Hi Matt,
For the equivalent of simple correlation, you want a design matrix
*with a constant* and with only one other column for the covariate you
are interested in.
You can then convert the t-value into an actual correlation
coefficient using "TASKS->Tools->Volume handling utilities->Stats
tools->Transform t maps" if you have the Volumes toolbox installed
(http://sourceforge.net/projects/spmtools)
With a design matrix that contains a constant, and two covariates,
then a test of [0 1 0] will report a t-value equivalent to the partial
correlation of the first covariate adjusted for the second. [0 0 1]
would test the second adjusted for the first. A test of [0 1 1] would
be testing the sum of the betas for the two covariates, which is not
equivalent to any standard correlation models (as far as I know) and
would e.g. ignore strong positive correlation of one covariate if
cancelled out by strong negative correlation of the other.
To the best of my knowledge, to test simple correlations of different
covariates requires fitting different design matrices (each with just
the constant and the single covariate of interest), but perhaps I'm
missing a trick here...
Hope that helps,
Ged.
Matthew Hughes wrote:
> SPMers,
>
> I have employed a multiple regression model for a second level analysis in
> SPM2 (the regression option without a constant), with two covariates (c1
> and c2), and an appropriate contrast image for each participant. I have
> three questions:
>
> 1. Does this model report standardised regression co-efficients?
>
> 2. To look at the effect of c1, without regressing out the effect of c2, I
> have entered a contrast vector of [1 0]. Should this contrast report the
> same effects as a simple correlation?
>
> 3. To obtain the effects of c1 with c2 regressed out, should I enter a
> contrast vector of [1 -1], or [1 1]?
>
>
> Thanks in advance,
>
> Matt.
>
|