how autism got counted

what “1 in 31” actually counts

The headline autism number comes from CDC's Autism and Developmental Disabilities Monitoring (ADDM) Network, and its construct label matters more than its value: records-based identification, not diagnosis. ADDM counts 8-year-olds meeting a surveillance case definition from health (and, where available, education) records at a non-representative, changing set of US sites. A child never told “you have autism” can be an ADDM case; a diagnosed child whose records aren't accessible can be missed [3][9].

CDC says this itself, in the limitations section of its own report [9]:

“Sites participating in the ADDM Network are selected through a competitive process, and the resulting catchment areas are not designed to be representative of the states in which the sites are located. Findings do not necessarily generalize to all children aged 8 years in the United States, and interpretations of temporal trends are complicated by changing catchment areas, case definitions, and diagnostic practices.”— CDC, surveillance year 2018 report, Limitations [9]

That sentence is the thesis of this page, written by the agency that publishes the number. The chart below adopts it structurally: the line is severed at every point where the agency's own record says the ruler changed.

not a national rate: the same year, site by site

The single best correction to the way this number is quoted is CDC's own site-range column. In surveillance year 2022, the combined figure was 32.2 per 1,000 — and the per-site figures ran from 9.7 per 1,000 in the Laredo, Texas area to 53.1 per 1,000 in California, a 5.5-fold spread inside the same year of the same surveillance system [11][12]. Charts of “the US autism rate over time” are charts of a network of places, and the places changed.

Note also what happened to the range itself: in 2016 the sites agreed almost perfectly (18.0–19.1); by 2022 they disagreed by a factor of 5.5. A number whose internal spread behaves like that is telling you about identification systems, not about children.

the three breaks on the record

2004 — the year CDC itself set aside. The 2004 data were published only as an appendix, and the network described the year in its own words: “a smaller-scale effort than other ADDM Network surveillance years… caution is needed when comparing 2004 results with those from other surveillance years” [2]. The later 2008 report went further and simply declined to compare against it, calling it “a smaller scale, optional effort based on available resources” [4]. So it is drawn as a disconnected point, exactly as the agency treats it.

2018 — the case definition changed, and CDC measured the step itself. Through 2016, trained clinician reviewers applied operationalized DSM criteria to abstracted record text — a child could be a case with no diagnosis anywhere in their records [3]. From 2018, a child is a case if the system ever labeled the child: a diagnostic statement, autism special-education eligibility, or an ASD ICD code [9][11]. That is a different construct — “records describe autism” became “the system ever said autism.” What makes this break unusually honest is that CDC published its own bridge study [9]:

“An analysis using data from ADDM Network surveillance years 2014 and 2016 compared the case definitions and found that, compared with the overall ASD prevalence using the previous case definition, ASD prevalence using the new case definition was approximately the same for 2014 and 7% lower for 2016. … Approximately 86% of all children who met either the previous or new case definition met both case definitions.”— CDC, surveillance year 2018 report [9]

So the annotation carries its own size: roughly zero to minus seven percent. The 2018 break is real, quantified by the agency, and too small to carry the 2016→2018 step by itself — both facts stay on the page.

2020 → 2022 — the network changed, on the record, for a stated reason. 2018 and 2020 used the identical 11 sites. 2022 has 16 — Indiana, Pennsylvania, Puerto Rico, and two Texas sites joined, so five of sixteen sites are new to the network [11]. The reason is not epidemiological and CDC states it plainly: “five additional sites were able to join in April because of 2023 Consolidated Appropriations Act funds for expansion” [11]. When the appropriation changed, the network changed; when the network changed, the number changed. The 27.6 → 32.2 step is a composition change, which is why 2022 is drawn as a disconnected point.

what DSM-5 did — the definition timeline, and the drop that never came

“Autism” as a diagnosis is younger than most people assume, and its definition has been rewritten repeatedly. The citable history [18]:

Here is the American Psychiatric Association describing its own 2013 merge [19]:

“Autism spectrum disorder is a new DSM-5 name that reflects a scientific consensus that four previously separate disorders are actually a single condition with different levels of symptom severity in two core domains. ASD now encompasses the previous DSM-IV autistic disorder (autism), Asperger's disorder, childhood disintegrative disorder, and pervasive developmental disorder not otherwise specified.”— APA, Highlights of Changes from DSM-IV-TR to DSM-5 [19]

