help
help > RE: Calculator p-value Negative Contrast
Jan 29, 2016 04:01 PM | Alfonso Nieto-Castanon - Boston University
RE: Calculator p-value Negative Contrast
Hi Kaylah,
Your analyses look perfectly fine. In the Calculator section the p-values described also look as expected, if you have a strongly negative association between performance and connectivity measures then the positive-contrast one-sided p-value will be close to 1 and the negative-contrast one-sided p-value will be close to 0 (and in both cases the two-sided stats will be exactly the same and also close to 0). To clarify, if the p-value of a positive contrast one-sided analysis is p, then the p-value of the associated negative contrast one-sided analysis will always be 1-p, and the p-value of the two-sided analysis will always be 2*min(p,1-p). So the values that you report simpy reflect that you are in fact finding a strong negative association between performance and connectivity measures (after controlling for age/gender/depression). Having said that, you would not typically report or use these statistics in your case. The analyses that you are performing in the Calculator tab are considered post-hoc analyses (they are based on the suprathreshold clusters obtained in your original analyses) so those statistics are always biased. What you typically report is the threshold used in your primary analyses (e.g. you may have used a combination of voxel-level uncorrected p-values <.001 and cluster-level FDR-corrected p-values <.05), and then if you want you may display those associations (like in the plots you attached) but always warning that these are post-hoc analyses and may not reflect the degree of association that may be expected in the general population.
Regarding the specific interpretation of your results, the first results (association between performance and connectivity in one group) show positive connectivity values that decrease with performance in this cluster (i.e. lower-performance subjects show positive connectivity while higher-performance subjects show weaker or no connectivity). Your second set of analyses (looking at differential associations in the two subject groups) are a bit more difficult to interpret from those plots. I believe they represent a positive association between performance and connectivity in groupB while a negative association between performance and connectivity in groupA, but I would suggest to plot instead the contrasts [0 0 1 0 0 0 0] (association in groupA) and [0 0 0 1 0 0 0] (association in groupB) separately to more easily interpret those results. In general, the sign of the [0 0 1 -1 0 0 0] contrast results indicate higher associations in groupA compared to groupB for positive results, and higher associations in groupB compared to groupA for negative results but one really needs to plot those individual associations in order to appropriately interpret the results (e.g. higher association in groupA could represent a positive association that is stronger in groupA, but it could also represent a negative association that is weaker -i.e. less negative- in groupA)
Hope this helps
Alfonso
Originally posted by Kaylah Curtis:
Your analyses look perfectly fine. In the Calculator section the p-values described also look as expected, if you have a strongly negative association between performance and connectivity measures then the positive-contrast one-sided p-value will be close to 1 and the negative-contrast one-sided p-value will be close to 0 (and in both cases the two-sided stats will be exactly the same and also close to 0). To clarify, if the p-value of a positive contrast one-sided analysis is p, then the p-value of the associated negative contrast one-sided analysis will always be 1-p, and the p-value of the two-sided analysis will always be 2*min(p,1-p). So the values that you report simpy reflect that you are in fact finding a strong negative association between performance and connectivity measures (after controlling for age/gender/depression). Having said that, you would not typically report or use these statistics in your case. The analyses that you are performing in the Calculator tab are considered post-hoc analyses (they are based on the suprathreshold clusters obtained in your original analyses) so those statistics are always biased. What you typically report is the threshold used in your primary analyses (e.g. you may have used a combination of voxel-level uncorrected p-values <.001 and cluster-level FDR-corrected p-values <.05), and then if you want you may display those associations (like in the plots you attached) but always warning that these are post-hoc analyses and may not reflect the degree of association that may be expected in the general population.
Regarding the specific interpretation of your results, the first results (association between performance and connectivity in one group) show positive connectivity values that decrease with performance in this cluster (i.e. lower-performance subjects show positive connectivity while higher-performance subjects show weaker or no connectivity). Your second set of analyses (looking at differential associations in the two subject groups) are a bit more difficult to interpret from those plots. I believe they represent a positive association between performance and connectivity in groupB while a negative association between performance and connectivity in groupA, but I would suggest to plot instead the contrasts [0 0 1 0 0 0 0] (association in groupA) and [0 0 0 1 0 0 0] (association in groupB) separately to more easily interpret those results. In general, the sign of the [0 0 1 -1 0 0 0] contrast results indicate higher associations in groupA compared to groupB for positive results, and higher associations in groupB compared to groupA for negative results but one really needs to plot those individual associations in order to appropriately interpret the results (e.g. higher association in groupA could represent a positive association that is stronger in groupA, but it could also represent a negative association that is weaker -i.e. less negative- in groupA)
Hope this helps
Alfonso
Originally posted by Kaylah Curtis:
Hi Alfonso,
We have run a few analyses and want to be sure we are interpreting our results correctly.
First, we are testing the correlation between a performance in a group controlling for covariates in a seed-to-voxel analysis (group A, performance*A, age*A, gender*A, depression*A) entering the following contrast: [0 1 0 0 0]. In the negative-contrast results explorer we have a significant cluster. We then imported the cluster values to Calculator. We selected the same contrast just mentioned in Predictor Variables and the cluster values as the Outcome variable. For this selection the statistics output shows a p-value = 1.0000000 (two-sided p = 0.000000). But when we switch the contrast to [0 -1 0 0 0], the p-value becomes 0.000000 (two-sided p = 0.00000000). How should these results be interpreted? See the first two examples in the attachment.
We also did another analysis comparing group performances between Group A and Group B, controlling for covariates. We selected Group A, Group B, performance*A, performance*B, age*AB, gender*AB, depression*AB with contrast [0 0 1 -1 0 0 0]. In the results explorer we found a few significant clusters in negative contrast, which we imported to the calculator. Again, we selected the same contrast just mentioned for predictor variables and a cluster for outcome variable. We have the same problem as above, in which the p-value = 1.0000000 (two-sided p = 0.000000) with the original contrast, but becomes 0.000000 (two-sided p = 0.00000000) when we switch the contrast to [0 0 -1 1 0 0 0]. An additional question here is how should we interpret predictors with negative values on the resulting graph in this contrast? See the last two examples in the attachment.
Thanks for your help!
We have run a few analyses and want to be sure we are interpreting our results correctly.
First, we are testing the correlation between a performance in a group controlling for covariates in a seed-to-voxel analysis (group A, performance*A, age*A, gender*A, depression*A) entering the following contrast: [0 1 0 0 0]. In the negative-contrast results explorer we have a significant cluster. We then imported the cluster values to Calculator. We selected the same contrast just mentioned in Predictor Variables and the cluster values as the Outcome variable. For this selection the statistics output shows a p-value = 1.0000000 (two-sided p = 0.000000). But when we switch the contrast to [0 -1 0 0 0], the p-value becomes 0.000000 (two-sided p = 0.00000000). How should these results be interpreted? See the first two examples in the attachment.
We also did another analysis comparing group performances between Group A and Group B, controlling for covariates. We selected Group A, Group B, performance*A, performance*B, age*AB, gender*AB, depression*AB with contrast [0 0 1 -1 0 0 0]. In the results explorer we found a few significant clusters in negative contrast, which we imported to the calculator. Again, we selected the same contrast just mentioned for predictor variables and a cluster for outcome variable. We have the same problem as above, in which the p-value = 1.0000000 (two-sided p = 0.000000) with the original contrast, but becomes 0.000000 (two-sided p = 0.00000000) when we switch the contrast to [0 0 -1 1 0 0 0]. An additional question here is how should we interpret predictors with negative values on the resulting graph in this contrast? See the last two examples in the attachment.
Thanks for your help!
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 | |