Dear Anderson,

Thank you very much for explaining that so clearly! Yes, it is very useful, and I'll give those set-ups a go. 

Many thanks,

On 5 Aug 2017, at 04:06, Anderson M. Winkler <[log in to unmask]> wrote:

Hi Lucia,

Please see below:

On 3 August 2017 at 19:19, Lucia M Li <[log in to unmask]> wrote:
Dear FSL users & experts,

I have 20 subjects, who were split into two intervention groups (n=8 drug, n=12 placebo), and underwent a task fMRI at two time points (pre and post intervention).

I was hoping to test for whether there is a:
- a main effect of intervention (the between subject factor)
- a main effect of time (the within subject factor)
- an interaction between the two factors

Can I suggest you change the way as the analysis is approached? The most interesting effect is the interaction time by group, in that it will show that the slopes over time differ when the subjects took the drug compared when they did not.

The main effect of time would collapse the two groups, but that isn't interesting because one of the groups took the drug, whereas the other did not. It's different than in an observational study where over time we may be investigating the progression of a disorder.

The main effect of group would collapse the two timepoints, but that isn't interesting either because in the first timepoint nothing was administered (either drug or placebo), such that there is no point in mixing the two.

You can still do the analysis and test the interaction (definitely the one you would want to report), and also the main effects of group and time, but if the interaction is significant, then these main effects are further less interesting, because the effect of one depends on the other, and vice-versa. In any case, there is a worked out example at this earlier post. Note that there are two designs, one for the within-subject effects and interaction, and another for the between-subject effects.

If the allocation of subjects into the two groups was random, and you'd like to show that there is no residual (incidental) difference between them after the randomisation, a simple two-sample t-test using the baseline is sufficient.

Also, if the subjects were randomised into treatments, you can test only the second timepoint, comparing the two groups, while including the first timepoint (baseline) as a continuous, voxelwise nuisance regressor in the design matrix, thus eliminating the need for a repeated measures design. This would be assembled as a two-sample t-test with additional covariate, following this example from the GLM manual.

Now trying to answer the questions below:
I thought that the 2-way mixed effect ANOVA might be the way to do it ( but am a little confused as to how exactly to set it up.

In the example given:
- does 'run' signify a factor with different levels or two runs done to acquire more data?

'Run' means timepoint. This is testing the main effect of time. It can also be seen as a within-subject factor with two levels. It doesn't mean the two runs were just for having more data (e.g., to improve power or SNR).
- if the former, am I right in thinking that C1 (run effect) would give me the effect of the within subject factor?

Yes, that's right.
- am I right in thinking that C2 would give me the effect of the interaction between of the two factors?

Yes, C2 is for the interaction group by time (or group by run), which is your most interesting contrast if you use this strategy, or if your subjects were not randomised (in which case you wouldn't use the baseline as nuisance).
i.e. there is no contrast which would give me the effect of the between subjects factor?

Exactly, in this design it isn't possible to test the main effect of group. That requires a different design, that is in the spreadsheet in the earlier post linked above.

Hope this helps!

All the best,



Many thanks in advance for your help!

Kind regards,