help > Task-related conditions as confound
Showing 1-7 of 7 posts
Jul 9, 2018 06:07 PM | Santiago Alcaide
Task-related conditions as confound
Hello experts and thanks in advance for your help.
My question is related to the Denoising step. Should i use my task-related conditions as confounds to be removed -regressed out?
I understand that by including them, conn will be removing task effects from the BOLD signal before computing connectivity measures to avoid task-related coactivations that can be misleading when being interpreted as region correlated activations. I am now quite confused, how can I explore task-related connectivity if i'm removing task-effects from the signal?
Thank you very much.
Santiago
My question is related to the Denoising step. Should i use my task-related conditions as confounds to be removed -regressed out?
I understand that by including them, conn will be removing task effects from the BOLD signal before computing connectivity measures to avoid task-related coactivations that can be misleading when being interpreted as region correlated activations. I am now quite confused, how can I explore task-related connectivity if i'm removing task-effects from the signal?
Thank you very much.
Santiago
Jul 9, 2018 07:07 PM | wzhong
RE: Task-related conditions as confound
I believe this post can help address your question: https://www.nitrc.org/forum/message.php?...
Jul 13, 2018 04:07 AM | Santiago Alcaide
RE: Task-related conditions as confound
Thank you for your answer wzhong. I don't seem to get it, how
to account for task-effect if they're regressed out? I'm currently
interested in task related connectivity, and connectivity
difference between diferent tasks.
https://www.nitrc.org/forum/message.php?...
In that message, Alfonso states
"This form of task-related responses are more readily identified using standard functional task-related analyses, so when analyzing condition-specific functional connectivity it is more common to focus on the remaining functional connectivity effects (not purely driven by common main task effect)"
Conn toolbox main paper identifies BOLD signal to be processed as the residual of regression model fit.
Must remaining signal be orthogonal to confound+noise list? Should that be interpreted? I can't understand where are task-effects expressed, in both conceptual and mathematical senses.
I hope i'm making myself clear, and thanks for your patience.
Santiago
https://www.nitrc.org/forum/message.php?...
In that message, Alfonso states
"This form of task-related responses are more readily identified using standard functional task-related analyses, so when analyzing condition-specific functional connectivity it is more common to focus on the remaining functional connectivity effects (not purely driven by common main task effect)"
Conn toolbox main paper identifies BOLD signal to be processed as the residual of regression model fit.
Must remaining signal be orthogonal to confound+noise list? Should that be interpreted? I can't understand where are task-effects expressed, in both conceptual and mathematical senses.
I hope i'm making myself clear, and thanks for your patience.
Santiago
Jul 13, 2018 01:07 PM | Alfonso Nieto-Castanon - Boston University
RE: Task-related conditions as confound
Hi Santiago,
Perhaps the gPPI framework helps to better understand the two different types of "task-effects" involved here:
1) the first "task-effect" in gPPI is the "main psychological effect". That refers to the main effect of the your task on the BOLD signal of a target region, i.e. changes in average BOLD signal in a target region covarying with the presence/absence of a given experimental task or condition over time
2) the second "task-effect" in gPPI is the "Pyscho-Physiological interaction effect". That refers to the interaction of the task and seed BOLD timeseries on the target region BOLD signal; i.e. changes in effective connectivity between two regions covarying with the presence/absence of a given task or condition over time
In functional connectivity analyses we typically care about (2), changes in connectivity between two regions covarying with an experimental task, while in functional activation analyses, one typically cares about (1), changes in average BOLD signal (activation) covarying with an experimental task. In CONN the "task-effects" in (1) are removed from the BOLD timeseries as part of the Denoising procedure, by means of including the "effect of task" regressors in the list of potential confounding effects. The "task-effects" in (2) can be investigated in CONN either by using gPPI or weighted GLMs, where both of these approaches allow you to study differences in connectivity between tasks/conditions (see http://www.conn-toolbox.org/measures/see... for additional math details about these approaches).
Hope this helps
Alfonso
Originally posted by Santiago Alcaide:
Perhaps the gPPI framework helps to better understand the two different types of "task-effects" involved here:
1) the first "task-effect" in gPPI is the "main psychological effect". That refers to the main effect of the your task on the BOLD signal of a target region, i.e. changes in average BOLD signal in a target region covarying with the presence/absence of a given experimental task or condition over time
2) the second "task-effect" in gPPI is the "Pyscho-Physiological interaction effect". That refers to the interaction of the task and seed BOLD timeseries on the target region BOLD signal; i.e. changes in effective connectivity between two regions covarying with the presence/absence of a given task or condition over time
In functional connectivity analyses we typically care about (2), changes in connectivity between two regions covarying with an experimental task, while in functional activation analyses, one typically cares about (1), changes in average BOLD signal (activation) covarying with an experimental task. In CONN the "task-effects" in (1) are removed from the BOLD timeseries as part of the Denoising procedure, by means of including the "effect of task" regressors in the list of potential confounding effects. The "task-effects" in (2) can be investigated in CONN either by using gPPI or weighted GLMs, where both of these approaches allow you to study differences in connectivity between tasks/conditions (see http://www.conn-toolbox.org/measures/see... for additional math details about these approaches).
Hope this helps
Alfonso
Originally posted by Santiago Alcaide:
Thank you for your answer wzhong. I don't
seem to get it, how to account for task-effect if they're regressed
out? I'm currently interested in task related connectivity, and
connectivity difference between diferent tasks.
https://www.nitrc.org/forum/message.php?...
