Dear experts,
We are using SPM to compare the PET scan of a single patient against a large database of normal subjects. This has been going on for a long time, and we have been updating a batch that was originally written in SPM99. Currently we use SPM5. I'd guess that the correct statistical test to perfom in this case would be a z-test, i.e.: seeing if a subject is significantly outside of the controls' distribution.
However, for reasons that nobody here seems to know, the "routine" batch uses a two-groups t-test with the variance set to "unequal". Obviously, I can't even try to guess how you could perform an unequal-variance t-test if one group has only a single data point in it. The thing is that SPM99 and SPM5 provide a result. SPM8 instead gives error. (when you pick "results" matlab prints red lines mentioning the inability to set axes to NaNs and Infs.) As you would expect!!! What's the variance of a single data point??
So, what I did to test everything out was taking an atlas, adding a random number with mean 0 and sigma 1 to the whole area and saving 100 such random "scans". Then I generated a scan adding 5 to this atlas. I first performed a two groups t-test with equal variance in spm5 and spm8. Both versions give a design matrix which is [1 0;0 1] and the spmT image is an uniform "5". Perfectly, as expected! If I select "unequal" instead I get an undisplayable design Matrix ("image CData cannot be complex") and I cannot even define contrasts to compute the con or spmT files in SPM8. In SPM5 I get a uniform spmT with a value of 16.19, and the design Matrix is [0.99 0;0 3.30]. I've briefly looked at the database of past scans and it initially appears that all of ours design matrixes have these values.
Does anybody have a clue about how SPM5 (and 99) were computing t-values under these (admittedly weird) assumptions?
Thank you,
Luca
|