PROJECT 1 AOMIC ID1000 INDEPENDENT REPLICATION
Analysis plan version 1.0
27 October 2026

Purpose and chronology
----------------------
Test whether relative allocation among five major brain-tissue compartments is
associated with general cognitive ability beyond total tissue volume, age and
sex, and whether the previously selected total-cerebellum to cerebral-white-
matter contrast replicates in AOMIC ID1000. This extends the existing HCP
discovery and DLBS confirmation study. It is not a new anatomical search.

Before this lock, the team knew the existing HCP and DLBS results and had
inspected AOMIC documentation, schema, missingness, ID matching, basic
demographics and positive-volume availability. No AOMIC brain-cognition
association, effect estimate, significance test or outcome-guided model choice
had been computed or inspected. A second methodological review preceded this
lock. The accompanying machine-readable receipt records the UTC lock time and
SHA256 hash. This is an internally timestamped prospective analysis lock for
the new AOMIC associations, NOT a public preregistration or Registered Report.

Data source and integrity
-------------------------
AOMIC ID1000, OpenNeuro ds003097, DOI 10.18112/openneuro.ds003097.v1.2.1.
Git commit 47f030747dbaf492acd9f9624528aef8d51d8bd9.
Source URLs, byte counts and SHA256 values are in source_manifest.json.
The four original source files are retained unchanged during analysis.
Recruitment was not described as family-enriched; genetic unrelatedness is
not verified, and no claim of genetically unrelated participants will be made.
FreeSurfer version differences across cohorts are a measurement limitation.

Population and exclusions
------------------------
Join participants.tsv participant_id to the volume-table Measure:volume key.
Require unique keys and a one-to-one merge. Primary inclusion requires finite
age; observed sex label male or female; complete finite IST_fluid, IST_memory
and IST_crystallised; and finite strictly positive values for the eight source
volumes used in the five compartments. Do not impute these variables.
Use all eligible released observations. Published release QC is retained.
No exclusion is based on a favorable association or on a cognitive outlier.
The availability audit found 928 matched participants and 926 cognitive
complete cases; the executed exclusion flow will determine final N.
Code sex as male=1 and female=0. The data dictionary says M/F but the actual
TSV uses male/female. Document this discrepancy rather than recoding blindly.

Anatomical mapping and scale
----------------------------
Ordered five-part vector, in cubic millimetres:
1 Cortical gray matter = CortexVol.
2 Cerebral white matter = CerebralWhiteMatterVol.
3 Subcortical gray matter = SubCortGrayVol, the supplied FreeSurfer aggregate.
4 Total cerebellum = Left-Cerebellum-Cortex + Right-Cerebellum-Cortex
  + Left-Cerebellum-White-Matter + Right-Cerebellum-White-Matter.
5 Brainstem = Brain-Stem.
Total analyzed tissue T is the sum of these five parts, not eTIV or whole
intracranial content. The primary scale term is ln(T), centered.
eTIV = EstimatedTotalIntraCranialVol, for fixed sensitivity analyses only.
Use all four orthonormal log-ratio coordinates from a Helmert basis. Taking
logs of the parts rather than of normalized proportions is algebraically
equivalent for this zero-sum basis. Record the exact basis matrix.

Cognitive scoring
-----------------
Primary outcome: first principal component of IST_fluid, IST_memory and
IST_crystallised after each is standardized using primary-sample mean and
population SD (ddof=0). Orient PC1 so the sum of its three eigenvector weights
is positive; standardize the resulting score to mean 0 and population SD 1.
Report all weights, correlations of domains with PC1 and explained variance.
Never drop a domain on the basis of its association with anatomy.
IST_intelligence_total is not included with its component domains in PCA.
Its standardized native score is a prespecified alternative-outcome sensitivity.

Primary models and confirmatory testing
--------------------------------------
Reduced OLS model: g = intercept + centered age in years + sex + centered ln(T).
Linear age is fixed for the narrow 19-26-year range. No age band or spline
choice will be made after seeing associations.
Full OLS model: reduced model + four standardized ILR coordinates.
Primary global test: four-df nested extra-sum-of-squares F test, alpha=0.05.
Report baseline and full R-squared, added R-squared, partial R-squared, F,
degrees of freedom and p. Partial R-squared = (RSS_reduced-RSS_full)/RSS_reduced.

