Chapter 1. Epidemiology

Starting with Kanner (1943), case definitions of autism have progressively broadened to include Rutter’s (1970) criteria and subsequently the ICD-10 (World Health Organization 1992) and DSM-IV-TR (American Psychiatric Association 2000). However, ASD in DSM-5 (American Psychiatric Association 2013) narrowed the concept.

When a population is identified for a survey, different strategies are employed to find individuals matching the study case definition. Some studies rely solely on service-provider databases or national registers for case identification; these have a common limitation of relying on samples already identified and diagnosed. Persons with the disorder who are yet undiagnosed are not counted as cases, leading to prevalence underestimation—a particular problem in communities with few available services. Other investigations rely on a multistage approach to identify cases. In a first screening stage, a wide net is cast to identify possible ASD cases, with the final diagnostic status being determined at subsequent stages. This process often consists of sending letters or screeners to school and health professionals, searching for possible cases of ASD. Systematic sampling techniques that would ensure near-complete coverage of the population are seldom used, and screening varies substantially in the data sources ascertained. Surveyed areas also often differ in terms of educational or health care systems. Finally, uneven participation rates in the screening stage can lead to variation in screening efficiency and bias in prevalence estimation, as illustrated in the hypothetical example of Figure 1–1.

The sensitivity of the screening methodology is difficult to gauge in surveys because the proportion of children truly affected with the disorder but not identified at the screening stage (false-negatives) remains generally unmeasured. Random sampling of screen-negative subjects to adjust estimates is not routinely performed because the low frequency of ASD makes such a strategy both imprecise and costly. For example, the surveys conducted by the Centers for Disease Control and Prevention (CDC) rely, for case ascertainment, on scrutinizing educational and medical records. Children without such records cannot be included. Although surveys that systematically screen the general school population have consistently detected pools of unidentified cases (Alshaban et al. 2019; Fombonne et al. 2016; Kim et al. 2011), the relative contribution of that group to the total prevalence pool varies.

Refinements to this component of autism survey methodology are needed in future studies. Of note, the CDC methodology identifies about 20% of previously undiagnosed ASD cases (Baio et al. 2018), suggesting that underidentification is a widespread phenomenon. Because recent prevalence studies suggest that ASD can no longer be regarded as rare, estimation of false-negatives has become a necessary strategy. Currently, the prevalence estimates of most surveys are likely underestimates of the “true” prevalence rates, with the magnitude of this underestimation remaining unknown.

After screening, screen-positive participants undergo a more in-depth diagnostic evaluation to confirm case status. The information used to determine diagnosis usually combines data from informants (e.g., parents, teachers, pediatricians) and other sources (e.g., medical records, educational sources), with direct assessment of the person with autism offered in some but not all studies. When subjects are directly examined, assessment methods vary from an unstructured examination by a clinical expert (but without demonstrated psychometric properties) to the use of batteries of standardized measures (e.g., Autism Diagnostic Interview–Revised [Lord et al. 1994], Autism Diagnostic Observation Schedule [Lord et al. 1989]). Many of these tools are also being adapted to be culturally appropriate for non-Western countries (Arora et al. 2018; Raina et al. 2015; Sun et al. 2019).

Obviously, surveys of very large populations (e.g., CDC surveys in the United States) or of national registers (e.g., Idring et al. 2012 in Sweden) cannot include direct diagnostic assessment of all participants. In those instances, the validity of diagnoses is evaluated in small, randomly selected subsamples who are given a more complete diagnostic workup (Rice et al. 2007).

These surveys were identified from previous reviews (Elsabbagh et al. 2012; Fombonne et al. 2011; Hill et al. 2014) and through systematic searches using major scientific literature databases (Medline, PsycINFO, Embase, PubMed). When multiple surveys were based on the same or overlapping populations, the most detailed and comprehensive account was retained. For example, surveys conducted by the CDC are included in the chapter appendix, although additional accounts for individual states are available elsewhere. Criteria for inclusion in this review were 1) full article published in English; 2) a target population size of at least 5,000; and 3) independent validation of caseness by professionals. From published articles, we abstracted information about country and area of the survey, population size, participant age range, number of children affected, diagnostic criteria used in case definition, and prevalence estimate (per 10,000). We also report, when available, sex ratios and the proportion of individuals with an IQ in the normal range.

