key: cord-0256726-vuztffdw
authors: Spechler, P. A.; Gutierrez, R. M.; Tapert, S. F.; Thompson, W. K.; Paulus, M. P.
title: The beneficial effect of sleep on behavioral health problems in youth is disrupted by prenatal cannabis exposure: A causal random forest analysis of ABCD data.
date: 2022-05-13
journal: nan
DOI: 10.1101/2022.05.12.22275012
sha: 2b2604a927b55469a6f173df5a1d9c50ee5b8129
doc_id: 256726
cord_uid: vuztffdw

Importance: Declining risk perceptions and touted medicinal claims about cannabis are likely related to the increasing prevalence of cannabis use during pregnancy. Yet, it remains unclear if prenatal cannabis exposure yields lasting neurodevelopmental alterations in children, and which facets of their behavioral health might be affected. Objective: To determine if prenatal cannabis exposure moderates the relationship between changing sleep hours on changing mood and behavioral problems in youth. Design: Causal random forest analysis of ABCD cohort data using baseline (ages 9-10) and year-1 follow up information. Setting: 11,875 children and their parents were assessed at 21 acquisition sites across the United States between July 2016 to October 2018. Participants: Participants with prenatal drug use data were included, and 94% of the baseline sample also had year-1 follow up data. Exposures: A change in sleep hours from baseline to year-1 was conceptualized as the dimensional treatment variable, and self-reported drug uses of the biological mother were conceptualized as moderators of that treatment. Main Outcomes and Measures: A change in internalizing (anxious/depressed mood) and externalizing (disruptive behavior) problems from baseline to year-1 were the two outcome measures. Sociodemographic and other prenatal drug exposures were included as covariates. Given the observational nature of the multi-site ABCD study, all hypotheses tested here were formulated after data collection. Results: There were n=9,826 children (4,663 female) included in analyses, of which n=605 (303 female) had any prenatal cannabis exposure. On average, internalizing problems largely remained stable from baseline (M=48.54, SD=10.56) to year-1 (M=48.75, SD=10.64)(t9,825=2.40, p=.016), as did externalizing problems (baseline M=45.64, SD=10.21; year-1 M=45.23, SD=10.09) (t9,825=5.44, p<.001). Significant average treatment effects indicated that participants with more sleep hours reported less internalizing (ATE=-.35, SE=.08, p=.003) and externalizing (ATE=-.28, SE=.07, p=.028) problems over time. However, a significant heterogeneous treatment effect by prenatal cannabis exposure (conditional-ATE=.92, SE=.36, p=.011) was found for the internalizing model, such that participants with this exposure did not show a beneficial effect of changing sleep on changing mood (B=.06, SE=.25). This finding was specific to cannabis, as no such effect was found for any sociodemographics or prenatal alcohol or tobacco exposures for the internalizing or externalizing models. Conclusions and Relevance: This study uncovers an actionable target (sleep) to improve mood and behavioral problems in typically developing children not prenatally exposed to cannabis. However, changing sleep may not have a similar effect for youth prenatally exposed to cannabis. Given the importance of the endocannabinoid system in regulating perinatal neurodevelopment and sleep, these findings suggest that cannabis exposure may interact with those processes to diminish the beneficial effects of sleep on mood in children, and thus calls into question the safety of cannabis use during pregnancy.

. This finding was specific to cannabis, as no such effect was found for any sociodemographics or prenatal alcohol or tobacco exposures for the internalizing or externalizing models.

This study uncovers an actionable target (sleep) to improve mood and behavioral problems in typically developing children not prenatally exposed to cannabis. However, changing sleep may not have a similar effect for youth prenatally exposed to cannabis. Given the importance of the endocannabinoid system in regulating perinatal neurodevelopment and sleep, these findings suggest that cannabis exposure may interact with those processes to diminish the beneficial effects of sleep on mood in children, and thus calls into question the safety of cannabis use during pregnancy.

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 13, 2022. ;  https://doi.org/10.1101/2022.05.12.22275012 doi: medRxiv preprint

Healthy sleep is an essential part of child and adolescent development. An extensive literature highlights the importance of sleep on physiological, neurological, and emotional development (see Tarokh and colleagues for review) 1 . At the same time, deficits in sleep have been associated with disadvantages in each of these domains.

For instance, poor sleep has been linked to the development of behavioral health issues like internalizing and externalizing problems 2 , with mounting evidence supporting the directionality that sleep disturbance is a precursor to future mood and behavioral disruption in youth 3, 4 .

