help >

**Is small cluster size problematic?**Showing 1-4 of 4 posts

Jul 6, 2018 08:07 AM | wzhong

Is small cluster size problematic?

Hi,

I have run some second-level seed-to-voxel analysis in conn. My model is quite simple: I am using multiple regression to find the voxels whose connectivity with a few ROI seed regions correlate with a behavioral variable of interest after covarying out mean FD, age and gender. I use a height extent of p<0.001 uncorrected and a cluster correction of p<0.05 FDR. I have a few significant results for multiple seed regions, but the size of the significant clusters are very small; the largest is about 50 voxels and many are between 15 and 30 voxels in size. I am wondering whether this would be problematic and if so what may have caused the problem.

Thanks!

I have run some second-level seed-to-voxel analysis in conn. My model is quite simple: I am using multiple regression to find the voxels whose connectivity with a few ROI seed regions correlate with a behavioral variable of interest after covarying out mean FD, age and gender. I use a height extent of p<0.001 uncorrected and a cluster correction of p<0.05 FDR. I have a few significant results for multiple seed regions, but the size of the significant clusters are very small; the largest is about 50 voxels and many are between 15 and 30 voxels in size. I am wondering whether this would be problematic and if so what may have caused the problem.

Thanks!

Jul 11, 2018 04:07 AM | Alfonso Nieto-Castanon -

*Boston University*RE: Is small cluster size problematic?

Hi,

That is strange, could you please give me more details about how you have specified&run your second-level model? This may have been caused by a problem in the estimation of the smoothing level of the residuals and/or in the assumptions of the RFT model used for cluster-level statistics, so perhaps a good idea would be to select the 'non-parametric stats' option in CONN's results explorer window in order to have CONN run permutation/randomization tests for cluster-level stats instead.

Hope this helps

Alfonso

That is strange, could you please give me more details about how you have specified&run your second-level model? This may have been caused by a problem in the estimation of the smoothing level of the residuals and/or in the assumptions of the RFT model used for cluster-level statistics, so perhaps a good idea would be to select the 'non-parametric stats' option in CONN's results explorer window in order to have CONN run permutation/randomization tests for cluster-level stats instead.

Hope this helps

Alfonso

*Originally posted by wzhong:*Hi,

I have run some second-level seed-to-voxel analysis in conn. My model is quite simple: I am using multiple regression to find the voxels whose connectivity with a few ROI seed regions correlate with a behavioral variable of interest after covarying out mean FD, age and gender. I use a height extent of p<0.001 uncorrected and a cluster correction of p<0.05 FDR. I have a few significant results for multiple seed regions, but the size of the significant clusters are very small; the largest is about 50 voxels and many are between 15 and 30 voxels in size. I am wondering whether this would be problematic and if so what may have caused the problem.

Thanks!

I have run some second-level seed-to-voxel analysis in conn. My model is quite simple: I am using multiple regression to find the voxels whose connectivity with a few ROI seed regions correlate with a behavioral variable of interest after covarying out mean FD, age and gender. I use a height extent of p<0.001 uncorrected and a cluster correction of p<0.05 FDR. I have a few significant results for multiple seed regions, but the size of the significant clusters are very small; the largest is about 50 voxels and many are between 15 and 30 voxels in size. I am wondering whether this would be problematic and if so what may have caused the problem.

Thanks!

Jul 11, 2018 07:07 AM | wzhong

RE: Is small cluster size problematic?

Hi Alfonso,

I should note that I performed the preprocessing steps including denoising and smoothing outside conn in a custom in-house pipeline, then I imported the data into conn using batch scripting skipping preprocessing and denoising steps. In my preprocessing I used 3dBlurToFWHM smooting with a kernel size of 6mm.

Would this have caused problems for smoothing when importing into conn? I also have non-smoothed, denoised image files, should I have used these in the conn analysis and used smoothing in conn? If these do not work I will try nonparametrics.

I ran the second level as follows:

Subject effects: [AllSubjects behav age_mean_centered male female mean_fd], contrast = [0 1 0 0 0 0]

Where behav is the outcome measure of interest (it is a continuous measures with a fairly small range between 1 and 10 and is not mean-centered); for the covariates, age has been mean centered, male and female are two binary indicators with values of 0 or 1 for being male or female (and vice versa), and mean_fd is the subject motion measure (not mean-centered since I think 0 is meaningful here indicating subjects without motion).

Conditions: Rest [1]

For Sources I entered my ROIs of interest and run a F-test contrast (eye(n)). I also tried running separate t-tests for each source ROI and bonferroni correct for the cluster FDR alpha-level, the resulting clusters are all small.

Thanks!

