I am trying to measure transfer entropy between time series of musical features and a time series of EEG features. These features are computed in a sliding window of length 3 s, with step size 0.25 s. As such, the time series we get are usually around 700 data points. This means we cannot really divide these time series into trials.
I have managed to use TRENTOOL's lower-level function transferentropy.m to get a raw TE value for a pair of time series, but as mentioned in the TRENTOOL paper, the TE values given by the estimator are not interpretable at face value. Are there ways other than taking multiple trials to 1) measure the statistical significance of TE for my time series; or, failing this, 2) at least render the TE values interpretable and comparable between subjects?
An idea that I had would be to use a procedure similar to the shift test, even though instantaneous mixing is not a problem for my data. Given the music feature time series X(t) and an EEG feature time series Y(t) and range of physiologically plausible lags u_1:u_n, I would create a number of shifted signals X_i(t+i) (where i > u_n) and compare TE(X->Y) to the distribution obtained from the various TE(X_i->Y). Would I be able to address either 1) or 2) with this?
A colleague also suggested to me to take segments of my time series and permute them - would this be another option? Any other suggestions from the community or developers would be very much appreciated.
Hi TRENTOOL community,
I am trying to measure transfer entropy between time series of musical features and a time series of EEG features. These features are computed in a sliding window of length 3 s, with step size 0.25 s. As such, the time series we get are usually around 700 data points. This means we cannot really divide these time series into trials.
I have managed to use TRENTOOL's lower-level function transferentropy.m to get a raw TE value for a pair of time series, but as mentioned in the TRENTOOL paper, the TE values given by the estimator are not interpretable at face value. Are there ways other than taking multiple trials to 1) measure the statistical significance of TE for my time series; or, failing this, 2) at least render the TE values interpretable and comparable between subjects?
An idea that I had would be to use a procedure similar to the shift test, even though instantaneous mixing is not a problem for my data. Given the music feature time series X(t) and an EEG feature time series Y(t) and range of physiologically plausible lags u_1:u_n, I would create a number of shifted signals X_i(t+i) (where i > u_n) and compare TE(X->Y) to the distribution obtained from the various TE(X_i->Y). Would I be able to address either 1) or 2) with this?
A colleague also suggested to me to take segments of my time series and permute them - would this be another option? Any other suggestions from the community or developers would be very much appreciated.
Thanks very much for your help, Soren