CCCPM Report

The result above is the headline cross-validated number for the main connectome model on the both network. The effect size (Pearson r for regression, AUC for classification) is computed on held-out test folds and averaged across folds — the conventional CPM headline number — and the p-value comes from permutation testing. The scatter plots below show every out-of-sample prediction (pooled across folds) for each network, and for the covariates-only model as a baseline the connectome should beat.

Confound control: edge selection and features. Confounds were accounted for both when selecting edges and in the connectome the model consumes.

Each model answers a different question:

• connectome — how well does connectivity predict once the covariate variance has been regressed out of it? (This run used model_input='residualized', so these are deconfounded network strengths.)
• covariates — the baseline from nuisance variables (age, sex, motion…).
• full — connectome + covariates combined (best prediction).
• increment — the added value of the connectome over covariates (full − covariates). A significant increment is your evidence the connectome carries information the confounds do not. It is shown as “—” for Pearson r and F1: subtracting two correlations is not a comparison of correlations (that needs Fisher z or Steiger's test), and F1 is a harmonic mean whose difference has no established interpretation. That is deliberate, not a failed computation.

Colours are the network the edges came from: red = positive, blue = negative, grey = both. Boxes span the outer CV folds; the table gives mean [SD] with permutation p-values (* p<0.05, ** p<0.01).

For each participant, CCCPM sums the connectivity of all selected positive edges (and, separately, the negative edges). These two scalar "network strength" values are the actual features the connectome model uses. The scatter shows how strength tracks the target; the histograms show the distribution of strength across participants.

• Connectivity matrix — node × node grid of stable edges; red/blue mark the positive/negative networks. Ordered by brain network when an atlas is given.
• Hub nodes — the nodes participating in the most stable edges.
• Network-summary matrix — the same edges collapsed to canonical networks (needs a network column in the atlas).
• Chord diagram — between-network connectivity as a ring; arc width is the aggregated stability.
• Glass brain — the selected edges drawn on a brain (needs node coordinates x, y, z); shown for the default subset.

Use the buttons to choose which edges the matrix, hub and chord views show: significant (stability p < 0.05, needs permutations), or the top 5% / 10% most stable edges.

The individual connections (region A — region B) that most reliably drive the prediction. Stability is how often an edge was selected across folds. Significance is assessed at the subnetwork level with the Network-Based Statistic: connected components of stable edges are tested against a permutation null of the largest component, so every edge in a significant subnetwork shares that subnetwork's p-value. A significant result means "this connected subnetwork is selected more consistently than chance" — not a claim about any single edge in isolation.

Significance method: Network-Based Statistic (component statistic: extent, stability threshold ≥ 0.5), 100 permutations, α = 0.05.
Positive network: largest component: 266 edges; 1 significant subnetwork(s).
Negative network: largest component: 252 edges; 1 significant subnetwork(s).