White matter reflects the childhood exposome

Replication Guide

• White matter reflects the childhood exposome
• Replication Guide
  • Abstract
  • Project Lead
  • Faculty Lead
  • Analytic Replicator
  • Collaborators
  • Project Start Date
  • Dataset
  • Github Repository
  • Respublica Project Directory
  • Slack Channel
• Directory Structure
• Code Documentation
  • Step 0: Required Software
  • Step 1: Prepare Dataset
    • ** Python Notebook:** Step1_Data_Prep.ipynb
    • Goal
    • Inputs
    • Feature exclusion
    • Outputs
  • Step 2: Covbat Harmonization
    • R Script: Step2_covbat_harmonization_ses.R
    • Goal
    • Inputs
    • Covariates protected
    • Outputs
  • Step 3: Subset data
    • Python Notebook: Step3_abcd_prep_harmonized_dfs
    • Goal
    • Inputs
    • Outputs
  • Step 4: Mass Univariate Analysis and PCA
    • Python Notebook: Figure_2_Mass_Univariate.ipynb
    • Helper module: partial_r.py
    • Goal
    • Inputs
    • Covariates
    • Mass-univariate model
    • PCA on bundle-by-metric partial r matrix
    • PCA outputs
  • Step 5: SA Axis Annotation
    • Python Notebook: Figure_3_SA_Axis_Annotation.ipynb
    • Goal
    • Inputs
  • Step 6: Multivariate Prediction of Exposome Scores from White Matter Features in Held-Out Data
    • Python Notebook: Figure_4_Multivariate.ipynb
    • Helper script: regression_functions.py
    • Goal
    • Inputs
    • Split-half structure and training/testing
    • Models
    • Feature handling and preprocessing
    • Outputs from rf.run_model_comparison(...)
    • Permutation testing
  • Step 7: Cognition vs. Exposome — Unique Multivariate and Mass-Univariate Effects
    • Python Notebook: Figure_5_Cognition.ipynb
    • Helper scripts: partial_r.py and regression_functions.py
    • Goal
    • Input data
    • Targets and confound structures
    • Multivariate Prediction (Ridge Regression)
    • Mass-Univariate Partial r
  • Step 8: Prep data for HBN Replication
    • Python Notebook: Figure_6_HBN_replication_data_prep.ipynb
    • Goal
    • Inputs
    • Output: ../output_data/cleanedinput_data/HBN/final_hbn_dataset_12_23.csv
  • Step 9: Covbat Harmonization of HBN data and Nuisance Regression
    • R markdown: Figure_6_HBN_harmonization.Rmd
    • Goal
    • Input data
    • Protected covariates (kept in signal during harmonization)
    • Nuisance regression on HBN (post-harmonization)
    • Output:../output_data/HBN_harmonized_refABCD16_then_nuisanceResid_12_23.csv
  • Step 10: HBN Replication
    • Python Notebook: Figure_6_HBN_replication_regression.ipynb
    • Goal
    • Input data
    • Multivariate prediction on held out data (ABCD → HBN)
    • Permutation testing (ABCD → HBN)
    • Within-HBN repeated cross-validation (sanity check)
    • HBN mass-univariate partial r (feature-level SES effects)

Abstract

The childhood environment is critical for brain development and contributes to cognitive outcomes. However, most neuroimaging studies examine a single environmental measure (e.g., socioeconomic status) or a limited set of exposures, obscuring how additive exposures jointly influence brain development. Here we investigated how white matter shape and tissue properties in youth are linked to the childhood exposome, a multidimensional measure capturing a wide variety of environmental exposures. Using multi-shell diffusion MRI from 8,183 children (ages 9-10) in the ABCD study, we quantified microstructural and macrostructural properties across 62 person-specific white matter tracts. The exposome showed widespread and highly replicable associations with both white matter microstructure and macrostructure. The strength of these effects was related to the degree a tract spanned the cortical hierarchy defined by the sensorimotor-association axis. Multivariate models demonstrated that patterns of white matter features explained 25% of the variance in the exposome in unseen individuals (out of sample r=0.50). Notably, white matter-based prediction of cognition was markedly reduced after accounting for the exposome (r=0.37 to r=0.16; an ~82% reduction in explained variance), indicating that brain-cognition associations overlap substantially with variance captured by the exposome. These findings generalized to an independent cohort, the Healthy Brain Network (n=869), which differs substantially from ABCD in MRI acquisition, participant selection, and childhood environments. Together, these results suggest that the environment is reflected in developing white matter architecture in childhood.

