Statistics for coherence in a within-subject experiment

[Robert Oostenveld]

Hi Jurij

The parametric tests that you suggest are not implemented in  
FieldTrip. In the clusterrandanalysis function we use a randomization  
approach for statistical testing. It is a non-parametrical approach,  
and the test statistic consists of a parametrtical statistic per time- 
frequency-channel, which is clustered and summed over neighbouring  
timepoints/frequencies/channels. Not the single-sample statistic, but  
the clustersum is used for the statistical test, thereby increasing  
sensitivity and controlling for  multiple comparisons. The single  
sample (time-frequency-channel) statistic has to be chosen to match  
with the experimental design (e.g. F or F) and the randomization/ 
permutation of the data between the subjects and conditions has to be  
adequate (in your case you want to shuffle data between conditions  
but only within each subject). That determines the statistic to use.  
I suggest that you go over the clusterrandanalysis tutorial at
http://www2.ru.nl/fcdonders/fieldtrip/doku.php? 
id=fieldtrip:documentation:tutorial:clusterrandanalysis

best regards,
Robert


On 19 Jun 2006, at 17:31, Jurij Dreo wrote:

> Hello!
>
> Thank you for your answers. Here are my replies to them:
>
> To avoid any misunderstandings here is again a summary of my study  
> design
> - I have 11 subjects.
> - Each of them performed 6 tasks related to working memory.
> - Each subject performed anywehere from 25 to 60 repeats (trial) of  
> each condition (the standard number of trials per task was 30 - but  
> some subjects performed the entire experiment twice (ergo 60  
> potential trials) .... and in some some of their trials are  
> unusable - artifacts ... ergo 25 min number of possible trials)
>
> Given my study design I thought something like a one-way ANOVA for  
> correlated samples would be in order.
> (as I am comparing 6 conditions within-subjects as every subject  
> performed all sets)
>
>
> -----------------
> > In your case you have observed a "value" in 11 subjects, over 6
> > conditions. That value happens to be coherence, but could as well
> > have been something else. Since you have manipulated the conditions
> > with each subject, you can use a dependent-samples test: the  
> observed
> > values depend on the subject (an example of a dependent-samples test
> > is a paired-t test).
> -------------------
>
> I´m not sure what you are saying here.... I agree with the  
> dependent samples idea... but I cannot preform a t-test ... ie.  
> more of them as that would inflate my p-rate overall.
>
> I thought that I would first do an ANOVA and then something like  
> the Tukey HSD Test to cimpare individual conditions
>
> --------------------
> > If the number of trials varies between subjects and/or sessions, it
> > will be preferable to transform your coherence values to z-scores
> > manually prior to submitting the values to clusterrandanalysis.
> --------------------
> Yes I agree that would need to be done manually.
>
> -----------
> > Now about the 6 conditions, am I right that you have a 1x6 factorial
> > design, or is there more strucure in the design? Do you have a
> > specific hypothesis about the different conditions? If you only want
> > to reject the null-hypothesis "the observed value (coherence) is the
> > same (more accurately: stems from the same distribution) in all my 6
> > conditions", you would use an omnibus F-statistic, i.e.
> > 'depsamplesF'. It might be that your hypothesis is more specific, or
> > also in case you find an omnibus effect, then you would probably  
> want
> > to perform explicit tests between two subsets of your 6 conditions
> > using a 'depsamplesT' test.
> >
> ----------
>
> I want to contrast all conditions at once (f -test anova) AND also  
> each combination of conditions (say a post- ANOVA Q-test)
>
> But Im still not sure which statistic to use and even if the one I  
> use will be able to deal with the fact that I have a different  
> number of conditions per condition per subject
>
> As far as I understand it... we must use some kind of ANOVA here...  
> A simple "paired" t-test will not do... as we are comparing more  
> than 2 conditions at once and I do not want to inflate my p-rate.
>
> Thank you for your help!
>
> Regards,
> Jurij Dreo
>