Autism Spectrum Disorder (ASD) is a lifelong neurodevelopmental disorder characterized by difficulties in the social-communication domain, and the presence of restricted and/or repetitive interests and/or behavior (1, 2) manifested before the age of three. ASD has encompassed a number of diagnoses that, in previous versions of medical classifications, were clustered in the group of Pervasive Developmental Disorders (PDDs). However, by being a spectrum, this condition is also characterized by a high level of heterogeneity of the phenotypic manifestations, associated with wide variability in levels of intellectual and language development, intra-individual discrepancies in cognitive profiles (3, 4), and frequently reported comorbidity with other developmental disorders and psychiatric conditions (5). This combination of specific deficits and high comorbidity places ASD among the most disabling developmental disorders, imposing an extremely high economic burden. For example, in the USA, by 2025, the forecast annual direct and productivity costs of ASD could reach $461 billion (6).

Investigations into the prevalence of autism began as early as the 1960s–1970s (7–9), even before the inclusion of ASD in the international diagnostic classification and before the unification of ASD diagnostic criteria. ASD prevalence in these initial studies was estimated to be between 0.5 (9) and 0.7 (8) cases per 10,000 people. Since the 1970s, prevalence studies have been conducted across different regions, covering at least 37 countries worldwide, although the data are still lacking in many low- and middle-income countries (10). Notably, these studies have shown a dramatic increase in ASD prevalence estimates since the late 1900s/early 2000s. This time trend is particularly evident if we compare data obtained in cross-sectional monitoring programs that were developed in some counties to track changes in the prevalence of ASD. Thus, according to the Autism and Developmental Disabilities Monitoring (ADDM) Network launched by the Centers for Disease Control and Prevention (CDC) in the USA, the prevalence of ASD among 8-year-old children in 2000 was 67 per 10,000 (11), 90/10,000 in 2006 (12), and 145/10,000 in 2012 (13). The most recent report, dated 2018, revealed that ASD affects approximately 230 per 10,000 children (14), indicating that its prevalence has increased by 243% since the first ADDM Network study in 2000 (14). Previously published systematic reviews demonstrated that these discrepancies in estimates over time probably are not caused by an increase in true prevalence but are instead associated with changes and improvements in diagnostic categories, methodology, and quality of research, increased access to diagnostic and intervention services, increased awareness of ASD among both professional and non-professional communities, and acceptance of the fact that ASD can coexist with other forms of developmental disorders [for example, in (15–17)].

In addition to time trends, there are other sources of heterogeneity, such as geographical region, country income, implemented study design, diagnostic criteria used, age of the studied population, and other socio-demographic factors (18–20). The complex interaction between these factors makes obtaining a reliable, unified worldwide prevalence estimate challenging. However, obtaining pooled worldwide prevalence could play a critical role in indicating the public health burden of the disorder, allocating budgets for ASD in countries where prevalence studies have not been conducted or are just about to be performed, and detecting problems in the identification and diagnostics of ASD in countries where the data are well below the world mean estimates, and for revealing potential environmental risk factors for ASD in some regions.

The data from the most recent studies of global prevalence estimates are somewhat variable. For example, the systematic review by Zeidan et al. (18) showed that 71 studies and 99 prevalence estimates from 34 countries published from 2012 to 2021 report a median prevalence of 100 per 10,000 children; in the meta-analysis of Salari et al. (19) showed that 74 studies, published from 2008 to 2021 the pooled prevalence comprised 60 per 10,000.

Given the great heterogeneity encountered in the prevalence of ASD over time and inconsistency among previously published estimates, this study aimed to conduct a new meta-analysis. Therefore, the first aim of this research was to update the previous meta-analyses and systematic reviews. The second aim was to identify and analyze potential moderators regarding methodological aspects, geographical locations, age groups, and related socio-economical indexes that may influence the observed heterogeneity in prevalence estimates. Considering the multiple data that could be extracted from a single study, the present research also contributes to the current knowledge by applying a state-of-the-art, three-level meta-analytical approach to estimating the mean pooled prevalence of ASD and previously existing diagnoses of Autistic Disorder (AD), Asperger Syndrome (AS), Atypical Autism (AA), and Pervasive Developmental Disorder - Not Otherwise Specified (PDD-NOS), along with subgroup estimates for ASD.