Project Lead

Briana Macedo

Faculty Lead

Theodore D. Satterthwaite

Analytic Replicator

Joëlle Bagautdinova

Collaborators

Joëlle Bagautdinova, Steven L. Meisler, Matthew Cieslak, Lucinda M. Sisk, Christos Davatazikos, Alexandre R. Franco, Meike D. Hettwer, Arielle S. Keller, Gregory Kiar, Audrey C. Luo, Allyson P. Mackey, Michael P. Milham, Tyler M. Moore, Valerie J. Sydnor, Kevin Y. Sun, Fang-Cheng Yeh, Ran Barzilay, Damien A. Fair, Russell T. Shinohara, Aaron Alexander-Bloch, Theodore D. Satterthwaite

Project Start Date

June 2025

Dataset

ABCD, with replication in HBN

Github Repository

https://github.com/PennLINC/macedo_wm_exposome

Respublica Project Directory

/mnt/isilon/bgdlab_hbcd/projects/macedo_wm/exposome

Slack Channel

#macedo_tractometry

Directory Structure

project_root/
├── code/
├── input_data/
│   ├── HBN/
│   └── ABCD (files at root)
├── output_data/
│   ├── cleaned/
│   │   ├── HBN/
│   │   └── ABCD (files at root)
│   ├── harmonized/
│   │   ├── HBN/
│   │   └── ABCD (files at root)
│   ├── model_outputs/
│   └── results/
│       ├── main_figures/
│       │   ├── figure2/
│       │   ├── figure3/
│       │   ├── etc
│       └── supplemental_figures/

Code Documentation

Step 0: Required Software

To access required Python software, you can download the macedo-em-exposome-env.yml file from GitHub: https://github.com/PennLINC/macedo_wm_exposome/blob/main/macedo-em-exposome-env.yml

Then, you can use anaconda/mamba to create an environment with the required software packages.

• conda env create -f macedo-em-exposome-env.yml

To access required R software, you can download install_r_packages.sh file from GitHub and then run

• chmod +x install_r_packages.sh
• bash install_r_packages.sh

To generate tracts, we use DSIStudio: https://dsi-studio.labsolver.org/download.html for MacOS15+.

Step 1: Prepare Dataset

** Python Notebook:** Step1_Data_Prep.ipynb

Goal

Create a single ABCD baseline arm dataframe that merges:

• white matter features
• split-half assignment
• exposome factor scores
• cognition principal components

Inputs

1. dMRI processed data from Steven: /mnt/isilon/bgdlab_hbcd/projects/meisler_abcd_dmri_new/data/raw_data/merged_data_meisler_analyses.parquet
2. ABCD Reproducible Matched Samples: participants.tsv from ABCD to determine demographically matched split halves using the matched_group variable. Also contains data for sensitivity analyses: namely the categorical parental_education and income variables. Sets values 777 (Don’t Know) and 999 (Declined to answer) to NA when creating the sample for the sensitivity analysis.
3. Exposome scores from Arielle: abcd_longitudinal_factor_scores_7f_bifactor.csv
   1. Keller AS, Moore TM, Luo A, et al. A general exposome factor explains individual differences in functional brain network topography and cognition in youth. Dev Cogn Neurosci. 2024;66:101370. doi:10.1016/j.dcn.2024.101370
4. Area Deprivation Index (ADI; for sensitivity analysis) from Steven, downloaded from LASSO
   1. /mnt/isilon/bgdlab_hbcd/shared_data/ABCD/LASSO_tabular/dairc/rawdata/phenotype/le_l_adi.parquet
5. Cognition data from Kevin: cognition_data.csv
   1. Thompson WK, Barch DM, Bjork JM, et al. The structure of cognition in 9 and 10 year-old children and associations with problem behaviors: Findings from the ABCD study’s baseline neurocognitive battery. Dev Cogn Neurosci. 2019;36:100606. doi:10.1016/j.dcn.2018.12.004
6. Combine all of this data using subject IDs from baseline arm (ages 9-10).

