help > Task-based gPPI a priori ROI-to-ROI questions
Showing 1-3 of 3 posts
Feb 18, 2017 06:02 PM | Kristina Gelardi
Task-based gPPI a priori ROI-to-ROI questions
Hi Alfonso and others,
I have 125 participants and am using task-based gPPI ROI-to-ROI, with a priori ROIs of 1 seed and 10 targets. I have a number of questions related to my analyses. I'd like to ensure I have interpreted previous forum conversations correctly, and to ask a few questions that I cannot find answers to.
1) Contrasts in task-based gPPI: It seems based on previous forum conversations (e.g., https://www.nitrc.org/forum/forum.php?th...) that I need to have a contrast between my condition of interest (a particular face expression) and another condition (fixation is increasingly used as neutral faces are not always perceived as neutral) [e.g., 1 -1], rather than just having my condition of interest [1] versus CONN determined baseline. Is this accurate?
2) Accounting for multiple comparisons: In results explorer I see that my seed-level (F-test) is significant in both the un-corrected and FDR columns, however, 3 of my seed to ROIs don't survive FDR. I am trying to interpret a comment on the FAQ page (http://www.alfnie.com/software/conn#TOC-...), where you indicate "Similarly, for ROI-to-ROI analyses, typically the analysis results are considered appropriately corrected for multiple comparisons (across all seeds/ROIs) as long as at least one of either the height (connection-level) or the extent (seed- or network- level) thresholds uses an analysis-wise false positive control method (either FDR- or FWE- corrected p-values)." I am wondering if this means that I can use the FDR F-stat with the uncorrected connection-level for the z-scores between my seed and the 3 targets.
3) I'd like to get the mean activation level for each of my ROIs in addition to the z-scores between ROIs, is that possible in CONN? If so, how do I do it? If not, what do you suggest as the best way to accomplish this?
4) Exporting 1st level for mediation analyses: Lastly, for one of my analyses which will be completed in CONN that is easy to do. However, I have a second analysis for which I would like to export the 1st level data to Mplus so that I can use the brain as a mediator in a structural equation model. How do I calculate the contrast from the 1st level data? For example, how do I calculate the contrast of condition A, with the z-scores for my amygdala seed to each of my targets, versus condition B, with the same ROI z-scores?
I know these are a lot of questions and I very much appreciate your help. Thank you for all the work you do to support us all in our connectivity analyses!
Kristina
I have 125 participants and am using task-based gPPI ROI-to-ROI, with a priori ROIs of 1 seed and 10 targets. I have a number of questions related to my analyses. I'd like to ensure I have interpreted previous forum conversations correctly, and to ask a few questions that I cannot find answers to.
1) Contrasts in task-based gPPI: It seems based on previous forum conversations (e.g., https://www.nitrc.org/forum/forum.php?th...) that I need to have a contrast between my condition of interest (a particular face expression) and another condition (fixation is increasingly used as neutral faces are not always perceived as neutral) [e.g., 1 -1], rather than just having my condition of interest [1] versus CONN determined baseline. Is this accurate?
2) Accounting for multiple comparisons: In results explorer I see that my seed-level (F-test) is significant in both the un-corrected and FDR columns, however, 3 of my seed to ROIs don't survive FDR. I am trying to interpret a comment on the FAQ page (http://www.alfnie.com/software/conn#TOC-...), where you indicate "Similarly, for ROI-to-ROI analyses, typically the analysis results are considered appropriately corrected for multiple comparisons (across all seeds/ROIs) as long as at least one of either the height (connection-level) or the extent (seed- or network- level) thresholds uses an analysis-wise false positive control method (either FDR- or FWE- corrected p-values)." I am wondering if this means that I can use the FDR F-stat with the uncorrected connection-level for the z-scores between my seed and the 3 targets.
3) I'd like to get the mean activation level for each of my ROIs in addition to the z-scores between ROIs, is that possible in CONN? If so, how do I do it? If not, what do you suggest as the best way to accomplish this?
4) Exporting 1st level for mediation analyses: Lastly, for one of my analyses which will be completed in CONN that is easy to do. However, I have a second analysis for which I would like to export the 1st level data to Mplus so that I can use the brain as a mediator in a structural equation model. How do I calculate the contrast from the 1st level data? For example, how do I calculate the contrast of condition A, with the z-scores for my amygdala seed to each of my targets, versus condition B, with the same ROI z-scores?
I know these are a lot of questions and I very much appreciate your help. Thank you for all the work you do to support us all in our connectivity analyses!
