This FAQ page contains questions we receive and answers we give users, as well as general tips we think it is important for users to know! Many other tips are available on the eeglablist archive. To search the list archive simply use Google, enter relevant keywords and add the “eeglablist” keyword.
Table of contents
- Basic Topics
- Files: Import/Export/Channel
- Basic Processing
- Advanced Topics
- Time Frequency
- Single Trial
Answer: We believe so. First, EEGLAB implements new algorithms for artifact rejection. Second, for the experienced MATLAB user EEGLAB is a fast and accurate way to start processing EEG and ERP data and to directly manipulate signal arrays. If you intend to use the ICA toolbox functions underlying EEGLAB, EEGLAB itself is a good starting point and introduction. EEGLAB also provides a full EEG structure to describe your data (signal, trials, channel location, reaction time, type of the trials, time limits and sampling rate) and allows you to use this structure either from the MATLAB command line or in MATLAB scripts.
Answer: Yes, to an extent, but… Because MATLAB sometimes uses large amount of RAM, we also took great care of inserting options in EEGLAB and in several processing functions to handle low memory conditions. On the other hand the MATLAB environment offers the advantage of stability and ease of use. Even the novice user under MATLAB can scale a data array by multiplying it by a scalar for instance (and in our software the data array is directly accessible to the user). MATLAB also offers the advantage of modularity. All of our functions, are stand-alone functions and most of them can be used independently of each other. Besides, MATLAB has grown much faster.
Answer: We cannot guarantee that we will provide full support for this software but we will be glad to help out if someone encountered any problem and to correct bugs when reported to us: Write to firstname.lastname@example.org.
When I ask the Mathworks salesman to sell me “only” a MATLAB license, he gave me prices for MATLAB, Simulink and Symbolic Math. Do I need all these to run EEGLAB? Answer: No, you do not need all that, you only need MATLAB. If possible, ask for an educational or even student version, as they are cheaper. For some spectral decompositions, you may also need the Signal Processing toolbox which has to be purchased separately.
Under Unix, I often get the following message ‘‘Warning: One or more output arguments not assigned during call to ‘XXX’.
??? Unable to find subsindex function for class char. ‘’
Answer: In most cases, this error indicates that MATLAB on Unix may experience problems. MATLAB might return this error when you or EEGLAB has defined a variable of the same name as any variable in your MATLAB workspace or .m file in the MATLAB path. To solve the problem, clear the variable or rename the function.
Answer: 2 solutions
- Buy more memory (RAM) for your computer
- Try the memory mapping scheme (in EEGLAB options) which will allow keeping the data on disk. Note that except for Neuroscan files, it is still necessary to import the full data file in memory.
Does it benefit to have a multi-core machine?
Answer: yes, it benefits in two ways. First, you may start in parallel several MATLAB sessions. Each of them is assigned one of the processors. Second, if you go to the MATLAB options, you may have the option to enable multi-core computation (General > Multithreading). This option is usually set by default. This is a very efficient option that will speed up your code usually linearly with the number of core (2 -> twice faster etc…).
Is it possible with EEGLAB to save an EEG data sorted in 10 epochs (for example) in 10 ascii files? Answer: you have to do this on the MATLAB command line:
>> epoch1 = EEG.data(:,:,1); >> save -ascii epoch1.txt epoch1 >> epoch2 = EEG.data(:,:,2); >> save -ascii epoch2.txt epoch2 >> epoch3 = EEG.data(:,:,3); >> save -ascii epoch3.txt epoch3
I work with recordings from 30 channels + 2 EOG channels. Naturally while exploring the data for artifact components, I want to have a look at the EOG too. As soon as I want to plot maps, I liked to leave the EOG out. But if I want to load an electrode position file with less electrodes than channels in the dataset, EEGLAB doesn’t accept that. Is there a more convenient method than creating two datasets (one without eog or without an electrode location file)? Answer: Simply blank out all the position fields for these channels using the channel editor after you have imported the channel location file (so they do not have an assigned position). Note: You may still include these channels in the ICA decomposition, even if their reference is different from the other scalp channels, though you should not attempt to plot them with a mixed reference. (We are working on a solution to this for other purposes).
