On Thu, Feb 19, 2009 at 3:31 PM, Michael Browning
<[log in to unmask]> wrote:
> . Using a whole brain standard cluster correction (Z 2.3, p<0.05) this
> reveals clusters in relevant areas.
> In order to demonstrate the form of the interaction I've used featquery to
> extract the mean % signal change for each task condition from the identified
> clusters. My concern is the analysis that I could justifiably perform on
> this extracted data. If I understand the non-independence argument (ie Vul
> et al) correctly then performing an ANOVA using (say) SPSS on the extracted
> data is likely to overestimate the interaction term (attention x emotion x
> group).
Unless you're explicitly interested in the effect size in terms of the
actual F value, then Vul's argument is irrelevant. You have already
demonstrated a significant 3-way interaction using properly corrected
voxelwise tests. There is no need at all to re-run the same 3-way test
on the extracted data.
Would I be justified in performing the standard explanatory analyses
> that would often be used in non-imaging data-- e.g. demonstrate that within
> a specific group the attention x emotion interaction is significantly
> positive, t-tests comparing activation to fear between groups etc?
Yes - standard posthoc tests (labelled as such) such as pairwise
comparisons can be used to determine the source of the interaction.
Because these are posthoc, Vul's arguments don't come into it. And
anyway, *none* of Vul's comments had any relevance to tests of
significance, nor of posthoc estimation of effect size. They were only
relevant to extrapolating estimates of population effect magnitude
based on cluster-mean effect magnitude, something that I'm not aware
of anyone actually doing or suggesting. Straw Man comes to mind...
> Or is
> this also likely to give inflated results?
As long as you describe the analyses on extracted cluster means as
posthoc tests to determine the source of the significant 3-way
interaction, you have no problem.
-Tom Johnstone
> If you don't think that a
> straightforward analysis of the extracted data is possible are there any
> other strategies?
>
> Thanks for the advice,
>
> Mike Browning
>
|