The results of the 71 surveys meeting these criteria are summarized in the chapter appendix. Most (54%) were published in 2011 or later. Studies were performed in 25 different countries, with several countries contributing multiple studies (including 15 studies in the United Kingdom and 17 in the United States, with 10 published by the CDC). Sample sizes ranged from 3,964¹ to 4.25 million (median 56,946), and participant ages ranged from 0 to 98 years (median 8 years). Two studies, both in England, specifically focused on adults and provided the only prevalence estimates thus far available for adults: 98.2 and 110 per 10,000 (Brugha et al. 2011, 2016). Surveys focusing on toddlers (Hoang et al. 2019; Nygren et al. 2012) and preschoolers (Christensen et al. 2019; Nicholas et al. 2009; Yang et al. 2015) together provided a mean estimate of 136 per 10,000.

The diagnostic criteria used reflected the reliance on modern diagnostic schemes (15 studies used ICD-10; 41 used DSM; 9 used both). Assessments were often performed with standardized diagnostic measures. Use of DSM-5 in recent surveys has resulted in lower prevalence than in surveys using DSM-IV or DSM-IV-TR (American Psychiatric Association 1994; Kim et al. 2014; Kulage et al. 2020; Maenner et al. 2014). In the most recent CDC study (Baio et al. 2018), there was an 86% overlap of diagnoses and only a 4% loss of ASD diagnosis between the case definitions in DSM-IV-TR and DSM-5, although this small effect could result from grandfathering existing pervasive developmental disorder diagnoses into the new DSM-5 criteria. This approach has been criticized by Fombonne (2018), who showed that, in that survey, use of DSM-5 criteria resulted in an 18% reduction in prevalence compared with DSM-IV.

In 32 studies in which IQ scores were reported, the proportion of subjects within the normal IQ range varied from 0% to 100% (median 55.0%). Males were overrepresented in the 58 studies reporting sex ratios, with male-to-female ratios ranging from 1.4:1 to 15.7:1 (median 4.3:1). There was a 189-fold variation in ASD prevalence, ranging from 1.4 to 264 per 10,000 (Figure 1–2). Substantial variation in confidence interval width also was found, reflecting variation in sample sizes and in each study’s precision. However, some consistency in ASD prevalence is found in the center of this distribution, with a median rate of 64.9 per 10,000 and a mean rate of 75.8 per 10,000 (interquartile range 35.8–104.0 per 10,000). Prevalence was negatively associated with sample size (Kendall’s τ =  0.18, P = 0.03), with small-scale studies reporting higher prevalence. There was also a significant positive correlation between ASD prevalence estimates and publication year (Kendall’s τ = 0.25, P = 0.003), with higher rates in more recent surveys. The average prevalence for surveys published in 2010 and later (n = 43) is 87 per 10,000. When considering only the surveys conducted in the United States, the average prevalence for surveys published in 2000 and later (n = 17) was 98 per 10,000 and for surveys published in 2010 and later (n = 9) was 130 per 10,000.

Eighteen studies since 2000 reported ASD prevalence estimates higher than 100 per 10,000, or 1%. Baird et al. (2006) and Kim et al. (2011) both employed proactive case-finding techniques, relying on multiple and repeated screening phases that surveyed different informants at each phase, which certainly enhanced the sensitivity of case identification. Multisource active surveillance techniques, as employed in the Stockholm Youth Cohort (Idring et al. 2012) and by the CDC’s Autism and Developmental Disabilities Monitoring (ADDM) Network, also improve identification of individuals with ASD. The most recent*² CDC prevalence estimate (study year 2016, children age 8 years; Maenner et al. 2020) of 170 per 10,000 reflects the highest estimate to date across all of the previous ADDM Network reports.

Overall, results of recent surveys (n = 71) agree on an average figure of 76 per 10,000 (equivalent to 7.6/1,000 or 0.76%), translating to 1 child out of 132 having an ASD diagnosis. However, this estimate represents an average and conservative figure, and substantial variability exists between and within studies and across sites or areas.

Increasing numbers of children referred to specialist services or known to special education registers has been taken as evidence for increased incidence of ASD. Such upward trends have been seen in many different countries and datasets (Gurney et al. 2003; Shattuck 2006), all occurring in the late 1980s and early 1990s. However, trends over time in referred samples are confounded by referral patterns, availability of services, heightened public awareness, decreasing age at diagnosis, and changes over time in diagnostic concepts and practices.

As an illustration, Figure 1–3 uses hypothetical data to contrast two methods for surveying ASD: one based on sampling from the total population, and the other relying solely on number of persons accessing services. The latter approach suggests a rise in disease occurrence that is not supported by population data. Such a pattern of results based on special education data was reported in Wisconsin (Maenner and Durkin 2010), where prevalence rates of ASD were stable between 2002 and 2008 in school districts with initially high baseline prevalence rates (approximately 120 per 10,000), whereas school districts with low baseline rates experienced significant increases in prevalence, as was seen in California (Fombonne 2001).