Feature exclusion

Remove:

• Bundles:
  • msmt_ProjectionBrainstem_DentatorubrothalamicTract-lr
  • msmt_ProjectionBrainstem_DentatorubrothalamicTract-rl
  • msmt_Commissure_AnteriorCommissure
  • msmt_ProjectionBrainstem_CorticobulbarTractL
  • msmt_ProjectionBrainstem_CorticobulbarTractR
• Metrics:
  • number_of_tracts
  • NODDI_directions

Outputs

1. Primary merged dataset (exposome + WM + cognition + matched_group) -> add your csv name here exposome_FINAL_11_3.csv
2. Primary merged cognition dataset (exposome + WM + cognition + matched_group)
   1. exposome_FINAL_cognition_11_10.csv
3. Sensitivity dataset (WM + cognition + matched_group + parental_education + income)
   1. parental_edu_income_sensitivity_df.csv

Step 2: Covbat Harmonization

R Script: Step2_covbat_harmonization_ses.R

Run this script in the terminal with:

srun -c 4 --mem=32G -t 24:00:00 ~/miniforge3/envs/macedo-wm-exposome-env/bin/Rscript /mnt/isilon/bgdlab_hbcd/projects/macedo_wm_exposome/macedo_wm_exposome/code/Step2_covbat_harmonization.R <analysis_type>

Analysis types are: main, cognition, and income. Each one takes a little while to run!

Goal

Harmonize white matter features across ABCD sites using CovBat with:

• Batch variable: batch_device_software (site, machine, software)
• Protected covariates: always age and sex, plus additional variables depending on analysis (see below)
• Leakage prevention: run CovBat twice:
  • once using split half A as the training subset (harmonization parameters learned from A)
  • once using split half B as the training subset (harmonization parameters learned from B)

These two harmonized datasets are later used depending on which split half is treated as the training set in downstream analyses.

Inputs

• Output of Step 1:
  • for main analysis: exposome_FINAL_11_3.csv
  • for cognition analysis: exposome_FINAL_cognition_11_10.csv
  • for SES-proxy sensitivity analysis: parental_edu_income_sensitivity_analysis_df.csv
• All columns beginning with bundle_ prefix are harmonized (WM features x subjects matrix).
• Other columns (demographics, exposome, cognition, etc.) are not harmonized and are re-attached afterward.

Covariates protected

Main analysis: we protect age, sex, and all of the exposome scores and subfactors:

model.matrix(~ age + sex + General_SES + School + 
							Family_Values + Family_Turmoil + 
							Dense_Urban_Poverty + Extracurriculars + 
							Screen_Time, data = exposome_df_filt)

Cognition analysis: we also protect for cognition variables:

model.matrix(~ age + sex + General_SES + School + 
							Family_Values + Family_Turmoil + 
							Dense_Urban_Poverty + Extracurriculars + 
							Screen_Time + neurocog_pc1.bl + 
							neurocog_pc2.bl + neurocog_pc3.bl,
							data = exposome_df_filt)

SES-proxy sensitivity analysis: we protect for parental education and income rather than exposome scores:

mod_exposome <- model.matrix(~ age + sex + parental_education + income + le_l_adi_addr1_national_prcnt, 
                              data = exposome_df_filt)

Outputs

Main analysis

• ../output_data/df_harmonized_exposome_cognition_A_12_10.csv
• ../output_data/df_harmonized_exposome_cognition_B_12_10.csv

Cognition analysis

• ../output_data/df_harmonized_exposome_cognition_A_12_10.csv
• ../output_data/df_harmonized_exposome_cognition_B_12_10.csv

