[Eeglablist] Statistical test

[cyril pernet]

Subject:
Re: [Eeglablist] Statistical test
From:
Makoto Miyakoshi <mmiyakoshi at ucsd.edu>
Date:
25/01/2017 01:53

To:
ali zahedi <ali.zahedi.bham at gmail.com>
CC:
EEGLAB List <eeglablist at sccn.ucsd.edu>


Dear Ali,

It is not totally clear, but the attached figure looks as if you have 2 
x 2 x 2 design. In this case, what you need is 3 way ANOVA. Of course, 
you can apply t-test from the beginning as planned test, but that can be 
recommended only when a small subset of the t-tests are of your interest 
(you don't want to repeat it too many times because of multiple 
comparison correction issues.)

So, why don't you try repeated measures 3 way, 2 x 2 x 2 ANOVA.
Makoto


--> just to add that this can easily be done using the LIMO EEG toolbox 
(model the 6 conditions per subject then do an ANOVA across subjects)

Cyril


On Mon, Jan 23, 2017 at 10:05 AM, ali zahedi <ali.zahedi.bham at gmail.com 
<mailto:ali.zahedi.bham at gmail.com>> wrote:

    Dear List,

    I am working on EEG signal processing and I need to apply
    statistical test on my findings (paired t-test). However, I am not
    confident in selecting the significance level which is usually set
    as 0.05 or 0.01, and I am not completely sure if the paired t-test I
    have used is correct or not.

    I have 20 EEG datasets collected from 20 individuals, and want to
    classify relaxation versus stress. For pre-processing the data I
    applied a chain of 3 different filters (F2,F3 and F4 as shown in the
    attached figure); which resulted in 8 different pre-processing
    modes. After filtering I used 2 different classification method (Red
    and Green bars in the attached figure) on the preprocessed data to
    see which classification method works best. Furthermore, I want to
    see if different filters that I have used have significant effect on
    the performance of the classification methods.

    As you can see I have three different factors which may affect
    performance of the 2 classification methods.
    In order to evaluate the obtained results I performed a paired
    t-test on the results.
    In the first step, I wanted to check if performance of the 2
    classification methods is different significantly. Therefore, I
    divided my data into 2 groups of each having 160 samples (20
    subjects * 8 modes) and performed a paired t-test.
    In the next step, I wanted to check if different filters have
    significant effect on the performance of the classification methods
    or not. Therefore, I divided the data into the following paired groups:

    Pair1= Group 1: F2 is applied (F21) Group 2: F2 is not applied (F20)
    Pair2= Group 1: F3 is applied (F31) Group 2: F3 is not applied (F30)
    Pair3= Group 1: F4 is applied (F41) Group 2: F4 is not applied (F40)

    Each of the groups in the pairs consist of 160 samples (20 subjects
    * 2 classification methods * 4 modes).

    In my test I considered p<0.05 as the significant level. I heard
    that 0.05 is not always correct to be selected and it might be
    divided by the number of measurements/repetitions due to
    Bonferroni correction. For example here 0.05/3.
    Is Bonferroni correction necessary to be applied? and what would be
    the significance level here? Is the paired t-test correct to use here?

    I really appreciate it if you could help me with this.

    Regards,
    Ali



-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20170127/9db5f5d6/attachment.html>
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: not available
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20170127/9db5f5d6/attachment.ksh>