Dear Kent,
> I (and perhaps the list) am hoping you could expand on your thoughts
>with respect to sources of 'global effects' in fMRI that should be adjusted
>with proportional scaling? My understanding of the origins of ANOVA and
>proportional scaling was for PET and SPECT imaging where radioactive counts
>could be dramatically different between successive scans - and since it was
>argued that noise and signal scaled proportionally - these within subject
>scaling procedures were employed. However, as PET camera's have become
>increasingly sensitive to local signal, they too have run into the
>local/global signal confound (as shown in Jesper Andersson's (1997),
>Neuroimage, 6(4), 237-44, when proportional scaling/ANCOVA models are used.
Absolutely. The initial issue was whether global effects were additive
or multiplicative. Quantitative rCBF-PET showed that they were additive
(calling for ANCOVA). However, this conclusion was restricted to global
effects
that were mediated by physiological changes. Global changes mediated by
scan-dependent changes in delivered radioactivity (SPECT) or sensitivity
(fMRI)
call for scaling.
You introduce a key issue, namely the "normalization-induced deactivations"
that
can occur with a "local/global signal confound", in the context of scaling:
>I am aware of some MR scanner related artefacts (namely rf) that can lead to
>occasionally 'bright' or 'dark' images (i.e., global effects) - but these
>are typically uncorrelated with the task making them less likely to be
>susceptible to the problems outlined by Aguirre et al in (1998) with
>proportional scaling. Also, I tend to see these artifacts less and less as
>MR equipment matures for fMRI protocols.
> I am familiar with your web site's description of proportional
>scaling (http://www.fil.ion.ucl.ac.uk/spm/doc/intro/#_E__Spatial_smoothing).
>In particular, the web site states that 'However, the issue of
>normalization-induced deactivations are better circumnavigated with
>experimental designs that use well-controlled conditions, which elicit
>differential responses in restricted brain systems'
> However, it is not always possible to design studies that have
>well-controlled conditions, such as the case of Go/No go paradigms (where
>the substantial motor activity from one condition (Go) is more highly
>correlated with the global signal than is the inhibitory activity associated
>with the 'No go' response. Thus, proportional scaling runs amuck in these
>situations.
> I guess my longwinded question comes down to this:
>If we assume we can control scanner related 'global' signal confounds, is
>there any reason to still use proportional scaling? (Even if the conditions
>in the paradigm are perfectly balanced)?
I think the principled response to this question is to reformulate it in
terms of model selection. In this context, the question is empirical; Is my
scaling model better or worse than my unscaled model? This would normally be
assessed with a likelihood ratio (i.e. p(y|scaling)/p(y|no scaling). I don't
think anyone has done this explicitly. Anecdotally, this could be done by
comparing the F statistics with and without scaling at a particular voxel
(although this is a ratio of transformed likelihood ratios). This approach
answers your question for each data set (y) and highlights the fact there is
no general answer. I suspect that with a high quality system and a well
designed
experiment, one could show no-scaling was better.
A key caveat here is drift removal. It is likely that many of the global
effects are low frequency. This means they can be removed by drift removal
without scaling. It would be interesting to pursue model selection with and
without low frequency confounds. In this case, I suspect scaling would
supervene in the absence of drift removal but not otherwise.
With very best wishes,
Karl
|