Income sensitivity analysis

• ../output_data/df_income_sens_A_11_30.csv
• ../output_data/df_income_sens_B_11_30.csv

NOTE: on clusters/Respublica, it is easier to run the R script from the command line rather than using the notebook

• When running the Rscript, you specify in the command line which analysis (main, cognition, income) you would like to run:
  • Step2_covbat_harmonization_ses.R main
  • Step2_covbat_harmonization_ses.R cognition
  • Step2_covbat_harmonization_ses.R income
• Run the following code in Respublica for all three analyses (e.g., main, cognition, income)
• srun -c 4 –mem=32G -t 24:00:00 ~/miniforge3/envs/macedo-wm-exposome-env/bin/Rscript /mnt/isilon/bgdlab_hbcd/projects/macedo_wm_exposome/macedo_wm_exposome/code/Step2_covbat_harmonization_ses.R main

Step 3: Subset data

Python Notebook: Step3_abcd_prep_harmonized_dfs

Goal

Create analysis-specific feature subsets from each harmonized dataframe (A-trained and B-trained), producing versions that include:

• macrostructure only
• NODDI microstructure only
• DKI microstructure only
• macro + NODDI
• macro + DKI

These subsets are used in downstream analysis.

Inputs

Outputs of Step 2.

Outputs

From dfA:

• dfA_macro_12_10.csv
• dfA_NODDI_12_10.csv
• dfA_macro_plus_NODDI_12_10.csv
• dfA_DKI_12_10.csv
• dfA_macro_plus_DKI_12_10.csv

From dfB:

• same pattern with prefix dfB_…

For cognition-harmonized versions we additionally output:

• dfA__cog_12_10.csv
• dfB__cog_12_10.csv

Step 4: Mass Univariate Analysis and PCA

Python Notebook: Figure_2_Mass_Univariate.ipynb

Helper module: partial_r.py

Goal

1. Run mass-univariate linear regression models testing associations between the General Exposome factor and each WM metric (bundle x metric), controlling for covariates.
2. Quantify association strength using partial r and compute FDR-corrected p-values.
3. Summarize the tract-by-metric association pattern using PCA applied to a tract-by-metric matrix of partial r values.

Inputs

From Step 3 (harmonized + subset dataframes).

• dfA_macro_plus_NODDI_12_10.csv
• dfB_macro_plus_NODDI_12_10.csv

Covariates

The main covariate set used in mass-univariate models:

covariates = ["age", "sex", "t1post_dwi_contrast",
							"General_SES", "School", "Family_Values", "Family_Turmoil",
							"Dense_Urban_Poverty", "Extracurriculars", "Screen_Time"]

Mass-univariate model

For each WM feature:

• Outcome: standardized feature: y = z-score(feature)
• Full model: y ~ z(covariates), OLS regression
• Reduced model: same model excluding General Exposome score
• Total models: 1,302 models (62 bundles x 21 metrics, 3 NODDI and 18 macrostructure)
• Significance testing: F-test (anova_lm(reduced, full))
• Multiple testing correction: Benjamini–Hochberg FDR across all features
• Partial r formula:

\[\text{partial } R^2 = \frac{R^2_{\text{full}} - R^2_{\text{reduced}}}{1 - R^2_{\text{reduced}}}\] \[\text{partial } r = \text{sign}(\beta_{\text{General\_SES}})\sqrt{|\text{partial } R^2|}\]

PCA on bundle-by-metric partial r matrix

PCA input matrix construction

1. Compute partial r results separately in split half A and split half B.
2. Average partial r values across split halves (A and B)
3. Z-score across bundles (scaling is applied so metrics are comparable before PCA).
4. Fit PCA

PCA outputs

  ../output_data/pca_ALL_LRkept_bundle_scores_PC12.csv
  ../output_data/pca_ALL_LRkept_metric_loadings_PC12.csv
  ../output_data/pca_ALL_LRkept_explained_var_PC12.csv

Step 5: SA Axis Annotation

Python Notebook: Figure_3_SA_Axis_Annotation.ipynb

Goal

Test whether tract-level PCA results (PC1 bundle scores from Step 4) relate to sensorimotor–association (S-A) axis range annotations for each tract.

