a bit confused as to what this result is telling me, especially in the context of my original contrast of 3 -1 -1 -1. If the original analysis is significant, but the 0 2 -1 -1 and the 0 0 1 -1 contrasts are not, could this be interpreted as the first variable having unique contributions to the pattern? Or have I completely misunderstood the point of running the other 2 contrasts?
Any help would be much appreciated, thank you!
Haeme
Hello Haeme,
Your interpretation is correct. The main reason for doing the other contrasts is for completeness.
Login to reply to this topic.