EEGLAB does not work when I try select the type of data file to import under >File > Import data > From ASCII …”.
Answer: Next to the importing text box there is a list box to indicate which type of data you want to import. Note that, not only you need to scroll the the list box but also to CLICK on the selected importing options so that they become selected.
We have been trying to input our polhemus 3-d file into EEGLAB. Displaying it in 2-d, the main problem is that “0 degrees” is towards the right ear in Polhemus and “90 degrees” is toward nasion, while in EEGLAB “0 degrees” is toward nasion and “90 degrees” is toward the right ear. In other words not only that everything is shifted 90 degrees, but Polhemus (Neuroscan) goes counter-clockwise and EEGLAB goes clockwise. We are trying to develop some conversion solution for this, but if
- you already have some experience w/ a situation like this and you may have some easy solution;
- you could advise us on how to get our Polhemus coordinates into EEGLAB some other (easier) way, we would appreciate.
Answer: It actually depends on how you reccorded your Polhemus coordinates. To fix this problem, under EEGLAB, in the channel editing window, there is a button “tranform axes”. Press this button and enter “theta = 90-theta;” and that will do the trick. Note: press the “auto shrink” button to visualize all your electrodes. You may also manipulate the XYZ coordinates and reconvert them to polar.
Remember that all EEGLAB graphic interface always process the current dataset. This means that if you use an interactive function (Data scrolling, Epoch rejection, Component rejection) and manually select or load another dataset from the EEGLAB interface, EEGLAB will apply all the changes or plot to the new dataset being scrolled. For instance, if you are scrolling data and select bad epochs, then load another data to look at it, then come back to your scrolling window to reject these epochs (and actually press the reject button), the epochs will be rejected in the New recently loaded dataset.
The maximum time does not correspond to the maximum time I specified. For instance I asked for epochs between 0 and 3 seconds at 125 Hz and end up with an interval of 0 to 2.992 s. There should be something wrong! Answer: Nothing is wrong. In your example, we must draw (125Hz * 3seconds = 375 points) and not 376 otherwise we would lose time linearity i.e. 2 epochs of 3 seconds would be 752, whereas if we draw 6 seconds of the data we would get 751 points !), but if we assign time 0 to the first point, then we must assign time 2.992 to the last point. Actually, the first point time is undetermined. Since the recording is made at 125 Hz, there is no possibility to know when the first point was recorded in between 0 or 0.08 seconds (otherwise one has to sample faster). By convention we take the first point to be at latency 0 and the last one at 2.992 (but we could choose latency 0.04 for the first point and 2.996 for the last one or 0.08 to 3…).
Could you help me to know why : I have 7 events, 3 being = 2 and 4 being = 1 in a “mother” file why, when I extracted epochs on event = 2, I only got 3 epochs (that’s right) and 6 events (that’s wrong, isn’t it?)?
Answer: Your 6 events are probably OK. It depends on the time window you used for epoching. Any event within the time window of each epoch will be taken into account. This means that some event outside of all epoch time window will be ignored and that some events within several time windows might be duplicated. You should look at the event latencies before and after epoching and you will see that it is consistent.
I have a problem to filter data: we see in our EEG data a component with the frequency of the electronics here, 50 Hz. I would like to remove only this frequency… but I don’t really know to do this with lowpass and high pass filtering. I think that this phenomenon of “pollution” by the electric frequency is something current in EEG? Answer: Yes, it is very common.
Answer: Check out firfilt plugin at EEGLAB Plugins page.
When using the ERP image menu item (‘Plot’ -‘Channel ERP image’), the top part of the output is the window with the sorted trials. The number of the trials seems to be off and it appears that not all trials are displayed. For instance, using a file with 17 trials the “Sorted Trials’ numbering goes from 6 to 12 with apparently 5 trials being displayed and two half trials on the top and at the bottom of the display (see attached jpg). This was also observed with another file that had more trials.
Answer: The erpimage() output is due to the smoothing window you selected. If you use a 5-trial average moving window, the first output should be the middle of this window (e.g., trial 2.5). Use a 0-smoothing window (width 1) and all the trials will be shown.
