Patterns of Social Determinants of Health and Child Mental Health, Cognition, and Physical Health

This cohort study identifies social determinants of health patterns and estimates their associations with US children’s mental health, cognition, and physical health.

This supplemental material has been provided by the authors to give readers additional information about their work.

eMethods 1. Social Determinants of Health Variables
We included 84 neighborhood-level SDOH variables across seven SDOH domains as below.

Construction of the Structural Bias/Stigma Variable
The bias variables are previously validated and modeled as indicators in factor analysis. 1 The end product was a factor score, which determined the structural stigma score for each state across various stigma domains.Except for anti-immigrant sentiment, where available indicators were limited, all indicators used had measurements for every state, including Washington, D.C (as opposed to limiting to the 17 states where the ABCD sample were recruited).The inclusion of all 50 states was crucial to situate the factor scores within the nation's overall distribution.Modelbased factor scores are centered around a mean of 0 in a normal distribution.
Every measure was coded to indicate that higher levels represented greater structural stigma.These measures were standardized to the average response value for all respondents, irrespective of their state of residence, and were then aggregated at the state level.As a result, each state's measure was the mean of the standardized individual responses for respondents living in that state.This rule did not apply to state-level variables, which were already at the state level and therefore required no further aggregation or standardization.All measures were included in the model selection if they represented residents from every state, including Washington, D.C.However, anti-immigrant sentiment, which depended heavily on state policy measures, did not include data from Washington, D.C., and was hence excluded from the models.
All responses at the state level were aggregated, regardless of the year of the survey.This process of averaging the responses enabled all states to have a substantial number of respondents, despite year-to-year sampling variation, thereby minimizing measurement error.Moreover, previous studies have indicated that while structural sexism and racism have seen a national decrease over time, the relative stigma levels of individual states (that is, rankings in relation to other states) have remained consistent.This suggests that a time-invariant measure is a valid method of operationalizing this construct.For more information, readers can refer to the original article describing the process 1 .

Child Mental Health Outcomes
We used the CBCL-generated summary scores for three sub-categories-internalizing, externalizing, and total problems-and scores for eight individual dimensions within these sub-categories.Raw scores of these scales were converted to t-scores using sex-and age-based norms from population-based studies.Higher scores indicated a greater prevalence of problems., with a t-score greater than 60 representing probable disorder.The analyzed CBCL scores include: The battery is psychometrically sound, incorporating item response theory and computerized adaptive testing, and is based on normative data from 5,000 participants as part of the NIH Blueprint for Neuroscience Research.The battery is psychometrically sound, incorporating item response theory and computerized adaptive testing, and is based on normative data from 5,000 participants as part of the NIH Blueprint for Neuroscience Research. 24][5][6] We used the seven NIH Toolbox CB individual items, crystallized intelligence and fluid intelligence composite scores, cognitive function total score (derived using scores from all seven test domains).
© 2023 American Medical Association.All rights reserved.
We also used the Wechsler Intelligence Scale for Children-Fifth Edition (WISC-V). 7e conducted three sensitivity analyses to assess the stability of our identified clusters (i.e., SDOH patterns).

Child Physical Health Outcomes
1. Sensitivity to dropping samples due to missing values.We re-analyzed the data by including children whose SDOH variables are with a missing rate of < 70%.Like the primary analysis, the SDOH variables were scaled based on z-score.K-nearest neighbors (KNN) imputation 10 was used to address missing values.After that, cluster analysis was conducted to re-identify SDOH patterns.
2. Sensitivity analysis at the neighborhood level.Since we don't have geographical information of the participants, we considered that children with the same SDOH profiles are from the same neighborhood.The 10,504 children included in the primary analysis are residing in 9,137 different neighborhoods.We remove duplicated samples within each neighborhood and re-conducted cluster analysis.
3. Sensitivity to the study sample.Specifically, following our previous study, 11 we randomly split the entire cohort into 5 folds.Then each time, we successively dropped 1 fold (20% of the sample) and used the remaining 4 folds (80% of the sample) to construct a subset.We then re-identified SDOH patterns in each subset.
Of note, in these sensitivity analyses, SDOH patterns were re-identified using the HAC algorithm, following the same criteria as in the primary analysis for determining the optimal number of clusters.We compared the reidentified SDOH patterns with those identified in the primary analysis presented in the main text.Cluster labels from the primary analysis and the sensitivity analyses were aligned based on a manual review of the SDOH profiles of the clusters.We then assessed for overlap between the SDOH patterns derived in the primary analysis with those derived in the sensitivity analyses.

eResults 1. Determination of the Optimal Number of Clusters (ie, SDOH Patterns) in the Primary Analysis
In the primary analysis, we took several steps to determine the optimal number of clusters, (i.e., SDOH patterns) based on the multidimensional, heterogenous SDOH data.Each SDOH pattern represents a cluster of children exposed to similar patterns of 84 SDOH indicators across the 7 domains.
Initially, the 'NbClust' algorithm suggested the presence of three potential clusters, as deduced by majority voting across 14 cluster measurements.However, to prevent a single SDOH pattern from dominating the study population, we took the largest cluster (which comprised over 64% of the children in our cohort) and split it into two separate clusters.This process led us to identify 4 distinct SDOH patterns or clusters.
Hence, the optimal cluster (SDOH pattern) number was 4 in the primary analysis.

