Service Use Classes Among School-aged Children From the Autism Treatment Network Registry

We analyzed secondary cross-sectional data from the ATN registry for school-aged children with ASD at their first follow-up visit. Data from 2008 to 2015 across 18 sites, including 2 Canadian and 16 American sites, were combined. The institutional review board at each ATN registry site approved this research.

The ATN registry is a multisite database that includes medical, behavioral, and quality of life (QoL) data on children with ASD collected by clinical assessment and parent report at the initial visit and thereafter during 3 routine visits (ie, first, second, and third follow-up visits). The ATN registry is briefly summarized here and has been previously described in greater depth.^27,28  The ATN sites represent a network of academic-affiliated autism centers of excellence in the United States and Canada committed to a comprehensive evaluation and care program for children with ASD. The ATN registry includes data on a subset of children with ASD who received care at a network clinical center. Individual sites enrolled a convenience sample of children on the basis of inclusion and exclusion criteria.^27,28  Site patients were eligible to enroll if they were aged <17.6 years and met diagnostic cutoffs for ASD as determined by the Autism Diagnostic Observation Schedule (ADOS) and the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition or the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition.^29,30  Exclusion criteria included a medical condition that precluded valid diagnostic testing (eg, blindness), and parents had to be fluent in English or Spanish.

This study included 1378 children aged ≥6 years with ASD and service use data (ie, no or some service use) from the first follow-up visit. We limited the sample to school-aged children with ASD so we could examine psychotropic medication use. We also chose data primarily from the first follow-up visit (9–15 months after baseline) versus enrollment (0–9 months after consent) so that service use might be better established. The median time between ATN registry enrollment and the first follow-up for the sample was 1 year (interquartile range: 10–14 months).

Thirteen service use indicators that were based on first follow-up visit data and that were applicable to the entire sample were examined (Supplemental Table 5). Educational and behavioral services included parents’ reported use of developmental therapy services (eg, speech and language therapy), behavioral interventions (eg, applied behavioral analysis), education support services (eg, learning center), and counseling services (eg, family therapy) for their child with ASD. Psychotropic medications were parent reported to ATN clinicians and included stimulants, selective serotonin reuptake inhibitors (SSRIs), antipsychotic medications, and α-agonists currently used by the child with ASD. CHAs included parents’ reported use of special diets (eg, gluten-free diet), melatonin, natural products (eg, vitamins), mind and body practices (eg, chiropractics), or other biomedical CHAs (eg, hyperbaric oxygen therapy) and off-label medication use (eg, antifungal medication). Melatonin could also be reported by parents to an ATN provider. Parents could additionally report “other” services used for the child with ASD. Three authors (O.J.L., S.E.L., and K.A.K.) reviewed the “other” services reported and categorized each as 1 of the 13 services examined.

We assessed the following variables: child’s age at first follow-up, sex, race (Asian American, black or African, white, and other race or multiracial), Hispanic ethnicity, insurance (public only, private only, or both public and private), and parent education level (high school education or less versus some college or more). Except for age, all sociodemographic data were from baseline. When available, family income above or below US $50 000 was determined from parent report. For 34% of the sample, US zip code (when available in the ATN registry) was used to estimate annual household income with median income as reported by the Population Studies Center Institute for Social Research at the University of Michigan.³¹ 

Developmental characteristics were examined by using intake or baseline data on ASD diagnosis (autism, Asperger syndrome, or pervasive developmental disorder not otherwise specified [PDD-NOS]) determined at intake from the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition²⁹  or ASD from the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition.³⁰  ASD severity was determined at intake by the ADOS clinical severity score, which ranged from 1 to 10; higher scores indicated more autism symptomatology.³²  Other developmental characteristics included attention-deficit/hyperactivity disorder (ADHD) determined from the first follow-up and IQ at baseline. Depending on age and testing availability at sites, IQ was primarily assessed by using the Stanford-Binet Intelligence Scale, Fifth Edition (full scale or the abbreviated IQ standard score).³³  Behavioral characteristics were determined at baseline by using Child Behavior Checklist (CBCL) externalizing and internalizing scores, with scores ≥70 indicating risk for behavior problems.³⁴  Health characteristics were assessed by using baseline or first follow-up data on the following measures: gastrointestinal (GI) problems (ie, any problems with vomiting, reflux, diarrhea, constipation, or stomach pain) reported at the first follow-up, child sleep problems measured at baseline by the Children's Sleep Habits Questionnaire (CSHQ) (with total scores >41 indicating clinically significant sleep problems),³⁵  and pediatric QoL measured at baseline by the Pediatric Quality of Life Inventory (PedsQL) total score (with higher scores indicating better health-related QoL).³⁶ 

