Back to PLS HelpLongitudinal data
Posted on 04/27/09 08:40:41
Number of posts: 22
Hi,
We have longitudinal fMRI data with 3 timepoints (i.e. runs). We are primarily interested in conducting a seed pls to examine whether functional connectivity alters with age. Unfortunately, however, after preprocessing we now have some subjects contributing to all 3 timepoints, some to 2 and some to just 1. Rather than treating the 3 timepoints as a between-subject factor, can we enter different numbers of runs for each subject - will PLS care that subject x's datamatrix contains run 1 and 3 data whilst subject y's datamatrix contains only run 1 data?
Any help would be grately appreciated.
Thanks in advance,
Dan.
Replies:
Untitled Post
I'm Online
Posted on 04/27/09 10:32:15
Number of posts: 291
Yes, you can enter different numbers of runs for each subject, as long as you do not merge data within each run. i.e. you will have to choose "Merge Data Across All Runs" for all subjects in the session window.
Untitled Post
Posted on 04/27/09 11:16:12
Number of posts: 22
Thank you for your very clear response.
One last question, if we select to "Merge Across All Runs" are we still taking into account the effects of time/visit/run - i.e. the longitudinal component of the data, which I see as the difference between runs?
Thanks again,
Dan.
Untitled Post
I'm Online
Posted on 04/27/09 11:19:53
Number of posts: 291
The difference between runs will not taking into account, since they are all averaged together.
Untitled Post
Posted on 04/28/09 05:19:42
Number of posts: 22
OK, thanks.
So what would you suggest if we did want to assess the within-subject effect of time/visit/run in a longitudinal analysis, where only very few subjects are contributing to all (three) timepoints?
What if we added age (say, in months) as an additional factor to be correlated to the fMRI data? Would this overcome the issue of missing runs?
Untitled Post
I'm Online
Posted on 04/28/09 10:29:52
Number of posts: 291
Lack of data can always pose limitation for your study. Adding another condition (like 'Age') will reveal the characteristic for that dimension; however, I don't see that it will help you to differentiate the effects between runs. Therefore, I suggest that you may have to get more data.
Untitled Post
Posted on 04/28/09 11:54:05
Number of posts: 394
Hi Dan,
I figured I should chime in on this. The decision on how to proceed will really depend on how many subjects are missing data points. There is no facility with our program to deal with missing data. If the number of subjects with missing data is small, I would suggest dropping them and proceeding so long as your sample size is reasonable.
As you suggest, you could also treat the three time points as a between subjects factor, but this runs into some problems for the permutation test as the assumption of exchangibility is violated.
If you want to give me a few more details of your study, I may be able to make some more suggestions
Randy
I figured I should chime in on this. The decision on how to proceed will really depend on how many subjects are missing data points. There is no facility with our program to deal with missing data. If the number of subjects with missing data is small, I would suggest dropping them and proceeding so long as your sample size is reasonable.
As you suggest, you could also treat the three time points as a between subjects factor, but this runs into some problems for the permutation test as the assumption of exchangibility is violated.
If you want to give me a few more details of your study, I may be able to make some more suggestions
Randy
Untitled Post
Posted on 04/28/09 12:25:56
Number of posts: 22
quote:
Hi Dan,
I figured I should chime in on this. The decision on how to proceed will really depend on how many subjects are missing data points. There is no facility with our program to deal with missing data. If the number of subjects with missing data is small, I would suggest dropping them and proceeding so long as your sample size is reasonable.
As you suggest, you could also treat the three time points as a between subjects factor, but this runs into some problems for the permutation test as the assumption of exchangibility is violated.
If you want to give me a few more details of your study, I may be able to make some more suggestions
Randy
Randy,Hi Dan,
I figured I should chime in on this. The decision on how to proceed will really depend on how many subjects are missing data points. There is no facility with our program to deal with missing data. If the number of subjects with missing data is small, I would suggest dropping them and proceeding so long as your sample size is reasonable.
As you suggest, you could also treat the three time points as a between subjects factor, but this runs into some problems for the permutation test as the assumption of exchangibility is violated.
If you want to give me a few more details of your study, I may be able to make some more suggestions
Randy
I really appreciate your input on this one, thanks.
We have longitudinal data of 65 adolescent subjects over 3 years. Due to the presence of braces at some point over the period, head motion etc., we have 44 subjects contributing to our time_1, 48 to our time_2 and 47 to our time_3. Unfortunately, however, we have only 25 contributing to all 3 timepoints, with 23 contributing to 2 timepoints.
Having read a few papers it would seem that the mixed-model regression will allow me capitalise on the within-subject factor of time whilst accounting for the missing data by also considering the cross-sectional component. As such, it seems the general effect of time can be assessed in this manner.
Primarily, however, we are interested in the development of functional connectivity and I had hoped that I could use a seed-PLS analysis to investigate this. In this context it makes no sense to merge the data across runs, thereby losing the within-subject effect of run on functional connectivity. It now seems that treating the different runs as 3 individual groups and correlating the data with age (in months) or visit/run is perhaps that only way to look at the effect of time.
Any suggestions would be greatly appreciated.
Many thanks in advance,
Dan.
Untitled Post
I'm Online
Posted on 04/28/09 12:38:15
Number of posts: 291
We usually run group test for different subject groups (e.g. Patient Group vs. Control Group). I am not sure if you can also try to run group test on time_1, time_2, and time_3. If it is possible, the difference between them can be obtained from group effects.
Untitled Post
Posted on 04/28/09 20:00:44
Number of posts: 394
quote:
Randy,
Dan,Randy,
I really appreciate your input on this one, thanks.
We have longitudinal data of 65 adolescent subjects over 3 years. Due to the presence of braces at some point over the period, head motion etc., we have 44 subjects contributing to our time_1, 48 to our time_2 and 47 to our time_3. Unfortunately, however, we have only 25 contributing to all 3 timepoints, with 23 contributing to 2 timepoints.
Having read a few papers it would seem that the mixed-model regression will allow me capitalise on the within-subject factor of time whilst accounting for the missing data by also considering the cross-sectional component. As such, it seems the general effect of time can be assessed in this manner.
Primarily, however, we are interested in the development of functional connectivity and I had hoped that I could use a seed-PLS analysis to investigate this. In this context it makes no sense to merge the data across runs, thereby losing the within-subject effect of run on functional connectivity. It now seems that treating the different runs as 3 individual groups and correlating the data with age (in months) or visit/run is perhaps that only way to look at the effect of time.
Any suggestions would be greatly appreciated.
Many thanks in advance,
Dan.
I think the best thing is to try and treat it like three groups, as Jimmy suggests, and see what comes out. You can still use the seed PLS in this context. The permutation test is slightly different for behavior/seed PLS than it is for activation analysis (task PLS). In the case of the seed PLS, the data from seed are randomly reassigned, while the rest of the data stay put, so there is "less" of a problem with exchangibility. If you wish to add both the seed and age in together, that might make for an interesting analysis.
Let me know if I can be of further help.
Randy
Login to reply to this topic.