Inputs

• Outputs from Step 4:
  • pca_ALL_LRkept_bundle_scores_PC12.csv
  • pca_ALL_LRkept_explained_var_PC12.csv
• Tract annotation files from Joelle:
  • abbreviations.xlsx
  • tract_sa_axis_ranges_thresh50.csv

We read in the data, ensure that bundle names are consistent across datasets, then compute a Pearson correlation.

Step 6: Multivariate Prediction of Exposome Scores from White Matter Features in Held-Out Data

Python Notebook: Figure_4_Multivariate.ipynb

Helper script: regression_functions.py

Goal

Test whether person-specific white matter features can predict General Exposome in held-out split-half data, using multivariate ridge regressions.

We evaluate prediction across three feature sets:

1. NODDI metrics only
2. Macrostructure metrics only
3. Macrostructure + NODDI (combined)

We then visualize predicted vs. actual performance in both split directions and quantify robustness using permutation tests.

Inputs

Harmonized feature datasets (from Step 3)

Each file is already harmonized and contains columns:

• covariates: age, sex, t1post_dwi_contrast, optionally estimated_brain_volume
• target: General_SES
• features: all columns starting with msmt_ (WM features)
• split label: matched_group (values 1/2 for the split halves)

Split-half structure and training/testing

Each harmonized dataset (A-harmonized, B-harmonized) contains both matched groups, but the harmonization was trained on one half only (Step 2).

Within rf.run_model_comparison(...), the split is:

• In A-harmonized dataframe (df_A):
  • Train, matched_group = 1/A
  • Test, matched_group = 2/B
• In B-harmonized dataframe (df_B):
  • Train, matched_group = 2/B
  • Test, matched_group = 1/A

Models

Main model: Ridge regression

• Model: sklearn.linear_model.RidgeCV
• Performance metric for reporting: Pearson correlation r(y_true, y_pred)

Supplementary model: Partial Least Squares (PLS)

• Model: sklearn.cross_decomposition.PLSRegression
• n_components = 17 selected from CV tuning

Feature handling and preprocessing

Train-only nuisance regression

Implemented in rf.prepare_direction_data(...):

1. Extract raw feature matrix X and nuisance matrix Z from train/test
2. Fit nuisance model on training only:
   • Linear regression: X_train ~ Z_train
3. Residualize:
   • X_train_resid = X_train - Ŷ_train
   • X_test_resid = X_test - Ŷ_test using the train-fit nuisance model

Train-only scaling

After residualization:

• StandardScaler() is fit on training only for X and y
• both train/test are transformed using the train-fit scalers

Outputs from rf.run_model_comparison(...)

Performance dictionary

• r_A and r_B, error metrics (mse, r2, medae), permutation p-values (p_val_A, p_val_B)

Haufe weights

• For interpretability, coefficients are Haufe transformed: haufe = cov(X_train) @ weights

Permutation testing

Goal: Quantify whether observed prediction performance exceeds chance.

Procedure

1. Compute test correlation r_true
2. Repeat n_perm = 1000 times:
   • permute y_train (training labels only)
   • refit model on permuted training labels
   • evaluate on the test set
3. Calculation of p-value from permutation test:

\[p = \frac{\#\{ r_{\text{null}} \ge r_{\text{true}} \} + 1}{n_{\text{perm}} + 1}\]

Note: permutation testing takes a while to run (about 16 hours)!