Is there any way to add lines above and below the average voltage line to indicate +/- 1 or 2 SD to show data dispersion across trials?
Answer: Use option ‘erpstd’ of erpimage() in the “More options” text box of the pop_erpimage() interactive window.
Do you have any good way to across subject/patient analyses with respect to power and coherence, especially making statistical comparisons between subjects/patients? Could we somehow use the data output of EEGLAB?
Answer: For cross-subject comparison, tftopo() is a powerful function that can summarize this information (it uses stored timef() and crossfO() function outputs). It has to be called from the command line though.
It is not very difficult to find components related to eyeblinks, etc. In my case, there are phases during the experiment, where people speak and/or move their eyes. I find it quite hard to determine which components are related to these artifacts and I already wonder if it is possible et all. Answer: To determine wich components are related to these artifacts, one approach is to isolate these trials (selecting them) and then use menu item “Plot > Component ERPs > With component maps” and select the time window where these event appear. This function will plot which components contribute to this type of artifact. However, you are correct in thinking that ICA cannot cleanly resolve ALL artifacts into one or a few components. For instance, “paroxysmal” artifacts (like the subject scratching their head during recording) would require a large number of ICA components to capture the variability of all the artifactual contributions in the data. Similar events and artifacts should be carefully pruned from the data before ICA decomposition.
Would muscle components be seen at high frequencies? Answer: Yes, that is typically what we observe. Muscle artifacts have a broadband high spectral amplitude at high frequencies (20-100 Hz or more). Also their spectrum does not look like the standard EEG (exponential decrease).
I am currently using ica to correct for artifacts. In the past I’ve visually inspected each single-trial epoch separately, indentified those trials with artifact activity, and then trained ICA on each trial separately to identify and remove artifactual components. As, you can imagine this process is extremely time consuming. Is it effective to train ICA on multiple or all concatenated trials at once, remove artifactual components, and then go back to visually inspect the corrected data for outlying artifacts?
Answer: You may not have grasped the nature of ICA, which is to extract data components whose activities (activations) are independent of the activities of other data sources. For this independence to be detected, it must be fully *expressed in the data* – and this normally requires many (>nchannels^2) data points. So separately decomposing individual epoch is just the wrong approach. We currently favor a seven-step approach:
- Visually reject unsuitable (e.g. paroxysmal) portions of the continuous data.
- Separate the data into suitable short data epochs.
- Perform ICA to derive independent components.
- Perform rejection on the derived components based on inspection of their properties.
- Visually inspect the raw data epochs selecting some for further rejection.
- Reject the marked components and data epochs.
- Perform ICA a second time on the pruned data - this may improve the quality of the ICA decomposition, revealing more independent components accounting for neural, as opposed to mixed artifactual activity.
These steps are made easier and more efficient by EEGLAB, which also includes functions suggesting to the user epochs, channels and components suitable for rejection.
While trying within EEGLAB to remove artifacts using ICA, I had trouble in recalculating an ICA decomposition after removing components. I tried to follow the guidelines in the tutorial, thinking that with fuzzy components it might work better to remove some clear artifact components first and the run a new ICA. When I tried to do that, the second ICA always took much longer and I got also some error message in the end telling me, that there was something wrong with the result.
Answer: The standard procedure we advise is first to perform ICA on the data and to remove bad trials using the ICA component activities. If you remove ICA components, the rank of the data will decrease (to <nchans). If the data have n channels, the rank of the data is (most probably) n. If you remove one component it will become n-1, and ICA will not be able to find n components in the pruned data). Thus, as a first step, you should only remove bad trials. This procedure will not alter the dimensionality of the data. As a second step, recompute ICA and remove bad components (the second run of ICA should result in clearer artifact components (for instance muscle at high frequencies), not contaminated by strong outlier trials. If you remove ICA components and want to re-run ICA, you must decompose the data with the ‘pca’ option to reduce the dimensionality of the decomposition to match the data rank (see below).
What is the rationale behind baseline zeroing the data before running it thru ICA, and is that always recommended?
