help
help > RE: Strong widespread connectivity
Jul 30, 2015 01:07 AM | Darren Yeo
RE: Strong widespread connectivity
Hi Alfonso,
Thanks so much for your detailed clarification. That really helped.
Best,
Darren
Originally posted by Alfonso Nieto-Castanon:
Thanks so much for your detailed clarification. That really helped.
Best,
Darren
Originally posted by Alfonso Nieto-Castanon:
Hi
Darren,
First, regarding the group-level average connectivity, yes, it is perfectly normal to observe extremely high stats for those results (for a sample size of ~50 as in your case). This simply reflects that many of the features observed in the seed-to-voxel connectivity patterns are not only very salient but also highly consistent across subjects.
Then, regarding the results when controlling for additional covariates, I would first double-check that all of your covariates are centered (for each covariate the mean value across all subjects is zero). When displaying main-connectivity effects while including additional covariate effects, the main-connectivity effect is always estimated at the zero-level of your covariates, so you typically want that zero-level to be some meaningful number (e.g. the average value of that variable across all subjects). If, for example, you are using a "gender" covariate with values 1's and 2's, then the zero-level of this covariate is not particularly meaningful (and the resulting main-connectivity effects will be reflecting simply an extrapolation of the expected connectivity for a gender=0 value...)
Last, and in general, when adding this sort of control covariates (assuming centering here), I would still expect to see either similarly high or even higher main-connectivity effects (since the effect of controlling for these covariates in this case will be mainly to decrease the residual inter-subject variability, and that typically will have a greater impact on the analysis sensitivity/power than the decrease in degrees of freedom resulting from the additional covariates in your second-level model).
Hope this helps
Alfonso
Originally posted by Darren Yeo:
First, regarding the group-level average connectivity, yes, it is perfectly normal to observe extremely high stats for those results (for a sample size of ~50 as in your case). This simply reflects that many of the features observed in the seed-to-voxel connectivity patterns are not only very salient but also highly consistent across subjects.
Then, regarding the results when controlling for additional covariates, I would first double-check that all of your covariates are centered (for each covariate the mean value across all subjects is zero). When displaying main-connectivity effects while including additional covariate effects, the main-connectivity effect is always estimated at the zero-level of your covariates, so you typically want that zero-level to be some meaningful number (e.g. the average value of that variable across all subjects). If, for example, you are using a "gender" covariate with values 1's and 2's, then the zero-level of this covariate is not particularly meaningful (and the resulting main-connectivity effects will be reflecting simply an extrapolation of the expected connectivity for a gender=0 value...)
Last, and in general, when adding this sort of control covariates (assuming centering here), I would still expect to see either similarly high or even higher main-connectivity effects (since the effect of controlling for these covariates in this case will be mainly to decrease the residual inter-subject variability, and that typically will have a greater impact on the analysis sensitivity/power than the decrease in degrees of freedom resulting from the additional covariates in your second-level model).
Hope this helps
Alfonso
Originally posted by Darren Yeo:
Hi Alfonso,
I am working on a resting-state analysis for a single group using several seed regions (spherical ROIs centered around peak coordinates from meta-analyses and anatomical atlas-based ROIs). I am getting pretty reasonable-sized clusters when I included several covariates of interest and no interest. For instance, if I wanted to find the relationship between the functional connectivity of left angular gyrus and a behavioral score A, while controlling for sex and age, I specified the following between-subjects contrast: [0 1 0 0]
However, out of curiosity, I also looked at the group-level connectivity without any covariates by specifying the following contrast:
All subjects [1], Rest [1], and Angular Gyrus Left [1]
Firstly, the surviving clusters are huge and span across many regions (similar to those shown in the manual). Secondly, the height statistics are very high such that even lowering the height threshold to p < 0.00000001 (and cluster p-FDR corrected at 0.05) still resulted in huge clusters. I have attached screenshots of the results. Is that supposed to be expected?
I then performed the following analyses to ascertain in the addition of covariates would make the cluster sizes more reasonable:
2) All subjects, Sex [1 0], Rest [1], and Angular Gyrus Left [1]
3) All subjects, Sex, Age [1 0 0], Rest [1], and Angular Gyrus Left [1]
The results did change rather drastically as shown in the screenshots.
If I would like to report the "pure" resting state connectivity without correlating with any covariates, in addition to my primary analyses, would such strong and widespread networks be ridiculous to report?
Thanks!
Best,
Darren
I am working on a resting-state analysis for a single group using several seed regions (spherical ROIs centered around peak coordinates from meta-analyses and anatomical atlas-based ROIs). I am getting pretty reasonable-sized clusters when I included several covariates of interest and no interest. For instance, if I wanted to find the relationship between the functional connectivity of left angular gyrus and a behavioral score A, while controlling for sex and age, I specified the following between-subjects contrast: [0 1 0 0]
However, out of curiosity, I also looked at the group-level connectivity without any covariates by specifying the following contrast:
All subjects [1], Rest [1], and Angular Gyrus Left [1]
Firstly, the surviving clusters are huge and span across many regions (similar to those shown in the manual). Secondly, the height statistics are very high such that even lowering the height threshold to p < 0.00000001 (and cluster p-FDR corrected at 0.05) still resulted in huge clusters. I have attached screenshots of the results. Is that supposed to be expected?
I then performed the following analyses to ascertain in the addition of covariates would make the cluster sizes more reasonable:
2) All subjects, Sex [1 0], Rest [1], and Angular Gyrus Left [1]
3) All subjects, Sex, Age [1 0 0], Rest [1], and Angular Gyrus Left [1]
The results did change rather drastically as shown in the screenshots.
If I would like to report the "pure" resting state connectivity without correlating with any covariates, in addition to my primary analyses, would such strong and widespread networks be ridiculous to report?
Thanks!
Best,
Darren
Threaded View
| Title | Author | Date |
|---|---|---|
| Darren Yeo | Jul 28, 2015 | |
| Alfonso Nieto-Castanon | Jul 29, 2015 | |
| Darren Yeo | Jul 30, 2015 | |