Regression-based RSA, p-values and noise ceiling

1

Hey everyone!

I aim to do RSA, but predicting the neural RDM by several predictor RDMs. The typical RSA approaches I found in the literature usually perform 1:1 correlations, and then compare the coefficients. I however want to use regression to disentangle the unique contribution of each predictor RDM to the neural RDM.

I haven’t really seen this in a paper, but maybe you can point me to one!

Two questions arise regarding the interpretability:

  • probably there would be problems in interpreting the p-values of my betas, as there are dependencies between the single elements of my predictor RDM (or the resulting vector of the lower triangle). What would be a good way to estimate the relative contribution of each predictor (besides the absolute beta estimates)?
  • How would I calculate the noise ceiling for each predictor? For the simple correlation, the methods are straightforward. But I can’t get the transfer to how to estimate it with multiple terms present.

Thank you so much for any hint in a direction.
Best,
Roman

2

Hi @kesslerr,

To perform regression RSA with multiple predictors, you can use cosmo_target_dsm_corr_measure from CoSMoMVPA.

For example:

    % Prepare measure arguments for RSA GLM
    measure     = @cosmo_target_dsm_corr_measure;
    measureArgs = struct();
    measureArgs.center_data = true;

    % Prepare regressor RDMs (convert to vector form)
    regressorRDMs          = cellfun(@(x) squareform(x)', {rdms.dsm}, 'UniformOutput', false);
    measureArgs.glm_dsm    = regressorRDMs;

    % Run the RSA GLM
    result = measure(dsMean, measureArgs);

This approach provides a regression coefficient for each regressor RDM.

While I am not an expert in this area, I usually calculate the noise ceiling of my RDMs using a reliability measure. For example, I correlate the lower triangle of a single subject’s RDM with the average RDM of all other subjects (computed iteratively for each subject) to estimate the noise in the RDMs. I assume you could adapt a similar reliability measure for regression.

I hope this helps!

Best regards,
Andrea