Dear John,
> 1) Could someone confirm this? I.e., is it legitimate to run
> analyses with
> unequal #s of trials through FEAT/FILM in this way?
yes.
>
> 2) Legitimacy aside, are there any important consequences to doing
> this?
> For example, would one predict that the condition with more trials
> would
> yield more stable/significant clusters of activity?
If I understand you correctly you will always bring contrasts of
correct vs incorrect trials to the second level. That means that the
precision of the cope will depend of the precision of the parameters
for both correct and incorrect trials. The resulting precision of the
cope will be highest when you have an equal number of correct and
incorrect trials (given a fixed number of total trials). The larger
the imbalance between correct/incorrect trials the poorer the
precision of the cope. It is easily realized if you consider the case
where we have only correct trials (i.e. no incorrect one). Then we
have no knowledge about the response to incorrect trials (i.e.
infinite variance) and hence we will infinite variance also for the
comparison to correct trials.
This kind of issues is exactly what FLAME is good at dealing with. It
ensures that you get an optimal weighting of the contrasts from the
first level (i.e. giving higher weight to first level copes with high
precision) and that your mixed effects variance (consisting of inter-
subject differences and of uncertainty of the first level estimates)
is correct.
> 3) Do you know of any citations from your group, or even from
> statistical
> texts on which FILM might be based, that I could reference in a
> publication
> to establish that it's valid to do this with FILM/FLAME?
I think that
Multilevel linear modelling for FMRI group analysis using Bayesian
inference
Woolrich et al.
NeuroImage
2004 vol. 21 (4) pp. 1732-47
would be the most relevant one. But Wooly might correct me on that
point?
Good Luck Jesper
|