By using the 13 service use indicators, latent class analysis (LCA) was performed to identify groups of children with ASD who had similar service use patterns. In LCA, item-response patterns are used on categorical variables reflecting an underlying construct (latent variable) to group similar individuals into mutually exclusive and exhaustive classes.^37,38  That is, individuals are assigned to one class for which they have the highest probability of membership. We fit the model for 1 to 8 classes and chose the final number of classes on the basis of model-fit statistics, including Akaike’s information criterion, consistent Akaike’s information criterion,³⁹  the Bayesian information criterion, and sample-adjusted values Bayesian information criterion.⁴⁰  We also compared item-response probabilities for the 3, 4, and 5 class models to determine if clinically meaningful groups were identified from each model, in addition to computing the bootstrapped parametric likelihood ratio test to compare these models. A maximum likelihood procedure with adjustment for missing-at-random missingness was used in the LCA.³⁸ 

Classes defined by the LCA were then compared by the sociodemographic, developmental, behavioral, and health characteristics. Continuous variables were compared across classes by using the Kruskal-Wallis test, and categorical variables were compared across classes by using the Fisher’s exact test extension.⁴¹  We then computed pairwise comparisons only for variables that had a statistically significant difference across all classes, similar to Fisher’s least significant difference approach to pairwise comparisons between groups after a significant overall analysis of variance.⁴²  Three multivariable models were then fit with pairs of defined classes as the dependent variable. The sex of the child was included in the multivariable models because it was considered clinically meaningful, along with other covariates that were statistically significant in bivariate comparisons across classes. These were logistic regression models with random effects for each site by using an unstructured covariance matrix. Given the small number of children from the 2 Canadian sites (n = 118), we did not directly compare children by country of origin.

All tests were 2 sided, and P < .05 was considered statistically significant. All analyses were performed in SAS version 9.4 (SAS Institute, Inc, Cary, NC) or R version 3.5. Procedures for LCA were used in SAS (SAS Institute, Inc),^43,44  and the R ggplot2 package⁴⁵  was used to create figures.

The mean child age at the first follow-up was 10.7 years (SD = 2.8). A plurality of the sample was male, non-Hispanic ethnicity, white race, privately insured or privately and publicly insured, and in a family with a parent with some college or more (Table 1). Of the sample, 61.6% lived in families whose annual income was ≥$50 000, 61.5% had an autism diagnosis, 20.3% had ADHD, and 32.7% had GI problems. Additional developmental, behavioral, and health characteristics are shown in Table 1.

An average of 2.5 (SD = 1.7) of the 13 services were used, and 9.9% of children did not report any service use. As shown in Supplemental Table 5, developmental therapy was the most commonly used service (77.4%). For psychotropic medication, stimulants were the most commonly used (31.9%). For CHAs, natural products were the most frequently used (21.5%).

LCA results revealed that a 4-class model best fit the data (Supplemental Fig 2). The 4-class solution included a class that had minimal service use labeled as limited services (LS) use (12.0%), a class likely to use multiple services across the 3 main service types labeled as multimodal services (MS) use (36.4%), a class that predominantly used educational and/or behavioral interventions labeled as educational and/or behavioral services (EBS) use (42.6%), and a class that primarily used special diets and/or natural products (SDNP) labeled as such (9.1%). The percentage of children who used each of the 13 services across these 4 service use groups is displayed in Fig 1.

Further examination of each individual service use group revealed that children in the LS group used the least number of services on average (mean = 1.0 ± 1.0), with stimulants being the most frequently used (29.9%) service among them. Children in the MS group used an average of 3.6 (SD = 1.5) services, with developmental therapy (94.7%) being the most commonly used service among them. All other services, except for special diets, other biomedical CHAs or off-label medication, and mind and body practices, were used by >20% of children. Children in the EBS group used an average of 1.6 (SD = 0.9) services, with developmental therapy (91.8%) being the most commonly used service among them, followed by educational support services (48.3%), stimulants (23.0%), and behavioral interventions (19.2%). All other services for the EBS group were used by <10% of children. Children in the SDNP group used an average of 4.1 (SD = 1.6) services, with special diets (83.9%) and natural products (81.6%) being the most commonly used among them, followed by developmental therapy (73.5%), behavioral interventions (43.6%), and educational support services (42.7%). All other services for the SDNP group were used by <20% of children.

