help
help > RE: Interpretation of results
May 29, 2014 08:05 AM | Alfonso Nieto-Castanon - Boston University
RE: Interpretation of results
Hi Merina,
First, and just to make sure, when testing the simple main effects in group A or B you typically want to have the other regressors centered (subtracting the mean). Otherwise the simple main effects in each group will be computed at the zero-level of all of your other regressors, which may be not particularly meaningful (e.g. average connectivity at age=0 or IQ=0 are not particularly meaningful levels to look at; it would be more meaningful to first center these variables so that the resulting age_centered=0 or IQ_centered=0 levels represent the average age and IQ within your sample). Could this explain why you are seeing some unexpected results in these analyses?
Best
Alfonso
Originally posted by Merina Su:
First, and just to make sure, when testing the simple main effects in group A or B you typically want to have the other regressors centered (subtracting the mean). Otherwise the simple main effects in each group will be computed at the zero-level of all of your other regressors, which may be not particularly meaningful (e.g. average connectivity at age=0 or IQ=0 are not particularly meaningful levels to look at; it would be more meaningful to first center these variables so that the resulting age_centered=0 or IQ_centered=0 levels represent the average age and IQ within your sample). Could this explain why you are seeing some unexpected results in these analyses?
Best
Alfonso
Originally posted by Merina Su:
Hi
Alfonso,
Many thanks for your reply! I've downloaded version 14c and tried your suggestions, but I'm still a little confused about the final step, please see below.
Originally posted by Alfonso Nieto-Castanon:
Works great, thanks for this! I had to get used to the shuffling of subjects when putting in contrasts, but it does display what I needed.
If so, how should positive and negative z-scores be interpreted? E.g. positive z-score is positive functional connectivity between regions, negative z-score indicates a negative functional connectivity?
Yes, exactly (again assuming that you were not testing any within-subject effects -e.g. between-conditions or between-sources)
Yes, but to actually assert that you need to perform additional analyses (you should not assert that from the post-hoc analyses alone, since those will certainly have some amount of bias). In particular I would run again seed-to-voxel analyses now looking at the simple main effects. For example select A, B, age, IQ, and gender and enter a [1 0 0 0 0] contrast to look at the connectivity in group A and a [0 1 0 0 0] contrast to look at the connectivity in group B. If the former is significant in region X as well then you can assert the 'positive connectivity in group A' portion of your phrase, and if the latter is significant (negative contrast) in region X as well then you can assert the 'negative connectivity in group B' portion. If you fail to assert some portion of this (e.g. you do not get significantly negative connectivity in region X for group B) then the negative values that you were seeing were likely due to selection bias (this bias is typical of post hoc analyses, and it will be greater if you select your values from the peak- response instead of averaged across the entire area of interest).
This is where I got a little confused. I ran the 1 0 0 0 0 and 0 1 0 0 0 contrasts and both gave me the same, but rather surprising result. Both analyses came up with non-significant contrast and lowering the threshold to cluster uncorrected, both came up with the same region, which is not even close to region X! What does this mean?
Many thanks for your help Alfonso, it's much appreciated!
Merina
Many thanks for your reply! I've downloaded version 14c and tried your suggestions, but I'm still a little confused about the final step, please see below.
Originally posted by Alfonso Nieto-Castanon:
Dear
Merina
Some thoughts on your questions below
Best
Alfonso
Originally posted by Merina Su:
I am not totally sure how the Y.mat file that you mention was generated, but assuming it contains the original values used in the second-level analysis (and given that in your case you are selecting a single condition and a single source) then yes, those should be the Fisher-transformed correlation coefficients for each subject between the selected source and region X or Y.
In case it is useful and in order to make these sort of explorations simpler, in the last version (CONN14c) I have added a small utility (in Tools->Calculator) that allows you to explore/display/analyze subject-level measures. In your case you would want to go to 'seed-to-voxel results explorer' for your original contrast (A, B, age, IQ, and gender with [1 -1 0 0 0] contrast) where you should see your region X, and the click on the 'import values' button. That will import the average connectivity within the X area for each subject as a new second-level covariate into the CONN toolbox. Importing the values averaged over the entire cluster is better than evaluating the values at the peak voxel since it will give produce considerably less bias. You can see what the corresponding values are in the Setup->Covariates->Second-level tab, and if you go to Tools->Calculator and select this new measure in the 'Measures' list (and then enter the same between-subject effects/contrast in the between-subject effects list as you did in the original analyses) it will display these values and their associated fitted effects.
