Hello all (particularly Tom and Jeanette)
thanks for sharing. This is most helpful!
Best regards,
Cornelius
On Thu, May 17, 2012 at 2:21 PM, Jeanette Mumford
<[log in to unmask]> wrote:
> Hi,
> see below....
>
>> I have a follow up question. If I understand you correctly, you wrote that
>> randomise would only be used for within subjects design. I found this on the
>> randomise website:
>>
>> Using randomise with a GLM that corresponds to one of the following simple
>> statistical models will result in exact inference:
>>
>> One sample t-test on difference measures
>> Two sample t-test
>> One-way ANOVA
>> Simple correlation
>>
>> For the two sample t-test, it is not specified whether they talk about the
>> paired or independent samples t-test
>
>
> That would be a non-paired data situation.
>
>>
>> Could it be that both are valid with randomise or, if only for within
>> subjects designs as you say, the paired samples t-test would be appropriate?
>
>
> As I said and as Tom explains in a recent email for a repeated measures
> ANOVA (within the last 2 weeks on the list), you should subtract the pairs
> of data and run a 1-sample t-test on the differences. This ensures that the
> permutation is done properly in randomise.
>
>>
>> They further say that using randomise for most other design will give
>> approximately exact statistics. Also, do you think that the exchangability
>> issue is equally relevant for non-timeseries data?
>>
> Exchangability is an issue whenever you have a covariance structure in your
> data and in other cases where permuting the data changes the distribution in
> some way. In this case the data within subject are correlated and you can't
> mess with that, which is why it is suggested that you do it this way.
>
> Cheers,
> Jeanette
--
Cornelius Werner
cornelius.werner<at>gmail.com
|