help
help > RE: NBS one sample-test with more negative than positive values/thresholding of connectivity matrices.
Feb 17, 2022 11:02 PM | Andrew Zalesky
RE: NBS one sample-test with more negative than positive values/thresholding of connectivity matrices.
Hi Ole,
1. I am not sure if I follow your concern about more negative than positive values. The one-sample test establishes a null distribution simply by flipping the signs of each sample. Many users wrongly select the one-sample test when they should be using the two-sample test. If you have multiple experimental conditions, I think that you may want to compare connectivity between the conditions, which would require a two-sample test. The one-sample test would be used to assess whether connectivity is different from a given fixed value (i.e. 0) in one particular condition.
2. Yes - it is valid to threshold the matrix prior to statistical inference. Of course a rationale is needed for the choice of threshold and the thresholding method. Any connections that are zero for all subjects will be ignored during statistical testing.
3. A directed version of the NBS is available for download from NITRC, although this version does not have a GUI and only includes very basic features. You may want to symmetrize the matrix to keep things simple in the first instance.
4. As I said, I am not sure if I understand the concern about negative and positive values. This should not be a barrier to statistical inferences, unless I am misunderstanding something.
best,
Andrew
Originally posted by Ole Jonas Böken:
1. I am not sure if I follow your concern about more negative than positive values. The one-sample test establishes a null distribution simply by flipping the signs of each sample. Many users wrongly select the one-sample test when they should be using the two-sample test. If you have multiple experimental conditions, I think that you may want to compare connectivity between the conditions, which would require a two-sample test. The one-sample test would be used to assess whether connectivity is different from a given fixed value (i.e. 0) in one particular condition.
2. Yes - it is valid to threshold the matrix prior to statistical inference. Of course a rationale is needed for the choice of threshold and the thresholding method. Any connections that are zero for all subjects will be ignored during statistical testing.
3. A directed version of the NBS is available for download from NITRC, although this version does not have a GUI and only includes very basic features. You may want to symmetrize the matrix to keep things simple in the first instance.
4. As I said, I am not sure if I understand the concern about negative and positive values. This should not be a barrier to statistical inferences, unless I am misunderstanding something.
best,
Andrew
Originally posted by Ole Jonas Böken:
Dear Prof. Zalesky,
We are running the NBS on task-based functional connectivity data and are facing several issues and maybe you can give us an advice on how to proceed.
1.) We conducted a one-sample t-test for each experimental contrast (e.g., one for the "alerting" contrast, one for "flanker" contrast etc.) and tried different thresholds, as recommended elsewhere in this forum, without getting any sign. results via the NBS.
I also read in the forum that it is a problem if we have more negative than positive values in the connectivity matrices (on a group level), if we want to apply a one-sample t-test, which is the case for all of our experimental conditions.
Just to be sure, is this really the case? Are our results therefore negatively affected by this mismatch of positive and negative connectivity values?
2.) Is it a "valid" approach to apply an a priori thresholding on the data, such as the "thresholding_absolute" from BCT to analyse positive and negative connections separately? Such as:
Because if we do so, we get significant results for both directions (positive and negative).
3.) To make things more complicated, our original data has directed (asymmetrical) connectivity matrices. Since the NBS is not suited for this kind of data we symmetrized our matrices beforehand.
4.) Do you have any other recommendations on how to work with connectivity matrices with more negative than positive connections while using the NBS and a one-sample t-test?
Many thanks in advance!
Best regards,
Ole
We are running the NBS on task-based functional connectivity data and are facing several issues and maybe you can give us an advice on how to proceed.
1.) We conducted a one-sample t-test for each experimental contrast (e.g., one for the "alerting" contrast, one for "flanker" contrast etc.) and tried different thresholds, as recommended elsewhere in this forum, without getting any sign. results via the NBS.
I also read in the forum that it is a problem if we have more negative than positive values in the connectivity matrices (on a group level), if we want to apply a one-sample t-test, which is the case for all of our experimental conditions.
Just to be sure, is this really the case? Are our results therefore negatively affected by this mismatch of positive and negative connectivity values?
2.) Is it a "valid" approach to apply an a priori thresholding on the data, such as the "thresholding_absolute" from BCT to analyse positive and negative connections separately? Such as:
- threshold_absolute(flanker, 0) and threshold_absolute(flanker*-1, 0).
Because if we do so, we get significant results for both directions (positive and negative).
3.) To make things more complicated, our original data has directed (asymmetrical) connectivity matrices. Since the NBS is not suited for this kind of data we symmetrized our matrices beforehand.
4.) Do you have any other recommendations on how to work with connectivity matrices with more negative than positive connections while using the NBS and a one-sample t-test?
Many thanks in advance!
Best regards,
Ole
Threaded View
| Title | Author | Date |
|---|---|---|
| Ole Jonas Böken | Feb 17, 2022 | |
| Ole Jonas Böken | Feb 28, 2022 | |
| Andrew Zalesky | Mar 3, 2022 | |
| Ole Jonas Böken | Feb 23, 2022 | |
| Andrew Zalesky | Feb 24, 2022 | |
| Andrew Zalesky | Feb 17, 2022 | |