My question is twofold:
1) Can we use subject scores (e.g. those in panel 3 of figure 1, and the bottom of figures 2 and 3 of McIntosh et al. 1996 NeuroImage) for cross-validation? i.e. can we leave one subject out, perform PLS for the remaining subjects, then multiply the resulting singular images (B1 in 1996 paper) with the excluded subject's brain images to arrive at subject scores per condition for the excluded subject, then perform some sort of sum-square differences calculation to predict which condition the subject was undergoing?
2) If the answer to (1) is yes, can we use the contents of the results .mat files and individual subject datamats to do this? Which variables correspond to the brain images and singular images?
Thank you!
Grace
In your second question, variables result.u and result.v in result.mat files are both singular images. Especially, result.u is singular image related to the brain, and we sometimes call it BrainLV or Salience. result.s is singular value. In result.mat, you cannot find brain image, unless you checked "save data" when you run PLS analysis from GUI window, which is unchecked by default. That's all I know.
Thank you! Do you know how I could save these if I am using batch_plsgui?
Also, how do you find the brain image for each individual subject? I was hoping this would be a variable in the datamat file.
Thanks again,
Grace
Thank you! Do you know how I could save these if I am using batch_plsgui?
Also, how do you find the brain image for each individual subject? I was hoping this would be a variable in the datamat file.
Thanks again,
Grace
When you use batch_plsgui, set "save_data" parameter to 1 in the analysis script.
Brain image for individual subject is in sessiondata.mat file, with variable "datamat".
Hi Grace - the answer to (1) is yes, but that essentially what the bootstrapping approach does. Instead of one subject, it resamples with replacement, so it may leave more than one person out in an given resample. It recalculates the SVD each time based on the new sample and thus we derive confidence intervals for the U and V side of the decomposition. The method you propose would work, but you need to go back to the datamat each time to recalculate the SVD.
Does the bootstrap approach we use suffice for your needs or do you need something more?
I'd like to be able to predict the condition based on brain activity for a novel subject, which I think is slightly different from the statistical significance that bootstrapping is used to show.. is that right?
The biggest problem I'm facing with the programming part of things is that the singular images from PLS contain different voxels from the left out subject. How was this dealt with in the bootstrapping approach?
Thanks,
Grace
I'd like to be able to predict the condition based on brain activity for a novel subject, which I think is slightly different from the statistical significance that bootstrapping is used to show.. is that right?
For a new subject, you wouldn't necessarily use bootstrap. Given you can get standard error estimates from the bootstrap, however, you can compute the new person's score and see where their scores sits in the distribution.
The biggest problem I'm facing with the programming part of things is that the singular images from PLS contain different voxels from the left out subject. How was this dealt with in the bootstrapping approach?
Sorry for all the questions, one more thing.. in the results.mat, how is st_datamat structured? Is it arranged by condition then by subject, such that if I had 3 conditions and 20 subjects, the first 20 rows would be the datamats for condition 1, subjects 1 through 20?
Thanks,
Grace
Sorry for all the questions, one more thing.. in the results.mat, how is st_datamat structured? Is it arranged by condition then by subject, such that if I had 3 conditions and 20 subjects, the first 20 rows would be the datamats for condition 1, subjects 1 through 20?
Thanks,
Grace
Sorry for all the questions, one more thing.. in the results.mat, how is st_datamat structured? Is it arranged by condition then by subject, such that if I had 3 conditions and 20 subjects, the first 20 rows would be the datamats for condition 1, subjects 1 through 20?
Thanks,
Grace
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