Answer: It is recommended because the EEG might have some electrical artifacts (slow trends) that you want to remove. If your data is perfectly flat (at very low frequencies), then you shouldn’t need to do that. You should baseline-zero each epoch, else use the continuous data and lowpass it if you are interested in focusing on e.g. 3-40 Hz activity. Also, you should prune the data of ‘messy’ patches (probably associated with a series of 1-of-a-kind, non-stationary maps). The data scrolling utility in EEGLAB makes this convenient, and if you perform this on the continuous data, records breakpoint events that guide subsequent epoching. Else, you can more severely prune the data to train the ICA model, then pass more of the data through the model (at the cost of somewhat higher SNR (signal to noise ratio) in the activation time series).
I don’t know how to make the independent component dimension reduced by PCA in EEGlab4.0. For example, I want to obtain 4 independent components from 32-channel EEG data.
Answer: To extract 4 components, in the second text box of the “Tools > Run ICA” interactive pop-up window, enter “ ‘pca’, 4 “ or “‘ncomps’, 4” and that will do it. Note we do not recommend using PCA (“ ‘pca’, 4 “) unless you have some good reason. Using first PCA components only will truncate the data (irrespective of components), and then ICA may not be able to find relevant components. The second possibility (“‘ncomps’, 4”) is more acceptable theoretically since it is a true ICA decomposition (that uses a rectangular matrix). In general, we advise finding as many components as possible (e.g. if you have enough memory on you computer to run ICA over all the channels).
You mentioned that we should use the extended ICA algorithm to extract subgaussian components (i.e., 60-Hz noise). For our MEG data (using the old binica() for Windows) we get several subgaussian components, but there is still leakage of 60 Hz onto many of the supergaussian components as observed with an FFT (mutlitaper). In the EEGLAB tutorial (1st_readme.txt) I noticed a question mark on whether binica for Windows (the old one) was stable. Is it unstable? Should we use the extended version? Can we use the old binica() for Windows? Answer: MATLAB seems to have speeded up running runica() tenfold from 5.3 to 6.x ! So binica() just gives us a 30% improvement in speed these days (although also a 2-4-fold decrease in process size, important for large datasets under 32-bit memory addressing). In the future, we will compile the improved version of binica() for Windows. The improvements concerned mainly the ‘pca’ option, which you may not need to use.
I have noticed that the runica() does not include a field for epoch size, so how does ICA recognize the epochs? Doesn’t this make a difference to the way ICA is handled? Answer: Epochs are concatenated before running ICA. In ICA, all time points of all epochs are shuffled so that epoch information is irrelevant.
I have some continuous 32-channel EEG data on which I would like to apply Infomax ICA. I am primarily interested in the 100 epochs from the data, which are 3000 frames each. There are only about 20-40 frames between the end of one epoch and the beginning of the next. Should I apply ICA to the continuous data, then epoch the ICs, or apply ICA to the concatenated epochs? Answer: You can apply ICA to either of them. Usually, we prefer to apply ICA to the concatenated epochs so ICA component are more likely to represent activity related to the task, but continuous data are fine too, especially if you have few epochs or few data points, since most of the same EEG and artifact processes are likely to be active ‘between’ epochs.
For multiple-epoch data, the scalp map obtained for the different epochs is the same for a particular component. Is this normal or is there some mistake that’s being done in the analysis? Answer: Because the ICA algorithm is applied in the electrode space domain, the same scalp maps are returned for all epochs. However the time course of one ICA component is different for each epoch (if its activation value is 0 at a given time, it means that this component is not expressed in the data at that particular time).
Is it not true that the ERP for a condition can be completely reconstructed from the timef() results, incuding ITC? One could write a function that takes the outputs of timef (‘times’,’freqs’,’ersp’,’itc’) and gives as it’s output the ERP. Could one then usefully manipulate the values in ERSP and ITC and see how the ERP would have looked without some aspect of the ERSP (like, take away the inter-trial coherence in the alpha band and leave all else the same). Finally, could one subtract the ‘ersp’ and ‘itc’ between two conditions and then reconstruct the difference ERP?
