Autism spectrum disorder reporting in lower socioeconomic neighborhoods

Study participants

The Autism and Developmental Disabilities Monitoring (ADDM) Network is a multi-state public health surveillance system for ASD and other developmental disabilities established by the CDC in 2000. This system was developed to monitor prevalence trends of ASD and other developmental disabilities in the United States through identification of 8-year-old children with various neurodevelopmental diagnoses. Case ascertainment is record-based, and data were obtained from healthcare and education service providers such as primary care clinics, hospitals, schools, and diagnostic and treatment centers. ASD case status is determined by the number and pattern of autism characteristics described in comprehensive reports, consistent with criteria for autistic disorder, Asperger’s disorder, and pervasive developmental disorder not otherwise specified (PDD-NOS), according to the Diagnostic and Statistical Manual of Mental Disorders 4th ed., text rev. (DSM-IV-TR) (APA, 2000). Although the fifth edition of the DSM was released in 2013 (APA, 2013), the use of DSM-IV-TR criteria in this study reflects the period of case ascertainment that preceded the use of DSM-5. Information on methodology and case ascertainment for this public health surveillance system has been previously published (Rice et al., 2007; Van Naarden et al., 2007).

Aggregate data at the census tract level for 8-year-old children with an ASD classification in surveillance years 2000, 2002, 2004, 2006, and 2008 were acquired from the following ADDM sites: Arizona, Maryland, New Jersey, South Carolina, and Utah. Data obtained were de-identified and aggregated by census tract, yielding a total of 2558 tracts. Census tract residence for ASD cases was based on residence at the time of surveillance (at 8 years of age). Tract-level variables provided for this study included the total number of ASD cases, the number of years each census tract was included in surveillance, percentage of male cases, percentage of cases that were White, Black, and other races (including Asian, Native American, Asian Pacific, and Alaskan Native), percentage of cases that were of Hispanic ethnicity, percentage of cases with an IQ ⩽ 70, and the percentage of cases with mothers having at least a bachelor’s degree (according to available birth certificates).

Information on tract-level area-based population characteristics was obtained from 2000 US Census data using American FactFinder^® (http://factfinder.census.gov). Variables obtained from census data included population of children 0–9 years of age, percentage of the census tract residents of White, Black, or other race in each census tract, percentage of tract residents with Hispanic ethnicity, percentage of the tract population living below poverty, MHI for each census tract, and census tract geographic classification (urban vs rural). To maintain confidentiality, these tract-level data were sent to ADDM sites, where aggregate surveillance data were linked based on tract numbers. Also, the ADDM sites rounded the count of children with ASD to the nearest 5 (e.g. tracts with 12 children were rounded to 10 while those with 24 children were rounded to 25) and MHI to the nearest $5000. ADDM and census data were then de-identified and returned for analysis.

Statistical analyses

The outcome variable was the number of ASD cases in each census tract and analyzed as count data. After examining the distribution of ASD cases by census tract, we analyzed ASD case counts using negative binomial regression. This model was determined to have the best fit according to the Akaike Information Criterion (AIC). Additionally, because the structure of the data is hierarchical, with census tracts (level 1) nested within counties (level 2), which are further nested into states (level 3), we used multi-level modeling (Bryk and Raudenbush, 1992). Due to aggregation of surveillance data for 5 years (2000, 2002, 2004, 2006, and 2008) with no indication of which surveillance year individual cases were ascertained in, we assumed a uniform distribution of cases ascertainment for this analysis. Therefore, we also assumed a uniform age distribution of children in each tract for our population calculation. We calculated population estimates by first dividing the total tract population of children 0–9 years of age, according to the 2000 census, by 10 to account for each age group (i.e. 0, 1, 2, … , 9 years old). This number was then multiplied by the total number of years each tract was included in ADDM surveillance to estimate the number of children ages 0, 2, 4, 6, and 8 years old during the 2000 census, and thus 8 years of age during ADDM surveillance years, as follows: (total population 0–9 years of age/10) × (number of surveillance years).

