An Informatics Analysis to Identify Sex Disparities and Healthcare Needs for Autism across the United States

Abstract

Autism is among the most common neurodevelopmental conditions. Timely diagnosis and access to therapeutic resources are essential for positive prognoses, yet long queues and unevenly dispersed resources leave many untreated. Without granular estimates of autism prevalence by geographic area, it is difficult to identify unmet needs and mechanisms to address them. Mining a dataset of 53M children using meaningful geographic regions, we computed autism prevalence across the country. We then performed comparative analysis against 50,000 resources to identify the type and extent of gaps in access to autism services. We find a steady increase in autism diagnoses from K-5, supporting delayed diagnosis of autism, and consistent under-diagnosis of females. We find a significant inverse relationship between prevalence and availability of resources (p < 0.001). While more work is needed to characterize additional trends including racial and ethnicity-based disparities, the identification of resource gaps can direct and prioritize new innovations.

Introduction

Autism spectrum disorder (ASD) is one of the most common developmental pediatric conditions, and is characterized by difficulties in social interaction and communication and restricted or repetitive interests and behaviors. For children with autism, early diagnosis and access to treatment results in improved outcomes¹–³. However, the median age of diagnosis in the U.S. is 51 months⁴ with wait times for diagnostic services or ongoing care visits ranging from 3-14 months⁵. Over the past ten years, the prevalence of autism has grown from 1 in 110 to 1 in 54⁴. This growth is likely to further stretch an already overburdened system of autism support services.

Current estimates of autism prevalence are produced by the Autism and Developmental Disabilities Monitoring (ADDM) Network which closely tracks the prevalence of autism among 8 year-old-children across 11 sentinel sites,⁴. Autism prevalence varies widely from site to site from 1.31% (Colorado) to 3.14% (New Jersey), suggesting that autism prevalence is also likely to vary widely across the country. However, estimates of autism prevalence are only useful if they successfully guide the investment of healthcare resources. While ADDM does provide an estimate of autism prevalence for a few states, the actual availability and type of autism healthcare resources relative to need for a given geographic area is understudied. To estimate the relative disparities between autism prevalence and autism healthcare resources between different geographic areas and demographic groups we combined two national datasets; the Civil Rights Data Collection (CRDC) and GapMap.

The Civil Rights Data Collection is collected bi-annually by the U.S. Department of Education from every school in the and encompasses 53 million students, including the total student enrollment and the number of students in each school receiving special education services for autism under the Individuals with Disabilities Education Act (IDEA). Incorporating data from educational sources into autism prevalence estimates has been found to reduce racial/ethnic differences in prevalence rates⁶ and to mitigate healthcare access disparities⁷. The IDEA criteria used by schools to identify autism is quite effective, with over 90% of autistic students identified by schools also meeting the criteria for an autism diagnosis via the Autism Diagnostic Observation Schedule (ADOS), a diagnostic standard⁸. For sites where educational records were available, ADDM found that between 85.1% and 93.6% of children with autism are receiving special education services⁴. This makes the number of students receiving autism services in school an excellent proxy for the total number of autistic children in the school. The national comprehensiveness of the CRDC and its geographic detail make it an excellent data set for estimating the relative autism prevalence between geographic areas.

In order to understand how diagnostic and therapeutic resources for autism vary from state to state, we compare autism prevalence rates estimated from CRDC to the autism resources available in GapMap. GapMap is an autism resource database and interactive online map containing 51,000 geolocated autism resources. GapMap was designed to simultaneously track autism epidemiology and help families find autism services, and has been used to study ASD resource availability in the U.S., using national-scale prevalence estimates in order to identify geographic areas with insufficient resources⁹. The average autistic individual is 20 miles (32 kilometers) away from the nearest autism diagnostic center⁹ and 83.86% of all US counties contain no autism diagnostic resources at all¹⁰. By using estimates of autism prevalence drawn from schools, rather than national averages, we can more accurately analyze and identify geographical gaps in autism services.

Using CRDC data, we estimate the prevalence of male and female autism across the U.S. We show that autism prevalence varies by grade, and that a larger percentage of male autism is diagnosed during elementary school ages than female autism. We replicate previous findings that the number of children receiving special education services are highly correlated across IDEA categories. Finally, using the GapMap autism resource database, we show that the availability of autism resources per autistic student varies dramatically by state, with states with higher autism prevalence providing fewer resources per child. Our work demonstrates that CRDC school-based data can help understand the geographical distribution of autism as well as ASD resource gaps in the U.S.

