# Analysis plan: intrinsic timescales and network-dominance dynamics of the default mode, salience and control networks in depression

Registered before any timescale, state-dynamics feature, or group or symptom comparison was computed on the real data.

## Hypotheses under test

A family of spatiotemporal accounts of depression makes the following directional predictions, which we test as stated, without endorsing them:

1. **Hierarchy.** Intrinsic neural timescales are ordered across large-scale networks: salience network (SN) shortest, control network (CEN) intermediate, default mode network (DMN) longest.
2. **Over-integration in depression.** DMN intrinsic timescales are longer in depression than in controls.
3. **Trajectory narrowing in depression.** Periods of DMN dominance last longer, and transitions among DMN-, SN- and CEN-dominant states are less varied (lower entropy), in depression.
4. **Rumination.** Within depression, brooding rumination rises with DMN timescale and with DMN dwell time.

Several recent studies report shorter, not longer, intrinsic timescales in major depression across networks, so prediction 2 is genuinely uncertain; all group tests are two-sided and report the direction found.

## Data

OpenNeuro ds002748 v1.0.5 (Bezmaternykh et al.; CC0): eyes-closed resting-state BOLD from 51 people with mild-to-moderate depression (ICD-10 F32.0, F32.1, F34.1, F34.9) and 21 healthy controls; TR 2.5 s, 100 volumes, 25 slices of 5 mm. `participants.tsv` gives group, age, sex, BDI, Zung, and Ruminative Responses Scale scores (RRS-sum, brooding, reflection). Files are fetched from the OpenNeuro S3 mirror by digest.

## What we saw before registering

Dataset description, participants.tsv column names and the count of non-missing values. We built and ran the preprocessing on all participants, blind to group, and inspected only quality measures (framewise displacement, maximum translation, registration Dice overlap, parcel coverage) and one registration overlay image (sub-01), which revealed a resampling-grid bug that was fixed before any feature was computed. We tested `code/analyze.py` on synthetic random features. No timescale, dwell time, entropy, or association with group or symptoms was computed or inspected on the real data before registration.

## Preprocessing (`code/preprocess.py`) and features (`code/features.py`)

Drop four volumes; rigid realignment to the mean volume (DIPY, mutual information); framewise displacement (50-mm radius); 12-parameter affine registration of the mean volume to the MNI152 template bundled with nilearn; resampling to 3 mm; nuisance regression of six motion parameters, their derivatives and five white-matter principal components, with linear detrending and a 0.01-0.1 Hz band-pass; averaging in the 100 Schaefer-2018 parcels (7-network labels).

Features: ACW-50 and ACW-0 (the lag, in seconds, at which a parcel's autocorrelation function first falls below 0.5 or zero, linearly interpolated) averaged within DMN (Default, 24 parcels), SN (SalVentAttn, 12) and CEN (Cont, 13); network-dominance states (at each volume, which of the three networks has the highest z-scored mean signal), giving DMN occupancy, mean DMN dwell time in seconds, and the entropy in bits of the empirical distribution of the nine ordered transitions between consecutive volumes.

## Quality exclusions (fixed before outcomes)

Exclude a participant with mean framewise displacement above 0.3 mm, more than 20% of volumes above 0.5 mm, maximum translation above 3 mm, or registration Dice below 0.80.

## Tests (`code/analyze.py`)

**Family A (hierarchy; Holm across three, one-sided as predicted):** paired Wilcoxon signed-rank on ACW-50: A1 DMN > SN, A2 DMN > CEN, A3 CEN > SN; Hodges-Lehmann estimates with bootstrap 95% CIs.

**Family B (depression; Holm across five, two-sided):** Mann-Whitney depression versus control for B1 DMN ACW-50, B2 DMN dwell time, B3 transition entropy (AUC = probability that a depressed participant's value exceeds a control's, bootstrap 95% CI); Spearman correlation within the depression group of RRS-brooding with B4 DMN ACW-50 and B5 DMN dwell time (bootstrap 95% CI).

**Secondary (unadjusted):** ACW-0 group contrasts for each network; ACW-50 group contrasts for SN, CEN and the whole brain; DMN occupancy; Spearman of BDI and of Zung with DMN ACW-50 across all participants; RRS-sum with DMN ACW-50 within depression; mean framewise displacement by group; rank regressions of DMN ACW-50 and DMN dwell time on group adjusting for mean framewise displacement, age and sex.

Seeds are fixed (bootstrap 10,000 resamples, seed 20261007).

## Power and limits

With about 51 versus 21 participants, a two-sided Mann-Whitney test at alpha 0.05 has about 80% power for a standardised difference near 0.75 (AUC about 0.70); Spearman within about 45 depressed participants with rumination scores has 80% power near rho = 0.40. Runs are short (96 analysed volumes, 4 minutes) with a slow TR, so ACW-50 is resolved only to a fraction of one 2.5-s lag and state statistics rest on 96 samples per person; null results rule out only moderate-to-large effects. Registration is affine only. Depression here is mild to moderate.

## Interpretation rules

A hypothesis is supported only by its registered test in its predicted direction after Holm adjustment within its family. A significant effect in the opposite direction is reported as evidence against the prediction. Null results are reported as such, with their CIs.