First, we investigated associations of demographic and SES distributions of ASD cases, including percentage of male cases, percentage of cases of White, Black, or other race, percentage of Hispanic ethnicity, percentage with an IQ ⩽ 70, and percentage of cases with mothers having at least a bachelor’s degree, and estimated tract ASD prevalence. We also examined area-based demographic and SES distributions of populations according to 2000 census data, including percentage of White, Black, or other race residents, percentage with Hispanic ethnicity, percentage of residents with at least a bachelor’s degree, MHI in the highest 25th percentile, percentage of families below the poverty line, and geographic classification (urban vs rural), of each census tract and their potential associations with tract-level ASD prevalence using multi-level negative binomial models. All race and ethnicity variables were scaled to 10% increments (i.e. 1–10%, 11–20%, …, 91–100%) along with education and percentage below poverty variables. Additionally, proportions for all three race categories were modeled as individual variables with neither serving as a reference for the other, as aggregation of data prevented categorization of race for independent subjects.

We examined differences in sources of records (school, healthcare, or both school and healthcare sources) for ASD ascertainment based on demographic and SES characteristics of ADDM surveillance subjects and residential tract populations from which ASD cases were drawn. Proportion of cases ascertained through records from school sources only and proportion of cases ascertained from both school and healthcare records were categorized into quartiles. Because only 46% of the tracts had any cases ascertained through healthcare sources only, this variable was categorized into a dichotomous variable indicating that at least one case was ascertained through healthcare sources only versus no cases were ascertained through healthcare sources only. The first stage examined associations of each predictive variable separately in univariable models with ascertainment outcomes modeled as follows: multi-level ordinal logistic regression to examine the proportion of cases ascertained through school records only, multi-level binary logistic regression to examine the binary outcome of any cases ascertained through healthcare records only, and multi-level ordinal logistic regression to examine the proportion cases ascertained through both school and healthcare records, the most favorable outcome. The second stage examined a combination of ADDM case characteristics as covariates and their associations with ascertainment record sources in multivariable ordinal hierarchical models. Additionally, percent male and percent with IQ ⩽ 70 were included in multivariable analyses of ADDM case characteristics, as these variables have been previously shown to be associated with identification of children with ASD even when disparities are apparent (Mandell et al., 2009). Third, we examined associations of area-based population distributions with ascertainment through school records only, healthcare sources, or both school and healthcare sources. Variables with a p value of 0.20 or less for univariable analyses were included in a final multivariable model (Hosmer et al., 2013). Rural tract classification was included in the multivariable model of area-based measures, as this has been found to be associated with access to health services and subsequent ASD identification (Kalkbrenner et al., 2011). Due to a high correlation between percentage of White and Black race variables, these were not modeled together as predictive variables to avoid multicollinearity. Additionally, percent of tract residents below the poverty line and tract MHI were also not modeled together for the same reason. All analyses were completed using SAS Version 9.3 statistical software (SAS Institute Inc., 2011).

Differences in race

When considering results of our analysis of race distributions of ASD cases in our study, our findings support those from prior literature that reported ASD prevalence is higher in Whites compared to non-Whites (Durkin et al., 2010; Windham et al., 2011; ADDM Network Surveillance Year 2008 Principal Investigators, 2012; ADDM Network Surveillance Year 2010 Principal Investigators, 2014). However, our results also showed prevalence was not lower in tracts with a greater percentage of Black residents, which supports the results of another study using data from the Metro Atlanta Developmental Disabilities Surveillance Program (MADDSP), a surveillance study of a five-county metropolitan area in Georgia, that reported higher prevalence of ASD in Blacks living in these areas (Bhasin and Schendel, 2007). Disparities observed in other studies may be due to the influence of various other factors. One possibility to consider is the potential for decreased ascertainment of ASD in these children, which was also demonstrated in the positive associations seen between Black race of ASD cases and presence of records from schools only, although our results were less than significant. Although in multivariable analysis, we observed no statistically significant association between proportion of Black cases obtained through schools only, in contrast to prior reports that have shown that Black children were more likely to be identified through schools only than through both schools and healthcare sources (Bhasin and Schendel, 2007; Yeargin-Allsopp et al., 2003) or through provision of state-funded services (Windham et al., 2011). However, similar to our observations, Bhasin and Schendel (2007) also reported no significant risk of under-ascertainment through health source only for Blacks. Considering that the federal government requires that school systems provide educational services to children with ASD at no charge in accordance with IDEA (1997), schools serve as a major source of ASD identification for low-income children, many of whom are not recognized until enrollment in school (Palmer et al., 2005).

