[Eeglablist] Grand Average

[Steve Luck]

I agree with Steve Politzer-Ahles about this.  Giving some subjects more weight than others could lead to bizarre results, whereas differences in number of trials (and hence differences in measurement error) are likely to simply reduce statistical power (and only modestly in typical situations).

More generally, the idea of using a sample of subjects to estimate the parameters of a larger population would be greatly distorted by giving some subjects greater weight than others.

Steve Luck

> From: Stephen Politzer-Ahles <politzerahless at gmail.com>
> Subject: Re: [Eeglablist] Grand Average
> Date: October 22, 2012 5:18:30 PM PDT
> To: Alberto Gonzalez V <vilanova5 at hotmail.com>
> Cc: <eeglablist at sccn.ucsd.edu>
> 
> 
> Hello Alberto,
> 
> There may be discussion of this issue in Luck (2005) and/or Handy (2004); if there is, you can ignore what I say and check those instead.
> 
> My assumption, though, is that the reason we typically average them the way we do, instead of using a weighted average, is that more epochs does not necessarily mean better data. It's true that an insufficient number of epochs (and/or subjects) will make the ERP noisy. But once you reach a certain point, adding more epochs does not make the data a lot better (see Luck's (2005) discussion of the signal-to-noise ratio). Each subject is meant to be one datapoint, so once a given subject reaches the threshhold after which she has "enough" trials to make a good ERP, then it's fair to make that subject a datapoint.
> 
> Also, of course, the characteristics of the ERP components you are interested in are likely to differ across subjects; some people may have a bigger P300 or N400 or whatnot overall. There is not necessarily a straightforward relationship between this and how clean their data are (i.e., it's not necessarily the case that someone who has a bigger/smaller P300 also happens to blink more/less during the experiment). Thus, by weighting subjects differently because of how many clean epochs they happened to have, you may be inadvertently biasing your grand averages towards certain individuals. At least when you treat all subjects equally, you are neutral as far as that is concerned.
> 
> Those are just my impressions; I don't know if there is published literature discussing this topic, and if there is then it of course is a better reference than my impressions!
> 
> Best,
> Steve
> 
> On Mon, Oct 22, 2012 at 7:51 AM, Alberto Gonzalez V <vilanova5 at hotmail.com> wrote:
> Hi to all,
> 
> I have a question about ERP methodology. Consider that we record the EEG during and task  in 3 subjects, then we do the averages  ( considering that the task has 60 epochs):
>         Subject 1  did a perfect task, so we did the average with 60 epochs.
>         Subject 2 had some problems during the recording, and the average was done with 40 epochs.
>         Subject 3 had only 20 epochs, but we think that it´s enough and did the average.
> 
> So the Subj 1 has all the epochs =1, Subj 2 has = 2/3 of the epochs, and Subj 3 has only =1/3. But in the grand averages we treat them as they had all the epochs (=1). Isn't better to give each subject a proportional value (considering it's number of epochs) in the grand average(something like:  ([Subj1*1]+[Subj2*2/3]+[Subj3*1/3])/2)?.
> 
> Thanks for your time!!!








-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20121023/c8575e34/attachment.html>