The decreasing age at diagnosis also results in increasing numbers of young children being identified in official statistics (Christensen et al. 2019) or referred to specialist medical and educational services. Earlier identification of children from the prevalence pool therefore may result in increased service activity, leading to a misperception by professionals of an epidemic.

Another possible explanation is that children presenting with the same developmental disability may receive one particular diagnosis initially and another diagnosis subsequently. Such diagnostic substitution (or switching) may occur when diagnostic categories become increasingly familiar to health professionals or when access to better services is ensured by using a new diagnostic category. The strongest evidence of diagnostic substitution contributing to ASD prevalence increase was shown in a complex analysis of U.S. Department of Education data in 50 U.S. states (Shattuck 2006) indicating that a relatively high proportion of children previously diagnosed with mental retardation were subsequently identified as having autism. Shattuck (2006) showed that the odds of receiving a diagnosis of autism increased by a factor of 1.21 per year between 1994 and 2003, whereas the odds of receiving a diagnosis of learning disability and mental retardation both decreased (0.98 and 0.97, respectively). Shattuck further established that the growing autism prevalence was directly associated with decreasing prevalence of learning disability and mental retardation within states.

Using individual rather than aggregate-level data, another study showed that 24% of the increase in the California Department of Developmental Services caseload was attributable to diagnostic substitution (from mental retardation to ASD) (King and Bearman 2009). Other types of diagnostic substitution are likely to have occurred as well for milder forms of ASD. For example, children with diagnoses of Asperger’s disorder may have been previously diagnosed with other psychiatric diagnoses (e.g., OCD, school phobia, social anxiety) in clinical settings before the developmental nature of their condition was fully recognized (Fombonne 2009).

Evidence that method factors could account for most of the variability in published prevalence estimates comes from a direct comparison of eight surveys conducted in the United Kingdom and the United States (Fombonne 2005). In each country, four surveys conducted around the same year in similar age groups exhibited an unexpected 4- to 5.5-fold variation in rates, with higher rates found when intensive population-based screening techniques were used, and lower rates reported in studies relying on passive administrative methods for case finding. Because no passage of time was involved, the magnitude of this gradient in rates is likely to reflect methodological differences.

Likewise, in a CDC survey of 346,978 children age 8 years in 2012, in which an average prevalence of 146 per 10,000 was reported across 11 U.S. states (Christensen et al. 2016), a threefold variation in prevalence by state was observed (range 82–246 per 10,000; see Figure 1–4). It would be surprising if there were truly this much inherent state-to-state variability in the number of children with autism in the United States. Thus, these differences likely reflect ascertainment variability across sites in a study that was otherwise performed with the same methods, at the same time, with children of the same age, and within the same country.

Studies conducted in Sweden (Gillberg et al. 1991) and in Toyota, Japan at different points in time (Honda et al. 2005; Kawamura et al. 2008) showed rises in prevalence rates that their authors interpreted as reflecting the effect of improved population screening of preschoolers, broadening of diagnostic concepts and criteria, and improved services. By contrast, two surveys of children born between 1992 and 1995 and between 1996 and 1998 in Staffordshire, United Kingdom (Chakrabarti and Fombonne 2001, 2005) and performed with rigorously identical methods for case definition and case identification suggested no upward trend in overall rates of ASDs, at least during the short time interval between studies.

Two large French surveys including birth cohorts from 1972 to 1985 (735,000 children, 389 of whom had autism) showed no upward trend in age-specific rates (Fombonne and du Mazaubrun 1992; Fombonne et al. 1997). Two recent Australian studies examined the effect of birth cohort on diagnosis. Randall et al. (2016) found that ASD diagnosis before age 7 was higher in the later-born cohort (birth year 2004–2005: 2.5%) than in the earlier-born cohort (1999–2000: 1.5%). A follow-up study of the later-born cohort found that both teacher- and parent-reported prevalence rates held the same pattern of higher ASD prevalence into school age (May et al. 2017). However, data assessing birth cohorts can be problematic, as illustrated in Figure 1–5, which shows an increase in the prevalence of ASD by year of birth across three hypothetical successive birth cohorts (a cohort effect; Figure 1–5A). Within each birth cohort, followed longitudinally, prevalence increases as children age (Figure 1–5B): for children in the 2000 birth cohort, based on previous ASD prevalence estimates, age 6 prevalence was 20 per 10,000, whereas at age 12 we may expect a prevalence of 80 per 10,000 for the same birth cohort. Birth cohort and age effects cannot be disentangled because they both share time as a common factor. Rather than signaling an increased incidence in successive birth cohorts, the increasing prevalence rates in later childhood and early adolescence within birth cohorts cannot reflect new onset of ASD and have no biological meaning. Observed increases in prevalence with age more likely reflect underdiagnosis in the preschool years and changes in public awareness, service availability, diagnostic concepts and practices, and perhaps overdiagnosis during school age (Fombonne 2018).

