sdm-help-list
77 Subscribers 
sdm-help-list > Meta-regression I have read all the forum.
Showing 1-2 of 2 posts
Dec 12, 2017 05:12 PM | Leidy Yurani Cubillos Pinilla - LMU
Meta-regression I have read all the forum.
Hi Joaquim Radua,
Thank you for this wonderful program.
1. I have done meta-regression using R to compare the "estimate" mean of the blobs from the EXTRACTION step with demographic and neuropsychological variables. Is this a good approach to do it, in order to answer the question: Does each of these variable independently affect the brain regions (Blobs) that were obtained when comparing patients and healthy controls?
2. Sometimes few articles 9/31 or even less have a average of 0 in the extraction of the blob file related to groups comparison (patients vs subjects), when I do the regression to other variables, the data looks fragmented with most of he data going to cero. My questions:
Should I still compare the confound variables with this Extraction file?
Should I do the analysis taking out the studies with the cero values or should I impute the values?
Should I do the regression if the number of studies to which the mean differ from cero are low?, is there a "rule" to delimit small amount of studies?
3. I have used the metaregression option of SDM, but this option, to my point of view, does not say specifically whether or not the variables (Specifically: demographic, neuropsychological variables) affect the same part of regions and the regions itself that are identified with the other analyses (difference between patients and healthy controls).
4. Some demographic variables are extreme across papers for example percent medicated tends to be 100% or close to 0%, to these variables I thought about using the meta-regression approach from SDM, then overlap the images with FSL and take out conclusions from there. Is this a good approach
Thank you very much,
Leidy
PRONIA - LMU.
Thank you for this wonderful program.
1. I have done meta-regression using R to compare the "estimate" mean of the blobs from the EXTRACTION step with demographic and neuropsychological variables. Is this a good approach to do it, in order to answer the question: Does each of these variable independently affect the brain regions (Blobs) that were obtained when comparing patients and healthy controls?
2. Sometimes few articles 9/31 or even less have a average of 0 in the extraction of the blob file related to groups comparison (patients vs subjects), when I do the regression to other variables, the data looks fragmented with most of he data going to cero. My questions:
Should I still compare the confound variables with this Extraction file?
Should I do the analysis taking out the studies with the cero values or should I impute the values?
Should I do the regression if the number of studies to which the mean differ from cero are low?, is there a "rule" to delimit small amount of studies?
3. I have used the metaregression option of SDM, but this option, to my point of view, does not say specifically whether or not the variables (Specifically: demographic, neuropsychological variables) affect the same part of regions and the regions itself that are identified with the other analyses (difference between patients and healthy controls).
4. Some demographic variables are extreme across papers for example percent medicated tends to be 100% or close to 0%, to these variables I thought about using the meta-regression approach from SDM, then overlap the images with FSL and take out conclusions from there. Is this a good approach
Thank you very much,
Leidy
PRONIA - LMU.
Jan 21, 2018 03:01 PM | Joaquim Radua
RE: Meta-regression I have read all the forum.
Dear Leidy,
First of all, my apologies for the late response...
1. I would suggest using the effect size of a voxel (e.g., the peak) rather than the mean of the effect sizes of a blob. You may do it with either R or SDM. P-values may be different because with R you usually assume normality, which may not be the case for SDM data, while with SDM you are testing for convergence, which is slightly different from what common tests do.
2. SDM assumes that voxels far from any peak have a null effect size, and that's why your data have zeroes. You cannot discard these studies. This may make data non-normal, for what SDM may be preferable.
3. I'm sorry to say that I don't see you point.
4. I think so.
Hope this helps,
Joaquim
First of all, my apologies for the late response...
1. I would suggest using the effect size of a voxel (e.g., the peak) rather than the mean of the effect sizes of a blob. You may do it with either R or SDM. P-values may be different because with R you usually assume normality, which may not be the case for SDM data, while with SDM you are testing for convergence, which is slightly different from what common tests do.
2. SDM assumes that voxels far from any peak have a null effect size, and that's why your data have zeroes. You cannot discard these studies. This may make data non-normal, for what SDM may be preferable.
3. I'm sorry to say that I don't see you point.
4. I think so.
Hope this helps,
Joaquim