Methods

Estimating autism prevalence

We use the CRDC to estimate autism prevalence at the sub-state level by aggregating school-level data into larger geographic units of Core-Based Statistical Areas (CBSA) as defined by U.S. Census Bureau. A CBSA is a geographic area chosen to delineate metropolitan or micropolitan boundaries associated commute regions, providing a natural way to group student enrollment across the US. We estimate overall prevalence rates using median prevalence across all CBSAs and empirically estimate 95% confidence intervals for prevalence across all CBSAs.

IDEA diagnosis correlated structure by multivariate ordinary least squares

We estimate the sex-specific correlation structure between autism and other IDEA categories using multivariate ordinary least squares; for each CBSA the total number of autistic students was regressed against total student enrollment and the totals for the other twelve IDEA categories of deaf-blindness, developmental delay, emotional disturbance, hearing impairment, intellectual disability, multiple disabilities, orthopedic impairment, specific learning disability, speech or language impairment, traumatic brain injury, visual impairment, and other health impairment. However, students can only be assigned one of these IDEA categories; this restriction has implications for correlation structure as discussed in Results.

Estimating per-grade autism prevalence

The CRDC dataset provides the number of autistic students per school, but not per grade. Grade-level estimates of prevalence are useful in order to estimate age at first diagnosis for the population. Using per-grade total enrollment estimates from CCD, we estimate autism prevalence per grade for each sex using regularized regression. Let X^e be a matrix where each entry $x_{ij}^e$ is the number of male students from race/ethnicity category e in grade j at school i. Our goal is to estimate β^e, a vector where entry ${\beta }_j^e$ is the prevalence of autism in males from race/ethnicity category e for grade j. We will do this using y^e, a vector where each entry $y_i^e$ is the number of autistic male students from race/ethnicity category e at school i. We model y^e as the sum of binomials with different probabilities, since each grade will have a different autism prevalence. Le Cam’s theorem states that the sum of binomials with different probabilities approximately follows a poisson distribution with parameter λ = X^eβ^e in our case. We use maximum likelihood estimation to estimate β^e

$$\begin{array}{l}\underset{\beta ,{\beta }^e\forall e\in E}{\max \text{imize}}{\displaystyle \sum _{e\in E}{y}^e}\log (X^e{\beta }^e)-1X^e{\beta }^e-\lambda \Vert {\beta }^e-\beta \Vert {}_2-\lambda \Vert {\beta }_{1:}-{\beta }_{:-1}\Vert {}_2\\ \text{subject}\hspace{0.17em}\text{to}\hspace{0.17em}0\hspace{0.17em}\text{<}\hspace{0.17em}=\hspace{0.17em}{\beta }^e\hspace{0.17em}<\hspace{0.17em}=\hspace{0.17em}1\hspace{0.17em}\forall e\in E\end{array}$$

The first two terms in the sum represent the log likelihood of our model, where 1 represents a vector of ones. The third term is a regularizer that encourages the autism prevalence estimates for each race/ethnicity category to be similar. The fourth term is another regularizer that encourages autism prevalence between consecutive grade levels to be similar. The regularization parameter λ was set to 0.1n where n is the number of schools included in the analysis. Finally, we constrain our autism prevalence estimates to be between 0 and 1. We fit our model in python using cvxpy¹¹ for male and female autism prevalence separately.

We note that our regularization scheme encourages race/ethnicity categories to have similar prevalence estimates. This makes it difficult to interpret prevalence differences across race/ethnicity categories.

Next, we use our grade-level autism prevalence estimates to estimate the age that autistic children begin receiving services in school. We do this by assuming that the increase in our autism prevalence estimates between grades preK-5th are primarily due to the delayed diagnosis of autism throughout these age ranges. This is supported by findings from ADMM that the median age of autism diagnosis is 51 months⁴. To estimate when autistic children begin receiving services in school, we calculate the difference in prevalence between neighboring grade levels, and divide this difference by the autism prevalence estimate for 5th grade (the grade with peak autism prevalence). This provides an estimate of the fraction of autistic children that begin receiving services in any given grade between preK-5th.

