Hi Gabor,
maybe I misunderstand your question, but if you already have
cluster-level significance, then you do not need to consider the
uncorrected significance in each voxel. Cluster-level significance is
corrected significance. The uncorrected significance that matters here
is the one that defined the cluster. All that needs be said about this
significance is that it should be chosen a priori, not that it should
be stringent. Hypothetically, you could have chosen this
cluster-defining uncorrected significance to be p=.20, and the cluster
significance would still be corrected, if computed appropriately (i.e.
if you have a good way to compute the null distribution of the cluster
statistic at this defining threshold).
There are discussions in the literature about using cluster extent and
cluster peak jointly to improve on the power of the test. These tests
usually rely on Montecarlo simulations to establish rejection
thresholds, and are available in packages such as AFNI.
Best wishes,
Roberto Viviani
University of Ulm, Germany
Quoting Gabor Oederland <[log in to unmask]>:
> Thanks for the responses so far! The thread Chris has linked to is
> actually the one I was refering to as well :-) The reason I asked is
> that there is no real discussion concerning uncorrected
> voxel-threshold IN COMBINATION with corrected cluster-threshold
> (corrected as provided by SPM output, so no "corrected" pre-definied
> value like k > 10).
>
> In the salmon poster, the authors "argue that relying on standard
> statistical thresholds (p < 0.001) and low minimum cluster sizes (k
> > 8) is an ineffective control for multiple comparisons." In a
> study of ~ 15 subjects and standard resolution (after normalization)
> the FWE-corrected cluster-threshold should be something to ~ 50 -
> 200 voxels, so much higher.
>
> The SCAN paper by Bennett and colleagues (2009) does provide any
> definite answer about corrected cluster-threshold either. They say
> "It is possible to use the combination of a P-value and a cluster
> size in a principled way, but it requires computing the proper
> values for each and every analysis. The cluster size criteria can
> change quite substantially from dataset to dataset. Further, it can
> be the case that required cluster sizes become so large that
> legitimate results with a smaller volume are missed."
>
> So, what about this combination of voxel-p = .001 and cluster-p =
> .05 FWE? Are the limitations (different cluster-threshold in
> different studies) acceptable or not?
>
>
> Best,
>
> Gabor
>
|