Skip to content

fmri_data_rsa_regression.md

Latest commit

 

History

History
99 lines (77 loc) · 4.69 KB

fmri_data_rsa_regression.md

File metadata and controls

99 lines (77 loc) · 4.69 KB
Raw

fmri_data.rsa_regression — representational similarity regression with bootstrap stats

← back to fmri_data methods · Object methods index · Recasting objects

Test whether a hypothesised grouping of images (a categorical model RDM) explains the dissimilarity structure of brain activity patterns across images. Builds a representational dissimilarity matrix (RDM) from obj.dat and regresses it on one or more model RDMs derived from binary design columns. Block-bootstrap stratified by study gives inference on each "generalisation index" — the canonical use case is testing whether a brain pattern generalises across constructs and studies (Kragel et al. 2018, Nature Neuroscience).

Code map

rsa_regression code map

Editable PowerPoint version

Usage

stats = rsa_regression(obj, design, study, varargin)

Inputs

Argument Type Description
obj fmri_data Image object — typically one image per subject (no repeated measures).
design [n_images × p] logical/binary Each column is a grouping variable; columns are converted into pairwise model dissimilarities (pdist of column → distances among images), normalised to sum to 1e5.
study [n_images × 1] int Block IDs for stratified bootstrap (e.g. study number).
'average_Euclidean' flag Brain RDM = Euclidean distance between mean voxel values per image (univariate).
'euclidean' flag Brain RDM = Euclidean distance using all voxels.
'cosine' flag Brain RDM = cosine distance using all voxels.
'correlation' flag Brain RDM = correlation distance, 1 - r (default).
'nobootstrap' flag Skip the 1000-iteration bootstrap and return point estimates only.

Outputs

stats is a structure with:

Field Type Description
gen_index [p+1 × 1] OLS regression slopes (the "generalisation indices") for each model RDM column predicting the brain RDM. The first row is the intercept.
bs_gen_index [1000 × p+1] Bootstrap distribution for each generalisation index (only without 'nobootstrap').
ste [p+1 × 1] Bootstrap standard error per index.
Z [p+1 × 1] Bootstrap normal-approx Z-scores.
p [p+1 × 1] Two-tailed p-values from Z.
sig [p+1 × 1] Significant at FDR q < .05 across the indices.
RDM square matrix Full brain RDM in square form (handy for plotting).

Notes

  • Each column of design defines a partition of images. The pdist of a binary column is 0 within a group and 1 between groups, so the resulting model RDM tests whether brain dissimilarity tracks that partition. Use condf2indic(...) to turn a categorical labelling into a column-binary matrix.
  • The bootstrap is stratified by study: within each study, images are resampled with replacement preserving the study membership of each row. This is the right null for "does the result hold within studies, after accounting for study identity?".
  • This implementation does not handle repeated measures — include only one image per subject.
  • Pairs of images with brain distance < 1e-5 are treated as missing (NaN) before fitting.

Example: testing pain / cognition / emotion generalisation in amygdala (Kragel 2018)

% 270 subject-level images systematically sampled from 18 studies in 3 domains
[data_obj, names] = load_image_set('kragel18_alldata');

% Restrict to amygdala
amy        = select_atlas_subset(load_atlas('Canlab2018'), {'Amy'});
masked_dat = apply_mask(data_obj, amy);

% Design: studies grouped 6-at-a-time → categorical study-cluster regressor
dsgn  = condf2indic(ceil(data_obj.metadata_table.Studynumber / 6));
study = data_obj.metadata_table.Studynumber;

% Regress brain RDM (correlation distance, default) on model RDM, with
% within-study bootstrap inference
stats = rsa_regression(masked_dat, dsgn, study);

disp([stats.gen_index stats.Z stats.p stats.sig])
imagesc(stats.RDM); axis square; title('Amygdala brain RDM');

See also