Generating K-12 ASD prevalence and resources maps and histograms

The state-level estimates of autism resources per thousand autistic student in Figure 6 was generated from GapMap and the CRDC. The data includes 49 states (excludes Iowa), Washington D.C. (denoted DC), and Puerto Rico (denoted PR), giving a total of 51 US state/territories.

Figure 6:

Autism resources exhibit a negative correlation with autism prevalence. On the left, we show a map of ASD resources per 1000 ASD students. On the right we show that higher rates of ASD prevalence correlate significantly to a lower number of autism resources per 1000 students (p < 0.001).

Because states with higher populations contain more resources than less populous states, and diagnosis rates differ from place to place, it is necessary to normalize for ASD population to determine the relative abundance of resources per state. Fig. 4, Resources per 1000 ASD Diagnosed Per State, was produced by dividing the total number of resources in each state by the total number of ASD diagnosed students in the CRDC dataset, then multiplying by 1000. To determine the prevalence of ASD K-12 students per state (Fig. 4, Prevalence of ASD K-12 Students per State), the total number of diagnosed students per state was divided by the total number of K-12 students enrolled in that state.

Results

Autism prevalence

Using CBSAs to aggregate school-level data provides consistent autism prevalence estimates across the U.S. as shown in Figure 2 yielding a prevalence estimate of autism among males to be 1.62% [0.86%, 2.58%], which is slightly lower than the current ADMM estimate of 2.94%⁴. We estimate the female prevalence of autism to be 0.33% [0.14%, 0.59%], which is also lower than the current ADMM estimate of 0.69%. Finally, we estimate the male:female sex ratio for autism to be 5.3:1 [3.4:1, 9.6:1] which is in range of the current ADDM estimate of 4.3:1.

Figure 2:

Autism prevalence is consistent across the Census Bureau’s core-based statistical areas (CBSAs). Even though CBSAs vary dramatically in total student enrollment, male and female autism prevalence and the autism male-female ratio remain consistent.

Our estimates of autism prevalence are lower than the most recent estimates from ADDM. It is possible that the 11 sentinel sites used by ADDM are not representative of the U.S. as a whole. However, it is more likely that children with autism are under-counted in the special education totals. Schools are required to choose a single IDEA category for each student, so autistic students that also qualify for other services may be included in a different IDEA category. In fact, ADDM found that for sentinel sites where special education information was fully available, an average of 32% of autistic students were reported under a different IDEA category.

Relationships between autism and other IDEA categories

To support our hypothesis that the number of autistic students are under-reported in the CDRC due to autistic students being reported in other IDEA categories, we examine the correlation between ASD and other IDEA categories in Figure 3. Three IDEA categories exhibit significant negative correlation with ASD counts: developmental delay, speech or language impairment, and multiple disabilities. This finding is supported by ADMM which found that among sites with access to school records for all students, children with autism are often included in other IDEA categories including developmental delay (0.7-23.2%), speech or language impairment (3.4-6.2%) and multiple disabilities (0.0-6.2%). Notably there is no significant correlation between the number of autistic students and the number of students with deaf-blindness, traumatic brain injury, hearing impairment, or visual impairment: this agrees with prior knowledge of the etiological distinction between these disabilities and autism.

Figure 3:

There are significant correlations between the number of ASD students in a CBSA and the number of students in other IDEA categories. Negative coefficients suggest the possibility that some students with ASD are being included in the other IDEA categories.

Autism prevalence by grade

Using grade-level enrollment, we are able to estimate grade-level autism prevalence as shown in figure 4. Autism prevalence increases from preschool through 5th grade as students are identified as qualifying for autism services. Interestingly, autism prevalence declines slightly from 5th-11th grade. There are several explanations for this decline. Changing diagnostic criteria for ASD between DSM-IV and DSM-V may have affected primary school students differently than secondary school students. Another possibility is that for some students, school-based autism services are only necessary in the younger grades.

Figure 4:

Autism prevalence varies by grade. Autism prevalence increases from preschool through 5th grade, then declines slowly. This pattern is consistent for both male and female students, and across every race/ethnicity category. The male-female ratio climbs from preschool through 11th grade ranging from 4x to 6x. Dotted black lines show most recent ADMM estimates. We exclude 12th grade due to data artifacts.

Using estimates for autism prevalence in each grade, and assuming that autism prevalence is constant across each grade level, we estimate the fraction of autistic students that begin receiving autism services in each grade, as shown in Figure 5. Approximately 60% of female autistic students are identified as requiring autism services before starting kindergarten as compared to 48% of male autistic students. This is supported by ADMM findings that female children with autism are more likely to be evaluated before 3 years old than male children⁴. By first grade, 97% of female autistic students will be identified, as compared to only 87% of male autistic students.

