help
help > RE: gPPI after denoising the Effect of the conditions
Apr 19, 2019 12:04 AM | Pedro Valdes-Hernandez - University of Florida
RE: gPPI after denoising the Effect of the conditions
Thank you Alfonso,
I guess my suspicions were correct.
I apologize for some typos I noted after I re-read the post. Especially the second term -beta2i*conv(hrf,stim2) which should be +beta2i*conv(hrf,stim2).
I guess then that, when doing the second level analysis of the gPPI results, the contrasts on the conditions, e.g. con1-cond2 will be contrast functions of the corresponding PPI interactions betas
In contrast, for the GLM, contrasts are functions of the betas estimated for each condition estimated from their concatenated signal intervals.
Which theoretical differences would you expect between these two appraches?
Originally posted by Alfonso Nieto-Castanon:
I guess my suspicions were correct.
I apologize for some typos I noted after I re-read the post. Especially the second term -beta2i*conv(hrf,stim2) which should be +beta2i*conv(hrf,stim2).
I guess then that, when doing the second level analysis of the gPPI results, the contrasts on the conditions, e.g. con1-cond2 will be contrast functions of the corresponding PPI interactions betas
In contrast, for the GLM, contrasts are functions of the betas estimated for each condition estimated from their concatenated signal intervals.
Which theoretical differences would you expect between these two appraches?
Originally posted by Alfonso Nieto-Castanon:
Hi
Pedro,
That is a good question. Your description of both effects (the effect of including the task effects during denoising and the equal-term part of the gPPI first-level model step) is perfectly accurate. The net effect is that, at least for gPPI analyses, whether one includes or not the task effects as part of denoising is irrelevant, as those main effects will effectively be removed either way during the gPPI first-level analysis step and both approaches will lead to exactly the same gPPI interaction estimates (beta#ik in your equations, which are the values that are then passed to the second-level analysis step). Depending on the approach (whether you include or not the task effects during denoising) the gPPI main effects of conditions (beta#i in your equations) will, of course, be different, but that is fine because those estimates are simply not used in CONN other as a way to ensure that we are controlling the interaction term estimate for typically-correlated main condition effects.
Hope this helps
Alfonso
Originally posted by Pedro Valdes-Hernandez:
That is a good question. Your description of both effects (the effect of including the task effects during denoising and the equal-term part of the gPPI first-level model step) is perfectly accurate. The net effect is that, at least for gPPI analyses, whether one includes or not the task effects as part of denoising is irrelevant, as those main effects will effectively be removed either way during the gPPI first-level analysis step and both approaches will lead to exactly the same gPPI interaction estimates (beta#ik in your equations, which are the values that are then passed to the second-level analysis step). Depending on the approach (whether you include or not the task effects during denoising) the gPPI main effects of conditions (beta#i in your equations) will, of course, be different, but that is fine because those estimates are simply not used in CONN other as a way to ensure that we are controlling the interaction term estimate for typically-correlated main condition effects.
Hope this helps
Alfonso
Originally posted by Pedro Valdes-Hernandez:
Hi CONN experts,
Suppose I have a set of conditions, say rest, stim1 and stim2
I've been wondering if it is correct to do gPPI using the task conditions after having used the Effect of stim1 and Effect of stim2 as confounds in the denoising step.
The way I see it, the Effect of these confounds are regressed out from the BOLD signal in an i-th region/voxel, by estimating:
yi = yi'+beta1i*conv(hrf,stim1)-beta2i*conv(hrf,stim2)
where yi' is the denoised signal
On the other hand, gPPI estimates the betas of the following model, given the target and seed regions/voxels i and k, respectively
yi' = beta1ik*conv(hrf,stim1)*yk'+beta2*conv(hrf,stim2)*yk'+ (PPI interactions)
beta1i*conv(hrf,stim1)+beta2i*conv(hrf,stim2)+ (main effect of conditions)
betak*yk' (main effect of seed)
which may seem to be controlling for the effect of the conditions for the second time.
Is this correct? Is so, is it acceptable? Is it irrelevant, i.e. after denoising, the main effect of the conditions in the gPPI model will not be significant (estimates beta1i=beta2i=0)? Or should I denoise without using the Effects of the conditions if gPPI is intended?
Thank you!
Suppose I have a set of conditions, say rest, stim1 and stim2
I've been wondering if it is correct to do gPPI using the task conditions after having used the Effect of stim1 and Effect of stim2 as confounds in the denoising step.
The way I see it, the Effect of these confounds are regressed out from the BOLD signal in an i-th region/voxel, by estimating:
yi = yi'+beta1i*conv(hrf,stim1)-beta2i*conv(hrf,stim2)
where yi' is the denoised signal
On the other hand, gPPI estimates the betas of the following model, given the target and seed regions/voxels i and k, respectively
yi' = beta1ik*conv(hrf,stim1)*yk'+beta2*conv(hrf,stim2)*yk'+ (PPI interactions)
beta1i*conv(hrf,stim1)+beta2i*conv(hrf,stim2)+ (main effect of conditions)
betak*yk' (main effect of seed)
which may seem to be controlling for the effect of the conditions for the second time.
Is this correct? Is so, is it acceptable? Is it irrelevant, i.e. after denoising, the main effect of the conditions in the gPPI model will not be significant (estimates beta1i=beta2i=0)? Or should I denoise without using the Effects of the conditions if gPPI is intended?
Thank you!
Threaded View
| Title | Author | Date |
|---|---|---|
| Pedro Valdes-Hernandez | Jan 9, 2019 | |
| Alfonso Nieto-Castanon | Apr 10, 2019 | |
| Pedro Valdes-Hernandez | Apr 19, 2019 | |
| Pedro Valdes-Hernandez | Jan 10, 2019 | |