Tianhao,




> I met the similar problem recently.
>
> *** My fMRI experiments: ( several subjects, only one group)
> Task1:  R1 H1 R1 H1 ...(rest1, high1)
> Taks2:  R2 H2 R2 H2 ...(rest2, high2)
> (rest1, rest2 = no stimulus in this session)
>
> *** I'd like to compare the difference between high1 and high2
> over all subjects

> *** Three methods are used to get the differences in the first-level.
> (1) I built the design matrix only by one condition for every task
> and every subject, because there is implicit rest in SPM99. Then I
> try to view (H2-H1) over all subjects by using conjunction of contrasts
> subject1: [-1 1 0 0..]
> subject2: [ 0 0 -1 1..]
> subject*: ...
> But the result looks incredible.

Conjunctions can be very powerful. I would recommend the use of this
technique if you have small populations.

>
> (2) I also use exclusive masking to see the difference.
> conjunction of contrasts [0 1 0 0..],[0 0 0 1...],.....
> exclusive masked by  conjunction of contrasts [-1 0..],[0 0 01 ..]..
> The result can make sense.

Masking can be a very powerful tool, particularly in conditions where
conjunctions are invalid such as when you have conditions that are not
orthogonal. However, this method does not generate the same descriptive
stats as a conjunction. This method identifies areas in common across two
SPMs. Conjunctions take both SPMs into account when performing the stats. I
hope I made this clear. If anyone else has a better explaination please
jump in.

>
> (3) I built the design matrix by two conditions for every task and
> subject. and then I use the following conjunction of contrasts to get
> the (H2-H1):
> subject1:[-1 1 1 -1 0 0 ...]
> subject2:[0 0 0 0  -1 1 1 -1..]
> ....
> The result looks totally different from the first method, but
> resonable.

I have personally performed the comparisons of using an implicit versus an
explicit rest model. The two methods have always given me the same resutls.
You may double check you contrast on this. Otherwise others may like to add
somethin here.

>
> So my questions are:
> (1) If the second method is right?

See above.

>
> (2) Why there is difference between the first method and the third.
> It seems they should be the same. So I wonder how the SPM99
> implement the implicit rest.

See above.

>
> (3) Are there other methods to view the difference?

There are several methods to address your question. I am sure that many
would debate which is the best. I would recommend that if you have enough
subjects (say greater than 12) you use a random effects model using a
paired T-test. This method would allow for the direct comparison of the two
study conditions. Additionally, but less desirable, you could just modal
all subjects in a fixed effects model and apply a contrast to look at H2>H1
across subjects rather than doing random effects or a conjunction analysis.
So, for 5 subjects entered into the same design matrix with H1 first and H2
second, your contrast would be:

1 -1 1 -1 1 -1 1 -1 1 -1

>
> (4)What is the masking threshold for?

This sets the threshold for the SPM used in the mask. Thus, the two SPMs
can be masked at different levels. Again, remember, this is just like
taking two SPMs and overlaying them to find the areas in common.

>
>
> Any input are appreciated.
>

I hope that others will put up their ideas as well. As for the correct or
best way to do this, I will look forward to what everyone has to say.


Paul Laurienti