Various exposures and environmental factors can have profound effects on sleep, which, in turn, might also affect mood and behavior. For example, exogenous cannabinoids like tetrahydrocannabinol (THC) readily cross the placenta and interacts with the fetal endogenous cannabinoid (endocannabinoid) system, which is heavily involved in orchestrating elements of neurodevelopment 5 . The endocannabinoid system is also implicated in regulating neural systems for sleep activity 6 , and may be vulnerable to interference by exogenous cannabinoids (e.g., THC) during development. Several studies suggest that, relative to children without prenatal cannabis exposures, those with exposures exhibit poor sleep features including more sleep disturbances (i.e., motility and arousal) and neural waveform differences during sleep-cycles as measured via neonatal electroencephalography (EEG) 7, 8 . This pattern of poor sleep following prenatal cannabis exposure appears to persist at 3-years 9 , and into late childhood 10 . As prenatal cannabis use may disrupt the child's prenatal brain development and influence their sleep systems, children with these exposures may be less equipped to achieve a similar level of sleep quality relative to non-exposed children.

A recent report by Paul and colleagues analyzed data from the Adolescent Brain Cognitive Development™ (ABCD) Study 11 and found that prenatal cannabis exposure was correlated with higher instances of internalizing and externalizing problems in children at baseline (age 9-10) 12 . Considering the literature highlighting the importance of sleep on mood and behavior in healthy children 1 , and other studies finding sleep disadvantages in children with prenatal exposure, it was therefore hypothesized that an . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted May 13, 2022. ; https://doi.org/10.1101/2022.05.12.22275012 doi: medRxiv preprint effect of sleep on mental health outcomes would vary by prenatal cannabis exposure status (i.e., moderation).

A rigorous examination of the causal effect of modifying sleep on child mental health outcomes would require a randomized controlled trial, which is impractical and prohibitively expensive. Alternatively, one can begin interrogating causal processes by applying statistical causal inference techniques to observational data 13 . Given the large sample and longitudinal nature of the ABCD Study 11 , this dataset is well suited for causal inference analyses. While inferring causality from observational data is challenging due to the possible influence of unobserved confounders, imperfect measurement of observed confounders, and reciprocal relationships/reverse causality 14 , one may begin to make causal inferences by analyzing data under the assumptions of the potential outcomes framework 13 . This framework infers causality by comparing outcomes between similar participants who either did or did not experience a treatment/exposure (e.g., active drug vs. placebo), and can be extended beyond categorical levels to study dimensional treatments or exposures.

In the present study, a change in sleep hours between baseline and year-1 follow up was conceptualized as a dimensional "treatment" measure to infer its effect on the degree of change in internalizing or externalizing problems (outcome measures). As the goal was to examine a modifiable behavioral feature (sleep) potentially influencing child health outcomes, prenatal cannabis exposure was not modelled as the treatment measure because its modification is impossible in this sample.

A causal random forest (CRF) model 15 was implemented to test our hypotheses.

This approach uses causal decision trees to find highly similar participants given a set of covariates (analogous to a matching procedure), and then examines differences in outcomes among those participants. Assuming that all variables influencing changes in sleep hours and mental health outcomes have been measured and accounted for ("unconfoundedness assumption"), the CRF approach has been shown to make valid statistical inferences 15 . Here, an average treatment effect (ATE) reflects the sample average predicted change in internalizing/externalizing problems given a change in sleep hours. As causal decision trees can efficiently model high-order interaction terms, this method also uncovers the features contributing to heterogeneity in treatment . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 13, 2022. ; https://doi.org/10.1101/2022.05.12.22275012 doi: medRxiv preprint effects. These conditional average treatment effects (CATE) reflect how treatment outcomes vary by observable characteristics of an individual. In doing so, we can parse heterogeneity to understand for whom a specific treatment would be more or less beneficial.

The ABCD study is a longitudinal project with 21 sites in the United States participating in data acquisition (abcdstudy.org). Starting at ages 9-10, biological, behavioral, and psychosocial information on children and their families are collected periodically as they advance into early adulthood (age [19] [20] . Parents provided written informed consent prior to participation, and children provided assent before data collection. Details on the ABCD study have been extensively described elsewhere 11, 16 .

Prior to running any analyses here, this study was partially pre-registered (from a larger study on prenatal cannabis exposures) on the open science framework platform (https://osf.io/a23s5).

Prenatal cannabis exposure was measured via a parent or caregiver self-report developmental history ('DevHx') questionnaire. Although individuals were asked about cannabis use before and after knowledge of their pregnancy, we collapsed across these exposure statuses to increase power to detect small effects. Likewise, a similar question was asked for any alcohol or tobacco use during their pregnancy. For all drug use questions, the parent or caregiver was asked to report about the biological mother's use of drugs during their pregnancy.

Disturbance Scale (SDS) for Children 17 . While sleep duration was collected on an ordinal scale (e.g., "5-7 hours"), the median at each interval was used for baseline and year-1 follow up. The difference between year-1 and baseline was calculated (so positive values signify an increase in sleep over time) and modelled as the treatment measure.

