Back to PLS Helpgroup comparison
Posted on 02/02/07 12:28:01
Number of posts: 100
Author: JC youn (211.46.97.---)
Date: 01-05-07 01:01
Dear Randy.
I have a FDG-PET data set, which consist of 10 AD patient with delusion and 10 age,sex and MMSE score matched AD control.
I want to identify the network of delusion in the patient with AD.
In this case, which method in PLS option should I select?
Please give me a some detail explanation for analyzing this dataset.
Best Regards.
Jong_chul, Youn.
Kyunggi Provincial Hospital for the Elderly
Date: 01-05-07 01:01
Dear Randy.
I have a FDG-PET data set, which consist of 10 AD patient with delusion and 10 age,sex and MMSE score matched AD control.
I want to identify the network of delusion in the patient with AD.
In this case, which method in PLS option should I select?
Please give me a some detail explanation for analyzing this dataset.
Best Regards.
Jong_chul, Youn.
Kyunggi Provincial Hospital for the Elderly
Replies:
Re: group comparison
Posted on 02/02/07 12:28:23
Number of posts: 100
Author: Randy (---.rotman-baycrest.on.ca)
Date: 01-05-07 11:20
JC youn wrote:
> Dear Randy.
>
> I have a FDG-PET data set, which consist of 10 AD patient with
> delusion and 10 age,sex and MMSE score matched AD control.
>
> I want to identify the network of delusion in the patient with
> AD.
>
> In this case, which method in PLS option should I select?
>
That depends on how you want to identify the "network" - based on activity differences, based on brain-behaviour correlations? That will indicate how to proceed with the analysis.
Date: 01-05-07 11:20
JC youn wrote:
> Dear Randy.
>
> I have a FDG-PET data set, which consist of 10 AD patient with
> delusion and 10 age,sex and MMSE score matched AD control.
>
> I want to identify the network of delusion in the patient with
> AD.
>
> In this case, which method in PLS option should I select?
>
That depends on how you want to identify the "network" - based on activity differences, based on brain-behaviour correlations? That will indicate how to proceed with the analysis.
Re: group comparison
Posted on 02/02/07 12:28:42
Number of posts: 100
Author: JC youn (211.46.97.---)
Date: 01-05-07 20:53
Dear.
Yes. Behavioral PLS could be used in this case.
But, I want to analyze my data without using behavioral PLS.
In this data set, I want to find out the regional differences of cerebral glucose metabolism between non-paranoid group and paranoid group.
Many Thanks.
Jong-Chul, Youn.
Date: 01-05-07 20:53
Dear.
Yes. Behavioral PLS could be used in this case.
But, I want to analyze my data without using behavioral PLS.
In this data set, I want to find out the regional differences of cerebral glucose metabolism between non-paranoid group and paranoid group.
Many Thanks.
Jong-Chul, Youn.
Re: group comparison
Posted on 02/02/07 12:29:03
Number of posts: 100
Author: Randy (---.dsl.bell.ca)
Date: 01-06-07 10:14
> In this data set, I want to find out the regional differences
> of cerebral glucose metabolism between non-paranoid group and
> paranoid group.
Ah, okay. The easiest way to do this is with the non-rotated PLS option. I assume you have one condition/image per subject within each group? If so you use the contrast manager window to assign '1' to one group and '-1' to the other, save that contrast file and read it in for the non-rotated PLS.
You could then take the areas identified from that analysis and extract one or more regions to look at their functional connectivity with the rest of the brain and compare that between groups. Essentially, you use the regional value as a behavior measure and run it as a behavior PLS.
cheers
Randy
Date: 01-06-07 10:14
> In this data set, I want to find out the regional differences
> of cerebral glucose metabolism between non-paranoid group and
> paranoid group.
Ah, okay. The easiest way to do this is with the non-rotated PLS option. I assume you have one condition/image per subject within each group? If so you use the contrast manager window to assign '1' to one group and '-1' to the other, save that contrast file and read it in for the non-rotated PLS.
You could then take the areas identified from that analysis and extract one or more regions to look at their functional connectivity with the rest of the brain and compare that between groups. Essentially, you use the regional value as a behavior measure and run it as a behavior PLS.
cheers
Randy
Re: group comparison
Posted on 02/02/07 12:29:22
Number of posts: 100
Author: JC youn (211.46.97.---)
Date: 01-08-07 02:03
Thank you for your kindness.
1. When I directly typed 1 to group 1 and 2 to group2, error massage (No contrast available to be save) was presented in the low part of window.
So, I typed 1 in constrast for group 1 and press add burton and type 2 in the contrast for group 2 and press add burton. After this, file could be saved as *_PETcontrast.txt. When I open this file, contrast was defined as 1.ooooooe+ooo -1.000000e+000. (group 1 is control and group 2 is paranoid delusion group). Is this process is correct?
2. After this, I got one LV. Should I -1 1contrast instead of 1 -1? In my opinion, it will produce essentially the same results.
