Stephan,
> The Friston et al. (1999) NeuroImage paper on using multiple conjunction
> analyses in a fixed-effects analysis suggests that such analyses, "might
> be a useful prelude to studies of large numbers of subjects that are
> required by random-effects analyses". However, couldn't one also use a
> regular fixed-effects analysis (not involving conjunctions) in the same
> manner, as a pilot study preceding a larger random-effects analysis?
I think the concern with a standard contrast in fixed effects analysis
is that the significance can be driven by a single subject. Using a
conjunction ensures that all subjects show a response.
> For example, one could do a pilot study with a small number of subjects,
> run a fixed-effects analysis, identify regions of activation, and then use
> these activation areas as regions of interest for a subsequent random
> effects analysis with a different group of subjects. The advantage would
> be that one could use a less stringent criterion for the random-effects
> analysis for the regions of interest (say, p < .005 vs. p < .001).
With SPM99's ability to get corrected p-values for subvolumes, I don't
see the need to muck about with uncorrected thresholds.
Do a pilot study, generate some hypotheses, specifically, generate a
mask image that defines your a priori region of interest; then get
corrected p-values for your subvolume... and your done. You've gained
power by focusing on a smaller region yet you're still controlling for
the multiple comparions within that region.
> One drawback I could see with using a fixed-effects pilot study followed
> up by a random-effects study is that the fixed-effects subjects could have
> been allocated to the random-effects study in the first place, which would
> increase the power of the random-effects analysis. It's unclear whether it
> is better to just increase the N for the random-effects analysis or do a
> two-stage approach with a fixed-effects hypothesis-generating stage
> followed by a random-effects analysis with an independent group -- any
> ideas?
This is a good point; I'm afraid I don't have a pat answer. The
issues include how much power is gained by the localization obtained
from the pilot study versus how much is lost from the final
random-effects study, also the size of the region will play a
role. (The smaller the region, the greater the power you can gain from
just focusing on that area.)
Hope this helps.
-Tom
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|