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 13, 2022. ; https://doi.org/10.1101/2022.05.12.22275012 doi: medRxiv preprint Child psychopathology was measured via the child behavior checklist (CBCL) 18 , which includes a summary T-score for total internalizing and externalizing problems, where higher scores reflect more problems. Likewise, the difference in problem scores between year-1 and baseline was calculated (i.e., positive values indicate an increase in problems over time) and used as the outcome measure. In light of the results from Paul et al. 12 that found that children with prenatal cannabis exposure exhibited elevations in both the internalizing and externalizing domain, we ran two separate models (change in internalizing or externalizing problems as the outcome).

To account for confounding effects, key variables thought to be related to both sleep duration and internalizing/externalizing problems were included in the model.

Specifically, sociodemographics like household income, parental education level, cohabitation status (i.e., married or separated), maternal age during pregnancy, child age, body-mass index, sex, and an indicator for white or non-white ethnicity were included. Other data from the mental health domain like maternal history of depression, and child baseline internalizing and externalizing scores were also included. Finally, any prenatal alcohol or tobacco use were also included as covariates.

Starting data were drawn from the baseline (N=11,873) and year-1 follow-up cohort contained in data release 3.0 from the ABCD study (abcdstudy.org). Participants were then filtered by having completed the developmental history questionnaire (n=11,533). While some siblings were included in ABCD, only one child from each family was randomly selected for inclusion to ensure independent sampling, thus yielding a sample of n=10,462. From there, only those with SDS and CBCL data available at baseline and year-1 were included for analysis (n=9,826). See figure 1 for inclusionary flow chart.

Data were analyzed using the 'grf' package in R via the 'causal forest' function with default parameter settings (https://grf-labs.github.io/grf/). A causal random forest 19 (CRF) is a data driven method designed to estimate the average and heterogeneous (e.g., moderated) treatment effects for a candidate causal pathway. Like a traditional random forest, a CRF constructs many decision trees to model a treatment effect on an . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 13, 2022. ; https://doi.org/10.1101/2022.05.12.22275012 doi: medRxiv preprint outcome measure given a set of covariates. This approach can identify features that contribute to heterogeneity in the treatment effect (called conditional average treatment effects, CATE). Summary statistics for the CATEs are obtained using a doubly-robust estimation procedure. CRFs produce asymptotically normal standard errors and meets other assumptions necessary to perform statistical inference 15, 19 . Moreover, the CRF uses "honest" tree splitting and out-of-bag model evaluation to enhance generalizability Any factors that were found to reflect significant conditional average treatment effects were probed further by plotting "simple slopes" 20 . For a binary factor (i.e., any prenatal cannabis exposure), the treatment effect of a change in sleep hours (x) on a change in internalizing/externalizing (y) was plotted at the two levels of the covariate (with vs. without exposure). Hence, this technique depicts the moderating effect of that covariate on the causal pathway and helps to convey the individual differences in treatment effects.

From the full ABCD baseline cohort of 11,873 children, there were n=9,826 who met inclusionary criteria (figure 1) for analyses. Within that sample, there were n=605 with any prenatal cannabis exposure. See table 1 for information on the analytic sample, separated by prenatal cannabis exposure status for descriptive purposes. 

This study examined whether the beneficial relationship between sleep and internalizing or externalizing problems was altered in children who were prenatally exposed to cannabis. Results indicated that an increase in sleep hours had a small causal effect on a decrease in internalizing and externalizing problems for the average child. This average treatment effect was moderated by prenatal cannabis exposure for internalizing problems, but not externalizing problems. Moreover, this finding appears to be specific to prenatal cannabis exposure, as heterogeneity was not found for prenatal alcohol or tobacco exposures. Hence, this study showcases how using a data driven causal inference framework can rigorously uncover substance use related individual differences in treatment effects.

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 13, 2022. The identified average treatment effect suggests that sleep is one modifiable behavioral domain that could be targeted to improve internalizing and externalizing problems for the average child. This finding is in line with other studies that found that sleep disturbances likely precede behavioral health problems in adolescence 3, 4 .

Therefore, building and maintaining healthy sleep behaviors could be a proactive way to protect against internalizing or externalizing problems in most typically developing children.

This study also suggests that certain subgroups of children (i.e., those with prenatal cannabis exposure) may not respond to sleep interventions in a similar fashion to their non-exposed peers. Hence, prenatal cannabis exposure might complicate treatment options as those children may require a different approach beyond sleep interventions to modify behavioral health problems. Nonetheless, these interpretations should be viewed considering the small effect sizes found here. And although small effects are to be expected 21, 22 , they underscore how sleep and prenatal cannabis exposure are two elements that are part of a constellation of factors influencing mental health and wellbeing during child development.

