Thanks Roberto and others so far - I should point out that I agree
that in many cases (e.g. below) you *could* do a statistical
comparison, but people often don't and yet still want to draw
conclusions as if they have . . .
On 03/08/2007, Roberto Viviani <[log in to unmask]> wrote:
> Hallo,
>
> : 1) Split patients into 2 groups on the basis of their cognitive score
> : (a "good" group and a "bad" group), compare each group with its own
> : matched control group, and present the SPMs side by side. Repeat for
> : other cognitive variables.
> :
> : This seems unsatisfactory to me because a) you don't directly
> : correlate the score with volume and b) more importantly you don't
> : directly compare the two patient groups, and the visual comparison of
> : the 2 SPMs is not a statistical comparison.
>
> I agree with you and Philip Saeman and Marko Wilke that this is not a good
> technique. The reason is that dichotomizing leads to loss of power, and
> possibly to other bad things. Some reviewers seem to like it however,
> possibly because they used it themselves, and want all the world to do the
> same, like certain drug addicts. Reference on this issue:
>
> MacCallum RC, Zhang S, Preacher KJ, Rucker DD (2002). On the practice of
> dichotomization of quantitative variables. Psychological Methods 7:19-40
>
> These authors conclude that "dichotomization is rarely defensible and often
> will yield misleading results". The practice is bad even if you pre-selected
> the participants from a large database so as to have very good and very bad
> ones before carrying out the scans. I do not agree, however, that you cannot
> do a statistical comparison: of course presenting the SPMs side by side
> isn't any such thing, but you could produce the SPM of the interaction
> between cognitive score group and patient status.
>
> <snip>
> : 3) Enter e.g. two cognitive scores as regressors in a single model,
> : and test whether either slope is significantly different to zero at
> : each voxel.
> :
> : This is asking a different question but the interpretation seems
> : limited: you can show regions in which score A correlates with volume,
> : having adjusted for score B (and vice versa), but you cannot conclude
> : that A is associated with those regions significantly more than B,
> : although I think this is also what people tend to assume.
>
> Why can't you? It is called a linear contrast (see Draper and Smith, Applied
> Regression Analysis, 3rd ed., section 9.1).
>
> If you are concerned with conclusions, remember that regressing on a
> non-experimental variable leads to inference that is conditional on the
> observed covariate.
>
> Roberto Viviani
> Dept. of Psychiatry III
> University of Ulm, Germany
>
|