Hi all,
as this is my first time working with fMRI data and SPM, there are a few things with the contrasts I am struggling with that i would like to get some advice on. I have tried to structure some of my questions below!
As i have received first level data, all my questions are regarding second-level (group) analysis, which contrasts make sense and specifically, how to interpret interaction effects.
The task i am analyzing is an emotion regulation task, for which i have set up two separate models to look at attending emotional pictures (model 1) and regulating emotions (model 2).
My aim is to investigate whether there are group differences in these tasks, comparing 3 groups: depression with childhood trauma (MDD+CT), depression without childhood trauma and healthy controls.
The two models have the following designs:
- a 3x3 factorial design with factor 'group' with 3 levels (MDD+CT, MDD-noCT, HCs) and a within-subject factor of 'valence' with 3 levels (positive, negative, neutral)
- a 3x2 factorial design with factor 'group' (3 levels again) and 1 within factor 'valence' with 2 levels (this time only positive and negative; there was no regulating neutral images)
For simplicity, lets consider the first model.
The design matrix looks like this: G1V1 G1V2 G1V3 G2V1 G2V2 G2V3 G3V1 G3V2 G3V3 (+ covariates)
I specified these using the full factorial option in SPM12; in which the within subject factor was labeled 'No' for Independence to indicate repeated measures.
Now, SPM automatically returns contrasts of main effects (F contrasts) and specific T comparisons as well as interactions.
This is where I am struggling.
SPM returns T contrasts comparing adjacent groups (e.g. MDD with and without childhood trauma).
1a) Why doesn't it also return group comparison of level 1 vs level 3 (MDD+CT > HC)? And why is it only testing in one direction?
I have specificied this contrast myself using the following contrasts: (T) 1 1 1 0 0 0 -1 -1 -1.
In my understanding, this is assessing differences of MDD+CT > HC (particularly, if MDD shows higher brain activation compared to HC), regardless of the valence. Is this correct or am I misunderstanding the contrasts alltogether?
1b) Also, would it make sense to define the T contrast the other way around: HC > MDD+CT: -1 -1 -1 0 0 0 1 1 1, to find brain regions that have a higher bold response in HC then MDD+CT?
1c) What if i would want to compare the MDD groups together against the HC group? Would I do that using -0.5 for the MDD regressors and 1s for the HC regressors ([-.5 -.5 -.5 -.5 -.5 -.5 1 1 1])?
2. I was wondering however, if maybe i am missing group differences because the different valence conditions could cancel each other out.
For example, if the MDD group has more activation during negative images compared to controls, but less during positive - would these effects "cancel" each other out, and show no difference between groups?
Is this something that we would catch with the interaction contrasts?
Regarding the interaction contrast, I also struggle to interpret these correctly.
3) When the F contrast, main interaction effect of group x valence, returns some significant clusters, how do you interpret this?
SPM also automatically returns some interaction T contrasts, but none of them return significant, so what drives the main effect?
What kind of post-hoc T-contrasts would need to be ran? And then how do you interpret these? In my experience with 'normal' data, plotting is the way to go, however i am not sure how to do this with fMRI data in SPM.
4) Lastly, I have a question regarding setting the p-value threshold.
I have played around in SPM with setting it at FWE<0.05 since this seems common for whole brain analyses. However, since my results were disappointing and nothing showed up, I also explored the results at a threshold of p<0.001 uncorrected. Then some clusters did show up, some of which said to have a FWE-corrected p-value of below 0.05. How is this possible?
Thank you in advance for helping me out! These may be very basic questions, but I have trouble applying examples that are out there to my own data.
Best,
Ploon
|