Regarding the potential influence of area-based racial distribution of tracts from which cases are drawn, another possible reason that lower ASD prevalence may be seen in areas with a greater percentage of Black residents in other studies could be access to care. Our results demonstrated that there was no statistical association between proportion of Black cases or Black residents and risk of under-ascertainment through healthcare sources only, which is in agreement with MADDSP results that also showed no significant association between Black race and healthcare only ascertainment (Bhasin and Schendel, 2007). On the other hand, a study from the National Survey of Children’s Health reported that Black children with ASD had less access to specialists (Liptak et al., 2008), which would reduce the probability of identification through healthcare sources. Other studies have demonstrated that Black children are also often diagnosed later than White children (Fountain et al., 2011; Mandell et al., 2002). Additionally, research has shown that Black children identified through ASD surveillance were more likely to have prior documentation of intellectual disability (ID) than ASD (Mandell et al., 2009). These differences in diagnosis times and classifications have contributed to ethnic differences in perceptions of abnormal behavior and subsequent recognition of ASD symptomatology (Tek and Landa, 2012).

Hispanic ethnicity

Our results indicated that ascertainment through school sources only or through the optimal ascertainment, both school and healthcare sources, decreased with proportion of Hispanic residents in tracts from which cases were drawn. This result coincides with those of a study in Texas school systems that reported that schools with more Hispanic students reported lower ASD prevalence in school systems (Palmer et al., 2005). Furthermore, Mandell et al. (2009) reported that surveillance data of Hispanic children with ASD were more likely to come from areas with access to school records, which has also been shown to be an important source of identification for low-income children (Palmer et al., 2005). In our study, prevalence was noticeably lower in areas with a greater Hispanic population (PR = 0.86). These differences seen in identification of Hispanics with ASD may contribute to the lower prevalence in Hispanics previously reported in other studies (Rosenberg et al., 2009; Windham et al., 2011). Specifically, Windham et al. (2011) reported that prevalence was lowest for Hispanics in the San Francisco Bay area, where Hispanics make up over 20% of the population according to the 2010 census. Additionally, a study in Philadelphia, a study in California, and another study in Wisconsin showed that Hispanics were also diagnosed at later ages than non-Hispanics (Fountain et al., 2011; Mandell et al., 2002; Magana, 2013). Some have suggested that lower rates of ASD seen in Hispanics may be due to language barriers (Zaroff and Uhm, 2012). Contrary to our results, others have shown that Hispanics have less access to healthcare (Liptak et al., 2008). Maternal education, which has been established as an important factor in diagnosis of children with ASD, was also reported to be lower in Hispanics (Mandell et al., 2009; Magana, 2013). These combined factors of language barriers, access to services, and maternal education may have contributed to a recently published report demonstrating that less than half of Hispanic children initially screened for ASD returned for follow-up visits to confirm diagnosis (Windham et al., 2014). Thus, deciphering what factors determine why ASD reporting in Hispanics is lower may require further investigation, and creating programs to target this population may be challenging.

Parental education

In terms of relationships between parental education and ASD diagnosis, our results echo prior reports. Our analysis showed that not only was the proportion of cases with college-educated mothers statistically lower in areas with more cases identified in schools only, but the proportion of college-educated mothers was higher in areas with more children obtained through both school and healthcare sources, the most favorable ascertainment outcome. This suggests that parents with more education may be more able to navigate care systems in order to secure services for their children (Windham et al., 2009). Several other studies have published results indicating that reported ASD prevalence is higher in children with more educated mothers (Liptak et al., 2008; Mandell et al., 2009). Additionally, studies have demonstrated that children with more educated parents received earlier ASD diagnosis (Bickel et al., 2015; Fountain et al., 2011). Bhasin et al. (2007) reported significantly higher maternal college education in children with ASD with no ID, but maternal education was not associated with identification through schools only (Bhasin and Schendel, 2007). On the other hand, another study of children in the Atlanta metropolitan area reported that mothers with less than 12 years of education were more likely to have children identified as having ASD through schools only (Yeargin-Allsopp et al., 2003), which is in accordance with our results. Our study also showed that areas with more educated residents in the population from which cases were drawn had a greater percentage of ASD cases identified through both school and healthcare sources, which also agrees with the results of a study using data from 12 ADDM sites in 2002 and 2004 showing more children with ASD in census blocks with higher-educated residents (Durkin et al., 2010).

