Interpreting Your Results

After cpm.run(...) finishes, everything CCCPM produces is written to the results_directory you specified. This page explains what each output means and how to read it — both the interactive HTML report and the underlying CSV/array files you can load for your own figures and statistics.

The same structure is produced for regression and classification; only the performance metrics differ (see Metrics).

Two concepts that appear everywhere

Almost every result is broken down by a model and a network.

Models — what each one is built from

For every cross-validation fold, CCCPM fits several models so you can see how much the connectome actually adds on top of your covariates:

Model	Predicts the target from…	Use it to answer…
`connectome`	connectome network strength only	How well does brain connectivity alone predict?
`covariates`	the covariates only (e.g. age, sex, motion)	What's the baseline from nuisance variables?
`full`	connectome + covariates	Best combined prediction
`increment`	the added value of the connectome over covariates alone (`full` − `covariates`)	Does the connectome explain anything beyond confounds?

The increment model is usually the most important for a confound-aware claim: a significant increment means the connectome carries predictive information that the covariates do not.

Two things to know when reading these rows:

increment is NaN for Pearson r and F1, by design. A difference of two correlations is not a correlation — comparing them needs Fisher z / Steiger's test — and a difference of two F1 scores is not an F1. Read the increment on explained variance, MSE, MAE, accuracy, balanced accuracy or ROC AUC. An em dash in those cells means "not a statistic", not "failed".
There is no residuals model. Whether the connectome behind connectome and full has been deconfounded is a property of the run (model_input), not a separate row — so one report is one cell of the confound 2×2. The report's headline says which cell it is; see Confound control.
Running without covariates leaves covariates, full and increment undefined. They are NaN-filled rather than omitted, and the report hides them.

Networks — which edges are used

Edge selection splits predictive edges into two sub-networks, and CCCPM evaluates each plus their combination:

positive — edges positively correlated with the target.
negative — edges negatively correlated with the target.
both — positive and negative networks combined.

So a single number like "Pearson r" is always reported per (model, network) pair.

The HTML report (`report.html`)

report.html is a self-contained, shareable summary — open it in any browser, print it to PDF, or send it to a collaborator (everything, including the figures, is embedded, so it works offline). It is a single scrolling page with a sticky sidebar, organised top-down from the headline result to the supporting detail:

Section	What it shows
Summary	The headline cross-validated result as a one-sentence verdict, key-stat chips (samples, nodes, edges, covariates, permutations, edge p-threshold), and predicted-vs-observed scatter plots for the positive, negative, and both networks plus the covariates-only baseline. Each scatter is annotated with its cross-validated effect size (Pearson r / AUC) and permutation p. Start here.
Model Comparison	One faceted figure comparing every model (`connectome`, `covariates`, `full`, `increment`) across metrics and networks, plus the APA results table (mean [SD] with permutation p-values). Foregrounds the `increment` model — your confound-control evidence.
Network Strengths	How the summed positive/negative network strength relates to the target, and the distribution of strength across participants.
Brain & Edges	The predictive edges in the brain: a connectivity matrix, a hub (node-degree) plot, and — when an atlas with a `network` column is supplied — a network-summary matrix and a chord diagram. With node coordinates (`x, y, z`) it also renders a glass-brain view. A button group switches the matrix/hub/chord views between the significantly stable edges (default, when permutations were run), the top 5% and the top 10% most stable edges; the glass brain shows the default subset. These figures come from the atlas you pass to `CPMAnalysis`.
Stable Edges	All significant edges (region A — region B) per network, the significance method and its diagnostics (e.g. largest NBS component), the permutation null-distribution plots, and a downloadable CSV of every selected edge with its stability and significance.
Data & Methods	Appendix: target distribution, covariate scatter matrix, the full run configuration, and (with an inner CV) per-fold hyperparameters.

Every section carries a short always-visible "what this shows" note so a reader unfamiliar with the internals can follow it.

You can view a full example report in the Simulated Data example.

Metrics

Metrics are computed out-of-sample (on held-out test folds) and then averaged across folds. Higher is better unless noted.

Regression

Metric	Meaning	Notes
`pearson_score`	Correlation between predicted and true values	The conventional CPM headline number
`explained_variance_score`	Fraction of target variance explained	1.0 is perfect; can be negative
`mean_absolute_error`	Average absolute prediction error	Lower is better; in target units
`mean_squared_error`	Average squared prediction error	Lower is better; penalises large errors

Classification

Metric	Meaning	Notes
`roc_auc`	Area under the ROC curve	Threshold-independent; robust headline metric
`balanced_accuracy`	Mean of sensitivity and specificity	Prefer over accuracy for imbalanced classes
`accuracy`	Fraction correctly classified	Misleading when classes are imbalanced
`f1_score`	Harmonic mean of precision and recall	Useful when the positive class is rare

Output files

Everything in the report is also available as plain files for your own analysis.

Performance

File	Contents
`cv_results_summary.csv`	Mean ± std of every metric, indexed by `(model, network)`. The numbers behind the Main Results page.
`cv_results_full.csv`	Per-fold metric values (before averaging) — use these for your own error bars / tests.
`cv_predictions.csv`	Out-of-sample predictions per subject: columns `sample_index, model, network, y_pred, y_true, fold, repeat`. With a `RepeatedKFold` each subject has one row per repeat; the report averages them so every subject is plotted once.
`cv_network_strengths.csv`	Summed positive/negative network strength per subject, with the target: `sample_index, y_true, network_strength, model, network, fold, repeat`.

Significance (only if `n_permutations > 0`)

File	Contents
`p_values.csv`	Permutation p-value for each `(network, model)` × metric. A small value (e.g. ≤ 0.05) means the observed performance is unlikely under the null of no brain–behaviour association.
`permutation/`	The full null distributions (`cv_results_*`) and permuted edge selections used to compute those p-values.

How the p-values are defined: the entire pipeline is re-run on many shuffled copies of the target to build a null distribution of performance; the p-value is the fraction of permutations that match or beat the real result. With n_permutations=100 the smallest possible p-value is ~0.01 — use 1000+ permutations for publishable significance.

Selected edges & stability

File	Contents
`edges.npy`	The edges selected in each fold.
`stability_edges.npy`	How consistently each edge was selected across folds (its stability). Edges chosen in most folds are the reliable ones.
`stability_edges_significance.npy`	Per-edge significance of edge stability (when permutations are run). With the default NBS method these are subnetwork-level p-values shared by every edge in a significant connected component; with TFCE they are per-edge FWER-corrected p-values.
`stability_edges_significance_meta.json`	Diagnostics for the significance test: method + parameters, the permutation null distribution, and (for NBS) the observed components and largest-component size. Drives the null-distribution plots in the report.
`data_insights/`	Cached input summaries and figures (target distribution, covariate scatter, feature/target/covariate names).

Logs

cpm_log.txt — a full, timestamped log of the run (configuration, per-step progress, final aggregated table). The first place to look if something behaved unexpectedly.
task_type.txt — regression or classification, recorded for the report.
plots/ — every figure from the report exported as PNG/PDF/SVG for direct use in papers.

A reading workflow

Open report.html → Summary for the headline result, then Model Comparison. Look at the full and increment models on the both network.
Check p_values.csv: is the increment model significant? That's your evidence the connectome adds information beyond confounds.
Inspect stability_edges.npy (or the Stable Edges section) to see which connections drive the prediction.
Use cv_predictions.csv / cv_results_full.csv to make your own figures or run additional statistics.