I would like to test, on the group level, whether each one of my regressors was significant, but I'm struggling with defining the contrast vectors. I have defined 9 regressors using a condition file and pmod, and in my glm file there are now 21 regressors: regressor 1 * bf1, regressor 1 * bf2, regressor 2 * bf1 etc, and three constants at the end.
My first question: when I now define the contrast vectors (I specify them on conditions), and I want to test for the significance of each regressor separately, how long should this vector be? Because in the manual I read the following:
*"Contrasts are usually specified on parameter estimates. For GLM, you can also choose to specify them on conditions or reconstructed responses per condition. In this case, your contrast vector needs to take into account only the first basis function." *
I took this to mean that the second element of the contrast vector in GLM would refer to "regressor 2 x bf1" rather than "regressor 1 x bf2". In other words, the contrast vectors should have 9 elements, rather than 18.
My second question is whether, if I want to test for the significance of all my regressors, I should define different contrasts like this:
regressor 1: [1 0 0 0 0 0 0 0 0]
Regressor 2: [0 1 0 0 0 0 0 0 0]
etc.
I have carried out the GLM and second-level analysis and got a resulting bar graph with mean parameter estimates that does not look unreasonable to me, but I would like to make sure that I am talking about the contrasts I think I am talking about.
Many thanks in advance.
Best wishes,
Jesse
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
In fact, I have some doubts about my analysis and my data. These doubts are caused by the fact that the three main events in the trial had a negative effect on the SCR, according to my second level analysis (see attachment). How should I interpret this, and what is the difference between red and grey bars in this graph?
I went back into the raw and imported data (see screenshots) and these caused some more doubts. The extremely high amplitude regular peaks that you see in the raw data are caused by pulses of the fMRI scanner (these SCR data were recorded during an fMRI experiment). Should I have filtered these out before importing the data into PSPM?
In the "predicted and observed" plot, I see very regular oscillations in the observed data of around 2 Hz (see screenshots). My supervisor said this might also be an artifact, but we're not sure what could have caused these.
Lastly, I have added the .mat file containing the data for one participant, one session. Could you tell me whether these data are usable? I lack experience with SCR data, so I could very much use the help of an expert eye.
you are absolutely right about the contrast definition and strategy.
As you said, the negative parameter estimates on the group level don't make biological sense. Although they contain noise and can be negative, the noise should have zero mean, so on average they should be zero or larger - and certainly not significantly below zero.
The noise examples you show are quite considerable. By the eye I can only identify a biological signal in the 22.30.56 screenshot. GLM filters the data, but with that amount of noise, the butterworth filters will not work well.
I think the example .mat file contains several sources of noise. There are large spikes that look like gradient switching artefacts, does the frequency correspond to your slice TR? You may get these out with a median filter - this can be very slow, so play around with the number of data points over which to filter.
You could then address the higher-frequency, lower-amplitude noise with suitable filters. I don't know whether you will be able to retain enough signal, but we had good results with our MRI sequences.
As a general point, I'd advise putting SCR electrodes on the foot to reduce gradient noise (there is a sketch in Boucsein 2012, and we have successfully tested this against hand and fingers in Bach, Flandin, Friston, Dolan 2010).
Hope this helps
Dominik
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I just reran the analysis on median-filtered data. I even tried two different window sizes, and they filtered out the large peaks in the signal very nicely (see screenshot). However, I still get negative parameter estimates. I am not sure what to do next. It seems to me that the data is either not very usable, or that something went entirely wrong in the analysis. However, I have carried the analysis out many times now, and I don't know what I could do differently.
It looks like you nicely filtered out the artefacts. The question is whether you also got rid of the physiological signal. The model-based approach may not be ideal for that kind of quality assurance. I don't know your trial design, but perhaps you can extract responses to a significant event or several events and just average them. This may tell you whether there is anything left to analyse.
If yes, a reason for the negative parameter estimates may also be the within-trial design if some events occur in very quick succesion; or otherwise a mismatch of modelled and real responses.
Hope this helps
Dominik
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Dear Dominik,
I would like to test, on the group level, whether each one of my regressors was significant, but I'm struggling with defining the contrast vectors. I have defined 9 regressors using a condition file and pmod, and in my glm file there are now 21 regressors: regressor 1 * bf1, regressor 1 * bf2, regressor 2 * bf1 etc, and three constants at the end.
