Dependent variables

Adequate sleep duration: Parent or guardians reported the average hours of sleep that their child received on most weeknights in the past week (< 6, 6, 7, 8, 9, 10 hours, or ≥ 11 hours). Based on age-appropriate sleep duration recommendations [28] that suggest 9–12 hours/day of sleep for ages 9–12 years and 8–10 hours for ages 13–17 years, responses were coded to group children into adequate sleep duration “Yes” (sleeps recommended hours) versus “No” (does not sleep recommended hours) categories.

Bedtime regularity: Parent or guardians answered the single item, “how often does this child go to bed at about the same time on weeknights?” Response options were on a 4-point scale: always, usually, sometimes, and rarely/never and dichotomized such that “always” and “usually” responses were considered as “Yes” (has regular bedtimes) and “sometimes” or “rarely/never” as “No” (does not have regular bedtimes).

Independent variables

Independent variables were grouped in accordance with the levels of the social ecological model: individual, social, and environmental.

Individual

During the screening questionnaire, parents or guardians reported their child’s age, sex (male/female), and race/ethnicity (Hispanic, White non-Hispanic, Black non-Hispanic, or other/multiracial non-Hispanic).

Family income level

Parents or guardians reported household income in the last calendar year. Responses were grouped as follows based on the percentage of the federal poverty level (FPL): (1) 0%–99% FPL, (2) 100%–199% FPL, (3) 200%–399% FPL, and (4) 400% FPL or greater. Higher scores indicate greater poverty.

Neurodevelopmental disorder (NDD) status

Guided by DSM-V criteria, children reported to currently have one or more of the following NDDs were classified as having a NDD: Autism, ADHD, Down syndrome, intellectual disability, learning disability, Tourette syndrome, or speech/other language disorder. Children not reported as having any of these conditions were categorized as typically developing (TD).

Physical activity (PA)

Parents/guardians reported how many days in the past week the child exercised, played sports, or was active for at least 60 min. Response options: (1) 0 days, (2) 1–3 days, (3) 4–6 days, (4) every day.

Screen time

Parents/guardians reported weekday hours spent watching TV/videos, playing video games, or using electronics (excluding schoolwork). Responses: (1) < 1 hour/day, (2) 1–2 hours/day, (3) ≥ 3 hours/day.

Ever and current depression:

Parents indicated whether a healthcare provider had ever diagnosed their child with depression (1 = yes; 0 = no). If yes, they were asked if the child currently had depression (1 = yes; 0 = no).

Social level variablesMaternal mental health

Mothers rated their mental/emotional health on a 3-point scale (1 = excellent/very good, 2 = good, 3 = fair/poor), with higher scores indicating poorer health.

Family meal frequency

Parents/guardians reported how many days in the past week household members ate a meal together. Responses: (1) 0 days, (2) 1–3 days, (3) 4–6 days, (4) every day.

Difficulty making/keeping friends

Parents rated the child’s difficulty making or keeping friends on a 3-point scale (1 = no difficulty, 2 = a little difficulty, 3 = a lot of difficulty).

Being bullied

Parents or guardians reported if their child was bullied, picked on, or excluded by other children in the last year (yes/no).

Parental aggravation

On a 5-point scale (Never–Always), parents answered three questions: (1) Is this child harder to care for than others their age? (2) Does this child often bother you? (3) Are you often angry with this child? If “usually” or “always” was selected for any item, parents were categorized as “usually/always aggravated” (1); all others as (2) “seldom aggravated.”

Environmental level variables

Someone smokes inside the home: Parents reported whether someone used tobacco products inside the home (yes/no).

Food sufficiency: Parents answered how well they could afford food over the past 12 months using a 4-point scale: (1) always enough nutritious food, (2) always enough but not the right food, (3) sometimes not enough, (4) often not enough.

Neighborhood safety: Parents rated the statement “Child is safe in this neighborhood” on a 3-point scale (1 = definitely agree, 2 = somewhat agree, 3 = somewhat/definitely disagree), with lower scores indicating greater safety.

Detracting neighborhood elements: Parents or guardians reported whether the following three items were present in their neighborhood: litter or garbage on the streets/sidewalk, poorly kept housing, or vandalism (broken windows/graffiti). Answers ranged from zero to three detracting elements.

Neighborhood social support: Parents indicated agreement with three statements: (1) neighbors help each other, (2) watch out for each other’s children, (3) know where to go for help. Respondents were considered to live in supportive neighborhoods if parents “definitely agreed” with at least one and “somewhat/definitely agreed” with the others.

Neighborhood amenities: This measure is a count of how many of the following four (parent/guardian reported) amenities are present in the child’s neighborhood: sidewalks/walking paths, park/playground, recreation center/boys’ and girls’ club, or libraries/bookmobiles.

Access to family centered care: Parents of children who had a health visit in the past year reported on five experience-of-care items: provider spends time, listens, respects values, provides information, and treats family as partner. “Usually” or “always” on at least one item indicated receipt of family-centered care (1 = yes).

Statistical analysis

Descriptive statistics were generated for all study variables. Unadjusted associations between independent variables and each of the study outcomes (i.e., adequate sleep and bedtime regularity) were calculated. To determine the most important variables related to each of the outcomes (adequate sleep duration, bedtime regularity), classification random forests (CRF) were used.

CRFs are a non-parametric ensemble supervised learning method that operate by constructing a multitude of decision trees and then using the individual tree predictions to produce a final outcome prediction through majority vote. CRFs offer several key advantages over traditional regression approaches: they can handle large numbers of variables without overfitting concerns, automatically detect complex associations between variables without requiring a priori specification and are robust to outliers and missing data. To ensure that the individual predictions do not become correlated, each tree utilizes random sampling without replacement to select a subset of variables before splitting a node. Random sampling (with replacement in this case) is also used to select a training set for each tree. Upon completion of the modeling process, it is possible to “rank” the variables used to grow a given CRF by using each variable’s predictive utility, often referred to as variable importance (VIMP). VIMP is a relative measure with higher values indicating a higher level of importance in producing an accurate prediction of the outcome. When a variable with a low VIMP is removed from its model, the veracity of the model’s final prediction will not be significantly impacted. Conversely, when variables with high VIMPs are removed, the model’s final predictions are much more noticeably (and negatively) impacted.

We selected CRFs over alternative ensemble methods (e.g., XGBoost) because our primary objective was variable selection rather than predictive accuracy optimization. CRFs provide interpretable variable importance measures, stable rankings across different hyperparameter settings, and robust performance with mixed data types, making them optimal for identifying intervention targets in explanatory research rather than developing clinical prediction models.

A total of four CRFs were generated—one for each of the sleep health outcomes for NDD and TD children. To compare values more easily across variables, raw VIMP values were first multiplied by 100 before presentation. Further, all VIMPs were reported using plots of 95% confidence intervals (CI), which were in turn approximated through simulations employing bootstrapping and approximately 160,000 individual decision trees per CRF. Since VIMPs do not provide a sense of directionality, we generated separate logistic regression models for each of the CRFs. Independent variables included in each of the logistic regression models were those identified by the regression random forest model as having a high VIMP, defined here as having one of the top 5 highest VIMPs for the respective outcome/cohort pairing. The odds ratios with accompanying 95% CI’s were reported in the logistic regression model results.

All analyses were performed using R v.4.3.1. For the logistic regression results, confidence intervals not containing 1 were statistically significant. The randomForestSRC package was used to fit the random forests and extract VIMPs. The function parameters used to grow the forests were set equal to the vignette default values as conceived by the package developers.