henhaobio
Sign in

bioRxivNeurosciencePreprint

A Guide for Population-based Analysis of the Adolescent Brain Cognitive Development (ABCD) Study Baseline Data

Steven G. HeeringaiD, Patricia A. BerglundiD

PDFDOI
In-depth full-text digest available · Jump to digest ↓

Abstract

ABCD is a longitudinal, observational study of U.S. children, ages 9-10 at baseline, recruited at random from the household populations in defined catchment areas for each of 21 study sites. The 21 geographic locations that comprise the ABCD research sites are nationally distributed and generally represent the range of demographic and socio-economic diversity of the U.S. birth cohorts that comprise the ABCD study population. The clustering of participants and the potential for selection bias in study site selection and enrollment are features of the ABCD observational study design that are informative for statistical estimation and inference. Both multi-level modeling and robust survey design-based methods can be used to account for clustering of sampled ABCD children in the 21 study sites. Covariate controls in analytical models and propensity weighting methods that calibrate ABCD weighted distributions to nationally-representative controls from the American Community Survey (ACS) can be employed in analysis to account for known informative sample design features or to attenuate potential demographic and socio-economic selection bias in the national sampling and recruitment of eligible children. This guide will present results of an empirical investigation of the ABCD baseline data that compares the statistical efficiency of multi-level modeling and distribution-free design-based approaches--both weighted and unweighted--to analyses of the ABCD baseline data. Specific recommendations will be provided for researchers on robust, efficient approaches to both descriptive and multivariate analyses of the ABCD baseline data.

In-depth summary

Claude Opus 5.5 · xhigh effort · 17 Oct 2026 · from the full text of version 1 (10 Feb 2020) · CC BY-ND 4.0

Unweighted ABCD baseline estimates lean toward higher-income, married, working-parent families, and independence-based standard errors are too small; use site-clustered ACS-calibrated weights for descriptives and three-level mixed models for regressions.

Analyses of the ABCD cohort that ignore its design risk biased population estimates and overconfident inference, because children are clustered within 21 non-randomly chosen sites, schools and families, consent may favour better-off families, and four sites add a twin oversample. The guide documents the study's propensity weights calibrated to the American Community Survey (ACS) and compares weighted design-based with mixed-model analyses of the baseline data to give concrete recommendations.

Key results

  1. ShownUnweighted, the ABCD baseline sample matches U.S. children on age, sex and household size but leans toward wealthier, two-working-parent, Western and non-rural families, and under-represents Asian, Native Hawaiian/Pacific Islander (NHPI) and American Indian/Alaska Native (AIAN) children combined, mainly Asian ones.

    Versus the 2011-2015 ACS: family income of $100,000 or more 40.7% vs 27.3%; two working parents 50.6% vs 40.8%; Census West 34.3% vs 24.1%; Asian, AIAN and NHPI children combined 3.2% vs 5.9%; expected share enrolled in rural school districts 12.5% vs 17.5% nationally per the National Center for Education Statistics. · 5. Properties of the ABCD Baseline Sample Cohort…

  2. ShownWeighted, design-based estimation with sites as clusters is the recommended approach for descriptive statistics: unweighted proportions stay biased toward married, working, higher-income families, and standard errors that assume independence are seriously too small.

    Raking makes weighted age, sex and race/ethnicity shares equal ACS exactly; weighted socio-economic shares come close. Standard-error underestimation grows with intra-cluster correlation, worst for traits concentrated in few sites (e.g. African American children). Weighted and unweighted means and percentiles of NIH Toolbox scores and emergency room (ER) visits differ only slightly. The tables are not in this text. · 7.A. Descriptive estimates-comparison of…

  3. SuggestedWith twins pooled into the sample, mixed models need a family level: a two-level (site, child) model attenuated estimates and seriously overstated standard errors, whereas the three-level model used by the ABCD Data Exploration and Analysis Portal (DEAP) closely matched ordinary least squares (OLS).

    Linear models of uncorrected NIH Toolbox Flanker and Reading scores, focusing on the age and top-income coefficients. A two-level model fitted only to the general-population sample (twin oversample excluded) came closer to the three-level results. Coefficient and standard-error tables are not in this text. · 7.B.2 LMMs for ABCD: Three vs. Two Levels?; 7.B.1 OLS vs. Linear Mixed Effects Models (LMM)

  4. SuggestedFor Flanker, the three-level mixed model and weighted design-based regression gave very similar age and income effects, with larger design-based standard errors; the main disagreement, for Reading and in the Poisson model, was the income effect, which the authors attribute to weights that vary with family income.