Bivariate results revealed statistically significant variation across the service use groups on age, insurance, income, ASD diagnosis, ASD severity, ADHD status, IQ, externalizing behavior problems, internalizing behavior problems, sleep problems, GI problems, and QoL (Table 2). Further pairwise comparison results revealed that age, insurance, income, ASD diagnosis, ASD severity, ADHD status, IQ, internalizing behavior problems, sleep problems, and GI problems had statistically significant associations with the LS group compared with one or more of the other service use groups. For the MS group relative to the EBS and/or SDNP group(s), age, income, ADHD status, externalizing problem behaviors, internalizing problem behaviors, sleep problems, GI problems, and QoL had statistically significant associations. Insurance, income, GI problems, and QoL each had a statistically significant association with being in the EBS group versus the SDNP group. All pairwise comparison results are detailed in Table 3.

Multivariable mixed-effect logistic regression model results revealed that greater ASD severity (P = .05) and more externalizing behavior problems (P = .018) were each significantly associated with being in the MS group relative to the EBS group (Table 4). In addition, older child age (P < .001) and lower ASD severity (P = .012) were each significantly associated with being in the LS group relative to the EBS group. Higher household income level (P = .007) and lower QoL (P < .001) were each associated with the SDNP group compared with the EBS group per multivariable analysis results.

In this study, we used LCA to identify 4 service use groups among school-aged children with ASD from a large multisite registry. In addition, this study’s findings provide new insights into how these service groups vary by sociodemographic, developmental, behavioral, and health characteristics. Together, these findings contribute new knowledge of observed service use patterns that may better inform family-centered shared decision-making for school-aged children with ASD. This study’s findings also provide a greater basis for future research in which service use is longitudinally examined by geographic area and in relationship to health outcomes among school-aged children with ASD.

The 4 distinct service use groups determined among school-aged children with ASD in this study were (1) LS, with the use of stimulants being the most likely (but still only used by <30% of children); (2) MS, with children using a broad range of the 13 services, especially developmental therapy, educational support services, stimulants, SSRIs, and melatonin; (3) EBS, with common use of developmental therapy, behavioral interventions, and/or education support services but infrequent CHA or psychotropic medication use; and (4) SDNP, which included the most likely use of special diets and natural products. The largest number of children with ASD were placed in the EBS class, followed by the MS, LS, and SDNP classes.

Regarding the EBS group, there has been little research examining if EBS use is associated with less psychotropic medication and/or less use of CHA among school-aged children with ASD. Past research has, however, revealed that behavioral intervention use is likely to occur in combination with use of other educational support and therapy services among school-aged children with ASD,²²  which is echoed to some extent by this study’s finding that children in the EBS class had similarly high probabilities of using developmental therapy services or educational support services. This study’s results also revealed that children in the EBS group were younger, were less likely to have ADHD, had less internalizing behavior problems, were less likely to have GI problems, and had higher QoL compared with children in other groups, suggesting that children in the EBS class may have less complex service needs.

Because both higher ASD severity and higher externalizing problem behavior scores were jointly predictive in the multivariable model of MS group membership compared with EBS group membership, the MS group may have more complex service needs. Additionally, children in the MS group had the highest probability of using some services with the greatest known risks (eg, antipsychotic medication).¹⁰  It, therefore, could be helpful for clinicians working with families of children in the MS group to provide extra support around treatment decision-making. More concretely, this might involve clinicians actively employing the safety and efficacy model^46,47  when providing additional counseling and decision-making support regarding all services being considered or used for a child’s ASD. Training on patient-clinician communication about health risks that has been shown to improve patient centeredness in medical practices^48–50  could be adapted for this more specific purpose in pediatric and family medicine practices as well as in autism clinics. Innovative approaches to integrated care delivery models, such as family navigation⁵¹  or more comprehensive care coordination across the larger system of care (eg, education and health care),⁵²  could also be used to better support families in their ASD service use.