I’ve been playing with the parameters of timef() and am surprised by the large differences in results dependent on the values for ‘cycles’, ‘winsize’, etc. This even makes we insecure about previous timef() results ….
Answer: Yes, this is possible (with some fuzziness regarding overlap-adding the overlapping spectral estimates, undoing the effects of tapering (windowing), etc. I haven’t focused on strictly “invertible” time/frequency transforms - which tend to be restrictive, since I am interested in analysis rather than synthesis.
However, in general the answers will be like the following:
Removing alpha ITC would be like filtering out the alpha in single trials and replacing it with random-phase noise. The effect on the ERP would be more or less exactly like filtering out the 10 Hz from the ERP, plus adding some noisiness…
The ERSP is a different matter, as ERSP changes can only produce a (stat. signif) ERP in conjunction with signif ITC. Adding an ampl. increase at alpha, say, to the ERSP and then going back to the time domain would scale up alpha band power in the ERP – if the ITC stays the same.
Changing both ERSP and ITC can be made to give any kind of mixed results summing those above.
Re timef() variability - think of it like the focus / f-stops on a camera - you cannot achieve focus on every plane at once. It is also like Heisenbug uncertainty, I hear - you cannot localize a ‘particle’ (wavelet) in time and frequency simultaneously. Or, as in cinema you, where one cannot record motion in a single instant, one cannot record an exact power estimate at an instant. Basically, a “good” time/frequency method can only give an estimate within a time/frequency area whose area is fixed, but not its shape. i.e. Do you want long-and-thin, or fat-and-short, or square?
So, one can make exact comparisons only across channels or components within the same t/f transform.
I am afraid that the time scale is slightly off for the time-frequency plots. E.g. a component time-frequency plot from a dataset where the epochs are from -2560 ms to +2046 ms and according to the plot’s time scale it appears that the epoch is from slightly before -2000 ms to slightly over +1500 ms. Do you know why? Answer: It is normal that the time limits are different from the original dataset, since the FFT (or wavelet) is applied over time windows and we consider the center of these windows. As a result, you lose half the window size on each edge of the plot (some hundred milliseconds depending on the window size and the sampling rate).
You should most likely use the pwelch method (implemented in the EEGLAB spectopo function). It is a windowed FFT (several FFT averaged).
The Thomson method (usually known as multitaper) is good too. It is first projecting the data onto an orthogonal base, then performing FFT.
They should all return similar results (FFT, pwelch, multitaper). I guess the Thomson method is less sensitive to noise but also more complex to use. I guess it would also be possible to use the welch method on top of multitaper. It is all a matter of preference. I would advised using the pwelch method which is easy (you just give as an option the length of the windows and the overlap). Multitaper would require you to select the number of basis vector in your othogonal base and this is much less intuitive (and also has consequences on the frequency resolution you can achieve).
I have been using the new EELAB toolbox for the past couple of weeks, especially timef() and crossf(). The multitaper method with bootstrap statistics has been giving me very nice stable results. Timef() with wavelets gives slightly different results, but also interesting. I noticed though that all the analysis has been designed to study coherence, phase-coherence, ITC, etc, for data organized as epochs. (e.g. inter-trial effects).
Is there a function for computing time-varying coherence between ‘independent’ activation functions for continuous spontaneous recordings? (i.e., spontaneous coherence for brief windows of time, 200-300ms). J.P. Lachaux (from Varela’s Lab) has several papers investigating this issue.
Answer: We also programed multitaper methods for timef(), but removed them from the current more flexible versions (not available yet on the Internet (May 23, 2003)). Note that we usually use neither the FFT (0) or N-cycle wavelet methods, but a compromise (0.5) setting that trades off frequency resolution (at low freqs) with stability at high freqs.
We have not yet had cause to program flexible visualization and handling of continuous coherence measures, though you can run timef()/crossf() on a continuous dataset (which is internally considered by EEGLAB to be 1 epoch). We’d welcome suggestions and/or code.
