help
help > RE: Calculator p-value Negative Contrast
Feb 4, 2016 02:02 PM | Alfonso Nieto-Castanon - Boston University
RE: Calculator p-value Negative Contrast
Hi Kaylah,
Regarding your first question, yes, the group-by-performance interaction effect in this case would appear to be driven by both of these effects (i.e. a significant difference in the association between performance and connectivity measures between groups, driven by positive associations in groupB and negative associations in groupA). The sizes of both the positive and negative associations seem to be comparable from your plots but it would be nice to quantify this explicitly (e.g. computing the regression coefficients and associated stats in each group separately)
Regarding your question about multiple comparisons, you are right that seed-to-voxel results are performed separately for each seed (so if you are testing multiple seeds you typically would need to apply and additional multiple comparison correction). The applicability of this additional correction is related to whether your 4 different analyses (one for each of your 4 seeds) are considered each testing a separate hypotheses (in that case you would not typically apply an additional correction), or whether those 4 different analyses are part of the same hypothesis (e.g. you were testing effects across any of those seeds, and in that case you would typically apply an additional correction). There are at least a couple of common ways to apply this additional correction if needed.
1) instead of performing 4 separate analyses (one per seed), perform a single analysis selecting all 4 seeds (and use a contrast eye(4) = [1 0 0 0; 0 1 0 0; 0 0 1 0; 0 0 0 1]). That will explicitly look for effects across any of your four seeds, so these analyses do not require additional multiple comparison correction. If you find a signifiant effect here, you can then perform post-hoc analyses to find out which of your 4 seeds might be driving the observed effect.
or 2) perform 4 separate analyses (one per seed) but now using a more conservative family-wise error threshold (e.g. use p-FWE<.05/4 to apply a bonferroni correction)
Hope this helps
Alfonso
Originally posted by Kaylah Curtis:
Regarding your first question, yes, the group-by-performance interaction effect in this case would appear to be driven by both of these effects (i.e. a significant difference in the association between performance and connectivity measures between groups, driven by positive associations in groupB and negative associations in groupA). The sizes of both the positive and negative associations seem to be comparable from your plots but it would be nice to quantify this explicitly (e.g. computing the regression coefficients and associated stats in each group separately)
Regarding your question about multiple comparisons, you are right that seed-to-voxel results are performed separately for each seed (so if you are testing multiple seeds you typically would need to apply and additional multiple comparison correction). The applicability of this additional correction is related to whether your 4 different analyses (one for each of your 4 seeds) are considered each testing a separate hypotheses (in that case you would not typically apply an additional correction), or whether those 4 different analyses are part of the same hypothesis (e.g. you were testing effects across any of those seeds, and in that case you would typically apply an additional correction). There are at least a couple of common ways to apply this additional correction if needed.
1) instead of performing 4 separate analyses (one per seed), perform a single analysis selecting all 4 seeds (and use a contrast eye(4) = [1 0 0 0; 0 1 0 0; 0 0 1 0; 0 0 0 1]). That will explicitly look for effects across any of your four seeds, so these analyses do not require additional multiple comparison correction. If you find a signifiant effect here, you can then perform post-hoc analyses to find out which of your 4 seeds might be driving the observed effect.
or 2) perform 4 separate analyses (one per seed) but now using a more conservative family-wise error threshold (e.g. use p-FWE<.05/4 to apply a bonferroni correction)
Hope this helps
Alfonso
Originally posted by Kaylah Curtis:
Hi Alfonso,
Thanks so much for your reply. It was very helpful. We went ahead and ran the associations separately, as you suggested. We believe we find a very strong positive correlation between performance and group B and a very strong negative association between performance and Group A (screenshots are attached). In this case would it be a correct interpretation to say that the negative contrast we find in the cluster is driven by both of these associations? Or would it be attributable more specifically to one than the other? We also included a screenshot of the main effect of the two.
Another question: we have read in the forum that the issue of multiple comparisons in still debated. In this case, we selected 4 seeds in a seed-to-voxel analysis, which we ran separately. We kept the default cluster threshold value of 0.05 and height threshold of 0.001 and considered only the FWE-corrected results. Would we need to correct again for multiple comparisons considering the 4 seeds? As in, FWE-corrected 0.05/4?
Thanks again for your help.
Kaylah
Thanks so much for your reply. It was very helpful. We went ahead and ran the associations separately, as you suggested. We believe we find a very strong positive correlation between performance and group B and a very strong negative association between performance and Group A (screenshots are attached). In this case would it be a correct interpretation to say that the negative contrast we find in the cluster is driven by both of these associations? Or would it be attributable more specifically to one than the other? We also included a screenshot of the main effect of the two.
Another question: we have read in the forum that the issue of multiple comparisons in still debated. In this case, we selected 4 seeds in a seed-to-voxel analysis, which we ran separately. We kept the default cluster threshold value of 0.05 and height threshold of 0.001 and considered only the FWE-corrected results. Would we need to correct again for multiple comparisons considering the 4 seeds? As in, FWE-corrected 0.05/4?
Thanks again for your help.
Kaylah
Threaded View
| Title | Author | Date |
|---|---|---|
| Kaylah Curtis | Jan 27, 2016 | |
| Alfonso Nieto-Castanon | Jan 29, 2016 | |
| Kaylah Curtis | Feb 3, 2016 | |
| Alfonso Nieto-Castanon | Feb 4, 2016 | |
| Kaylah Curtis | Feb 3, 2016 | |