Tom,
Thanks for your reply.
Even if I use Puncorr < 0.001 as threshold to generate mask, all clusters
obtained with explicit mask t-test analysis have exactly the same fdr
probability. ie. in my case, all clusters had fdr probability of 0.043.
However, without the explicit mask the clusters had different fdr
probabilities ranging from 0.07 to 0.159.
Anusha
> Anusha,
>
>> I have a question regarding FDR probability.
>>
>> Inorder to study the FA value differences between two groups I perform
>> two-sample t-test under basic models in SPM2. The tmaps look good but
>> non of
>> the clusters were significat at FDR/FWE < 0.05. So to reduce the
>> multiple
>> comparisons,
>>
>> 1. I generate t map with Puncorr < 0.05 and k = 50 and save it as an
>> image.
>> 2. I re-do the basic model-two-sample ttest, with the t-image generated
>> as
>> explicit mask.
>>
>> The tmap and the t/z values are exactly the same as before and the
>> FDR/FWE
>> have lower probality values, as expected. My question is;
>>
>> 1. using a mask as in (1) is it acceptable ? My reason for using such
>> a mask is that I am only interested in testing those volxels with
>> Puncor < 0.05.
>
> No, I'm afraid it's totally invalid. You are in essense throwing away
> all the data w/ Pucorr<0.05 and pretending like it never existed. Put
> another way, you are using your statistic image itself to define what
> regions to make inference on, which is a no-no; by assessing the
> Puncorrs you have gone and done many tests and you have to account for
> this.
>
>> 2. all clusters listed obtained with explicit mask as in (1) have
>> same FDR values. why is this?
>
> I would have thought they changed some; do they not change a bit, even
> if you increase your Puncorr thresh? (Say to 0.001)?
>
> That they didn't change at all could be attributed to FDR's
> adaptiveness; FDR inferences are determined by the emperical
> distribution of P-values; by throwing out big P-values you've
> truncated the distribution but possibly haven't changed the shape of
> the tail of the distribution that much.
>
> However, it is possible that you've found a bug. If you see
> absolutely no change in FDR P-values with ever smaller masks, then
> there would appear to be a problem. Let me know if that seems to be
> the case.
>
> -Tom
>
>
> -- Thomas Nichols -------------------- Department of Biostatistics
> http://www.sph.umich.edu/~nichols University of Michigan
> [log in to unmask] 1420 Washington Heights
> -------------------------------------- Ann Arbor, MI 48109-2029
>
|