The interesting part is what the field expected this to do to the numbers, because the expectation ran downward:

• Predicted: a 2014 meta-analysis projected DSM-5 would cut ASD diagnoses by a pooled 31% (CI 20–44), with PDD-NOS down 70% [20].
• Projected onto surveillance records: applying DSM-5 criteria to ADDM records suggested prevalence would fall from 11.3 to 10.0 per 1,000, with 81.2% of DSM-IV-TR surveillance cases meeting the new criteria [22].
• Measured, five years later: the same group's follow-up meta-analysis found a pooled decrease of 20.8% (16.0–26.7) in study samples — smaller than predicted [21]. In clinic samples, DSM-5 criteria identified 91% of children with DSM-IV diagnoses [23].

And the surveillance series itself? It kept rising through 2013 with no downward step — the predicted drop never appeared in the counts [20][21]. That is why the DSM-5 marker on the chart is a thin line the data runs through. Two claims this supports, and one it does not: DSM-5 merged four diagnoses into one; the predicted shrinkage failed to materialize in surveillance data; and — because the rise long predates 2013 and the measured DSM-5 fingerprint is a decrease smaller than predicted — “DSM-5 caused the rise” is not a claim this record supports, and this page does not make it.

the second series: when the questionnaire changed, the number doubled

The National Health Interview Survey measures something different from ADDM: parents are asked whether a doctor or health professional ever told them their child has autism, ages 3–17. No records are checked. It is the cleanest measurement-artifact exhibit in the entire autism dataset, because of what happened in 2014:

No population changes that fast. NCHS's own authors say so — the paper is literally titled “…Following Questionnaire Changes in the 2014 National Health Interview Survey,” and it states that “the revised question ordering and new approach to asking about developmental disabilities in the 2014 NHIS likely affected the prevalence estimates” — previously, “some parents of children diagnosed with ASD reported this developmental disability as other DD instead of, or in addition to, ASD” [13]. The children did not change in 2014. The form did.

A precision point on the 2019 band. NCHS's own language is a hedge, not a prohibition. A flat NCHS sentence saying 'do not compare 2019+ to 1997-2018' does not exist; it was searched for and not found. The blunt 'should not be directly compared' phrasing belongs to CDC's ADHD programme page, not to NCHS survey documentation. We attribute each quote to the right agency and do not upgrade the hedge. Within its segments, NCHS reports the movements were not statistically significant — neither 2014→2016 nor 2019→2021 [14][15].

three constructs, three numbers, one era

Around the mid-2010s, three defensible, primary-sourced measurements of childhood autism coexisted — and they differ by design, not by error. This is the cheapest demonstration that “the autism rate” is a construct-dependent quantity:

One more construct completes the picture, from outside the US. The one large study that screened an entire population — every child in a South Korean community, 2005–2009 fieldwork, screening plus direct assessment rather than records — found 2.64% (CI 1.91–3.37), with “two-thirds of ASD cases in the overall sample… in the mainstream school population, undiagnosed and untreated” [17]. At the time, US records-based figures were around 1%. Different country, different service system, different era, wide interval — it is a construct exhibit, not a trend point, and it appears on no axis here. What it teaches survives all those caveats: when you look for autism directly instead of reading records, you find more of it, and most of what you find was undiagnosed.

how much of the rise is counting?

This is the question the whole page has been building to, and it has a peer-reviewed literature that almost never gets quoted with its confidence intervals attached. Three quantified decompositions [24][25][26]:

The best single exhibit in the literature does not produce a percentage at all. A Swedish group measured the autism symptom phenotype in year-by-year population samples of twins for a decade, and compared it with registered diagnoses in the full national child population [27]:

“The annual prevalence of the autism symptom phenotype was stable during the 10 year period (P=0.87…). In contrast, there was a monotonic significant increase in prevalence of registered diagnoses… (P<0.001)” — with the authors concluding that “administrative changes… rather than secular factors affecting the pathogenesis, are important for the increase.”— Lundström et al., BMJ 2015 (twins n=19,993 for the phenotype; all 1,078,975 Swedish children born 1993–2002 for diagnoses) [27]

Symptoms flat, diagnoses rising — same decade, same country. The thing being measured did not move while the count of it climbed.