3. I can not easily understand why one or more regions should be extracteda and analyzed further. Could you exxplain more detail?
Best Regards.
Date: 01-08-07 02:03
Thank you for your kindness.
1. When I directly typed 1 to group 1 and 2 to group2, error massage (No contrast available to be save) was presented in the low part of window.
So, I typed 1 in constrast for group 1 and press add burton and type 2 in the contrast for group 2 and press add burton. After this, file could be saved as *_PETcontrast.txt. When I open this file, contrast was defined as 1.ooooooe+ooo -1.000000e+000. (group 1 is control and group 2 is paranoid delusion group). Is this process is correct?
2. After this, I got one LV. Should I -1 1contrast instead of 1 -1? In my opinion, it will produce essentially the same results.
3. I can not easily understand why one or more regions should be extracteda and analyzed further. Could you exxplain more detail?
Best Regards.
Re: group comparison
Posted on 02/02/07 12:29:44
Number of posts: 100
Author: Randy (---.dsl.bell.ca)
Date: 01-08-07 21:41
JC youn wrote:
> 1. When I directly typed 1 to group 1 and 2 to group2, error
> massage (No contrast available to be save) was presented in the
> low part of window.
> So, I typed 1 in constrast for group 1 and press add burton and
> type 2 in the contrast for group 2 and press add burton. After
> this, file could be saved as *_PETcontrast.txt. When I open
> this file, contrast was defined as 1.ooooooe+ooo
> -1.000000e+000. (group 1 is control and group 2 is paranoid
> delusion group). Is this process is correct?
perfect
> 2. After this, I got one LV. Should I -1 1contrast instead of 1
> -1? In my opinion, it will produce essentially the same
> results.
you are correct. You will get the same result, but the signs on the LV will be flipped. You only need to run one contrast.
> 3. I can not easily understand why one or more regions should
> be extracteda and analyzed further. Could you exxplain more
> detail?
The analysis of mean differences will identify a distributed pattern that differentiates groups. This does not guarantee that these areas are interacting with each other, which would be the characteristic of a network. So to identify the network, you need to follow up the analysis of activity differences with something like a seed-voxel PLS.
Klaas Stephan has a very nice example of how common activity changes does not indicate a network in his paper:
Stephan KE (2004) On the role of general system theory for functional neuroimaging. Journal of Anatomy 205:443-470.
Date: 01-08-07 21:41
JC youn wrote:
> 1. When I directly typed 1 to group 1 and 2 to group2, error
> massage (No contrast available to be save) was presented in the
> low part of window.
> So, I typed 1 in constrast for group 1 and press add burton and
> type 2 in the contrast for group 2 and press add burton. After
> this, file could be saved as *_PETcontrast.txt. When I open
> this file, contrast was defined as 1.ooooooe+ooo
> -1.000000e+000. (group 1 is control and group 2 is paranoid
> delusion group). Is this process is correct?
perfect
> 2. After this, I got one LV. Should I -1 1contrast instead of 1
> -1? In my opinion, it will produce essentially the same
> results.
you are correct. You will get the same result, but the signs on the LV will be flipped. You only need to run one contrast.
> 3. I can not easily understand why one or more regions should
> be extracteda and analyzed further. Could you exxplain more
> detail?
The analysis of mean differences will identify a distributed pattern that differentiates groups. This does not guarantee that these areas are interacting with each other, which would be the characteristic of a network. So to identify the network, you need to follow up the analysis of activity differences with something like a seed-voxel PLS.
Klaas Stephan has a very nice example of how common activity changes does not indicate a network in his paper:
Stephan KE (2004) On the role of general system theory for functional neuroimaging. Journal of Anatomy 205:443-470.
Re: group comparison
Posted on 02/02/07 12:30:02
Number of posts: 100
Author: JC youn (211.46.97.---)
Date: 01-08-07 22:48
Thank you.
I have another data set, which was analyzed using behavioral PLS (I mentioned this dataset in previous quesitons).
In that analysis, I had several behavioral measures including psychotic symptom. depresseion etc. I got several statisticallly significant LVs that were correlated with behavioral measures and stopped the analyses.
(This analytic process were same as Mentis et al. 2002)
In this dataset, should I extract mean CMRglc in the regions that were significantly corrrelated with certain behavioral measure and subsequently analyze data using seed-voxel PLS to identify the network?
Best Regards.
Date: 01-08-07 22:48
Thank you.
I have another data set, which was analyzed using behavioral PLS (I mentioned this dataset in previous quesitons).
In that analysis, I had several behavioral measures including psychotic symptom. depresseion etc. I got several statisticallly significant LVs that were correlated with behavioral measures and stopped the analyses.
(This analytic process were same as Mentis et al. 2002)
In this dataset, should I extract mean CMRglc in the regions that were significantly corrrelated with certain behavioral measure and subsequently analyze data using seed-voxel PLS to identify the network?
Best Regards.