Kristina
Feb 20, 2017 08:02 PM | Alfonso Nieto-Castanon - Boston University
RE: Task-based gPPI a priori ROI-to-ROI questions
Hi Kristina,
Some thoughts on your questions below
Best
Alfonso
Originally posted by Kristina Gelardi:
Not necessarily. This depends on your design as well as on which "task" conditions you select when defining your gPPI analyses. If your design is an event related design then the implicit baseline is typically not modeled as a condition and it represents the baseline activity in the absence of stimuli / experimental conditions. In this case you typically select the individual "task" conditions, and you can safely interpret the individual-condition effects (e.g. when selecting just the "task1" condition) as connectivity differences between this "task1" condition and the implicit baseline. If, on the other hand, you have a block design and you have 'task1', 'task2', and 'fixation' conditions (where fixation corresponds to the times between the task conditions, i.e. all times are modeled by one explicit condition), then you have two choices: 1) when defining your gPPI analyses you include all 'task1', 'task2', and 'fixation' conditions. In this case you should not look at the individual single-condition effects, but instead use between-condition contrasts to evaluate connectivity differences; and 2) when defining your gPPI analyses you include only your task conditions -'task1' and 'task2', but not 'fixation'-. In this case you can again look at the individual 'task1' or 'task2' single-condition effects, and those will already represent differences in connectivity between the corresponding task1 or task2 conditions and the implicit baseline (now 'fixation'). If you do have a block design and wish to be able to evalute "within-condition" connectivity effects (i.e. what is the strength of conenctivity during this condition block), then simply use weighted-GLM analyses for that instead (weighted-GLM and gPPI analyses are almost exactly equivalent in the case of long block designs). Hope this helps clarify
2) Accounting for multiple comparisons: In results explorer I see that my seed-level (F-test) is significant in both the un-corrected and FDR columns, however, 3 of my seed to ROIs don't survive FDR. I am trying to interpret a comment on the FAQ page (http://www.alfnie.com/software/conn#TOC-...), where you indicate "Similarly, for ROI-to-ROI analyses, typically the analysis results are considered appropriately corrected for multiple comparisons (across all seeds/ROIs) as long as at least one of either the height (connection-level) or the extent (seed- or network- level) thresholds uses an analysis-wise false positive control method (either FDR- or FWE- corrected p-values)." I am wondering if this means that I can use the FDR F-stat with the uncorrected connection-level for the z-scores between my seed and the 3 targets.
If you are selecting just a single seed, then there is no need to use any form of seed-level correction (uncorrected F-test values can be used as a omnibus test of any effect between this seed and any of the target ROIs). On the other hand, if you are simultaneously selecting multiple seeds and want to evalute whether there is any effect between any of these seeds and any of the target ROIs, then yes, you need to use FDR- or FWE- corrected values for that. If your F-stats survive FDR correction, then you can conclude that each of those seeds show an effect across some of the target ROIs (and you can then use uncorrected connection-level thresholds to effectively perform post-hoc analyses to learn which target ROIs may be responsible for these significant omnibus effects)
3) I'd like to get the mean activation level for each of my ROIs in addition to the z-scores between ROIs, is that possible in CONN? If so, how do I do it? If not, what do you suggest as the best way to accomplish this?
sorry but mean BOLD signal levels are disregarded after denoising (either by band-pass filtering, or regression or session-effects, or detrending). I would suggest to look in the ROI_Subject*_Session*.mat files (in your project conn_*/data folder) which contains the BOLD timeseries prior to denoising within each ROI (for each subject/session) and computing the average of those timeseries directly.
4) Exporting 1st level for mediation analyses: Lastly, for one of my analyses which will be completed in CONN that is easy to do. However, I have a second analysis for which I would like to export the 1st level data to Mplus so that I can use the brain as a mediator in a structural equation model. How do I calculate the contrast from the 1st level data? For example, how do I calculate the contrast of condition A, with the z-scores for my amygdala seed to each of my targets, versus condition B, with the same ROI z-scores?
If you use the 'import values' option, that will create a set of second-level covariates containing the average z-score values for each selected seed/target ROI pair and condition. You could then simply compute the between-condition differences directly within CONN (e.g. if you have two covariates named 'z_AB1' and 'z_AB2' representing the z-scores between regions A and B during condition 1 and 2, respectively, you can simply define a new covariate in Setup.Covariates.SecondLevel, name it 'dz_AB' and enter in the 'values' field the string "z_AB2-z_AB1" without quotes to compute those differences). You can then export those covariates into a .txt or .csv file that Mplus may read.
Hope this helps
Alfonso
Some thoughts on your questions below
Best
Alfonso
Originally posted by Kristina Gelardi:
Hi Alfonso and others,
I have 125 participants and am using task-based gPPI ROI-to-ROI, with a priori ROIs of 1 seed and 10 targets. I have a number of questions related to my analyses. I'd like to ensure I have interpreted previous forum conversations correctly, and to ask a few questions that I cannot find answers to.
1) Contrasts in task-based gPPI: It seems based on previous forum conversations (e.g., https://www.nitrc.org/forum/forum.php?th...) that I need to have a contrast between my condition of interest (a particular face expression) and another condition (fixation is increasingly used as neutral faces are not always perceived as neutral) [e.g., 1 -1], rather than just having my condition of interest [1] versus CONN determined baseline. Is this accurate?
I have 125 participants and am using task-based gPPI ROI-to-ROI, with a priori ROIs of 1 seed and 10 targets. I have a number of questions related to my analyses. I'd like to ensure I have interpreted previous forum conversations correctly, and to ask a few questions that I cannot find answers to.