eResults 2. Sensitivity Analysis to Validate the Stability and Reproducibility of SDOH Patterns
In all sensitivity analyses, the HAC algorithm still detected the 4-cluster structure (i.e., 4 SDOH patterns), and the same as what were identified in the primary analysis.The re-identified clusters (i.e., SDOH patterns) were highly overlapped with that identified in the primary analysis.(See eFigures 6-8.)

eMethods 1 . 2 . 3 . 1 . 1 . 2 . 3 . 4 . 5 . 6 . 7 . 8 . 1 . 2 .
Social Determinants of Health Variables eMethods Detailed Measures of Study Outcomes eMethods Determination of the Optimal Cluster Number eMethods 4. Sensitivity Analyses eResults Determination of the Optimal Number of Clusters (ie, SDOH Patterns) in the Primary Analysis eResults 2. Sensitivity Analysis to Validate the Stability and Reproducibility of SDOH Patterns eReferences eFigure An Illustration of Missing Values in the Child  SDOH Variable Matrix eFigure Dendrogram for SDOH Pattern Identification in the Primary Analysis eFigure Visualization of SDOH Patterns in 2D t-SNE Space eFigure Characteristics of the Identified Social Determinants of Health (SDOH) Patterns (With Detailed Information) eFigure Association Between SDOH Patterns With Marijuana State Law Types and Census Tract Urban Classification eFigure Results of Sensitivity Analysis to Dropped Samples Due to Missing Values eFigure Results of Sensitivity Analysis at the Neighborhood Level eFigure Results of Sensitivity Analysis to Include Data Samples eFigure 9. Forest Plot Showing Associations Between the Identified Social Determinants of Health (SDOH) Patterns and Mental Health and Suicidal Behavior Outcomes of Children at 1-Year Follow-Up eFigure 10.Forest Plot Showing Associations Between the Identified Social Determinants of Health (SDOH) Patterns and Mental Health and Suicidal Behavior Outcomes of children at 2-Year Follow-Up eFigure 11.Forest Plot Showing Associations Between the Identified Social Determinants of Health (SDOH) Patterns and Physical Health Outcomes of Children eTable The SDOH Variables Used for SDOH Pattern Identification eTable Characteristics in SDOH Variables by the Identified SDOH Patterns eTable 3. P values of associations between the identified SDOH Patterns and mental health, cognition, and physical health outcomes.

Step 1 .Step 5 .
Data collection and linkage ABCD cohort (>11,000) • 84 SDoH variables from 7 SDoH domains (based on WHO, AHRQ, CDC, literature) of children in ABCD cohort Step 2. Data pre-processing Step 3. Clustering analysis to identify SDoH pattern Step 4. Sensitivity analyses to validate the reproducibility of the identified SDoH patterns • Data cleaning (missing values, etc.) • Data transformation (rescale higher values as poorer SDoH, etc.) Mental Health • CBCL Syndrome Scales (Internalizing) • CBCL Syndrome Scales (Externalizing) • CBCL Syndrome Scales (Summary Scores) Associations between SDoH patterns and child development outcome Cognitive Health • NIH Toolbox Cognitive Battery (Crystallized Intelligence, Fluid Intelligence, Total Score) • WISC-V Matrix Reasoning Physical Health • Exercise 60 minutes • BMI • Sleep Disorders SDoH Patterns A SDoH pattern = a subg of children with similar exposures to SDoH indic

eFigure 7 .
Results of Sensitivity Analysis at the Neighborhood Level(A) Dendrogram of hierarchical clustering analysis.(B) Visualization of identified SDOH patterns in the 2D t-SNE space.(C) Confusion matrix for comparing the SDOH patterns identified by the primary analysis and sensitivity analysis.Density of color indicates the proportion of overlapped children within SDOH patterns identified by the primary analysis and sensitivity analysis.

eMethods 4. Sensitivity Analyses
We used children's general physical health outcomes and sleep disorder outcomes measured by the Sleep Disturbance Scale for Children [SDSC] 8 sub-items and composite score, including:

Characteristics in SDOH Variables by the Identified SDOH Patterns
Health and Environment Domain Access to green space: Percentage households without a car located further than a half-mile from the nearest supermarket, transformed to z-scores and multiplied by -1.Comparisons across all 4 SDOH patterns were performed using analysis of variance (ANOVA) test for continuous variables and χ2 test for categorical variables.Two-tailed Pvalues smaller than 0.05 were considered as the threshold for statistical significance.
Educational Domain Early childhood education enrollment: Percentage 18-24 year-olds enrolled in college within 25-mile radius, transformed to z-scores.reshist_addr1_coi_zed_schpov0.08 0.21 [-0.77-1.10]EducationalDomain Teacher experience: Percentage students in elementary schools eligible for free or reduced-price lunches, transformed to z-scores and multiplied by -1.reshist_addr1_coi_zed_teachxp 0.08 -0.15 [-0.91-0.40]EducationalDomain Early childhood education centers: Percentage teachers in their first and second year, transformed to z-scores and multiplied by -1.© 2023 American Medical Association.All rights reserved.a© 2023 American Medical Association.All rights reserved.eTable3.

P values of associations between the identified SDOH Patterns and mental health, cognition, and physical health outcomes.
© 2023 American Medical Association.All rights reserved.