ROI-based RSA

[wenjiaz...@gmail.com]

Dear Nick,

I am trying to do a ROI-based (not ROI searchlight) RSA with COSMO these days.

Specifically, I defined a ROI which contains about 100 voxels.

The activation pattern RDM of this ROI was constructed by calculating the anti-correlation of activation pattern from each pair of items.

This activation pattern RDM was compared with theoretical model RDM and then resulted only one similarity value.

Last, I perfomed a t tset of these similarity values from all participants aganist zero to investigate whether this ROI could represent the information from theoretical RDM.

I am wondering whether this analysis method make sense.

And, is any Csomo script or syntax helpful for doing this? 

Especially, how to construct the activation pattern RDM of a specific ROI.

Thank you.

Wenjia

[Nick Oosterhof]

> On 14 Jul 2017, at 16:35, wenjiaz...@gmail.com wrote:
>
> I am trying to do a ROI-based (not ROI searchlight) RSA with COSMO these days.
> Specifically, I defined a ROI which contains about 100 voxels.
> The activation pattern RDM of this ROI was constructed by calculating the anti-correlation of activation pattern from each pair of items.

Bij anti-correlation you mean one-minus-correlation, or something else?

> This activation pattern RDM was compared with theoretical model RDM and then resulted only one similarity value.

You correlated the model RDM (in vectorized form, e.g. using squareform) with the neural RDM (also in vectorized form)? If yes, this seems appropriate. Note that this can also be done using cosmo_target_dsm_corr_measure.

> Last, I perfomed a t tset of these similarity values from all participants aganist zero to investigate whether this ROI could represent the information from theoretical RDM.
>
> I am wondering whether this analysis method make sense.

It does, this seems fine (assuming you correlated a vectorized model RDM with a vectorized neural RDM).

> And, is any Csomo script or syntax helpful for doing this?
> Especially, how to construct the activation pattern RDM of a specific ROI.

You could use cosmo_dissimilarity_matrix_measure to compute an RDM, or cosmo_target_dsm_corr_measure to compute agreement between a model RDM and neural patterns.
Also see the RSA tutorial: http://cosmomvpa.org/ex_rsa_tutorial.html
Hope that helps!

[wenjiaz...@gmail.com]

Thank you. I will try this method.

在 2017年7月14日星期五 UTC+8下午11:37:30，n.n.oosterhof写道：

[张文嘉]

Dear Nick,

I have tried 2 methods to do this analysis.

1, I used cosmo_target_dsm_corr_measure to compute the agreement between a model RDM and neural patterns RDM.

 measure=@cosmo_target_dsm_corr_measure;

        measure_args=struct();

        measure_args.target_dsm=target_dsm.Noun;

        measure_args.type = 'Spearman';

        

       Results=measure(ds_roi, measure_args);

2, I used cosmo_dissimilarity_matrix_measure to compute the neural patterns vectors,

Also, I changed the model RDM to vectors using squareform function.

Lastly, I computed the agreement between these two kinds fo vectors (neural pattern and model).

 Dissimilarity_AP = cosmo_dissimilarity_matrix_measure(ds_z);

 Dissimilarity_AP_samples=Dissimilarity_AP.samples;

 

        model={'Noun'};

        RDM_fn=['F:\Exp4_RSA\RDM\' char(model) '.mat'];

        target_RDM=load(RDM_fn);

        Dissimilarity_model_RDM =target_RDM.Noun;

        Dissimilarity_model=squareform (Dissimilarity_model_RDM);

        Dissimilarity_model= Dissimilarity_model';

 

        Results = cosmo_corr(Dissimilarity_AP_samples, Dissimilarity_model);

I found the results of these two methods are slightly different. For example:

1, 

0.009722702

0.034425945

-0.014866409

-0.01035037

0.043782142

0.01178115

2,

0.007152219

0.020538755

0.001774223

-0.006357882

0.045372291

0.007670262

I am wondering why the difference exist and which one is more reasonable.

Thank you. 

在 2017年7月14日星期五 UTC+8下午11:37:30，n.n.oosterhof写道：

[Nick Oosterhof]

In the first analysis you are using the Spearman correlation type, but in the second one the default - which is Pearson. Could that be the reason for the difference?

[张文嘉]

Yes, the different correlation type is the reason. Thank you.

在 2017年8月28日星期一 UTC+8下午3:40:34，n.n.oosterhof写道：