1) Contrasts in task-based gPPI: It seems based on previous forum conversations (e.g., https://www.nitrc.org/forum/forum.php?th...) that I need to have a contrast between my condition of interest (a particular face expression) and another condition (fixation is increasingly used as neutral faces are not always perceived as neutral) [e.g., 1 -1], rather than just having my condition of interest [1] versus CONN determined baseline. Is this accurate?
Not necessarily. This depends on your design as well as on which "task" conditions you select when defining your gPPI analyses. If your design is an event related design then the implicit baseline is typically not modeled as a condition and it represents the baseline activity in the absence of stimuli / experimental conditions. In this case you typically select the individual "task" conditions, and you can safely interpret the individual-condition effects (e.g. when selecting just the "task1" condition) as connectivity differences between this "task1" condition and the implicit baseline. If, on the other hand, you have a block design and you have 'task1', 'task2', and 'fixation' conditions (where fixation corresponds to the times between the task conditions, i.e. all times are modeled by one explicit condition), then you have two choices: 1) when defining your gPPI analyses you include all 'task1', 'task2', and 'fixation' conditions. In this case you should not look at the individual single-condition effects, but instead use between-condition contrasts to evaluate connectivity differences; and 2) when defining your gPPI analyses you include only your task conditions -'task1' and 'task2', but not 'fixation'-. In this case you can again look at the individual 'task1' or 'task2' single-condition effects, and those will already represent differences in connectivity between the corresponding task1 or task2 conditions and the implicit baseline (now 'fixation'). If you do have a block design and wish to be able to evalute "within-condition" connectivity effects (i.e. what is the strength of conenctivity during this condition block), then simply use weighted-GLM analyses for that instead (weighted-GLM and gPPI analyses are almost exactly equivalent in the case of long block designs). Hope this helps clarify
2) Accounting for multiple comparisons: In results explorer I see that my seed-level (F-test) is significant in both the un-corrected and FDR columns, however, 3 of my seed to ROIs don't survive FDR. I am trying to interpret a comment on the FAQ page (http://www.alfnie.com/software/conn#TOC-...), where you indicate "Similarly, for ROI-to-ROI analyses, typically the analysis results are considered appropriately corrected for multiple comparisons (across all seeds/ROIs) as long as at least one of either the height (connection-level) or the extent (seed- or network- level) thresholds uses an analysis-wise false positive control method (either FDR- or FWE- corrected p-values)." I am wondering if this means that I can use the FDR F-stat with the uncorrected connection-level for the z-scores between my seed and the 3 targets.
If you are selecting just a single seed, then there is no need to use any form of seed-level correction (uncorrected F-test values can be used as a omnibus test of any effect between this seed and any of the target ROIs). On the other hand, if you are simultaneously selecting multiple seeds and want to evalute whether there is any effect between any of these seeds and any of the target ROIs, then yes, you need to use FDR- or FWE- corrected values for that. If your F-stats survive FDR correction, then you can conclude that each of those seeds show an effect across some of the target ROIs (and you can then use uncorrected connection-level thresholds to effectively perform post-hoc analyses to learn which target ROIs may be responsible for these significant omnibus effects)
3) I'd like to get the mean activation level for each of my ROIs in addition to the z-scores between ROIs, is that possible in CONN? If so, how do I do it? If not, what do you suggest as the best way to accomplish this?
sorry but mean BOLD signal levels are disregarded after denoising (either by band-pass filtering, or regression or session-effects, or detrending). I would suggest to look in the ROI_Subject*_Session*.mat files (in your project conn_*/data folder) which contains the BOLD timeseries prior to denoising within each ROI (for each subject/session) and computing the average of those timeseries directly.
4) Exporting 1st level for mediation analyses: Lastly, for one of my analyses which will be completed in CONN that is easy to do. However, I have a second analysis for which I would like to export the 1st level data to Mplus so that I can use the brain as a mediator in a structural equation model. How do I calculate the contrast from the 1st level data? For example, how do I calculate the contrast of condition A, with the z-scores for my amygdala seed to each of my targets, versus condition B, with the same ROI z-scores?
If you use the 'import values' option, that will create a set of second-level covariates containing the average z-score values for each selected seed/target ROI pair and condition. You could then simply compute the between-condition differences directly within CONN (e.g. if you have two covariates named 'z_AB1' and 'z_AB2' representing the z-scores between regions A and B during condition 1 and 2, respectively, you can simply define a new covariate in Setup.Covariates.SecondLevel, name it 'dz_AB' and enter in the 'values' field the string "z_AB2-z_AB1" without quotes to compute those differences). You can then export those covariates into a .txt or .csv file that Mplus may read.
Hope this helps
Alfonso
Feb 20, 2017 08:02 PM | Kristina Gelardi
RE: Task-based gPPI a priori ROI-to-ROI questions
Thank you so much Alfonso for your help. I really appreciate your
thorough feedback and for answering my many questions.
Kristina
Kristina