Back to PLS HelpPLScmd group analysis
Posted on 02/02/07 12:13:00
Number of posts: 100
Author: Brian Luus (---.vc.shawcable.net)
Date: 11-27-06 10:14
How is the pls_analysis.m program to be used for group analysis of single-trial data? Is it possible to construct a single datamat of all the single-trial data and run a group PLS in one shot?
We have extracted fMRI timecourses for a number E of ROIs ("elements"), at T timepoints, for Tr trials, for K conditions, for S subjects. What is the best approach for constructing the datamat and finding a reliable trend for our group?
Is it reasonable to make an Sx1 cell array datamat in which each subject's data is in it's own Tr x K row by E x T column matrix, treating each subject as their own group? For this, we have specified that there are K conditions, and Tr subjects per condition. Or are we instead to run PLS on each subject' single-trial data, extract the timecourses for the LVs, and run a second PLS on the single-subject LV timecourses?
Also, does it matter whether we run a Tr x K row by E x T column datamat all at once, or could we just run one Tr x K row by T column datamat separately for each E elements?
Thanks kindly. -Brian
Date: 11-27-06 10:14
How is the pls_analysis.m program to be used for group analysis of single-trial data? Is it possible to construct a single datamat of all the single-trial data and run a group PLS in one shot?
We have extracted fMRI timecourses for a number E of ROIs ("elements"), at T timepoints, for Tr trials, for K conditions, for S subjects. What is the best approach for constructing the datamat and finding a reliable trend for our group?
Is it reasonable to make an Sx1 cell array datamat in which each subject's data is in it's own Tr x K row by E x T column matrix, treating each subject as their own group? For this, we have specified that there are K conditions, and Tr subjects per condition. Or are we instead to run PLS on each subject' single-trial data, extract the timecourses for the LVs, and run a second PLS on the single-subject LV timecourses?
Also, does it matter whether we run a Tr x K row by E x T column datamat all at once, or could we just run one Tr x K row by T column datamat separately for each E elements?
Thanks kindly. -Brian
Replies:
Re: PLScmd group analysis
Posted on 02/02/07 12:13:29
Number of posts: 100
Author: Jimmy (---.rotman-baycrest.on.ca)
Date: 11-27-06 13:24
Hi Brian:
> How is the pls_analysis.m program to be used for group analysis
> of single-trial data?
The pls_analysis.m program is based on (but not compatible with) the PLSgui program below:
http://www.rotman-baycrest.on.ca/pls/UserGuide.htm
As such, the construction of datamat is also inherited from the format we used there.
In addition, the concept of "group" we discussed usually applies to 2 or more groups of subjects. e.g. control group vs. patient group etc.
> Is it possible to construct a single
> datamat of all the single-trial data and run a group PLS in one
> shot?
I am not able to answer this question.
> We have extracted fMRI timecourses for a number E of ROIs
> ("elements"), at T timepoints, for Tr trials, for K conditions,
> for S subjects. What is the best approach for constructing the
> datamat and finding a reliable trend for our group?
I am not quite sure what "Tr trials" exactly indicates. Let's put it in this way. You have extracted fMRI timecourses for a number E of ROIs at T onset timepoints for all S subjects, and you have already known which timepoints corresponds to which one of those K condition. For each onset, there are a number W of TRs for Hemodynamic Response temporal window. Here is the way to construct the datamat:
1. You need to put each subject of each condition in 1 row, and the order must be subjects in conditions. i.e.
Row 1: Condition 1 Subject 1
Row 2: Condition 1 Subject 2
... ...
Row S: Condition 1 Subject S
Row S+1: Condition 2 Subject 1
Row S+2: Condition 2 Subject 2
... ...
Row 2S: Condition 2 Subject S
... ...
Row K*S: Condition K Subject S
2. Within each row, it is always the onset timepoints of each condition followed by Hemodynamic Response window for each onset timepoint. Let's assume that condition1 corresponds to onset timepoints 0, 12, 18, 32, 35, and here is what for Condition 1 Subject 1:
T0 T0+TR T0+TR*2 ... T0+TR*(W-1) T12 T12+TR T12+TR*2 ... T12+TR*(W-1) ... T35 T35+TR T35+TR*2 ... T35+TR*(W-1)
The time sequence will be the same for other subjects of Condition1, but of cause with different image data.