These findings stand in contrast to the low risk perceptions 23 Future studies are also needed to determine the biological mechanisms that explain why and how prenatal cannabis exposure disrupts with the effect of sleep on mood. As the endocannabinoid system helps regulate neurodevelopment and sleep architecture 6 , there is a strong likelihood that prenatal cannabis exposures could interfere with those systems. This hypothesis is supported by the many studies . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 13, 2022. ; https://doi.org/10.1101/2022.05.12.22275012 doi: medRxiv preprint reporting disadvantageous associations between prenatal cannabis use and child sleep features [7] [8] [9] [10] . Hence, children with these exposures may be at a functional disadvantage to experience the same beneficial effects of sleep on their mood relative to their peers without such exposure. However, the precise mechanism linking endocannabinoid interference to mood-related outcomes remains unclear. Nonetheless, this study adds to the mounting evidence that prenatal cannabis use may impact the child's brain and behavioral development and should thus motivate individuals to avoid using cannabis while pregnant.

Limitations of this study include the self-report nature of the treatment and outcome measures, which are not as reliable as objective measurement of sleep hours collected via wearable technology or biochemical verification of cannabis use during pregnancy. Additionally, the observational nature of the ABCD study, and the small effect sizes reported here should be considered when interpreting the possible causal nature of these findings. Despite using a rigorous causal inference framework, these results would benefit from being reproduced by a study using an experimental design that manipulates sleep hours. And although a set of confounders were included in the model, it is possible that there are omitted variables that could explain differences in sleep and behavior. Additionally, the use of two time points makes it difficult to parse out the temporal ordering of causal effects 25 . In future ABCD data releases, studies examining trajectories of sleep, mood, and behavior throughout adolescence will be possible by leveraging multi-timepoint data.

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 13, 2022 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) Point estimates obtained from a doubly robust estimation procedure that regressed the outcome measure (internalizing/externalizing problems) onto the treatment measure (sleep hours) and covariates (tabulated here). Bars reflect two standard errors of the mean. The only significant conditional average treatment effect was found for changing sleep on changing internalizing problems by prenatal cannabis exposure. he re he ng . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 13, 2022. ; https://doi.org/10.1101/2022.05.12.22275012 doi: medRxiv preprint . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted May 13, 2022. ; https://doi.org/10.1101/2022.05.12.22275012 doi: medRxiv preprint Tables: Table was divided by prenatal cannabis exposure ("CB Exposure") status for descriptive purposes. P-values from two-sample t-tests (or Mann-Whitney U tests for categorical data). Note, given the large sample, small differences are significant at p<.05. All features except baseline sleep hours were included as covariates in causal random forest analyses.

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted May 13, 2022. ; https://doi.org/10.1101/2022.05.12.22275012 doi: medRxiv preprint

Sleep in adolescence: Physiology, cognition and mental health

Trajectories of sleep problems in childhood: associations with mental health in adolescence

A Systematic Review Assessing Bidirectionality between Sleep Disturbances, Anxiety, and Depression

A meta-analysis and model of the relationship between sleep and depression in adolescents: Recommendations for future research and clinical practice

Marijuana Use in Pregnancy and While Breastfeeding

The Effects of Prenatal Alcohol and Marijuana Exposure: Disturbances in Neonatal Sleep Cycling and Arousal

Effect of prenatal marijuana exposure on sleep wake cycles and amplitude-integrated electroencephalogram (aEEG)

A Longitudinal Study of Prenatal Marijuana Use: Effects on Sleep and Arousal at Age 3 Years

Prenatal cannabis exposure and sleep outcomes in children 9-10 years of age in the adolescent brain cognitive development SM study

The conception of the ABCD study: From substance use to a broad NIH collaboration

Associations Between Prenatal Cannabis Exposure and Childhood Outcomes: Results From the ABCD Study

Causal inference in statistics, social, and biomedical sciences

The Taboo Against Explicit Causal Inference in Nonexperimental Psychology

Estimation and Inference of Heterogeneous Treatment Effects using Random Forests

The Adolescent Brain Cognitive Development (ABCD) study: Imaging acquisition across 21 sites

The Sleep Disturbance Scale for Children (SDSC) Construct ion and validation of an instrument to evaluate sleep disturbances in childhood and adolescence

The Child Behavior Checklist and related instruments

Generalized random forests

Moderation in Management Research: What, Why, When, and How

Meaningful associations in the adolescent brain cognitive development study

Recalibrating expectations about effect size: A multi-method survey of effect sizes in the ABCD study

Women's perspectives about cannabis use during pregnancy and the postpartum period: An integrative review

Rates of Prenatal Cannabis Use Among Pregnant Women Before and During the COVID-19 Pandemic

How to Deal With Reverse Causality Using Panel Data? Recommendations for Researchers Based on a Simulation Study

BSL Internalizing Problems (mean (SD))

BSL Externalizing Problems (mean (SD))

BSL Sleep Hours (mean (SD))