Back to PLS HelpPosted on 09/16/16 08:36:02
Number of posts: 14
Hello Randy,
I would like to perform voxel-wise statistics on the permuted data to identify significant clusters (instead of identifying reliable clusters using bootstrapping). For example ( in case of 1000 permutations), lets assume that PLS has identified that LV1 is significant (p<0.05).
1. I would like store one unpermuted and 1000 permuted brain salience corresponding to LV1.
2. Then perform voxel-wise enhancement using TFCE on each of the one unpermuted and 1000 permuted brain saliencesto obtain their corresponding tfce maps.
3. Using these TFCE maps (one unpermuted and 1000 permuted) I would like to calculate significant clusters by using family wise correction (FWE) for multiple comparisons. This would give a p-value to each voxel corresponding to the significance with which it relatesto the latent variable structure.
Can you please tell me, if I am allowed to do the above mentioned processing steps on the brain salience maps (i assumed here that brain salience maps are statisticam maps calculated for unpermuted and permuted data). My main intention is identify significant clusters with a corresponding p-value.
Thank you for your time.
Best wishes
Jay
Replies:
Posted on 09/17/16 10:53:20
Number of posts: 394
Hello Randy,
I would like to perform voxel-wise statistics on the permuted data to identify significant clusters (instead of identifying reliable clusters using bootstrapping). For example ( in case of 1000 permutations), lets assume that PLS has identified that LV1 is significant (p<0.05).
1. I would like store one unpermuted and 1000 permuted brain salience corresponding to LV1.
2. Then perform voxel-wise enhancement using TFCE on each of the one unpermuted and 1000 permuted brain saliencesto obtain their corresponding tfce maps.
3. Using these TFCE maps (one unpermuted and 1000 permuted) I would like to calculate significant clusters by using family wise correction (FWE) for multiple comparisons. This would give a p-value to each voxel corresponding to the significance with which it relatesto the latent variable structure.
Can you please tell me, if I am allowed to do the above mentioned processing steps on the brain salience maps (i assumed here that brain salience maps are statisticam maps calculated for unpermuted and permuted data). My main intention is identify significant clusters with a corresponding p-value.
Thank you for your time.
Best wishes
Jay
Hi Jay,
The calculation you want to do is possible but you will need to modify the code to save the singular vectors (saliences) for each permutation. This wil take up a lot of RAM so make sure you have a machine that has max RAM available.
BUT - the question is why would you want to do this? If you take a look at the McIntosh & Lobaugh 2004 PLS review (particularly Fig 1) you will see the bootstrapping is a better way of capturing reliable voxels than permutation p-values. The permutation is good for assessing the overall significance of the relationships, but I would urge you to consider what specific question you want to address at the voxel level before you try to calculate p-values for individual voxels.
the paper link is here: https://www.dropbox.com/s/pwszup8m0f92us3/pls_review.pdf?dl=0
cheers
Randy
Posted on 09/20/16 05:23:21
Number of posts: 14
Posted on 09/20/16 06:25:32
Number of posts: 394
Hello Randy,
I would like to perform voxel-wise statistics on the permuted data to identify significant clusters (instead of identifying reliable clusters using bootstrapping). For example ( in case of 1000 permutations), lets assume that PLS has identified that LV1 is significant (p<0.05).
1. I would like store one unpermuted and 1000 permuted brain salience corresponding to LV1.
2. Then perform voxel-wise enhancement using TFCE on each of the one unpermuted and 1000 permuted brain saliencesto obtain their corresponding tfce maps.
3. Using these TFCE maps (one unpermuted and 1000 permuted) I would like to calculate significant clusters by using family wise correction (FWE) for multiple comparisons. This would give a p-value to each voxel corresponding to the significance with which it relatesto the latent variable structure.
Can you please tell me, if I am allowed to do the above mentioned processing steps on the brain salience maps (i assumed here that brain salience maps are statisticam maps calculated for unpermuted and permuted data). My main intention is identify significant clusters with a corresponding p-value.
Thank you for your time.
Best wishes
Jay
Jay - I would suggest you carefully consider the paper I sent you in the last exchange. If your goal is to focus your inference on the voxel level, then PLS is not the appropriate analysis as it is a multivariate approach covering patterns of voxels rather than voxels per se. If you want to look at whether a voxel/cluster is significant there are other packages that can do this (FSL, SPM, etc).
I'm Online
Posted on 09/20/16 08:14:57
Number of posts: 229
1. In structural analyses, the cluster localization is important to be able to interpret the strength of multivariate pattern associated with the anatomic location. Permutation method makes it possible to perform statistical assessment in identifying significant clusters (or voxels) which are corrected for multiple comparisons (by accepting 5% false positievs) and associate significantly with the latent structure.
Jay - the same logic applies to structural analyses - if you were to take your favourite flavour of tissue segmentation, and run those maps in a PLS analysis, the statistical test is at the pattern level. No FDR, etc would be required at the voxel level to identify those voxels that are reliably contributing to the pattern.
Nancy
Login to reply to this topic.
