It is definitely appropriate to do a time-frequency analysis on this experiment. If the inter-stimulus interval (ISI) is a constant 500 ms, you will get a nice steady-state visual evoked potential (also called frequency tagging or the flicker effect) at 2 Hz. Depending on the purpose of the task, it might be interesting to look at activity at this frequency.

Because the inter-stimulus-interval is 500 ms, it will be difficult to investigate any transient frequencies below 2 Hz. But because of temporal smearing, it might be good to stay above ~4 Hz or perhaps even higher. Of course, the duration of the frequency response to the stimulus depends on how long the brain continues processing that stimulus.

Also, I'd recommend using wavelet convolution instead of a short-time FFT. Because the intervals are so short, you will probably want to maximize the temporal resolution of the time-frequency result. This would be very computationally expensive for a short-time FFT because of all the FFT's you'd need to compute (once per time point). Wavelet convolution, in contrast, would give you a high temporal resolution TF result and requires only one FFT of the data and one FFT per wavelet frequency.

Regarding stationarity, analyses like short-time FFT and wavelet convolution assume that the data are stationary for the duration of the time window, which is generally a few hundred ms, depending on the parameters. There will likely be transient violations of stationarity immediately after stimulus onset, but that this is not a fatal concern; it will just produce some frequency smearing.

For more, consider: http://mitpress.mit.edu/9780262019873

You might find this work useful.

Advanced Temporal and Spatial Analysis of Electroencephalographic Signals Walter Freeman, Rodrigo Quian Quiroga.

Thank you very much for the suggestion. I would definitely take a look at it.

It is definitely appropriate to do a time-frequency analysis on this experiment. If the inter-stimulus interval (ISI) is a constant 500 ms, you will get a nice steady-state visual evoked potential (also called frequency tagging or the flicker effect) at 2 Hz. Depending on the purpose of the task, it might be interesting to look at activity at this frequency.

Because the inter-stimulus-interval is 500 ms, it will be difficult to investigate any transient frequencies below 2 Hz. But because of temporal smearing, it might be good to stay above ~4 Hz or perhaps even higher. Of course, the duration of the frequency response to the stimulus depends on how long the brain continues processing that stimulus.

Also, I'd recommend using wavelet convolution instead of a short-time FFT. Because the intervals are so short, you will probably want to maximize the temporal resolution of the time-frequency result. This would be very computationally expensive for a short-time FFT because of all the FFT's you'd need to compute (once per time point). Wavelet convolution, in contrast, would give you a high temporal resolution TF result and requires only one FFT of the data and one FFT per wavelet frequency.

Regarding stationarity, analyses like short-time FFT and wavelet convolution assume that the data are stationary for the duration of the time window, which is generally a few hundred ms, depending on the parameters. There will likely be transient violations of stationarity immediately after stimulus onset, but that this is not a fatal concern; it will just produce some frequency smearing.

For more, consider: http://mitpress.mit.edu/9780262019873

I agree with Michael. A long time ago (in the 'stone age' of Psychophysiology) we did FFT on succesive heart beats (R-peaks of the ECG). To find out which frequency components were responsible for a phenomenon called sinus arhythmia (the natural variation in successive heart beats). We found two major peaks in the frequency spectrum: a respiratory peak (around 0.3. Hz) and a much slower peak around 0.1 Hz which was sensitive to cognitive task load. But if you register the heart beat during a fixed paced reaction time task ( e.g. ISI 1 sec) requiring a button press repons, you will find a third peak, the task rhythm, in your hear beat spectrum. We called it 'entrainement ', very much like the effect you get with flicker in the EEG.

.

Thank you very much for all your answers and suggestions.

The aim of the analysis we were considering was simply comparing power at some canonical frequencies between two groups, so initially we were not interested in the time evolution of the signal in such a short time window. Indeed, in the question I refer to FFT decomposition rather than to wavelet convolution because I am concerned about the edge artifacts of the latter method (it might require longer buffer zones than the FFT). That is, if the epochs last the 500 ms between successive stimulus presentations, there would be edge artifacts spanning nearly 200 ms in each edge of the epoch (two cycles of alpha activity at 10 Hz) using wavelet convolution. This leaves only 100 ms of reliable data to be further analyzed. I know there are some methods to overcome this problem, such as reflecting data at each edge of the epoch or using overlapping epochs and then trimming off the redundant parts of the epoch. However I have never used them before. Therefore it would be really appreciated if anyone could give me further information in order to make the most reasonable choice between FFT and wavelet convolution and also between the reflection or the overlapping data approaches.

Finally, regarding the epoch, I am also concerned about which baseline period might be selected, since there is not enough time before the stimulus onset to set a suitable baseline period. Consequently, I wonder if you have any suggestions on alternative methods to normalize data in order to be able to compare frequency power between two groups.

Thank you in advance

PS. I would like to thank Michael X Cohen for his suggestion, although I must say I have already started the book a week ago, thus I am still at the very beginning. Nevertheless, I think is a really interesting and useful book. So thank you again for the effort of synthesize all that information in such a clear and accessible way.

If you are not primarily interested in the temporal dynamics, but rather, comparing two groups (patients vs. controls, I guess?) in terms of oscillatory responses, then I would go for FFTs and not wavelet convolution. The frequency precision will be much higher, and you wouldn't have to worry about edge artifacts. You could treat the data as if it were resting-state data: perform FFTs over successive time windows, average the power spectra together, and then compare across groups.