My first question: when I now define the contrast vectors (I specify them on conditions), and I want to test for the significance of each regressor separately, how long should this vector be? Because in the manual I read the following:
*"Contrasts are usually specified on parameter estimates. For GLM, you can also choose to specify them on conditions or reconstructed responses per condition. In this case, your contrast vector needs to take into account only the first basis function." *
I took this to mean that the second element of the contrast vector in GLM would refer to "regressor 2 x bf1" rather than "regressor 1 x bf2". In other words, the contrast vectors should have 9 elements, rather than 18.
My second question is whether, if I want to test for the significance of all my regressors, I should define different contrasts like this:
regressor 1:
[1 0 0 0 0 0 0 0 0]
Regressor 2:
[0 1 0 0 0 0 0 0 0]
etc.
I have carried out the GLM and second-level analysis and got a resulting bar graph with mean parameter estimates that does not look unreasonable to me, but I would like to make sure that I am talking about the contrasts I think I am talking about.
Many thanks in advance.
Best wishes,
Jesse
Dear Dominik,
In fact, I have some doubts about my analysis and my data. These doubts are caused by the fact that the three main events in the trial had a negative effect on the SCR, according to my second level analysis (see attachment). How should I interpret this, and what is the difference between red and grey bars in this graph?
I went back into the raw and imported data (see screenshots) and these caused some more doubts. The extremely high amplitude regular peaks that you see in the raw data are caused by pulses of the fMRI scanner (these SCR data were recorded during an fMRI experiment). Should I have filtered these out before importing the data into PSPM?
In the "predicted and observed" plot, I see very regular oscillations in the observed data of around 2 Hz (see screenshots). My supervisor said this might also be an artifact, but we're not sure what could have caused these.
Lastly, I have added the .mat file containing the data for one participant, one session. Could you tell me whether these data are usable? I lack experience with SCR data, so I could very much use the help of an expert eye.
With very best wishes, and many thanks,
Jesse
Last edit: Jesse Geerts 2016-02-01
Dear Jesse
you are absolutely right about the contrast definition and strategy.
As you said, the negative parameter estimates on the group level don't make biological sense. Although they contain noise and can be negative, the noise should have zero mean, so on average they should be zero or larger - and certainly not significantly below zero.
The noise examples you show are quite considerable. By the eye I can only identify a biological signal in the 22.30.56 screenshot. GLM filters the data, but with that amount of noise, the butterworth filters will not work well.
I think the example .mat file contains several sources of noise. There are large spikes that look like gradient switching artefacts, does the frequency correspond to your slice TR? You may get these out with a median filter - this can be very slow, so play around with the number of data points over which to filter.
You could then address the higher-frequency, lower-amplitude noise with suitable filters. I don't know whether you will be able to retain enough signal, but we had good results with our MRI sequences.
As a general point, I'd advise putting SCR electrodes on the foot to reduce gradient noise (there is a sketch in Boucsein 2012, and we have successfully tested this against hand and fingers in Bach, Flandin, Friston, Dolan 2010).
Hope this helps
Dominik
Dear Dominik,
Thanks for the advice. I am now going to median filter the data, try the analysis again and I will get back to you.
Best,
Jesse
Hi Dominik,
I just reran the analysis on median-filtered data. I even tried two different window sizes, and they filtered out the large peaks in the signal very nicely (see screenshot). However, I still get negative parameter estimates. I am not sure what to do next. It seems to me that the data is either not very usable, or that something went entirely wrong in the analysis. However, I have carried the analysis out many times now, and I don't know what I could do differently.
Best wishes,
Jesse
Hi Jesse.
It looks like you nicely filtered out the artefacts. The question is whether you also got rid of the physiological signal. The model-based approach may not be ideal for that kind of quality assurance. I don't know your trial design, but perhaps you can extract responses to a significant event or several events and just average them. This may tell you whether there is anything left to analyse.
If yes, a reason for the negative parameter estimates may also be the within-trial design if some events occur in very quick succesion; or otherwise a mismatch of modelled and real responses.
Hope this helps
Dominik