Previous research reveals that ∼15% of school-aged children with ASD are reported to not use any psychotropic medication or educational and behavioral services.¹⁷  This study’s findings revealed that >20% of school-aged children with ASD were in either the LS class or the SDNP class. In each group, there was some probability of using certain psychotropic medications (eg, stimulant use in the LS group) or other services. Children in the LS group were more likely to be older and have less severe ASD than those in the EBS group, suggesting that older children with ASD and those with less severe symptomatology may use fewer services generally. Although we did not examine age of ASD diagnosis or diagnostic delays per se, some research suggests that older age at ASD diagnosis and longer diagnostic delays may contribute to less use of behavioral interventions or school-based therapy.⁵³  Children in the SDNP group were more likely to have higher household income and lower QoL compared with those in the EBS group, suggesting that better-resourced families of children with ASD who may experience poorer overall health status are more inclined to use CHA. Because most special diets and natural products are not covered by insurance plans,⁵⁴  having greater financial means may, indeed, contribute to greater use by children in the SDNP class. Given limited evidence of efficacy and safety for use of special diets and natural products in children with ASD,¹⁰  it may additionally be helpful for clinicians to bolster their communication and efforts to engage in shared treatment decision-making with this particular group of children with ASD and their families.

This study has important limitations. First, generalizability is limited by the nature of the clinical sample drawn from the ATN registry insofar as children had an ASD diagnosis, were seen at an ATN site that was a large academic medical center in a metropolitan area, had one parent proficient in English or Spanish, had insurance, and did not have certain other conditions (eg, blindness). Especially vulnerable ASD subgroups (eg, those living in rural areas and/or without a formal ASD diagnosis) may, therefore, not be well represented in this study. This may bias findings toward children with ASD who tend to be higher services users. Nevertheless, a strength of the ATN registry is that ASD diagnosis was clinically confirmed. The ATN registry additionally captures data on many services used for ASD symptom management. Other large-scale data sources that include children with ASD do not provide this depth of information on ASD-relevant service use. Still, it is possible that not all services used were reported and that some services reported may have been used for conditions other than ASD.

Although the ATN registry provides extensive data on the health and developmental characteristics of children, it does not include data on certain factors, such as aspects of health care quality⁵⁵  (eg, shared treatment decision-making) and support networks,⁵⁶  that may affect service use patterns. In light of this study’s findings showing limited statistically significant associations of sociodemographic, developmental, behavioral, and health characteristics with service use patterns, future research is needed to determine service use groups by using databases with richer information on health care quality and/or qualitative study to explore what additional factors may contribute to the 4 groups identified in this study. In addition, further study on how service use patterns differ by geography (particularly between the United States and Canada) and health systems (eg, school- versus community-based services) is needed. Given limited sample sizes for many ATN sites, we were unable to directly compare sites on service use, despite the importance of understanding geographic variation in autism service use. Lastly, we cannot infer any temporal precedence regarding the relationships examined. Given substantial loss to follow-up after the first follow-up visit, we were unable to examine change between the first, second, and third follow-up visits in service use patterns. Future researchers may, therefore, employ latent transition analysis using other databases to examine how children may move between service use groups over time and latent class growth analysis to determine how the service use groups may longitudinally contribute to health outcomes.

Four service use groups among school-aged children with ASD were identified in this study. Service use groups differed by certain sociodemographic, developmental, behavioral, and health characteristics. Study findings contribute new knowledge on ASD service use and may be used to inform both clinical practice and further study regarding ASD service use.

We thank Frances Lu for her initial assistance with the data analysis for this project.

ADHD

attention-deficit/hyperactivity disorder

ADOS

Autism Diagnostic Observation Schedule

ASD

autism spectrum disorder

ATN

Autism Speaks Autism Treatment Network

CBCL

Child Behavior Checklist

CHA

complementary health approach

CSHQ

Children's Sleep Habits Questionnaire

EBS

educational and/or behavioral services

GI

gastrointestinal

LCA

latent class analysis

LS

limited services

MS

multimodal services

PDD-NOS

pervasive developmental disorder not otherwise specified

PedsQL

Pediatric Quality of Life Inventory

QoL

quality of life

SDNP

special diets and/or natural products

SSRI

selective serotonin reuptake inhibitor