Hi Jona,
that makes sense. I have another couple of questions, if you don?t mind me asking. 
1) I tried to use linear_regression_raw on my data, but I got the weird result attached. It looked like my data needed to be baseline-corrected (which isn?t apparently applied within the actual function), but I wouldn't know how to do it. Is it the apply_baseline method that I should use for that purpose?
Yes, you can call .apply_baseline on evoked objects. Or possibly filter the raw data harder.
2) A way around the problem in 1) would be to use linear_regression on already epoched and baseline-corrected data as shown here <Page Redirection. Is there a plotting results other than the topomaps exemplified there?
The two are entirely different, but if linear_regression covers your use case, you should go for it.
Both return evoked objects (one per condition for _raw, and one each for t value, log p value, and beta for the epochs version per predictor). Try plotting the evoked objects with .plot_joint(), or maybe .plot_image.
3) I don?t seem to understand what some of the arguments of linear_regression_raw really say. In particular, which argument takes the predictors to run the regression against - covariates or solver? What does each of them actually do?
solver takes the solver - i.e., the algorithm to solve the linear system. covariates is one of two options for specifying predictors.
Many thanks for all the help, and apologies for the amount of questions!
-Roberto
----------
Roberto Petrosino
Ph.D. Student in Linguistics
CT Institute for the Brain and Cognitive Sciences
University of Connecticut
Hi Jona,
many thanks for your reply.
So, linear_regression_raw will take predictors from the covariates argument, which will have to refer to an array having the same length as the event file - am I correct? What is the reasoning behind having two regression functions (one for continuous and one for non-continuous data)?
(The suggestion is to use pandas data frames for linear_regression_raw)
They work completely differently. One fits independent models at each time point, the other fits one model for the whole data set.
I was looking at the script here <Page Redirection, and I was a bit unclear about some lines of code - I just started migrating from eeglab, so I apologize if I sound completely ignorant about everything here.
names = ['intercept', 'trial-count']
intercept = np.ones <numpy.ones — NumPy v1.26 Manual), dtype=np.float)
design_matrix = np.column_stack <numpy.column_stack — NumPy v1.26 Manual([intercept, # intercept
np.linspace <numpy.linspace — NumPy v1.26 Manual(0, 1, len(intercept))])
So, if I?m understanding correctly, here the code is defining one of the two predictors for the next regression analysis. `intercept` is an array as long as the variable `epochs`, and filled with ones. Such an array is ?glued? with another array of the same length, filled with evenly spaced numbers between 0 and 0. Everything is then put in the linear regression model:
# also accepts source estimates
lm = linear_regression <mne.stats.linear_regression — MNE 1.7.0.dev17+g20174f448 documentation(epochs, design_matrix, names)
My question is - what is `trial_count` in the first place? It is defined as follows:
trial_count = lm['trial-count']
but I don?t seem to understand what that really is. Basically what I was trying to do is plotting rERP waveforms with the epoched regression function, similarly to what is showed for the non-epoched regression function here <Page Redirection.
trial_count is the 2nd column of the design matrix. It?s, as you said, evenly spaced numbers from 0 to 1, so it?s 0 for the first and 1 for the last epoch. It gets its name from the ?names? list (defined further up) fed as the 3rd argument to linear_regression.
So it?s just a regressor for how late you are in the experiment.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.nmr.mgh.harvard.edu/pipermail/mne_analysis/attachments/20170902/2daa6fb2/attachment-0001.html