Re: group comparison
Posted on 02/02/07 12:30:22
Number of posts: 100
Author: Randy (---.dsl.bell.ca)
Date: 01-09-07 10:43
>
> In this dataset, should I extract mean CMRglc in the regions
> that were significantly corrrelated with certain behavioral
> measure and subsequently analyze data using seed-voxel PLS to
> identify the network?
yes, that would be an excellent idea!
Cheers
Randy
Date: 01-09-07 10:43
>
> In this dataset, should I extract mean CMRglc in the regions
> that were significantly corrrelated with certain behavioral
> measure and subsequently analyze data using seed-voxel PLS to
> identify the network?
yes, that would be an excellent idea!
Cheers
Randy
Re: group comparison
Posted on 02/02/07 12:30:39
Number of posts: 100
Author: JC youn (211.46.97.---)
Date: 01-11-07 03:26
Dear Randy.
1. In the previous question for control and paranoid group, I and You wrote as belows
> > 2. After this, I got one LV. Should I -1 1contrast instead of
> 1
> > -1? In my opinion, it will produce essentially the same
> > results.
>
> you are correct. You will get the same result, but the signs
> on the LV will be flipped. You only need to run one contrast.
>
After performed analysis, I found the results were not identical.
Although almost all locations in the results of one contrast (1 -1) were found in the results of another contrast (-1 1), slight differences on locations,, voxel sizes and BSR were detected. How can I explain this result?
2. When I performed seed voxel analysis, how can I determine the number of seed voxels? Is there any standard method or limitation of number?
3. For the seed voxel analysis, can I select the the voxel value, which was not detected in non-rotated Task PLS?
Many Thanks.
Date: 01-11-07 03:26
Dear Randy.
1. In the previous question for control and paranoid group, I and You wrote as belows
> > 2. After this, I got one LV. Should I -1 1contrast instead of
> 1
> > -1? In my opinion, it will produce essentially the same
> > results.
>
> you are correct. You will get the same result, but the signs
> on the LV will be flipped. You only need to run one contrast.
>
After performed analysis, I found the results were not identical.
Although almost all locations in the results of one contrast (1 -1) were found in the results of another contrast (-1 1), slight differences on locations,, voxel sizes and BSR were detected. How can I explain this result?
2. When I performed seed voxel analysis, how can I determine the number of seed voxels? Is there any standard method or limitation of number?
3. For the seed voxel analysis, can I select the the voxel value, which was not detected in non-rotated Task PLS?
Many Thanks.
Re: group comparison
Posted on 02/02/07 12:31:02
Number of posts: 100
Author: Randy (---.rotman-baycrest.on.ca)
Date: 01-11-07 10:08
JC youn wrote:
> >
> > you are correct. You will get the same result, but the signs
> > on the LV will be flipped. You only need to run one
> contrast.
> >
>
> After performed analysis, I found the results were not
> identical.
> Although almost all locations in the results of one contrast (1
> -1) were found in the results of another contrast (-1 1),
> slight differences on locations,, voxel sizes and BSR were
> detected. How can I explain this result?
you will get slightly different statistical maps b/c of the resampling with bootstraps. if you select 100 random resamples twice, it is quite unlikely that the two sets of 100 will be the same, so you will get slightly different answers. If you run 1000 bootstrap resamples twice, you are more likely to get substantial overlap, but this takes time. If you have a large sample size (e.g., more than ~ 20 subjects per group), I would suggest doing more than the prescribed # of bootstraps to get closer to asymptotic estimation - say 500 to 1000 resamples.
> 2. When I performed seed voxel analysis, how can I determine
> the number of seed voxels? Is there any standard method or
> limitation of number?
no upper or lower limit - for this I usually suggest you go with what is of interest.
> 3. For the seed voxel analysis, can I select the the voxel
> value, which was not detected in non-rotated Task PLS?
yes, of course!
Date: 01-11-07 10:08
JC youn wrote:
> >
> > you are correct. You will get the same result, but the signs
> > on the LV will be flipped. You only need to run one
> contrast.
> >
>
> After performed analysis, I found the results were not
> identical.
> Although almost all locations in the results of one contrast (1
> -1) were found in the results of another contrast (-1 1),
> slight differences on locations,, voxel sizes and BSR were
> detected. How can I explain this result?
you will get slightly different statistical maps b/c of the resampling with bootstraps. if you select 100 random resamples twice, it is quite unlikely that the two sets of 100 will be the same, so you will get slightly different answers. If you run 1000 bootstrap resamples twice, you are more likely to get substantial overlap, but this takes time. If you have a large sample size (e.g., more than ~ 20 subjects per group), I would suggest doing more than the prescribed # of bootstraps to get closer to asymptotic estimation - say 500 to 1000 resamples.
> 2. When I performed seed voxel analysis, how can I determine
> the number of seed voxels? Is there any standard method or
> limitation of number?
no upper or lower limit - for this I usually suggest you go with what is of interest.
> 3. For the seed voxel analysis, can I select the the voxel
> value, which was not detected in non-rotated Task PLS?
yes, of course!
Login to reply to this topic.