Step 7: Cognition vs. Exposome — Unique Multivariate and Mass-Univariate Effects

Python Notebook: Figure_5_Cognition.ipynb

Helper scripts: partial_r.py and regression_functions.py

Goal

Evaluates whether white-matter features explain unique variance in:

• General cognition beyond the general exposome (SES) and
• General exposome (SES) beyond cognition

using both:

1. multivariate ridge regression and
2. mass-univariate partial r (feature-level effects)

Input data

Harmonized split-half ABCD datasets, cognition and SES variables (Step 3)

e.g. dfA_macro_plus_NODDI_cog_12_10.csv)

Targets and confound structures

Analysis	Target	Confounds	Interpretation
Cognition (base)	neurocog_pc1.bl	age, sex, QC	WM → cognition
Cognition	SES	neurocog_pc1.bl	age, sex, QC, General_SES	WM explains cognition unique of SES
SES (base)	General_SES	age, sex, QC	WM → SES
SES	Cognition	General_SES	age, sex, QC, neurocog_pc1.bl	WM explains SES unique of cognition

Multivariate Prediction (Ridge Regression)

Procedure

• WM features are nuisance-regressed and scaled using training data only
• Ridge models are trained on split A and tested on split B, and vice versa.
• Performance is summarized using Pearson r between predicted and observed outcomes.

Mass-Univariate Partial r

Covariate sets

cog_covariates      = [age, sex, t1post_dwi_contrastqc_prediction, neurocog_pc1.bl]

ses_covariates      = [age, sex, t1post_dwi_contrastqc_predictiont1post_dwi_contrastqrediction, General_SES]

combined_covariates = [age, sex, t1post_dwi_contrastqc_prediction, neurocog_pc1.bl, General_SES]

Note: this takes a while to run (about 40 minutes).

Step 8: Prep data for HBN Replication

Python Notebook: Figure_6_HBN_replication_data_prep.ipynb

Goal

Assemble an HBN dataset containing:

• bundle-level NODDI metrics
• bundle-level macrostructure metrics
• QC measure (t1_neighbor_corr)
• demographics (age, sex, site)
• General SES exposome score

Inputs

NODDI bundle scalarstats (subject-level TSVs)

Local folder: ../input_data/HBN/noddi/

Generated from CUBIC zipped derivatives with:

bash unzip_files.sh \
/cbica/projects/pennlinc_rbc/datasets/LINC_HBN/derivatives/QSIRECON-1-1-0_BUNDLE-STATS_zipped \
/cbica/projects/macedo_long_wm/data/noddi \
"qsirecon/derivatives/qsirecon-wmNODDI/sub-*/ses-1/dwi/sub-*_bundles-DSIStudio_scalarstats*"

Refer to Tien’s A12D guide for downloading HBN data:

• https://pennlinc.github.io/AI2D/docs/datasets/HBN/
• https://pennlinc.github.io/AI2D/docs/get_data#31-get-data-without-datalad

We keep only variables {icvf, isovf, od} and take the mean value per bundle per metric (the mean column)

Macrostructure bundle stats (subject-level CSVs)

Local folder: ../input_data/HBN/macro/

Generated from CUBIC zipped derivatives with:

bash unzip_files.sh \ 
/cbica/projects/pennlinc_rbc/datasets/LINC_HBN/derivatives/QSIRECON-1-1-0_BUNDLE-STATS_zipped \
/cbica/projects/macedo_long_wm/data/macro \
"qsirecon/derivatives/qsirecon-MSMTAutoTrack/sub-*/ses-1/dwi/sub-*msmt_bundlestats*"

QC files (subject-level TSVs)

Local folder: ../input_data/HBN/qc/

Generated from:

bash unzip_files.sh \
/cbica/projects/pennlinc_rbc/datasets/LINC_HBN/derivatives/QSIPREP-1-0-1_zipped \
/cbica/projects/macedo_long_wm/data/qc \
"qsiprep/sub-*/ses-1/dwi/sub-*desc-image_qc*"

We extract t1_neighbor_corr from each file.

Demographics / site metadata

Local folder: ../input_data/HBN/participants.tsv

Location on CUBIC: /cbica/projects/pennlinc_rbc/datasets/LINC_HBN/BIDS/participants.tsv

Columns: participant_id, Sex, Age, site

Exposome score for HBN

Local folder ../input_data/HBN/HBN_General_SES.csv

Location on CUBIC: /cbica/projects/thalamocortical_development/sample_info/HBN/HBN_General_SES.csv