    OLS shows age-9 children about 2.7 units below 10-year-olds on Flanker and Reading, and $200K+ families 4.5-5.5 points above the lowest income group. For Reading, the mixed-model income coefficient was considerably lower than the pooled design-based one, though not significantly once uncertainty is considered. Poisson models of lifetime ER visits showed only minor relative-risk differences. · 7.B.3 Three-level LMM vs. Design-based…; 7.C Generalized Linear Model: Poisson…; 7.B Linear Regression analyses—comparison of…

  5. SuggestedPooling the twin oversample with the general sample barely changed weighted descriptive estimates, while moderately lowering their standard errors, and, once family clustering was modelled, gave regression results similar to the general sample alone.

    Twin sites enrolled 150-250 extra twin pairs each from state registries. Pooled and general-population-only weighted estimates differed only slightly. After demographic and socio-economic controls, test scores and ER visits did not appear to differ substantially between twins and single births; the authors caution this may not hold for genetic or familial analyses. · 7.A. Descriptive estimates-comparison of…; 8. Summary: Recommendations for research…; Special twin supplement

Methods, data and reproducibility

Propensity weights came from a frequency-weighted logistic regression of ABCD versus ACS membership on 8 shared demographic and socio-economic variables, inverted, trimmed at the 2nd/98th percentiles and raked to ACS age, sex and race/ethnicity counts. Unweighted, weighted design-based and two- or three-level mixed-model estimates were then compared for descriptives and for linear and Poisson regressions.

  • ABCD baseline cohort: U.S. children aged 9-10 recruited mainly through probability samples of schools at 21 sites from 2016 to 2018, including a twin oversample at four sites · human · n = 11,874 children (11,184 in the weighting model)
  • American Community Survey 2011-2015 Public Use Microsample, the benchmark for U.S. children aged 9-10 · human · n = 376,370 children from over 8,000,000 households
  • National Center for Education Statistics urban/rural school classification and 2014 enrollment data for the site catchment areas
Methods & tools
R survey package, R GAMM4, R MICE, R rake (iterative proportional fitting), SAS PROC SURVEYMEANS / SURVEYREG, Stata svy, ABCD Data Exploration and Analysis Portal (DEAP), NIH Toolbox (Flanker, Reading), SUDAAN, SPSS Complex Samples, Mplus, WesVar, IVEware (listed survey software)
Reagents & environment
Appendices print example syntax for SAS V9+ (PROC SURVEYMEANS, PROC SURVEYREG), Stata (svyset, svy) and the R survey package, declaring site as the cluster and the propensity weight; package versions are not stated.
Reproducibility gaps
No code or data links are given; ABCD data sit in the NDA/DEAP environment, with access terms not described here. The coefficient and standard-error tables for the method comparisons are not included in this text.
Limitations (6)
  • The recommendations rest on a few example analyses (two NIH Toolbox scores and ER visits) with no gold standard, and the authors warn against over-generalizing them. · stated by the authors
  • Weights calibrate only on variables measured identically in ABCD and ACS, so bias tied to unmeasured selection such as school and parental consent can remain and can rarely be verified. · stated by the authors
  • Population weighting of multi-level models was not evaluated here; preliminary trials are unreported and work is ongoing. · stated by the authors
  • Twin and singleton samples looked exchangeable for these outcomes, which may not hold for studies of genetic or familial effects. · stated by the authors
  • The text is internally inconsistent: the weighting model uses n=11,184 while the cohort is n=11,874, and the base sampling fraction is printed as both 0.00145 and 0.000145. · noted in this summary
  • Only cross-sectional baseline data are examined, so the guidance does not cover attrition or weighting for later longitudinal waves. · noted in this summary

Read this if you analyse ABCD baseline data and must choose between weighted design-based estimation and mixed models, or decide how to handle site clustering and the twin oversample.

Citations
247

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISDick '21Hagler '19Palmer '21Feczko '21Hussain '23Weigard '21
83% semantic matchNeuroscience
Anthony Steven Dick, Daniel A. Lopez, Ashley L. Watts+21 more·NeuroImage·Jun 17, 2021· 244 citations