In our experience, multitaper decomposition are not very useful because we are interested in low frequency activities (i.e. 5 Hz) at which we want the maximum time resolution. So we had to use 1 multitaper only (using several multitapers reduces noise but also reduces the time/frequency resolution) which is equivalent to a standard FFT. Using several multitapers in the gamma band would make sense though because the time resolution is much higher and not critical. Multitaper has been removed from the newtimef() function but it is still possible to use it in the timef() function.
We normally display our chrono-spectrograms with the lower frequencies at the bottom of the plot, just the opposite how Eagle does it. Is there an easy way to change this?
Answer: Right now the only way to change this is to click on the figure and type
>> set(gca, ‘ydir’, ‘normal’);
I guess you could edit the timef() and crossf() MATLAB functions to change all the plots to this format. We will introduce this as an option in the future.
How can I sort single trial ERP in erpimage() based on their amplitude at a determined latency. Answer: use the ‘ampsort’ option of erpimage() (not separately queried in GUI, you have to put it into the “More options” text box at the bottom right of the pop_erpimage() interactive window). Erpimage() option ‘ampsort’ sorts on spectral amplitude; if you want to sort on potential value at some epoch time point, then use option ‘valsort’.
For reasons too complex to describe, I am trying to epoch with respect to various latencies RELATIVE to the events in my dataset. For example, epoch onset = event latency + some value, so I can calculate the set of latencies and use them in a call to epoch.m (the second parameter). I couldn’t specify the latencies directly using pop_epoch function.
Answer: There are 2 solutions for this problem:
1) modify event latencies or create new events. For instance
for index = 1:length(EEG.event) EEG.event(index).latency = EEG.event(index).latency + 10; % +10 sample points end; % then store data [ALLEEG EEG CURRENTSET] = eeg_store(ALLEEG, EEG, CURRENTSET); % then epoch from the menus
2) use the epoch() function instead of the pop_epoch() function. The epoch() function processes latencies directly. You will then obtain a MATLAB array that you may import in EEGLAB.
The lines in erpimage() (e.g. ‘X’ axis on the average) are too thick. How can I control their thickness?
Answer: To change the figure aspect for publication, you can go in the figure menu and use the MATLAB menu item “Tools > Edit”. Then you can select any object in the figure. The second button will display a contextual menu where you will be able to change line thickness, color, font aspects…, or even draw additional lines or add text. We also export figures as Postcript files and open them with Adobe Illustrator in vector format to fine tune it. See also this web page.
Could you tell me how to export figures from EEGLAB (to include in a “word.doc” for example, or to export to Excel)? This for the scroll channel data eegplot(), and for channel spectra and maps…and other plots….
Answer: For most EEGLAB figures, simply use menu item “File > Export”.
For the scrolling channel data function (eegplot()), first use menu item “Figure > Edit figure” to restore the default MATLAB menu. From the command line, you may also use the command
>’>’ print -djpeg scrollfigure.jpg %to print in jpeg
>’>’ print -depsc scrollfigure.eps %to print in postcript color
Note: use the command “>’>’ set(gcf, ‘paperpositionmode’, ‘auto’)” first to print the figure in the same aspect ratio as it is shown on screen.
When I tried to plot the channel spectrum, the axis labels and tick labels did not appear clearly on the screen and I’ve got the following error messages from the MATLAB command line: “Plotting scalp distributions: Warning: Unrecognized OpenGL version, defaulting to 1.0” Answer: You may solve the problem by changing the OpenGL version on the MATLAB command line by typing: “feature(‘UseGenericOpenGL’,1)”.
Sometimes MATLAB crashes when I try to print a figure. If I save the figure on disk, first, some parts are missing. Do you know how to fix this problem? Answer: You may solve the problem by changing the OpenGL version on the MATLAB command line by typing: “feature(‘UseGenericOpenGL’,1)” For the printing error, we also experience this; it is a MATLAB problem which is not consistent between Windows and Linux. We always print or save to files (.jpg or .eps Postscript), then print the files. For instance, use the software FreeRawPrint to send the Postscript file to the printer under windows. Even with this strategy, some parts of complex figures may disappear, but this is rare (then we use screen captures, or use a Windows machine, since printing seems to be more reliable under Windows OS). We hope MATLAB will become better at this in the future (Is MATLAB listening?).