Some thoughts on your questions below
Best
Alfonso
Originally posted by Merina Su:
Dear Alfonso and others,
I have finished some preliminary second-level analyses and want to explore some of the results further. However, I want to be sure that I have understood the analyses steps to properly interpret some of the results:
Data:
I have 2 groups (A and B), 3 covariates of no interest (age, IQ, gender), and 3 covariates of interest (I, II, III)
Analyses (seed-voxel):
1. I have looked for between subject effects with a GLM with A,B, age, IQ, gender and the contrast (1 -1 0 0 0). This gave a region X
2. Separate regression analyses with all, age, IQ, gender, covariate I and the contrast (0 0 0 0 1). This gave region Y.
Post-hoc analyses:
I then used SPM8 to explore and plot the fitted responses in regions X and Y. If I understand correctly, these values stored in Y.mat would correspond to the z-scores that were fed into the second-level analyses - is this correct?
I have finished some preliminary second-level analyses and want to explore some of the results further. However, I want to be sure that I have understood the analyses steps to properly interpret some of the results:
Data:
I have 2 groups (A and B), 3 covariates of no interest (age, IQ, gender), and 3 covariates of interest (I, II, III)
Analyses (seed-voxel):
1. I have looked for between subject effects with a GLM with A,B, age, IQ, gender and the contrast (1 -1 0 0 0). This gave a region X
2. Separate regression analyses with all, age, IQ, gender, covariate I and the contrast (0 0 0 0 1). This gave region Y.
Post-hoc analyses:
I then used SPM8 to explore and plot the fitted responses in regions X and Y. If I understand correctly, these values stored in Y.mat would correspond to the z-scores that were fed into the second-level analyses - is this correct?
I am not totally sure how the Y.mat file that you mention was generated, but assuming it contains the original values used in the second-level analysis (and given that in your case you are selecting a single condition and a single source) then yes, those should be the Fisher-transformed correlation coefficients for each subject between the selected source and region X or Y.
In case it is useful and in order to make these sort of explorations simpler, in the last version (CONN14c) I have added a small utility (in Tools->Calculator) that allows you to explore/display/analyze subject-level measures. In your case you would want to go to 'seed-to-voxel results explorer' for your original contrast (A, B, age, IQ, and gender with [1 -1 0 0 0] contrast) where you should see your region X, and the click on the 'import values' button. That will import the average connectivity within the X area for each subject as a new second-level covariate into the CONN toolbox. Importing the values averaged over the entire cluster is better than evaluating the values at the peak voxel since it will give produce considerably less bias. You can see what the corresponding values are in the Setup->Covariates->Second-level tab, and if you go to Tools->Calculator and select this new measure in the 'Measures' list (and then enter the same between-subject effects/contrast in the between-subject effects list as you did in the original analyses) it will display these values and their associated fitted effects.
Works great, thanks for this! I had to get used to the shuffling of subjects when putting in contrasts, but it does display what I needed.
If so, how should positive and negative z-scores be interpreted? E.g. positive z-score is positive functional connectivity between regions, negative z-score indicates a negative functional connectivity?
Yes, exactly (again assuming that you were not testing any within-subject effects -e.g. between-conditions or between-sources)
Also, if group A has mainly positive z-values in
region X, and group B has mainly negative values, does this mean
that the functional connectivity difference was driven by the
increased functional connectivity in group A vs negative
connectivity in group B?
Yes, but to actually assert that you need to perform additional analyses (you should not assert that from the post-hoc analyses alone, since those will certainly have some amount of bias). In particular I would run again seed-to-voxel analyses now looking at the simple main effects. For example select A, B, age, IQ, and gender and enter a [1 0 0 0 0] contrast to look at the connectivity in group A and a [0 1 0 0 0] contrast to look at the connectivity in group B. If the former is significant in region X as well then you can assert the 'positive connectivity in group A' portion of your phrase, and if the latter is significant (negative contrast) in region X as well then you can assert the 'negative connectivity in group B' portion. If you fail to assert some portion of this (e.g. you do not get significantly negative connectivity in region X for group B) then the negative values that you were seeing were likely due to selection bias (this bias is typical of post hoc analyses, and it will be greater if you select your values from the peak- response instead of averaged across the entire area of interest).
This is where I got a little confused. I ran the 1 0 0 0 0 and 0 1 0 0 0 contrasts and both gave me the same, but rather surprising result. Both analyses came up with non-significant contrast and lowering the threshold to cluster uncorrected, both came up with the same region, which is not even close to region X! What does this mean?
Many thanks for your help Alfonso, it's much appreciated!
Merina
Threaded View
| Title | Author | Date |
|---|---|---|
| Merina Su | Apr 28, 2014 | |
| Alfonso Nieto-Castanon | May 8, 2014 | |
| Merina Su | May 15, 2014 | |
| Alfonso Nieto-Castanon | May 29, 2014 | |
| Veronique DT | May 25, 2021 | |