> Is it reasonable to make an Sx1 cell array datamat in which
> each subject's data is in it's own Tr x K row by E x T column
> matrix, treating each subject as their own group? For this, we
> have specified that there are K conditions, and Tr subjects per
> condition. Or are we instead to run PLS on each subject'
> single-trial data, extract the timecourses for the LVs, and run
> a second PLS on the single-subject LV timecourses?
What we did is just to enclose the above datamat matrix into the cell array (single element for 1 group), and run pls_analysis.m
> Also, does it matter whether we run a Tr x K row by E x T
> column datamat all at once, or could we just run one Tr x K row
> by T column datamat separately for each E elements?
The result could be different. What we did is just run all E elements together, instead of run it separately for each element.
Date: 11-27-06 13:24
Hi Brian:
> How is the pls_analysis.m program to be used for group analysis
> of single-trial data?
The pls_analysis.m program is based on (but not compatible with) the PLSgui program below:
http://www.rotman-baycrest.on.ca/pls/UserGuide.htm
As such, the construction of datamat is also inherited from the format we used there.
In addition, the concept of "group" we discussed usually applies to 2 or more groups of subjects. e.g. control group vs. patient group etc.
> Is it possible to construct a single
> datamat of all the single-trial data and run a group PLS in one
> shot?
I am not able to answer this question.
> We have extracted fMRI timecourses for a number E of ROIs
> ("elements"), at T timepoints, for Tr trials, for K conditions,
> for S subjects. What is the best approach for constructing the
> datamat and finding a reliable trend for our group?
I am not quite sure what "Tr trials" exactly indicates. Let's put it in this way. You have extracted fMRI timecourses for a number E of ROIs at T onset timepoints for all S subjects, and you have already known which timepoints corresponds to which one of those K condition. For each onset, there are a number W of TRs for Hemodynamic Response temporal window. Here is the way to construct the datamat:
1. You need to put each subject of each condition in 1 row, and the order must be subjects in conditions. i.e.
Row 1: Condition 1 Subject 1
Row 2: Condition 1 Subject 2
... ...
Row S: Condition 1 Subject S
Row S+1: Condition 2 Subject 1
Row S+2: Condition 2 Subject 2
... ...
Row 2S: Condition 2 Subject S
... ...
Row K*S: Condition K Subject S
2. Within each row, it is always the onset timepoints of each condition followed by Hemodynamic Response window for each onset timepoint. Let's assume that condition1 corresponds to onset timepoints 0, 12, 18, 32, 35, and here is what for Condition 1 Subject 1:
T0 T0+TR T0+TR*2 ... T0+TR*(W-1) T12 T12+TR T12+TR*2 ... T12+TR*(W-1) ... T35 T35+TR T35+TR*2 ... T35+TR*(W-1)
The time sequence will be the same for other subjects of Condition1, but of cause with different image data.
> Is it reasonable to make an Sx1 cell array datamat in which
> each subject's data is in it's own Tr x K row by E x T column
> matrix, treating each subject as their own group? For this, we
> have specified that there are K conditions, and Tr subjects per
> condition. Or are we instead to run PLS on each subject'
> single-trial data, extract the timecourses for the LVs, and run
> a second PLS on the single-subject LV timecourses?
What we did is just to enclose the above datamat matrix into the cell array (single element for 1 group), and run pls_analysis.m
> Also, does it matter whether we run a Tr x K row by E x T
> column datamat all at once, or could we just run one Tr x K row
> by T column datamat separately for each E elements?
The result could be different. What we did is just run all E elements together, instead of run it separately for each element.
Re: PLScmd group analysis
Posted on 02/02/07 12:13:50
Number of posts: 100
Author: Randy (---.rotman-baycrest.on.ca)
Date: 11-29-06 10:27
Hi Brian,
Jimmy answered some of your questions about data formats for what our PLS code supports, but it seems that you are try to do something a little different. If I understand your post, you want to do a sort two-level analysis where you do each subject first, keep their single trial information intact, and then do a group analysis to identify a robust trend across the sample. This essentially the random effects analysis that is part of SPM, FSL, etc., except you wish to look at single trial info. Is this correct?
If so, let me present a couple of issues:
1. You don't need to do the random effects analysis with PLS, since the resampling (bootstrap & permutation) procedure is a mixed-effects analysis.
