Back to PLS HelpPosted on 05/20/20 12:34:05
Number of posts: 9
Hello,
I ran a mean centered PLS and a non-rotated PLS using Matlab 2014a. I am using the latest version of the PLS software, available online. Both the mean centered and non-rotated PLS have the issue of having means outside the lower and upper bounds.
As a sample, jus two LVs so it doesn't get overwhelming:
| PLS – Mean Centered | |||||
| LV_1 | |||||
| orig_usc | llusc | llusc_adj | ulusc | ulusc_adj | |
| First condition | -24.36 | -34.52 | NaN | -27.43 | NaN |
| Second condition | -21.25 | -29.61 | -20.95 | -22.98 | -20.95 |
| Third condition | 20.78 | 21.40 | 19.07 | 28.88 | 20.76 |
| Fourth condition | 24.84 | 27.42 | 24.81 | 36.56 | 24.81 |
| PLS – Non-rotated contrasts |
| LV_1 | |||||
| orig_usc | llusc | llusc_adj | ulusc | ulusc_adj | |
| First condition | -7.40 | -9.19 | -11.97 | -1.94 | -4.81 |
| Second condition | 24.24 | 14.19 | 24.64 | 22.47 | 24.64 |
| Third condition | 40.37 | 25.59 | NaN | 35.36 | NaN |
| Fourth | 42.97 | 27.32 | 43.02 | 37.74 |
43.02 |
A few more pieces of info inspire from this https://www.rotman-baycrest.on.ca/index.php?action=view_thread&id=2198&module=bbmodule&src=%40random45c35fcb17881 , where there was a similar issue except non-rotated PLS seemed to work:
Here are pics of the distributions of the 500 bootstraps for the second PLS (non rotated; https://drive.google.com/open?id=19pc-2vgd2Cv11AHAwPr8iWh4bfx3bBoC):
I tried removing outliers (participants) and seeing if disregarding extreme values for the bootstrap could help. It doesn't. I ran the PLS on two subsets of participants (28 vs. 20, instead of 48) and get the same issue.
Please let me know what you think and if you have any questions. How could we fix this CI issue?
Thank you.
Annick
Replies:
Posted on 05/21/20 09:16:19
Number of posts: 1
Hi,
I just wanted to follow this post as we are having a similar issue. We ran a mean-centered task-PLS analysis and have found that the orig_usc does not fall within the ulusc-llusc bounds. Below is an example (I have highlighted the problem condition. I've not run into this issue before but I can see from the forum that this isn't the first time the issue has come about. We run this analysis on the latest verison of PLS (I downloaded it from the downloads page a few days ago to make sure it was correct). Interestingly, the CI look accurate when I look at mean brain scores with CI within PLS. It's only an issue when trying to recreate these plots outside of PLSGUI.
I hope this helps with fixing the issue.
Sam
Posted on 05/21/20 11:18:06
Number of posts: 9
Hi Sam,
Interesting. Yes, the CI look fine in the PLSGUI but not when looking at orig_usc, llusc, etc.
I hope there's a solution.
Annick
Posted on 05/21/20 12:10:15
Number of posts: 394
Thanks for posts. We're working on a fix but it will take some time to get it posted since we're all offsite these days. The main issue is that the scores in the bootstrap field are mean-centred at each iteration, which can mess up the distribution. There is a better way to do that, but as I said, it will take a bit to get it in place.
Alternatively you can try to version where there is no meancentereing from here:
https://www.dropbox.com/sh/fmxqxsjs3rvmphz/AADAD3mUAhvPZ5LewdZv-4d8a?dl=0
Can one of you send me your results file so that I can check the difference in the CI's in the result field versus what gets plotted in the GUI?
thanks
Randy
Posted on 10/28/20 05:28:10
Number of posts: 16
Thanks for posts. We're working on a fix but it will take some time to get it posted since we're all offsite these days. The main issue is that the scores in the bootstrap field are mean-centred at each iteration, which can mess up the distribution. There is a better way to do that, but as I said, it will take a bit to get it in place.
Alternatively you can try to version where there is no meancentereing from here:
https://www.dropbox.com/sh/fmxqxsjs3rvmphz/AADAD3mUAhvPZ5LewdZv-4d8a?dl=0
Can one of you send me your results file so that I can check the difference in the CI's in the result field versus what gets plotted in the GUI?
thanks
Randy
Hi Randy, all -- I am having a similar issue with out of bounds CIs. It's a non-rotated task PLS design with four conditions, and there are two conditions for which the CIs don't include the orig_usc values.
I have tried using the PLS version in the dropbox link you posted, and this doesn't solve the issue.
I've also tried all the different mean-centering types.
Do you have any suggestions on how to proceed? Would it be totally invalid to simply calculate a bootsrapped CI for the brain scores in orig_usc?
Cheers,
Reece
Posted on 10/28/20 09:30:50
Number of posts: 394
Hi Randy, all -- I am having a similar issue with out of bounds CIs. It's a non-rotated task PLS design with four conditions, and there are two conditions for which the CIs don't include the orig_usc values.
I have tried using the PLS version in the dropbox link you posted, and this doesn't solve the issue.
I've also tried all the different mean-centering types.
Do you have any suggestions on how to proceed? Would it be totally invalid to simply calculate a bootsrapped CI for the brain scores in orig_usc?
Cheers,
Reece
Hi Reece - try two things
1) in the boot_result array there is an ulusc_adj and llusc_adj, which shifts the boostrap distribution to centre it around the point estimate. If it has NaN then the point estimate is on the extreme of the boostrap distrubtion,
2) Calcuate the standard error from the boostrap distrubiton directily. The distribution is saved as "distrib" so you can calculte the standard deviation for each element and that is the standard error,
Posted on 10/28/20 17:09:55
Number of posts: 16
Hi Randy -- thanks for your quick reply!
Re option 2, just to be sure I'm on the right page, would a 95% confidence interval for a condition be generated by running this command:
std(squeeze(result.boot_result.distrib(1,1,:)))*1.96
Thanks in advance,
Reece
Posted on 10/28/20 17:24:06
Number of posts: 394
Hi Randy -- thanks for your quick reply!
Re option 2, just to be sure I'm on the right page, would a 95% confidence interval for a condition be generated by running this command:
std(squeeze(result.boot_result.distrib(1,1,:)))*1.96
Thanks in advance,
Reece
yep - perfect
Login to reply to this topic.