As an example, an analysis of special education data from Minnesota showed a 16-fold increase in children identified with ASD from 1991–1992 to 2001–2002 (Gurney et al. 2003). However, during the same period, a 50% increase was observed for all disability categories (except severe intellectual deficiency), especially ADHD. The large sample size allowed the authors to assess age, period, and cohort effects. Prevalence increased regularly in successive birth cohorts; for example, among 7-year-old children, prevalence rose from 18 per 10,000 among those born in 1989, to 29 per 10,000 among those born in 1991, to 55 per 10,000 in those born in 1993. Age effects were also apparent within the same birth cohorts; for example, among children born in 1989, the prevalence increased with age from 13 per 10,000 at age 6, to 21 per 10,000 at age 9, to 33 per 10,000 at age 11. As argued by Gurney et al. (2003), this pattern is not consistent with the natural etiology of ASD, which first manifests in early childhood. The analysis by Gurney and colleagues also showed a marked period effect because rates started to increase in all ages and birth cohorts in the 1990s. The authors noted that this phenomenon coincided closely with the inclusion of ASD in the federal Individuals with Disabilities Education Act (1990) in the United States. A similar interpretation of upward trends had been put forward by Croen et al. (2002) in their analysis of the California Department of Developmental Services data and by Shattuck (2006) in his analysis of trends in U.S. Department of Education data.

SES can be defined by parental education, income, parental occupation, or some combination of these factors. More than 20 studies have investigated associations between these factors and ASD prevalence. Recent U.S.-based studies suggest an association between higher SES and higher ASD prevalence. Using CDC ADDM Network data states, Durkin et al. (2017) found a dose-response relationship between SES (defined as parental education) and ASD prevalence in all recent survey years, in white, Black, and Hispanic children. This difference remained present regardless of sex, prior ASD diagnosis, and source of records (i.e., medical only versus medical and educational). However, no significant difference was found among those with co-occurring intellectual disability, and the difference appeared to lessen in non-Hispanic children over time. Similarly, Dickerson et al. (2017) used U.S. Census tract data to show a negative association between neighborhood poverty level (as defined by median household income or proportion in poverty) and ASD prevalence at ADDM Network sites.

Many studies of racial/ethnic minorities show lower rates of ASD compared with white or European populations, although these differences appear to be narrowing in more current studies. Recent data from the CDC ADDM Network (Baio et al. 2018) suggest an overall lower rate of ASD among non-Hispanic Black (160 per 10,000), Hispanic (140 per 10,000), and Asian/Pacific Islander (135 per 10,000) children compared with white children (172 per 10,000) in the United States. Similar trends have been noted in 4-year-olds in the ADDM Network; however, most recent waves of this survey suggest that racial/ethnic differences may be narrowing or even disappearing, particularly for non-Hispanic Black children (Christensen et al. 2019; Nevison and Zahorodny 2019). However, some of these differences may be explained by ascertainment in the ADDM Network because nonwhite children were more likely to be excluded due to missing information (Imm et al. 2019).

In studies outside of the United States, reports about racial/ethnic differences in ASD prevalence have been mixed, and most studies are not adjusted for SES, which makes it difficult to assess the unique effect of race/ethnicity from other confounders. In addition, what constitutes a minority race or ethnicity is quite variable by country. In Israel, Davidovitch et al. (2013) and Jaber et al. (2018) both conducted studies showing a lower prevalence of ASD in ultra-Orthodox Jewish subjects than in the general Israeli population. Davidovitch et al. (2013) also found a lower prevalence in Israeli Arab subjects, but Jaber et al. (2018) found no difference. Levaot et al. (2019) found lower prevalence in Bedouin-Arab compared with Jewish children in southern Israel. Findings from a 1999–2003 census report in Stockholm, Sweden (Barnevik-Olsson et al. 2010), revealed that the prevalence rate of pervasive developmental disorder with learning disability was higher in Somali compared with non-Somali Swedish children. Finally, in Western Australia, children of Indigenous mothers were less likely and children with East-African Black mothers were more likely to carry an ASD diagnosis compared with children of white mothers (Fairthorne et al. 2017).

Overall, the research findings related to low SES and minority status primarily point to problems of underdiagnosis due to problems in access to health care services and health literacy. To obtain an accurate depiction of ASD prevalence in underserved populations, investigators will need to specifically reach out to these populations to ensure equal participation and oversample these groups so that sample sizes are adequate. In addition, validated screening and diagnostic tools in multiple languages are needed to ensure that diagnoses, when they occur, are accurate. Finally, key variables in these analyses, such as parental education, income, and race/ethnicity, need to be directly and routinely measured.