The oldest study in this literature is also the one most often over-quoted: Croen 2002 found California's autism increase almost mirrored by a decrease in intellectual-disability-without-autism — but its own text hedges (“whether there has also been a true increase in incidence is not known”), the near-1:1 offset was contested in print the following year, and King & Bearman later quantified that pathway at only about 26% [28][25]. A related California analysis found changing age at diagnosis could explain a 12% increase and inclusion of milder cases a 56% increase — multiplicative factors on a several-fold rise, not additive shares, and its conclusion is that the increases are not fully explained [29]. Globally, the most recent systematic review attributes the rise in reported prevalence to awareness, public-health response, case identification and definition progress, and community capacity — as combined factors, with no percentage attached [30].

Both flat verdicts are therefore off the table. “It's all better counting” is unsupported — no study attributes 100% to ascertainment. “The rise is real and unexplained” is equally unsupported — the quantified artifact share starts at a quarter and may reach most of it. The same discipline governs the folklore: the 1998 paper that launched the vaccine story was retracted in full by The Lancet in 2010 [31], and that one clause of history is this page's entire engagement with it — the decomposition literature above is what actually explains where the rise comes from, mechanism by measured mechanism.

the closing exhibit: who gets identified changed too

One more passage from CDC's own 2022 report, because it reframes what a rising count can mean [11]:

“Before 2016, the highest ASD prevalence was observed among White children and in children from neighborhoods with higher socioeconomic status… In 2020, higher ASD prevalence was observed for the first time among historically underserved groups including non-Hispanic Black and Hispanic children.”— CDC, surveillance year 2022 report [11]

An identification pattern inverted. The children who were hardest for the system to see are now identified at the highest rates — which is a statement about access and ascertainment, in CDC's own framing, not about etiology. The 2018 case-definition change points the same direction: CDC noted the new definition “could be more likely… to include children of lower SES,” because the old one excluded children whose records lacked detail [9].

The adult side of that same access story — why the best English estimates find most autistic adults undiagnosed, what “on the spectrum” refers to clinically, and what an actual evaluation involves — is a different kind of page, and it lives in the lexicon entry on “on the spectrum”. This page is about how a number is made; that one is about what the words mean and who never got counted at all.

what this page does not show

• Whether there is or isn't an “autism epidemic.” The series measure identification by systems, every method change moved the number, and the decomposition studies leave an explicit residual. The page declines the verdict in both directions because the data cannot deliver one.
• That people are overdiagnosed. The decomposition studies attribute specific fractions of the rise to specific non-etiologic mechanisms; none of them measures “overdiagnosis,” and none licenses a claim about any individual's diagnosis.
• That DSM-5 caused the rise. The rise long predates 2013, and DSM-5's measured fingerprint in the impact literature is a decrease smaller than predicted.
• A single joined “autism rate over time.” Every panel here keeps its own construct, its own axis, and its own breaks. Splicing ADDM to NHIS, or 2016 to 2018, or 2020 to 2022, would manufacture trends the agencies' own records contradict.
• Anything about self-diagnosis rates or accuracy. No such measurements exist — that finding, with its evidence, lives on the lexicon page.
• Anything about you or your child. This is a page about a statistic.

how this page is built

Every series is a checked-in JSON file under data/system-education/ carrying its source citations, pull date, measurement-type label, and the agencies' own comparability warnings as text; both series also download as CSV (linked in the structured data). Every value was verified against a primary source — the MMWR reports via PubMed E-utilities, CDC's own archived data table, and NCHS's publication PDFs read page-by-page — in two independent verification chains on the date stamped above. One citation trap is worth naming for anyone reproducing this: the 2012 ADDM value circulates in two versions because CDC corrected it in a 2018 republication [6]; this page carries the corrected 14.5, and any source quoting 14.6 is quoting a superseded number. The charts are server-rendered inline SVG with no charting library; nothing in the chart component can join two segments, which makes the no-splice rule structural rather than editorial.