Figure 5:

Female autistic students begin receiving services at younger ages than male autistic students. Female children are more likely to be identified as needing autism services before school begins than male students.

There are several explanations for this finding. Female autism may be more severe than male autism, resulting in younger age of diagnosis. Alternatively, schools may be more effective at identifying male autism than female autism, resulting in missed autism diagnoses for females during elementary-school.

Comparing autism prevalence to resources across the United States using GapMap

Next, we compared autism prevalence to resources using GapMap, as shown in figure 6. There is a significant inverse relationship (p < 0.001) between autism prevalence among K-12 students and available autism healthcare resources available as indexed by GapMap; a percentage point increase in autism prevalence translates to approximately 90 fewer resources per thousand autistic students. The notable outliers are Montana, North Dakota, and South Dakota, with autism prevalence of 0.589%, 1.003%, 0.903% and resources per thousand autistic students of 260, 313, 243, respectively. In contrast, New York state has similar autism prevalence of 0.685% but roughly half (163) the autism healthcare resources per autistic student.

It is interesting to note that New York has similar characteristics to states in the Northeast or West – California, Massachusetts, or Washington – since their overall population is dominated by high density population areas. Thus, it would make sense if they had similar ASD provider infrastructures and followed similar data trends of high ASD prevalence but low number of resources available. However, this was not the case. In fact, New York was the opposite because it has low ASD prevalence (0.685%) and a relatively high number of resources available (163.1 resources per 1000 diagnosed). In future research, it would be imperative to explore why New York differs from similarly populated states. This would enable us to better understand how states with varying population density can effectively provide accessible ASD resources across their state.

Classification of resources into service categories

To analyze resource frequency by state in context, we need to cross reference the frequency of types of resources (7) with prevalence data (6) and population growth trends¹².

Resource frequency shortcomings in Western states are likely due to recent population growth, which entailed an increase in total ASD K-12 diagnosed cases. Western states have undergone significant population growth from 2008-2018. All states have grown significantly more than the 50-state median of 0.63% growth (California 0.78%, Oregon 1.07%, Arizona 1.34%, Nevada 1.35%, Idaho 1.35%, Utah 1.73%)¹². With this quick population growth, many of these states are now under-equipped in resources. Nevada has a 1.31% diagnosed over total enrolled which is above the 51 state/territory average of 1.11%. It also has the lowest resources per 1000 diagnosed across ABA, therapy, and diagnostic resources among all US states (7). Thus, with a rapidly growing population (2.09% from 2017-2018, highest in the US) and high childhood ASD prevalence, Nevada has shortcomings in autism resources. It is recommended that more resources of all categories be opened across the state.

Similarly, Utah’s resource shortcomings and low prevalence rate are due to the recent population spike. Utah has the 4th lowest resources per 1000 for ABA and diagnostic resources and the 7th lowest resources per 1000 for therapy resources among the 51 US state/territories (7). While Utah’s childhood ASD prevalence is low at 8.95%, it had the largest growing population in the US from 2008-2018 (¹²). Utah’s childhood ASD undiagnosed cases in the coming years has likely grown from the population spike. Therefore, it is recommended that Utah opens more resources, especially diagnostic resources, across the state. California and Arizona also consistently fall in the lower bins across all categories of resources (7). Despite lower population growth from 2008 to 2018 than NV and UT, these states contain a wide variety of population density by geographical location. California and Arizona should devote attention to increasing total resources available to better serve their growing population.

Montana (74.37), New York (59.19), Louisiana (29.08), and Colorado (32.22) all have high diagnostic resource frequency per 1000 diagnosed, and low K-12 ASD prevalence rates (MT 0.589%, NY 0.685%, LA 0.721%, CO 0.733%).

North Dakota and South Dakota are outliers above upper thresholds for resources per 1000 diagnosed across all categories (7). Even so, both states have mild prevalence rates of ASD K-12 cases (1.00% and 0.903% for ND and SD respectively) that are slightly lower than the nationwide average of 1.11%. Both states have low population density, so it may be necessary for these states to have more resources per 1000 autistic students in order for resources to be geographically accessible to every student.