Output: ../output_data/cleanedinput_data/HBN/final_hbn_dataset_12_23.csv

Step 9: Covbat Harmonization of HBN data and Nuisance Regression

R markdown: Figure_6_HBN_harmonization.Rmd

Run this with:

srun -c 4 –mem=32G -t 24:00:00 ~/miniforge3/envs/macedo-wm-exposome-env/bin/Rscript /mnt/isilon/bgdlab_hbcd/projects/macedo_wm_exposome/macedo_wm_exposome/code/Figure_6_HBN_harmonization.R

Goal

Harmonize HBN dataset to the ABCD feature space for replication analyses, by:

1. combining ABCD and HBN into a single harmonization dataset
2. harmonizing features with CovBat (site as batch), anchored to ABCD site 16
3. retaining only HBN rows after harmonization
4. nuisance regression HBN features for age (natural spline, df=3) and sex

Input data

From HBN_replication_regression.ipynb and Step1_Data_Prep.ipynb, respectively

hbn_raw  <- read.csv("../input_data/HBN/final_hbn_dataset_12_23.csv")
abcd_raw <- read.csv("../output_data/exposome_FINAL_cognition_11_10.csv")

We combine these datasets and only include features that are in both ABCD and HBN. Then we harmonize, with reference site set to ABCD site 16 (the site with the largest sample), with software and scanner versionanon0b7.Siemens_VE11B (harmonization procedure is otherwise the same as in Step 2).

Protected covariates (kept in signal during harmonization)

• ns(age, df=3) + sex + General_SES

Nuisance regression on HBN (post-harmonization)

After harmonization, run nuisance regression in HBN only:

• nuisance design: ~ ns(age, df=3) + sex + qc_prediction
• fit all features simultaneously via limma::lmFit

This produces harmonized and residualized features suitable for replication analyses.

Output:../output_data/HBN_harmonized_refABCD16_then_nuisanceResid_12_23.csv

Step 10: HBN Replication

Rmd: Figure_6_HBN_GAM.Rmd

Python Notebook: Figure_6_HBN_replication_regression.ipynb

Helper scripts

• regression_functions.py (prediction of held out data + permutation testing)
• partial_r.py (mass-univariate partial r)

Goal

1. Train a multivariate model in ABCD and test out-of-sample on HBN (external generalization).
2. Characterize feature-level HBN associations with SES with mass-univariate analysis.

Input data

• HBN harmonized dataset (from Step 9):

  ../output_data/HBN_harmonized_refABCD16_then_nuisanceResid_12_23.csv

• ABCD harmonized split-half datasets (from Steps 2–3):

  ../output_data/dfA_macro_plus_NODDI_12_10.csv

  ../output_data/dfB_macro_plus_NODDI_12_10.csv

Multivariate prediction on held out data (ABCD → HBN)

Target: General_SES

Model: ridge regression

Confounds in ABCD model training: ['age', 'sex', 't1post_dwi_contrastqc_prediction']

Feature set: bundle_msmt_cols (macrostructure + NODDI features, columns starting with msmt_)

External testing dataset: HBN: df_HBN (already harmonized + nuisance regressed in Step 9)

Note: HBN features are already nuisance regressed in Step 9, so the external testing call uses:

test_features_already_residualized=True. This prevents redoing nuisance regression.

Permutation testing (ABCD → HBN)

Same procedure as in Step 6. Permutes training labels only.

Saved results: Stored as .npz bundles containing r_true, null, and p_val:

• ../figures/perm_results_external/ABCD_A_to_HBN_macroNODDI_perm.npz
• ../figures/perm_results_external/ABCD_B_to_HBN_macroNODDI_perm.npz

Within-HBN repeated cross-validation (sanity check)

Goal: Estimate how well SES is predictable within HBN using the same feature set.

Method: Repeated 5-fold CV (20 repeats) within HBN, fits ridge regression in each fold.

Output: Fold-level and repeat-level performance.

HBN mass-univariate partial r (feature-level SES effects)

Goal: Identify which bundles/metrics in HBN show significant SES associations.

Target covariate: General_SES

Covariates in model: ["General_SES", "age", "sex", "qc_prediction"]

Multiple-comparisons control: FDR across all features