The EEGLAB News #6


From the eeglablist

Q. Chiara Terzo: We are doing an ERP study on the Implicit Association Test (IAT), and specifically we look at N2 (cognitive conflict) and LPP (allocation of the attention towards motivationally relevant stimuli). Rereferencing does indeed affect ICA, but it does also affect the estimate of the evoked potentials (carried on the already cleaned data), especially affecting the late components. Indeed, when referencing to the right mastoid I get the LPP from 400 ms on. When referencing to the common average, the LPP disappears and I get instead a negative pattern from 300 ms on. The literature on ERPs and the IAT does not have a consistent reference, but on average they do not use the common average as a reference. Thus, what I am wondering is whether using the right mastoid as a reference would bias too much my ICA.

A: Arnaud Delorme: You do not need to re-reference the data before running ICA. There is no best reference for running ICA decomposition - I prefer to run it after converting the data to average reference but the benefit has never been clearly demonstrated to me. Changing the reference will of course change the independent component or raw data topographies you are observing, but not the location of the component process nor of its equivalent dipole. When interpreting results at the scale channel level, it is important to compare your results to published figures that use the same reference you have used in your analyses. This is one limitation of working with channel data. Re-referencing also affects the ICA component scalp topographies. Thus, when you re-reference your data in EEGLAB, the ICA scalp topographies (scalp maps) are also re-referenced by the EEGLAB reref() function. In the “envtopo” plot (EEGLAB menu item Plot > Component ERPs > With component maps), the ICA component map polarities also depend on the latencies at which they are plotted; the component scalp map polarity is inverted when the activity (activation) of the component is negative at the latency at which its contribution is indicated. See also comments on ICA and polarity from Joe and Makoto below (with some expansions for instructional use by Scott).

Joe Dien: .. the sign of the loadings ('mixing matrix', in ICA speak) is arbitrary and only relates to the sign of the original voltages once you multiply them by the ICA component activations and add together their scalp projections at each channel to backproject the data. This is because the same set of loadings (i.e., the columns of the ICA mixing matrix, EEG.icawinv) accounts both for the time points where the activations go negative and for the ones where they go positive. So it is only the product of the two [mixing_matrix X component_activations] that gives you the data with its original sign for each time point and channel …

Makoto Miyakoshi: ICA component maps have inherent uncertainty on the polarity of the decomposed data. ICA models the scalp data as the (matrix) product of the component activation time courses (rows of EEG.icaact) and the activation maps (columns of EEG.icawinv). For example, just as 1 = 1*1 = (-1)*(-1), so your P300 at scalp channel Pz (to reference) may be represented by a positive ERP value in a 'red' scalp topo (map) or by a negative ERP value in a 'blue' scalp topo, depending on whether the IC activation at that time point is positive or negative. Did you stumble on this point? For example, if you want the IC that most explains the largest part of the P300 peak at channel Pz to show a positive P300, we can use that prior knowledge to correct the IC map polarity without reference to the actual component activation at 300 ms. Else (as in EEGLAB functions envtopo() and pop_envtopo(), see above), you can backproject that component to the scalp; the polarities of the channel values then correctly represent the voltage contributed by the IC source process to the channel signal. At the group-level analysis, EEGLAB STUDY ERP functions align IC polarities within each cluster. That does not guarantee that when plotting the scalp maps in this cluster, you will always see positive ('red') topo values at Pz, but at least the component maps are then maximally consistent across subjects. The polarities of the same component activations are also polarity-reversed, so in each case the back-projected scalp channel data for the component will have its correct polarity. If you don't use the EEGLAB STUDY functions to perform group-level analysis, I'm afraid you will need to implement some solution to align the IC map polarities. But again, this only is a matter of concern when you attempt to run group statistics on the component maps or activations themselves in isolation, and not on the component backprojected channel signals (i.e., the products of the component maps and component activations) and when you polarity-reverse only the component maps (or the component activations).