Regression-based baseline correction

💬 Support & Discussions
1

Hi there,

I hope you are all well :slight_smile: I am using regression-based baseline correction and, as I have a couple of questions about my results, I was wondering if someone might have been able to help double-check my interpretation of the data.

I wish you a nice day!
Sabia

2

Hello Sabia,

can you post the questions including plots? I think other users would be interested as well and therefore it migh be useful if we track and record for others.

Best,

Carina

2 Likes
3

Hi Carina,

Thank you so much for your reply! Sure, I think this is a great idea but I apologise in advance for my silly questions! :slight_smile:

I have run the regression-based baseline correction on one participant only. This is because I was unsure as to whether this is something that needs to be done on all participants at the same time or just on one to then inform the actual baseline correction (e.g.: choice of baseline window) during pre-processing. How should one run regression-based baseline correction?

The first plot shows the estimated effect of the baseline period.
beta_values_predictor

Although it is clear that earlier time points should be more baseline corrected than later points, I am struggling to define a precise baseline window from this plot. Can the plot inform in any way the choice of baseline window? (e.g.: by looking at this plot, can I determine whether I should use -1000 0 or -2000 0?) I am aware of Alday’s suggestion in his 2019 paper (i.e.: “a baseline window on the
order of a few hundred milliseconds may be the sweet spot for many experimental designs under typical laboratory conditions without large high‐frequency artifacts”), however, I am struggling to determine the baseline window from the plot.

The second plot shows the interaction term from the regression model (i.e: one condition*baseline predictor).
beta_value_condition*baseline

I have the strong suspect that there are differences in the effect of the baseline period across condition trials, since the beta value does not seem to fluctuate around 0, but it rather mostly fluctuates above 0. By looking at the second plot, do you also think this is the case? If there are differences, what would be a natural next step?

I thank you for your help and time! :slight_smile:
Sabia

4

Hello Sabia,

  • estimated effect of baseline plot:
    Ideally you would fit a hierachichal mixed linear model, where you add random slopes and intercepts for each participant. If you don’t feel comfortable using a hierachichal model I guess you can fit one model per participant and then do some stats on the beta values. For the participant you plotted I would increase the baseline window, as it looks like there is something going on before 200ms prestimulus. Of course this depends on your design (e.g. inter stimulus interval?).

  • interaction plot:
    Regarding the interaction plot, it does indeed look like there are differences between your conditions. However, the beta values are rather small, so I am not sure if the differences are that important. How you continue your analysis depends on your further analysis steps. Ideally you add the regressors of interest (conditions etc.) to your model and stick with regression based EEG analysis. If you want to simply average over trials and plot ERPs, I would suggest to work with the data that remains after regressing out the baseline and the interaction per participant. This would be the beta coefficients for your conditions in the regression model e.g. reg_model[‘auditory’].beta. You can simply treat them as your signal and continue with your planned analysis.

I hope this helps?

Carina

2 Likes
5

Hi Carina,

Thank you so much for your detailed explanation! I will try to implement your suggestions now!

All the best,
Sabia

2 Likes