I am learning to work with EEG data in young and elderly subjects. The goal is to find differences in P300 (amplitude & latency) between both age groups during learning. I analyzed the data based on Grand Averages of both groups. As a confirmatory analysis I've tried to find the P300 peak for each subject individually and I performed analysis based on this peaks. Now I am interested in the variability in P300 latency in each trial, especially in elderly subjects, as the decreased P300 amplitude might be partially explained by averaging of more variable peaks. The single trial data looks very noisy and based on the fixed time window for searching P300 peaks, I am getting very different latencies for each trial. The topographies of single trial data also does not help - some looks ok, but others are just noise. Does someone know, how to handle P300 in single trial EEG data? Are classification methods for BCI going to work for this project?
I remember trying using single-epoch measurements with electroretinogram (ERG) recordings. Basically, ERG responses are electrical potential recordings from the neural activity in the retina. ERGs are usually also obtained using ensemble averages.
The single-sweep ERG amplitude measures were, as expected, also really messy due to random noise and eye movement artifacts that are normally averaged out when using ensemble averages. However, when using all 250 available individual epochs I had at the time, I could still show that ERG amplitudes slowly declined along the duration of the recording. So I took advantage of all available recordings by plotting all available amplitudes against time and performing linear regression. Linear regression is a particularly robust statistical method when a lot of data is available.
In your case, where you basically wish to do an analysis of variance on populations of people, I would try to pool all the data available per subject and determine the standard deviation, or a related variance parameter. Then doing that for all 20 normal subjects will yield a collection of variabilities. This could then be compared to the older population. I can't advice you on the most appropriate statistical test unfortunately, but for a starters you can simply plot the 2 point clouds and compare them visually. If it looks promising, you can do a t-test (which is likely not the most appropriate test, but it will give you a fair indication). If it all looks promising, you have to find a proper analysis of variance to compare the two.