CoSMo MVPA negative z values

[Katrin]

Hi Nick,

When I  run the t-test on the group data or when I run cosmo_montecarlo_cluster_stat on the same data then I consistently get many negative z values. 

I would like to ask how should I interpret the negative z-values? 
My idea is that these clusters  have anticorrelation between the neural  similarity matrix and the target similarity matrix. I haven't seen studies reporting this, so I don't know how this should be interpreted. 

[Nick Oosterhof]

This can happen with representational similarity analysis (RSA) with a target (model) matrix, for example when using cosmo_target_dsm_corr_measure. Negative z-scores from cosmo_montecarlo_cluster_stat indicate negative correlations between the model and neural similarity matrices.
The interpretation is that the model (target) neural similarity matrix is consistently 'wrong' in predicting neural similarity. In other words, pairs of items that are very similar according to the model matrix are very dissimilar according to the neural dissimilarity matrix. More generally, the more similar a pair of items is according to the model matrix, the more dissimilar the pair is according to the dissimilarity matrix.

So what coudl be causing this? An potentially interesting explanation is that the neural dissimilarity matrix is really different than the model matrix.
A less interesting explanation is that the researcher made a mistake and accidentally provided a similarity (not *dissimilarity) matrix as the 'target' option in cosmo_target_dsm_corr_measure. This less interesting scenario is only possible if the target matrix is presented in vector form (see cosmo_squareform), as cosmo_target_dsm_corr_measure will raise an error if a square similarity matrix is used as the 'target'.