I post this question for some statistical discussion on the nature of PPI (psychophysical interaction analysis).
During PPI analysis, the task regressor was also entered as a covariate to tease apart the connectivity originating co-activation during task.
The task regressor was modeled using hymodynamic functions such as GAMMA.
As the hymodynamic function can not fit our brain response very well, the task regressor can not explain all the co-activation variance.
Thus, the PPI we got may still be a by-product of task-coactivation.
So how about do the PPI in two steps?
First, we use a TENT function to best fit the activation in every TR during a trial to best explain the variance induced by task condition and all other covariates.
Second, we use the residual time-course in the first step to do PPI analysis.
Of course, statistically, the two step can be done within one GLM.
The main point is the necessity of using TENT to best explain the task-induced activation.
I would appreciate if you can shed some lights to me.
That seems reasonable, except it leads to the question,
why you are using GAM instead of TENT in the first place?
We generally suggests applying PPI on top of the current
analysis, but not changing the analysis to generate a PPI
So might it be reasonable to use TENT in your original
Thanks very much for your reply!
I am doing some analysis on the classical false-belief localizer task. In the task, false belief question (10s) and the answer phase (4s), false photo question (10s) and the answer phase (4s) was separated by 12 seconds fixation. The order of false belief and false photo condition was randomized, of course.
In the initial analysis, I used a GAM to get the contrast of belief-photo. Some researchers used a BOXCAR function instead of a GAM. However, I found that the GAM give me best fit in my sample.
And then I run a PPI analysis by adding the PPI terms to the orignial GLM.
Although I get some interesting findings, I am anxious that what I found might be a residual of task-coactivation.
So I come here and ask the question to make sure if it is preferable to fit all the task-relevant variance before doing PPI analysis.
I think using TENT can best capture the task-relevant activations.
If I understand it correctly, you agree with my point that it should be better to use TENT to explain all task-relevant variance as we can, at least in the case of a classical event-related design.
I am not sure about the block design and the rapid event-related design. Theoretically, we can also use TENT, right?
thanks very much!
If the events are 10s and 4s long, the GAM functions will
do a very poor job of modeling them (unless you use a
boxcar-convolved version of GAM). So your basis functions
should be either:
GAM(8.6,0.547,10) and GAM(8.6,0.547,4)
BLOCK(10) and BLOCK(4)
or maybe UBLOCK instead of BLOCK
or, if question and answer are never compared, it would
be fine to use BLOCK(10,1) and BLOCK(4,1).
It seems like many options, but if question and answer
are never compared, the BLOCK functions would all lead
to identical group stats, with only a different scaling
of the betas.
The I would say not to use TENT, but to replace GAM with
something more appropriate. TENT would be more work,
and would only be appropriate if you expect shape differences
across subjects or the brain (or if these basis functions should
fail to model the main response, too).
Thanks very much!
A recent paper (Di et al., 2017, human brain mapping 38:1723–1740) states that “imperfect decovolution may introduce spurious psychophysical interactions and how to avoid it”.
I noticed that you are one of the authors of this paper.
I feel that the centering of psychological variable was mean to minimize the correlation between the main effect term and the PPI term in the GLM model.
The best method I think , also suggested in the paper, is “… In addition, AFNI suggests removing psychological effects from the physiological variable before calculating PPI, which is an effective step to minimize collinearity between the PPI term and main effects. …”
(1) I think the procedure (removing psychological effects from physiological variable) should be done in every PPI analysis. Centering the psychological variable is seems like to hide the problem by changing the meaning of the task regressor. I suggest using a full model to explain all task related variance and noise variance such as head emotion in the seed time course and then using the residual timecourse as the physiological variable.
(2) Another thought is it might be unnecessary to use deconvolution-convolution approach because the difference between the two are trivial, at least in the case of a event related design with long trial duration.
What’s your opinion?