I should note that I performed the preprocessing steps including denoising and smoothing outside conn in a custom in-house pipeline, then I imported the data into conn using batch scripting skipping preprocessing and denoising steps. In my preprocessing I used 3dBlurToFWHM smooting with a kernel size of 6mm.

Would this have caused problems for smoothing when importing into conn? I also have non-smoothed, denoised image files, should I have used these in the conn analysis and used smoothing in conn? If these do not work I will try nonparametrics.

I ran the second level as follows:

Subject effects: [AllSubjects behav age_mean_centered male female mean_fd], contrast = [0 1 0 0 0 0]

Where behav is the outcome measure of interest (it is a continuous measures with a fairly small range between 1 and 10 and is not mean-centered); for the covariates, age has been mean centered, male and female are two binary indicators with values of 0 or 1 for being male or female (and vice versa), and mean_fd is the subject motion measure (not mean-centered since I think 0 is meaningful here indicating subjects without motion).

Conditions: Rest [1]

For Sources I entered my ROIs of interest and run a F-test contrast (eye(n)). I also tried running separate t-tests for each source ROI and bonferroni correct for the cluster FDR alpha-level, the resulting clusters are all small.

Thanks!

Jul 12, 2018 03:07 PM | Alfonso Nieto-Castanon -

*Boston University*RE: Is small cluster size problematic?

Hi,

The details of your second-level analysis specification all seem perfectly correct, but yes, it is possible that the issue might be related to the custom smoothing process. Generally if the images are not sufficiently spatially smoothed, and/or if the level of smoothing is not homogeneous across the entire volume, that might cause problems with some of the Random Field Theory assumptions underlying the parametric cluster-estimation procedure (again pointing to either applying additional spatial smoothing to your data or using non-parametric stats in order to validate your results)

Hope this helps

Alfonso

The details of your second-level analysis specification all seem perfectly correct, but yes, it is possible that the issue might be related to the custom smoothing process. Generally if the images are not sufficiently spatially smoothed, and/or if the level of smoothing is not homogeneous across the entire volume, that might cause problems with some of the Random Field Theory assumptions underlying the parametric cluster-estimation procedure (again pointing to either applying additional spatial smoothing to your data or using non-parametric stats in order to validate your results)

Hope this helps

Alfonso

*Originally posted by wzhong:*Hi Alfonso,

I should note that I performed the preprocessing steps including denoising and smoothing outside conn in a custom in-house pipeline, then I imported the data into conn using batch scripting skipping preprocessing and denoising steps. In my preprocessing I used 3dBlurToFWHM smooting with a kernel size of 6mm.

Would this have caused problems for smoothing when importing into conn? I also have non-smoothed, denoised image files, should I have used these in the conn analysis and used smoothing in conn? If these do not work I will try nonparametrics.

I ran the second level as follows:

Subject effects: [AllSubjects behav age_mean_centered male female mean_fd], contrast = [0 1 0 0 0 0]

Where behav is the outcome measure of interest (it is a continuous measures with a fairly small range between 1 and 10 and is not mean-centered); for the covariates, age has been mean centered, male and female are two binary indicators with values of 0 or 1 for being male or female (and vice versa), and mean_fd is the subject motion measure (not mean-centered since I think 0 is meaningful here indicating subjects without motion).

Conditions: Rest [1]

For Sources I entered my ROIs of interest and run a F-test contrast (eye(n)). I also tried running separate t-tests for each source ROI and bonferroni correct for the cluster FDR alpha-level, the resulting clusters are all small.

Thanks!

I should note that I performed the preprocessing steps including denoising and smoothing outside conn in a custom in-house pipeline, then I imported the data into conn using batch scripting skipping preprocessing and denoising steps. In my preprocessing I used 3dBlurToFWHM smooting with a kernel size of 6mm.

Would this have caused problems for smoothing when importing into conn? I also have non-smoothed, denoised image files, should I have used these in the conn analysis and used smoothing in conn? If these do not work I will try nonparametrics.

I ran the second level as follows:

Subject effects: [AllSubjects behav age_mean_centered male female mean_fd], contrast = [0 1 0 0 0 0]

Where behav is the outcome measure of interest (it is a continuous measures with a fairly small range between 1 and 10 and is not mean-centered); for the covariates, age has been mean centered, male and female are two binary indicators with values of 0 or 1 for being male or female (and vice versa), and mean_fd is the subject motion measure (not mean-centered since I think 0 is meaningful here indicating subjects without motion).

Conditions: Rest [1]

For Sources I entered my ROIs of interest and run a F-test contrast (eye(n)). I also tried running separate t-tests for each source ROI and bonferroni correct for the cluster FDR alpha-level, the resulting clusters are all small.

Thanks!