CanlabCore for MVPA

[fmri2...@gmail.com]

Hi, CANLab experts,
My colleagues and I are attempting to use your tools for the MVPA. We have some questions, as we are beginners with this kind of data analysis. In detail, our design is like this: large reward, small reward; larger loss, small lose, and neutral events. This is an event related design. We are doing the preprocessing with SPM12. Our preprocessing steps go as following.

1. time series realignement [Realign: Estimate & Reslice]
2. coregistration between fMRI and sMRI [Coregister: Estimate]
3. Segment
4. normalisation to a standard space [Normalise: Estimate & Write]
5. fMRI model specification
6. Model estimation

Now we are supposed to do the step "Contrast manager" step(in fact, t-test) to begin the final MVPA work. As said above, we have two contrasts: "larger reward vs. small reward"; "larger loss vs. small loss" for the MVPA. Our aim is to see whether the MVPA can discriminate the comparison between them. To reiterate, the comparison means: "larger reward vs. small reward" COMPARED TO "larger loss vs. small loss".

So I was wondering whether I can directly do this t-test: (larger reward- larger loss) vs. (small reward - small loss). From the mathematics perspective, this t-test is equal to the comparison as I described above. This opens a further question, can we do the MVPA with only one t-test, say "larger reward vs. small reward" in my case?

Any comments or examples for our analysis are much appreciated.

Thank you very much,
Simi Luck

[Stephan Geuter]

HI Simi,

you can compute the interaction contrast (larger reward- larger loss) vs. (small reward - small loss) on the subject-level and then test the resulting contrast images using a t-test on the group level, as you described it below. If you want to use SVM to discriminate between conditions, you would need two contrast images to separate. Most fMRI textbooks will have extensive discussions of these topics. We recently published a chapter dealing with these questions that might be helpful; http://wagerlab.colorado.edu/files/papers/steph.pdf

https://www.coursera.org/learn/functional-mri

Best,
Stephan

> --
> You received this message because you are subscribed to the Google Groups "WagerlabTools" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to wagerlabtool...@googlegroups.com.
> For more options, visit https://groups.google.com/d/optout.

[Stephan Geuter]

Hi Simi,

MVPA is a general term for various multivariate methods. if you’re conditions reward and neutral are between subjects, then you would have one con-image per person. Otherwise, you would take one image per condition and label it 0 or 1, respectively. a simple contrast on each of the conditions  (eg. c=[1 0 0 0…] and c=[0 1 0 0 0 ..]  will give you the con-image for the respective conditon. Our predict function can classify conditions using SVM or predict continuous outcomes (within and between subjects) using various (penalized) regression methods.

i presume your question about leave-one-out refers to the cross-validation method. you can specify arbitrary folds including LOO. 

Best, 

Stephan

On Feb 20, 2017, at 1:01 PM, Simi Luck <fmri2...@gmail.com> wrote:

Hi, Stephan,

Thanks for your message. I will check your paper. 

To be honest, I was a little confused with the statement that " you would need two contrast images to separate" in the MVPA. For instance, if I only want to use the MVPA to predict the reward vs. neutral, I do not have two contrast images. Does it mean that I can NOT do the MVPA for this case?

According to my understanding, in general, when we do the MVPA, we need provide two data columns: one is the category (e.g., 0, 1 1 0 1 0 0 0  1 1 0 1....), then the 2nd column is the data. The idea is to train a model to see whether the machine can discriminate and sort the data to two categories (i.e., 0, 1 1 0.....).  Am I correct?  My understanding is that with Tor Wager's toolbox for MVPA, people need use one group data (leave one out method) to train a model, and then use this model to predict the remained data. One issue here is that here we do not have the **category column here. Is because of this reason, that users must have two contrast images to separate?

If users must have two contrasts, it perhaps also means that Tor Wager's toolbox has some limitations in the MVPA work. As apparently, people can not do the MVPA if he only has one contrast. Am I incorrect?

> To unsubscribe from this group and stop receiving emails from it, send an email to wagerlabtools+unsubscribe@googlegroups.com.