Targeted balance: b = ln(total cerebellum / cerebral white matter) / sqrt(2).
Rotate to a four-coordinate orthonormal ILR basis whose first coordinate is b.
Include all three remaining orthogonal coordinates, and the same covariates.
Standardize each coordinate with its primary-sample population SD. Report the
coefficient per one SD of b, SE, 95 percent t-based CI and two-sided p.
The prespecified direction is positive. Do not substitute cerebellar cortex
for total cerebellum in the primary contrast.

The two tests use fixed-sequence hierarchical interpretation: first global
composition at alpha=0.05, then the targeted positive balance at two-sided
alpha=0.05. Report the focal estimate regardless, but do not declare full
targeted replication when the global gate is not passed. Global association
without a positive significant focal result is global-only replication.

Robust uncertainty and fixed sensitivity analyses
------------------------------------------------
Compute HC3 covariance for the full primary model and its rotated equivalent.
Report the four-df HC3 Wald F approximation and focal HC3 t-based CI and p.
If conventional and HC3 tests disagree, explicitly qualify the inference; do
not select the favorable result. No sensitivity can replace a failed primary
test. Primary g weights and all primary predictor standardizations remain
fixed across sensitivity subsets, except the explicitly different total-score
outcome and the explicitly full-reestimation bootstrap below.

Run this complete list once:
S1 Native standardized IST_intelligence_total instead of PC1.
S2 Add centered age squared to the original baseline.
S3 Replace ln(T) with centered ln(eTIV).
S4 Include both ln(T) and ln(eTIV).
S5 Add squared centered ln(T) and sex-by-centered-ln(T).
S6 Add standardized ln(1+SurfaceHoles), requiring finite nonnegative holes.
S7 Add education_level as two indicators, low reference, medium and high.
   Interpret as conditional sensitivity, not a mandatory confound adjustment.
S8 Exclude anatomical extremes defined without cognition: for any of the five
   log-compartments or ln(eTIV), absolute deviation from the primary median
   greater than 5 times 1.4826 times the median absolute deviation. If a MAD is
   zero, that variable contributes no outlier flag. Keep original scaling.
S9 Exclude observations with Cook's distance greater than 4/N from the primary
   full model. Label as an influence diagnostic, never a revised primary result.
Report all sensitivities, sample counts, global added/partial R-squared and
tests, and the focal coefficient with conventional and HC3 uncertainty.

Bootstrap
---------
Use 2000 participant-pairs resamples, random seed 20261027. Re-estimate the three
cognitive means/SDs, PCA weights and g scaling and the anatomical-coordinate
means/SDs inside each resample. Preserve deterministic orientation. Refit the
same focal full model. Report its 2.5th and 97.5th coefficient percentiles.
This propagates scoring and standardization uncertainty. It is a supplementary
CI, not another hypothesis-testing gate. Record any failed rank-deficient
resamples rather than replacing them silently.

Diagnostics and independent numerical verification
---------------------------------------------------
Check matrix rank, finite results, positive source volumes, join uniqueness,
basis orthonormality and zero sums, equality of the Helmert and rotated-model
fitted values/RSS, native-total versus subscale-sum consistency, and primary
source hashes. Record software versions and execution time.
A separate numerical implementation will rebuild inputs and the primary
models without importing the primary analysis functions. Compare N, PCA
variance, R-squared, F, focal coefficient, conventional and HC3 uncertainty.
This is computational verification, not an independent human review.

Reporting and boundaries
------------------------
Report null or opposite-direction findings as fully as positive findings.
Do not search other balances, participant subgroups, IQ outcomes, age models
or outlier thresholds for a favorable result. Any later analysis must be
explicitly dated and labeled post hoc, without changing this locked file.
Do not treat naïve full-HCP standard errors as family-corrected or use them
for an apparently precise combined HCP-DLBS-AOMIC inferential estimate.
Do not hide the existing weaker S900 unrelated HCP result.
New replication evidence does not guarantee acceptance by any journal.

Primary sources
---------------
Snoek et al. 2021. The Amsterdam Open MRI Collection, a set of multimodal MRI
datasets for individual difference analyses. Scientific Data 8, 85.
https://doi.org/10.1038/s41597-021-00870-6
https://lukas-snoek.com/thesis/aomic.html
https://openneuro.org/datasets/ds003097/versions/1.2.1
https://github.com/OpenNeuroDatasets/ds003097/tree/47f030747dbaf492acd9f9624528aef8d51d8bd9