Overall, 85 articles with 317 extracted prevalence estimates were included in the quantitative synthesis, covering the period from 1994 to 2019 and a population of 120,130,596 individuals from 29 countries. Compared to previous studies aimed at assessing worldwide prevalence, we have tried to select potentially the most reliable data by including in the analysis only those studies in which case identification was based on formal diagnoses according to ICD and DSM classifications or on the results of so-called “gold standard” methods (ADOS and ADI). Thus, this study lacks the results of a large pool of data from studies based on national surveys and studies that applied diagnoses according to classifications other than ICD and DSM, or those, without information about case identification. Particularly, this study also did not include data from studies in which case identification was based solely on the educational diagnosis of ASD, as in many systems, it is often based on the most appropriate service delivery approach, so children with ASD may be qualified under another disability category, and indeed children without ASD with severe socio-communication impairments could be assigned under ASD for receiving an appropriate educational plan, as in case of classification in the framework of The Individuals with Disabilities Education Act (IDEA) (50).

Among the included studies, a substantial number was presented by high-income countries (83.5%), mainly from the USA (23.5%). Analysis of raw data found that the average percentage of individuals classified as having an intellectual disability in our sample was 34.9 ± 13.5, 60.0 ± 23.8, 3.4 ± 4.5, 24.1 ± 21.6 for ASD, AD, AS, and AA/PDD, respectively. However, it is rather difficult to compare found results with those available in the literature, as they vary considerably, and no one comprehensive review or meta-analysis was not identified.

The mean male-to-female ratio was 4.3 (Med = 4.2, SD = 1.5) for ASD and 4.2 (Med = 4.2, SD = 1.6) for AD. For other subgroups, the ratio was more variable, and the gender ratio increased to 9.7 (Med = 6.2, SD = 8.0) for AS and 7.0 (Med = 4.3, SD = 3.0) for AA. For the whole group of ASD, the obtained estimate is equal to the proportions found in the systematic review of the ASD prevalence performed by Zeidan et al. (18) and in the meta-analysis of Loomes et al. (93), focusing specifically on the male-to-female ratio. However, when the low-quality studies were eliminated in the latter meta-analysis, the ratio was decreased to 3.32.

The prevalence estimates, produced by a three-level meta-analysis, comprised 72 per 10,000 (95% CI = 61–85) for ASD, 25 (95% CI = 18–33) for AD, 13 (95% CI = 7–20) for AS, and 18 (95% CI = 10–28) for the combined group of AA and PDD-NOS. However, as expected, a large amount of variation in prevalence across studies was found due to within- and between-study heterogeneity, which substantially reduces the reliability of these estimates.

Univariate analysis of moderators, associated with study location, revealed that pooled prevalence of ASD was significantly higher in North America (1.01% [95% CI = 0.79–1.25]), compared with estimates obtained for Europe (0.73% [95% CI = 0.57–0.91]) and Asia (0.41% [95% CI = 0.26–0.59]), where among countries that were included in the subgroup analysis the estimate from the USA (1.12% [95% CI = 0.92–1.33]) was the highest, while the lowest prevalence was identified in Taiwan (0.11% [95% CI = 0.02–0.29]). Similar results were identified in the systematic review and meta-analysis of ASD prevalence by Salari et al. published in 2022: for America, the prevalence was 1%, and for Asia–0.4%. However, for Europe, the prevalence was lower–0.5%. As in the cited study, the most prominent estimate was obtained for the region of Australia; it should be noted that in our analysis, the whole region of Oceania (along with Africa) was not included in the subgroup analysis due to the lack of data, eligible for this type of analysis.

Our findings also confirm a clear trend associated with country income and HDI: countries with higher income and HDI tend to report higher prevalence estimates. It may be assumed that this difference could be attributed to the detection gap, associated with lower access to appropriate diagnostics, lower awareness among parental and professional communities about autism, and specific cultural attitudes to the developmental and mental disorders that may influence pathways of families to seek care (51). The effect of socio-economic factors on the prevalence of ASD is evident not only in the case of between-countries comparisons but also it is reported for differences between different communities within one study. As an example, a study conducted in Texas found that the state’s high-income community had a six times higher prevalence of ASD than the lowest-income group, as per the administrative data (52). On the other hand, the effects of country income and HDI in our analysis may also be confounded by the fact that high-income countries are much more represented in our sample compared with low-income and are based on larger sample sizes.

As was expected, the subgroup analysis also revealed a significant moderator effect of study time, reported in the scope of previous research: the global pooled point-prevalence of ASD consistently increased from 0.25% (95% CI = 0.12–0.42) during the period from 1994 to 1999 to 0.99 (95% CI = 0.73–1.28) in the 2015–2019 period.