Discussion

The CRDC is a valuable resource for estimating autism prevalence in the U.S. It encompasses 53M students and provides a fine-grained geographical distribution of students receiving autism services. However, autism prevalence rates estimated from this data are lower than other studies. Our estimates are likely underestimating the prevalence of autism in the U.S. This discrepancy is due in part to the CRDC requirement that students be included in only one IDEA category. Our analysis shows that IDEA categories are significantly correlated across CBSAs, suggesting that some students with autism are recorded in other IDEA categories including developmental delay, emotional disturbance, intellectual disability, orthopedic impairment, or specific learning disability. Another explanation for lower prevalence rates in school data is parents choosing not to report their child’s autism diagnosis to the school. For ADMM sites where educational records were available for all students, 6.4-14.9% of autistic students did not have special education records. More work is needed to understand the concordance of medical records and school records for autism. Finally, autism may be under-diagnosed in some regions of the U.S., which could result in lower overall prevalence estimates.

Despite this limitation, we were able to estimate changes in autism prevalence by grade level and use these estimates to understand the age at which autistic children begin receiving services. Male autistic students are identified throughout the primary grades, while female autistic students are most often identified before entering school. This suggests that female autism may be easier to identify at younger ages as compared to male autism, or alternatively that female autism is under-diagnosed in grades K-5.

Since states are so large and include a diverse array of urban/suburban/rural areas, in the future, we plan a more granular analysis at the CBSA, county, or school level. Doing so would enable us to recognize gaps in resources within states. Furthermore, by using GapMap’s resource data of precise location (longitude and latitude), we could calculate the average driving time to resources from the closest educational provider. Availability is likely constrained by excessive wait times and/or difficulty in reaching these resources (driving 20 miles in a rural area would take far less time than driving 20 miles in Los Angeles or New York). Along with other factors like traffic congestion and public transit availability, it may be possible to define an accessibility index that would give us a better grasp on where resources should be opened such that they can consistently serve families before or after school.

Another source of further analysis is GapMap’s future capability of recording capacity and wait times for all resources in the database. The analysis executed using GapMap counts the number of resources without any ability to gauge capacity of that resource. This could explain why northern mid-west states seem saturated in resources as compared to coastal states. If a therapy resource in California is a two story clinical building with 10+ therapists while a therapy resource in North Dakota is an office with one therapist, GapMap is currently unable to account for the difference in throughput for children with autism. With capacity data, we could better evaluate where resource gaps are more or less prevalent. Moreover, GapMap does not assess the current capability or wait time of a resource. In certain urban areas, diagnostic service centers may not have availability for many months, rendering them unavailing for families who are seeking a timely diagnosis for their child. Timely diagnosis is vital to early intervention which in turn significantly ameliorates symptoms of ASD. Therefore, with the knowledge that the average wait times for diagnostic resources in a location is high, we would be able to better pinpoint where alternative diagnosis capabilities are needed to reduce the congestion of resource demand in an area.

Finally, in the northern parts of the United States (namely North Dakota, Montana, and South Dakota), there was low prevalence of ASD, but a relatively high number of resources available per autistic student. This may indicate that these Northern states have better autism resource coverage than other states. However, these particular states are geographically large with areas of low population density. It may be that many resources are required in order to be geographically accessible to students. Further research looking into precise geographical locations of clinics, clinic size and capacity, and average wait times at clinics would be needed to better understand the landscape of autism resources in relation to where autism rates are lower. Our map analysis suggests that there may be an increasing nearly nationwide shortage of resources for the growing ASD K-12 population. More research is needed to better understand where and why there are scarcities of resources in relation to ASD prevalence.

Our work shows that CRDC data is a valuable resource for analyzing the geographical distribution of autism. In concert with autism resource databases like GapMap, it can help us understand where more resources are needed to support autistic students in the U.S. CRDC is an incredibly comprehensive resource, containing data on 53M students. We encourage other researchers to consider working with this resource.

Figures & Table

Figure 1:

GapMap is an autism resource database containing the locations of diagnostic and therapeutic autism resources across the U.S.

Figure 7:

Diagnostic resources per thousand autistic students by state. There were a median of 24 ABA resources, 36 therapy resources and 27 diagnostic resources per thousand autistic students across the 50 states. South Dakota and North Dakota have the largest number of resources per autistic student across all categories while Puerto Rico, Nevada, and California consistently have the fewest across all categories.