Dear Christopher,
> I am in the perhaps unenviable postion of identifying CBF deficits in a
> clinical group in which the location of these abnormalities is not
> uniform. Therefore, I cannot compare this group as a whole to a
> normative database, but need to assess the extent of abnormality in
> each individual. In preparing to do this with SPM, several decisions
> arise:
>
> These are SPECT [HMPAO] data using very high-resolution, fan beam
> collimated gamma cameras. We were careful to administer the same
> activity of tracer to every subject, and to acquire a comparable number
> of counts for each scan. The point resolution in air for this system is
> 7-8mm FWHM. While this resolution is comparable to most PET systems
> for CBF imaging, my sense is that SPECT data are nevertheless
> considered inferior for the pruposes of modeling with SPM. The general
> recommendation is to use proportional scaling for SPECT-perfusion data.
> My understanding is that this reflects more "significant" global
> confounds with these data, which means a larger range of global mean
> values AND regional variance which is confounded with that global
> variability. I beleive I can test the degree to which this is the case
> in my dataset, but not in native SPM. I am not a matlab-maven, nor do I
> have one in my group [but I do know some]. Is there a relatively
> straightforward way for me to assess the appropriateness of ANCOVA for
> my data?
I would stick to proportional scaling. AnCova has only been validated for
parametric rCBF data. There is no need to confirm or disconfirm the AnCova
model for SPECT if you use scaling normalization.
> Less esoteric are the following: this study is designed to test the
> recovery of flow deficits in treatment. We have three scans for each
> subject over the period of treatment. We thus designed a databse of
> controls scanning each three times, so as to atake into account normal
> within-subject variability over time with this scanning system.
>
> The research questions are 1) how severe are the flow deficits for a
> given subject, and 2) how much do they change [if at all] over the
> course of treatment?
>
> To answer the first, I had originally thought to use only the baseline
> scan[pre-treatment]. This however leaves me with the problem of
> comparison of a single measure to a databse which contains repeated
> measures within subjects.
>
> My first option in SPM96 is, I believe, "compare groups:single scan per
> subject". My plan is to use all normal scans as one group, allowing
> that the repeated measures within some of these subjects will add to
> the population variance AS IF the normal database represented random
> sampling allowing for replacement of previously sampled subjects.
> However, this does not make the best use of the data as collected. If I
> am to account for as many confounds as possible, I would like to add in
> data regarding age, gender, perhaps extent of smoking, and model the
> within-subject variability separately from across-subjects.
>
> The next option would be "Multi-study: replication of conditions". In
> this, I would have two studies, one for each group. For all subjects
> with three scans, I am all set. Two questions:
>
> I believe I am required to eliminate subjects who do not have all
> datapoints with this model; for the case of controls in which I have
> only two of three repeated measures, I would propose to use a mean
> image of these [or perhaps enter one twice]. Is this valid?
> Second, the problem of using all three scans from a patient, in which
> we predict there is a time- and/or treatment related change in regions
> which are abnormal at baseline. One option is to use only the first
> scan, allowing the model to partition within-subject repeated
> variability differently from the comparison between these two groups.
> Another is to use all three time points, given that the change in
> abnormal regions, "normalization", is incomplete-- thus, even the last
> time point contains disease-related variance in some regions from the
> normal group. This appears to be an attractive option, *particularly*
> if the "plotting" option will, as it seems, allow me to discern the
> effect size at each timepoint separately for any region which is
> returned as significantly abnormal for the group of three scans. Can
> anyone tell me, a priori, if I am going to lose more in sensitivity by
> including all three scans [since the effect size, in some cases at
> least, will decrease with each scan], than I would gain in specificty
> by testing all three as opposed to a single, "diagnostic" image?
I think the crtical thing here is to ensure that the d.f. are less than
the number of subjects analyzed. This ensures that you are not conflating
variance due to repeated measures with inter-subject variability. In other
words it ensures you are using a proper random effects analysis. The
best way to do this is for:
(i) [simple] main effect of group
Analyze the 1st scan per subject (or the sum of two scans per subject),
using - compare groups:single scan per subject
(i) group x time interaction (i.e. time- and/or treatment related change)
Analyze 2 scans per subject using 'Multi-subject: different conditions'
(i.e. 2 conditions). Interactions would then be tested for using
contrasts like [-1 1 1 -1]. Note that here the subject or block effect
(i.e. effect of interest in the first analysis) is treated as a
confound so that each subject only contributes one 'difference'.
> The next order of complexity would be to include the covariates I
> mentioned. In the case of "Multi-study: conditions and covariates", I
> would again have two studies [normals and patient]. I would then
> include as covariates of no interest [confounds] age, gender, smoking,
> and time [1, 2, or 3], to specify regions which differ, as much as
> possible, only on the clinical status of interest.
This is absolutely right (although it would not model confound x time
interactions i.e. if you think age attenuates non-specific scan order
effects then you might want to include this interaction as another
confound)
> My question here to the framers of SPM, or anyone else who can tell
> me: how does this model differ from, for example, using
> "Multi-subject:conditions and covariates" in which I enter the same
> covariates immediately above BUT include clinical status [control or
> patient] as a covariate of interest?
There is no difference - as you probably guessed (unless you are
looking at interations - see above).
With best wishes,
Karl
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|