Dear Katya,
>
>Thanks very much for your helpful reply. I have still one doubt, though,
>whether we are talking about teh same thing. I agree, it is obviously
>better, if I have an area of interest from previous work and use those
>coordinates to search for significance in the area.
Sure. I suspect that you are perhaps right that we might not be
talking about the same thing!
>But what I meant with
>using the nearest cluster approach in SVC in SPM, is whether it is
>legitimate to take a cluster which came up in the same analysis as n.s.
>corrected (with a generous p-value, of say p < .05),
The first thing to say is that, if by n.s. corrected, you mean not
exceeding significance after correction for multiple comparisons over
the whole brain, then you should probably just ignore this column of
'corrected' results given by SPM. If you have a prior anatomical
hypothesis, then such a harsh correction is really not appropriate,
and there may be lots of important results which you should be
reporting but which do not exceed this high threshold. I think that
this is probably what Federico Turkheimer meant in his message (not
that no correction at all is required!).
> then click on this
>particular cluster which I know is subthreshold (and which is more or less
>in an area which I would hypothesise, for example somewhere in PFC) and do
>an SVC with the nearest cluster option to those same coordinates.
If you have a prior anatomical hypothesis, then it is this hypothesis
that you need to reduce the search volume, and thus reduce the
multiple comparisons problem. You can't just click on any old
cluster, and then use the peak coordinates here as the coordinates in
a 'nearest cluster' analysis (this is not a proper 'nearest cluster'
analysis). If you are going to use this type of analysis, you are
going to have to be pretty sure of exactly what you are doing and
why. In particular, you must decide on the coordinates for the
nearest cluster analysis a priori, rather than just using the
coordinates of a cluster which you know, a posteriori, to be
singificant. Even having done this correctly, you can only use the
cluster-level statistics, not the voxel-level statistics.
> In this
>case, the p-value changes to significant because I have restricted the
>search to the cluster itself which I know is there, even if not significant
>if corrected for the entire volume. Would this be legitimate?
>
No (if I understand what you are proposing correctly).
The cluster itself does not constitute your prior hypothesis about
what regions you expect to be active. The fact that a cluster 'is
there, even if not significant' (i.e. not exceeding the threshold for
correction for the whole brain) cannot be used as an independent way
of selecting a small search volume, because it is not independent.
You know that these voxels will show an effect, even if the null
hypothesis is true, because it is on the basis of such an effect that
they have been selected. The underlying principle is that your
method for selecting voxels to search through must not bias you
towards observing an effect even under the null hypothesis.
>thanks;
You are welcome. I hope that we are now talking about the same thing.
Best wishes,
Richard.
>
>At 15:02 06/06/01 +0000, you wrote:
>>Dear Katya,
>>
>>>If I don't get any suprathreshold voxels which survive teh corrected p
>>>values for any p < .001, .005, .05., but I have sub-threshold voxels in
>>>apriori hypothesised anatomical brain regions, can I then use the SVC option
>>>of SPM 99 in precisely those coordinates for "nearest cluster"(that gives me
>>>always significance), is that legitimate?
>>
>>The most conventional approach here would certainly be to couch your
>>a priori hypothesis in terms of a sphere centered on a given voxel
>>(of which you already know the coordinates) and with a radius of,
>>say, 5 or 10mm depending on how large you expect your area to be and
>>how sure you are of its location in a new subject group. SPM then
> >allows you to enter this sphere in the new analysis. The p values
>>for height are now appropriately corrected for this small volume. It
>>might be even better if you can make an image of your region of
>>interest (from the previous study) and then use this for the small
>>volume correction.
>>
>>However, the nearest cluster method is theoretically legitimate,
>>although understanding what the inference really means can, under
>>some circumstances, be a little difficult. The cluster level
>>statistics tell you the probability with which a given cluster would
>>equal or exceed its observed size, given the null hypothesis. If you
>>have used some independent method to choose a single cluster, then
>>you could quote the uncorrected cluster level statistics. After all,
>>under the null hypothesis, there is no reason why the nearest cluster
>>to your chosen voxel should exceed the expected size any more than
>>any other cluster does, i.e. the your method for choosing the cluster
>>shouldn't bias the size of the cluster under the null hypothesis.
>>
>>The complication of inference comes with considering a thought
>>experiment in which the nearest cluster lies rather obviously outside
>>the general area that you are interested in. Adhering strictly to
>>your statistical method, this has to be reported as a positive
>>result. After all, your hypothesis wasn't set up for a given area
>>of interest, it was set up for the nearest cluster to a given voxel.
>>If you are tempted to say 'oh well, in this case the nearest cluster
>>obviously falls outside my area of interest, so I won't report it'
>>then you would appear to be entering murky water statistically,
>>because you are not really prepared to accept the 'nearest cluster'
>>under all circumstances. Perhaps a real statistician might comment
>>on this.
>>
>>Obviously it goes without saying that you can't report the
>>uncorrected statistics for height in your 'nearest cluster'. These
>>are bound to be above threshold, even under the null hypothesis,
>>because that's how a cluster is defined!
>>
>>Personally I would stick to the more conventional SVC outlined
>>earlier, unless there was some very strong reason for using the
>>'nearest cluster' approach. That's partly because I am in the habit
>>of quoting statistics for height. On the other hand, Karl expressed
>>disappointment recently that the 'nearest voxel' approach isn't used
>>more frequently.
>>
>>Best wishes,
>>
>>Richard.
>>--
>>from: Dr Richard Perry,
>>Clinical Lecturer, Wellcome Department of Cognitive Neurology,
>>Institute of Neurology, Darwin Building, University College London,
>>Gower Street, London WC1E 6BT.
>>Tel: 0207 679 2187; e mail: [log in to unmask]
>>
--
from: Dr Richard Perry,
Clinical Lecturer, Wellcome Department of Cognitive Neurology,
Institute of Neurology, Darwin Building, University College London,
Gower Street, London WC1E 6BT.
Tel: 0207 679 2187; e mail: [log in to unmask]
|