2. Keeping single trial info for each subject and trying to do a group analysis is tricky for the resampling procedure as we have implemented it, unless you treat each subject as their own group (as you suggest). This might be a neat way to "cluster" subjects in terms of similarities of effects. The issue would be whether your single trial data are clean enough.
3. As for data configuration, I would go with Trials X conditions (Tr X K) in rows and ROI X Time (E X T) in columns. Breaking it down by ROI (E) does not really take advantage of the distributed signal and effectively makes it a univariate analysis (except for the time domain).
Date: 11-29-06 10:27
Hi Brian,
Jimmy answered some of your questions about data formats for what our PLS code supports, but it seems that you are try to do something a little different. If I understand your post, you want to do a sort two-level analysis where you do each subject first, keep their single trial information intact, and then do a group analysis to identify a robust trend across the sample. This essentially the random effects analysis that is part of SPM, FSL, etc., except you wish to look at single trial info. Is this correct?
If so, let me present a couple of issues:
1. You don't need to do the random effects analysis with PLS, since the resampling (bootstrap & permutation) procedure is a mixed-effects analysis.
2. Keeping single trial info for each subject and trying to do a group analysis is tricky for the resampling procedure as we have implemented it, unless you treat each subject as their own group (as you suggest). This might be a neat way to "cluster" subjects in terms of similarities of effects. The issue would be whether your single trial data are clean enough.
3. As for data configuration, I would go with Trials X conditions (Tr X K) in rows and ROI X Time (E X T) in columns. Breaking it down by ROI (E) does not really take advantage of the distributed signal and effectively makes it a univariate analysis (except for the time domain).
Re: PLScmd group analysis
Posted on 02/02/07 12:14:13
Number of posts: 100
Author: Jennifer Labus (149.142.243.---)
Date: 11-29-06 18:08
Interesting question.
A few questions about your response.
How is "the resampling (bootstrap & permutation) procedure a mixed-effects analysis"?
What do you mean by "single trial data are clean enough"?
Date: 11-29-06 18:08
Interesting question.
A few questions about your response.
How is "the resampling (bootstrap & permutation) procedure a mixed-effects analysis"?
What do you mean by "single trial data are clean enough"?
Re: PLScmd group analysis
Posted on 02/02/07 12:14:33
Number of posts: 100
Author: Randy (---.rotman-baycrest.on.ca)
Date: 11-30-06 17:08
Jennifer Labus wrote:
> Interesting question.
> A few questions about your response.
>
> How is "the resampling (bootstrap & permutation) procedure a
> mixed-effects analysis"?
More properly, I should have said "equivalent to" a mixed effects analysis. Fixed effects are estimated, but by resampling subjects, you get an estimate of the robustness of the effects within the sample, which is a random effects question. Stephen Strother also makes this point about his NPAIRS resampling framework:
Strother SC, Anderson J, Hansen LK, Kjems U, Kustra R, Sidtis J, Frutiger S, Muley S, LaConte S, Rottenberg D (2002) The quantitative evaluation of functional neuroimaging experiments: the NPAIRS data analysis framework. Neuroimage 15:747-771
> What do you mean by "single trial data are clean enough"?
In the sense that the evoked response is somewhat detectable in each trial without a whole lot of filtering. Remember that most fMRI analyses are done on trial-averaged data.
Date: 11-30-06 17:08
Jennifer Labus wrote:
> Interesting question.
> A few questions about your response.
>
> How is "the resampling (bootstrap & permutation) procedure a
> mixed-effects analysis"?
More properly, I should have said "equivalent to" a mixed effects analysis. Fixed effects are estimated, but by resampling subjects, you get an estimate of the robustness of the effects within the sample, which is a random effects question. Stephen Strother also makes this point about his NPAIRS resampling framework:
Strother SC, Anderson J, Hansen LK, Kjems U, Kustra R, Sidtis J, Frutiger S, Muley S, LaConte S, Rottenberg D (2002) The quantitative evaluation of functional neuroimaging experiments: the NPAIRS data analysis framework. Neuroimage 15:747-771
> What do you mean by "single trial data are clean enough"?
In the sense that the evoked response is somewhat detectable in each trial without a whole lot of filtering. Remember that most fMRI analyses are done on trial-averaged data.
Login to reply to this topic.
