I was wondering if you ever came across negative betas at the second level in the GLM model?
I have a set of results that seems to make relatively sense, for example, there is a difference in between acquisition vs extinction, safe cue vs threatening cue, high anxious vs low anxious, etc, however the direction doesn't always make sense, most betas tend to be negative.
This is a fear conditioning study with shocks and relatively short ISIs (1-2s) and ITIs (2-3s) and visual inspection of the data shows that a SCR elicited by a shock on previous trial interferes with the next trial (even on a grand average plot and after jitter). So, I am worried that since there is a trough at the cue onset the GLM estimates a negative beta. I model all events and the actual shock time has 0.25s pre and 0.75 post removed.
The first level contrasts were all mean per condition, e.g. [1 0 0 1 0 0] for the intercept of 1st and 4th regressor - is that ok or should the contrast sum to 0?
More conceptually, what sort of data would produce negative betas in this case?
Thank you,
Ondrej
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
consistently negative SCR are physiologically implausible and so I think it is always a good idea to check what is wrong. They are more often encountered in standard (not model-based) analysis, when a new SCR occurs during the decay of the preceding one; in PsPM they usually only occur when the timing is wrong.
However, in your case it is true that any analysis method will have difficulties disentangling SCR to stimuli with less than 2 s SOA. There is a biophysical limit here - the system becomes strongly non-linear below 0.6 Hz (Gerster et al. 2018 Psychophysiology).
You may still be able to interpret condition differences on an ordinal scale, though.
Best
Dominik
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
thank you for your answer. I did a little digging and GLM results for another data set with larger ITI (3-5 s) and ISI (2-4s) seems to have the same issue: also negative betas but, interestingly, consistent results with study 1 (it's the same task). While I tripple checked the timings previously I went back to try and find any bug. The onset files seem fine but when I plot the events via PSPM (GLM review) the number of events doesn't match (there are way less events, approx. 1/4).
I plotted my raw timings and PSPM output in the attached timing.mlx file. Are the PSPM-plotted events the ones used in the regression? If so, then that could explain the issue with negative betas.
I attach one session of one participant (onset+data, and the live script mentioned above).
Dear Ondrej
The reason not all events were shown, when reviewing the glm model, was because of a bug in the 'review model' functionality. To fix this bug you need to replace line 146 in pspm_rev_glm: m = floor((j/cl);
with
m = floor((j-0.1)/cl);
The events used in the regression are from the onsets file, i.e., all events, and not only the events in your plot.
Hope this helps a little.
Best,
Laure
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
thank you. So, the glm got estimated using all regressors and it's only the visualisation that's inaccurate. Therefore the negative betas aren't likely a result wrong timing file.
Dominik, by ordinal differences do you mean that I can look at the results using non-parametric statistics?
Thank you,
Ondrej
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
If there is no error, and the stimulus frequency is not below 0.5 per second, then I'd think the assumptions of the evoked model are not fulfilled because (a) either SCR have variable onset, as is common in fear conditioning or during long (> 1 s) stimulus presentations, and/or (b) the latency is different because of anticipation (resulting in shorter latency) or prolonged stimulus presentation (resulting in longer latency wrt stimulus onset).
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I was wondering if you ever came across negative betas at the second level in the GLM model?
I have a set of results that seems to make relatively sense, for example, there is a difference in between acquisition vs extinction, safe cue vs threatening cue, high anxious vs low anxious, etc, however the direction doesn't always make sense, most betas tend to be negative.
This is a fear conditioning study with shocks and relatively short ISIs (1-2s) and ITIs (2-3s) and visual inspection of the data shows that a SCR elicited by a shock on previous trial interferes with the next trial (even on a grand average plot and after jitter). So, I am worried that since there is a trough at the cue onset the GLM estimates a negative beta. I model all events and the actual shock time has 0.25s pre and 0.75 post removed.
The first level contrasts were all mean per condition, e.g. [1 0 0 1 0 0] for the intercept of 1st and 4th regressor - is that ok or should the contrast sum to 0?
More conceptually, what sort of data would produce negative betas in this case?
Thank you,
Ondrej
Hi Ondrej
consistently negative SCR are physiologically implausible and so I think it is always a good idea to check what is wrong. They are more often encountered in standard (not model-based) analysis, when a new SCR occurs during the decay of the preceding one; in PsPM they usually only occur when the timing is wrong.
However, in your case it is true that any analysis method will have difficulties disentangling SCR to stimuli with less than 2 s SOA. There is a biophysical limit here - the system becomes strongly non-linear below 0.6 Hz (Gerster et al. 2018 Psychophysiology).
You may still be able to interpret condition differences on an ordinal scale, though.
Best
Dominik
Dear Dominik,
thank you for your answer. I did a little digging and GLM results for another data set with larger ITI (3-5 s) and ISI (2-4s) seems to have the same issue: also negative betas but, interestingly, consistent results with study 1 (it's the same task). While I tripple checked the timings previously I went back to try and find any bug. The onset files seem fine but when I plot the events via PSPM (GLM review) the number of events doesn't match (there are way less events, approx. 1/4).
I plotted my raw timings and PSPM output in the attached timing.mlx file. Are the PSPM-plotted events the ones used in the regression? If so, then that could explain the issue with negative betas.
I attach one session of one participant (onset+data, and the live script mentioned above).
Best wishes,
Ondrej
Last edit: Ondrej Zika 2019-04-02
Dear Ondrej
The reason not all events were shown, when reviewing the glm model, was because of a bug in the 'review model' functionality. To fix this bug you need to replace line 146 in pspm_rev_glm: m = floor((j/cl);
with
m = floor((j-0.1)/cl);
The events used in the regression are from the onsets file, i.e., all events, and not only the events in your plot.
Hope this helps a little.
Best,
Laure
Dear Laure and Dominik,
thank you. So, the glm got estimated using all regressors and it's only the visualisation that's inaccurate. Therefore the negative betas aren't likely a result wrong timing file.
Dominik, by ordinal differences do you mean that I can look at the results using non-parametric statistics?
Thank you,
Ondrej
If there is no error, and the stimulus frequency is not below 0.5 per second, then I'd think the assumptions of the evoked model are not fulfilled because (a) either SCR have variable onset, as is common in fear conditioning or during long (> 1 s) stimulus presentations, and/or (b) the latency is different because of anticipation (resulting in shorter latency) or prolonged stimulus presentation (resulting in longer latency wrt stimulus onset).
Thank you. If they have variable onset would estimating betas for each trial be a solution?