In that message, Alfonso states
"This form of task-related responses are more readily identified using standard functional task-related analyses, so when analyzing condition-specific functional connectivity it is more common to focus on the remaining functional connectivity effects (not purely driven by common main task effect)"
Conn toolbox main paper identifies BOLD signal to be processed as the residual of regression model fit.
Must remaining signal be orthogonal to confound+noise list? Should that be interpreted? I can't understand where are task-effects expressed, in both conceptual and mathematical senses.
I hope i'm making myself clear, and thanks for your patience.
Santiago
https://www.nitrc.org/forum/message.php?...
In that message, Alfonso states
"This form of task-related responses are more readily identified using standard functional task-related analyses, so when analyzing condition-specific functional connectivity it is more common to focus on the remaining functional connectivity effects (not purely driven by common main task effect)"
Conn toolbox main paper identifies BOLD signal to be processed as the residual of regression model fit.
Must remaining signal be orthogonal to confound+noise list? Should that be interpreted? I can't understand where are task-effects expressed, in both conceptual and mathematical senses.
I hope i'm making myself clear, and thanks for your patience.
Santiago
Jul 13, 2018 10:07 PM | brainconn
RE: Task-related conditions as confound
Hi Alfonso,
Sorry for the naive question, but it might be related to this topic. I understand that gPPI is based on regression, but how exactly does it compare to the difference in bivariate correlations between two conditions. Let's say I want to compute functional connectivity between region X and Y and have 2 conditions (A and B). In conn, I would select "bivariate correlation" and in the contrast manager type [1 -1] for condition A and B respectively. This will basically compare the slopes between condition A and B that represent the correlation between region X and Y. Isn't this a kind of basic PPI?
Thanks!
Ben
Sorry for the naive question, but it might be related to this topic. I understand that gPPI is based on regression, but how exactly does it compare to the difference in bivariate correlations between two conditions. Let's say I want to compute functional connectivity between region X and Y and have 2 conditions (A and B). In conn, I would select "bivariate correlation" and in the contrast manager type [1 -1] for condition A and B respectively. This will basically compare the slopes between condition A and B that represent the correlation between region X and Y. Isn't this a kind of basic PPI?
Thanks!
Ben
Jul 14, 2018 11:07 AM | Alfonso Nieto-Castanon - Boston University
RE: Task-related conditions as confound
Hi Ben,
That is a good question, and yes you are right that those analyses attempt to quantify the same construct (differences in connectivity between conditions) and in fact are both conceptually and numerically very similar. If I recall correctly the relationship holds that for a block design with large-enough block-lengths the two analyses should lead to exactly the same results when both are using bivariate regression measures.
When one of them uses correlation measures instead there will be some subtle differences induced by the different scaling factors across the two types of measures, so the relationship is not as straightforward (e.g. you could have two conditions that have the same regression but different correlation connectivity measures, and vice versa). Last, when the design is event-related then typically weighted GLM bivariate correlation measures become less robust (as they start using more -but considerably shorter- segments of the BOLD timeseries) and gPPI is typically preferred in those cases as it will better handle/model the increased interaction between adjacent events.
Just for reference, the math details of all of these models can be found in www.conn-toolbox.org/measures
Hope this helps
Alfonso
Originally posted by brainconn:
That is a good question, and yes you are right that those analyses attempt to quantify the same construct (differences in connectivity between conditions) and in fact are both conceptually and numerically very similar. If I recall correctly the relationship holds that for a block design with large-enough block-lengths the two analyses should lead to exactly the same results when both are using bivariate regression measures.
When one of them uses correlation measures instead there will be some subtle differences induced by the different scaling factors across the two types of measures, so the relationship is not as straightforward (e.g. you could have two conditions that have the same regression but different correlation connectivity measures, and vice versa). Last, when the design is event-related then typically weighted GLM bivariate correlation measures become less robust (as they start using more -but considerably shorter- segments of the BOLD timeseries) and gPPI is typically preferred in those cases as it will better handle/model the increased interaction between adjacent events.
Just for reference, the math details of all of these models can be found in www.conn-toolbox.org/measures
Hope this helps
Alfonso
Originally posted by brainconn:
Hi Alfonso,
Sorry for the naive question, but it might be related to this topic. I understand that gPPI is based on regression, but how exactly does it compare to the difference in bivariate correlations between two conditions. Let's say I want to compute functional connectivity between region X and Y and have 2 conditions (A and B). In conn, I would select "bivariate correlation" and in the contrast manager type [1 -1] for condition A and B respectively. This will basically compare the slopes between condition A and B that represent the correlation between region X and Y. Isn't this a kind of basic PPI?
Thanks!
Ben
Sorry for the naive question, but it might be related to this topic. I understand that gPPI is based on regression, but how exactly does it compare to the difference in bivariate correlations between two conditions. Let's say I want to compute functional connectivity between region X and Y and have 2 conditions (A and B). In conn, I would select "bivariate correlation" and in the contrast manager type [1 -1] for condition A and B respectively. This will basically compare the slopes between condition A and B that represent the correlation between region X and Y. Isn't this a kind of basic PPI?
Thanks!
Ben
Jul 17, 2018 12:07 AM | brainconn
RE: Task-related conditions as confound
Thanks Alfonso, that was very helpful.
Best,
Ben
Best,
Ben