As for age groups, the prevalence of ASD was significantly higher for children aged between 6 and 12 years (0.82% [95% CI = 0.67–0.98]) compared to the subgroup of children under the age of 5 (0.60% [95% CI = 0.45–0.76]) and those, older than 13 years (0.57% [95% CI = 0.40–0.76]). As it is mentioned by Fombonne et al. (10), by ages 6–10, diagnoses can be assigned more robustly, while at lower ages, some children could be missed up to primary school entry or later when social demands from a child increase and specific autistic impairments become more evident. At the same time, at older ages, some improvements in milder forms of ASD can pose challenges for both identification and diagnostic confirmation.

Regarding the moderator analysis for the methodological aspects of studies, it was shown that the lowest estimates were obtained from studies based on information from health insurance (0.35% [95% CI = 0.10–0.77]) and different types of administrative databases (0.48% [95% CI = 0.31–0.68]). Substantially higher estimates were revealed from studies where cases were identified based on records-review surveillance (1.22% [95% CI = 0.91–1.56]), followed by studies with mixed designs (0.80% [95% CI = 0.52–1.14]), direct surveillance (0.75% [95% CI = 0.51–1.03]), and registers (0.66% [95% CI = 0.44–0.93]). These findings are predictable, as they substantially vary in their complexity, reliability, and consistency of instruments used, the professional level of clinicians participating in the case identification, feasible sample sizes, and cost (10, 53). Thus, studies that use existing databases with routinely collected health information have the lowest cost and provide large sample sizes but do not control effects associated with the professional level of clinicians and reliability of diagnostic methods and are unable to count individuals without an official diagnosis, which could underestimate the prevalence of ASD. At the same time, methods based on records-review surveillance also rely on available clinical and educational information obtained in routine practice. However, these records are scrutinized through two phases by trained experts. The case finding phase involves systematic screening of multi-source records by trained record abstractors aimed at revealing the “red flags” of ASD. In the case of screening-positive results, all available information is extracted and subjected to in-depth evaluation by trained clinicians specializing in the diagnostics of ASD (53). So, even if a child does not have an official diagnosis of ASD, he/she could be classified under this category, and new cases could be established. Nevertheless, since this design does not involve a direct assessment of a child and a direct interview with a child’s caregivers, this design could rely on mechanical rules in the interpretation of available records, lacking clinical judgment and, as a result, overestimate prevalence estimates of ASD (54, 55).

Although all the moderating variables examined in this study showed a significant effect on the pooled prevalence estimate, none of them showed a significant effect on reducing remaining heterogeneity. To get more reliable prevalence estimates, we performed additional steps. Based on the assumption that more recent data (in relation to the improvements in the unification of diagnostic criteria of ASD and the development of more precise diagnostic instruments) from middle- and higher-income countries in Europe and North America are presumably more reliable for ASD prevalence (because they probably have a better case identification policy and have more resources for training of clinical experts and access to reliable diagnostic instruments), we narrowed our sample to estimates from 2015 to 2019 and these regions, using age of participants and study design variables as covariates for the final model. Thus, the final prevalence rate was 1.18% (95% CI = 0.92–1.48), which seemed closer to the up-to-date mean prevalence and the consensus estimate of 1%, accepted by the World Health Organization (56). However, the heterogeneity for this estimate was still very high due to within studies effects, so this result should be interpreted with caution.

This study, to our knowledge, is one of the most comprehensive attempts to address the issue of worldwide prevalence of ASD and previously existing relative diagnoses using meta-analytical statistics and the first that employed the three-level approach to deal with multiple estimates from a single study. Inclusion criteria were stricter than other systematic reviews, which could result in capitalizing on more reliable data.

However, there are a number of limitations. First, only those studies published in English were included, so several studies were missed due to language limitations. Second, partly because of limited data in the literature, partly due to the stricter inclusion criteria, the meta-regression analyses might be under-powered (as we applied a minimal 5 data-point cut-off per level of covariate), so obtained results of the source of heterogeneity should be considered as an exploratory and require future confirmation. Finally, very substantial heterogeneity remained unaccounted for across all subgroups and models examined, even when between-study variance was eliminated and more complex models were applied. As mentioned above, this fact substantially limits the reliability and generalizability of the obtained results.

The original contributions presented in this study are included in the article/Supplementary material, further inquiries can be directed to the corresponding author.

OIT and EG developed the concept and design of the study. DK, ES, OIT, OST, PS, RR, and TD with the exclusion of EG, participated in the data collection. OIT and RR wrote the manuscript. All authors contributed to and approved its final version.