Access to care

Our analysis showed that proportion of tract residents living below the poverty line was negatively associated with ASD ascertainment through both school and healthcare records (optimal ascertainment), suggesting that access to care in census tracts may also play a role in ASD prevalence estimates. We attempted to control for access to care by adjusting for urban geographic classification, which has previously been indicated as influencing administrative prevalence (Palmer et al., 2005) and access to care (Kalkbrenner et al., 2011). After this adjustment, neighborhood income, as measured by percent below the poverty line, was no longer related to optimal ascertainment. The positive association we observed between increased tract MHI and ASD prevalence is similar to that reported from MADDSP data indicating higher ASD prevalence in neighborhoods with higher MHI in metropolitan Atlanta and higher proportion of children with ASD ascertained through both school and non-school sources in tracts with higher MHI (Bhasin and Schendel, 2007). Fountain et al. (2010) also reported earlier diagnosis in children from neighborhoods with higher property values. These results support reports that children from low-income families have decreased access to ASD services (Kogan et al., 2008; Liptak et al., 2008). Notably, a study in California showed that areas with higher income also had a greater number of pediatricians and ASD diagnostic clinics (Mazumdar et al., 2013). Our results also indicated that ascertainment through healthcare records only was higher in areas with a greater percentage of tract residents living below poverty. This could, however, be attributed to provision of Medicaid in this population. The literature has shown that Medicaid is commonly used to fund services for children with ASD (Semansky et al., 2013; Shaw et al., 2013). A recent study showed that poverty-eligible Medicaid patients were diagnosed earlier than disability-eligible patients; however, the same study showed earlier diagnosis for children from counties with a greater MHI (Daniels and Mandell, 2013). Additionally, Mandell et al. (2002) also demonstrated that Black and Hispanic children receiving services through Medicaid were diagnosed later than White children. Therefore, factors associated with access to care may be more complex than a simple relationship with income. Data on private insurance of children in our study were not provided, but previous research has shown that a lack of health insurance is associated with reduced access to care (Kogan et al., 2008). However, deciphering how income directly or indirectly affects access to services and healthcare may be difficult.

Limitations

These study results are subject to several limitations. First, in order to prevent identification of cases from the ADDM study, aggregate data were provided for a total of five surveillance years (2000, 2002, 2004, 2006, and 2008). Because this is aggregate data, no assumptions can be made for characteristics of individual children. Additionally, population estimates used to account for prevalence were based on 2000 census data under the assumption of equal age distributions, and not the exact population of 8-year-old children for each surveillance year, as details on which years surveillance was conducted in census tracts were not provided. Thus, there may be a potential for measurement error. Data for other races, including Asian, Asian Pacific, Native American, Native Alaskan, and any race identified as other than Black or White was condensed into one group due to lower frequency of data; however, we recognize that language barriers and cultural differences may influence ASD ascertainment between these racial groups, and that further investigation of these groups is needed. Furthermore, data used for surveillance in the ADDM study were drawn from clinical and education sources; thus some information of symptomology of children may have been missing and could have led to misclassification and missing information on manifestation of symptoms that may have been useful in differentiation of racial and ethnic variances. We also acknowledge that education level of mothers of cases was missing for approximately 25% of ASD cases. Moreover, we had no information on availability of private insurance or provision of Medicaid, which may have influenced cost and provision of services for children and subsequently confounded the results. However, future analysis at the city and county level could utilize Area Resource File data as proxies for this information.