What’s the Deal With Autism Rates?

One of the major priorities of the new Secretary of Health and Human Services (HHS), Robert F. Kennedy Jr. (RFK Jr.), is to figure out why rates of autism have been increasing over time.

But is this a good priority? To determine why autism rates have increased, we first have to know if autism rates have increased. There’s some evidence that speaks to that. For example, diagnosed rates of autism have clearly increased over time across every major data collection effort. If we look at recent snapshot data from one such effort conducted by the California Department of Developmental Services (CDDS), we can see that autism diagnoses have stunningly increased since the 1930s, from applying to 0.001% of those in 1931 to 1.2% of five-year-olds in 2021.

Despite the striking and shocking nature of the data on this graph, interpreting it as evidence that autism has increased over time is hard.

For starters, the ages in the most recent cohorts are young—they’re children—but in the oldest cohorts, we’re talking about adults whose data wasn’t gathered by the CDDS when they were kids, but instead much later, long after these people had made it to adulthood, when diagnosis is frequently neglected. Autism is also associated with shorter lifespans, which are even shorter for the people with the most severe manifestations of the condition. Accordingly, many of the oldest autistics would have already died off by the time the CDDS got around to surveying. These biases deflate the rate of autism the further back you go, exaggerating the subsequent increase.

The other, more pressing issue is that the constellation of traits categorized as “autism” wasn’t named until Leo Kanner’s descriptions of “abnormal behaviour” in the 1940s, and it would be another nearly forty years before American psychiatry provided criteria for autism diagnoses in the third edition of the Diagnostic and Statistical Manual of Mental Disorders (DSM-III). Before the DSM-III came out in 1980, autism diagnoses were usually ad hoc, based on the personal views of different clinicians and researchers about what autism is and how to diagnose it. Because autism was regarded as particularly severe prior to the DSM treatment of it and those whose condition goes unnoticed by adulthood likely aren’t severe cases, and because there was practically no incentive to diagnose until the most recent generations, that meant it was diagnosed rarely, making subsequent increases that much more exaggerated.

The earliest data on autism rates is severely biased and greatly underestimates the prevalence of autism. In fact, all intertemporal comparisons of autism rates are biased for a simple but important reason: the definition of autism has changed!

Starting with Kanner’s criteria, autism diagnoses were exceptionally rare with good reason. The rarity of a diagnosis using Kanner’s criteria has to do with the peculiar conditions the criteria demanded to provide a diagnosis. For example, a child had to show both social aloofness and elaborate repetitive routines. The former symptom is usually only observed in the profoundly mentally retarded—children with IQs that tend to be below 35—and the latter seemingly requires substantial cognitive ability.

After the DSM-III became available, clinicians started diagnosing more and more children with the condition: an autism diagnosis was a shiny new tool in the psychiatric arsenal. To qualify for a diagnosis, a child had to be judged as having all of six different criteria, including “Onset before 30 months of age”, “pervasive lack of responsiveness to other people”, “gross deficits in language development”, “if speech is present, peculiar speech patterns such as immediate and delayed echolalia, metaphorical language, pronominal reversal”, “bizarre responses to various aspects of the environment”, and “absence of delusions, hallucinations, loosening of associations, and incoherence as in schizophrenia.”

These criteria were replaced fourteen years later when the DSM-IV came out. With the new criteria, it became far easier to diagnose someone with autism. Now, instead of having to meet six high bars, one only had to meet six of sixteen lower bars. Instead of having to demonstrate a “pervasive lack of responsiveness to other people”, the clinician using the DSM-IV only had to note a “lack of spontaneous seeking to share enjoyment, interests, or achievements with other people”; instead of “gross deficits in language development”, children could merely show difficulty “sustain[ing] a conversation” or a “lack of varied spontaneous make-believe play or social imitative play”. The clinician’s judgment of what’s “bizarre” was even replaced with the far more lax criteria of children showing a “persistent preoccupation with parts of objects”. What’s worse, some of the DSM-IV criteria were practically duplicates of one another, such as “failure to develop peer relationships appropriate to developmental level” and “lack of social or emotional reciprocity”, thus making it even easier to provide someone with an autism diagnosis.

The two diagnoses contained in the DSM-III (infantile autism and childhood onset pervasive developmental disorder) had also multiplied into five. The new diagnoses included the now-familiar Asperger’s and the less familiar Pervasive Developmental Disorder Not Otherwise Specified, the latter of which first appeared in the 1987 DSM-III-R, and was defined by non-severe, subthreshold symptoms. The mild variants of autism introduced between the DSM-III and the DSM-IV quickly came to comprise the majority of autism diagnoses and continue to do so.

Non-severe autism that was easier to get kids diagnosed paired with increasing efforts to identify kids as having autism. Around 2000, the practice of co-diagnosing autism with other conditions like Down’s, Tourette’s, intellectual disability, and cerebral palsy took root in the psychiatric profession. At the same time, widespread recognition that autism could exist at every level of intelligence also gained traction and efforts to identify the condition in younger children kicked off.

All of this happened in an environment that demanded the identification of ever more people with all sorts of disabilities. With the passage of the Rehabilitation Act of 1973, Congress required schools to make special accommodations for disabled students. Two years later, Congress passed the Education for All Handicapped Children Act (known as the Individuals with Disabilities Education Act or “IDEA” from 1990 onwards), leading to the “Child Find” mandate, whereby schools were required to actively identify, locate, and evaluate all children with disabilities, regardless of severity. With the Child Find mandate in place, each disabled child found led to additional funds for schools, incentivizing them to classify ever more children as having disabilities.

Given the known loosening of diagnostic criteria, the efforts to identify people who fit the new criteria, novel incentives to diagnose and over-diagnose, the advent of public awareness campaigns and consequently growing public awareness of autism and other disabilities, the question of “If” autism rates have really increased should be thought of as at least unsettled, and more likely the default hypothesis should be that changes in autism diagnosis rates reflect changes in diagnostics, laws, and culture rather than real growth in the prevalence of autism in our society.

One of the most prominent arguments that there is a “real” epidemic of autism comes from the a misinterpretation of data by the University of California Medical Investigation of Neurodevelopmental Disorders (M.I.N.D.) Institute.

In 1999, the CDDS reported that autism had increased by almost 273% between 1987 and 1998. The scale of this increase was enough to throw California’s legislature into high alert and they commissioned M.I.N.D. to ascertain the causes of the increase. In 2002, M.I.N.D. reported there was “no evidence that a loosening in the diagnostic criteria has contributed to the increase number of autism clients served by [CDDS] Regional Centers”. The basis of this conclusion was M.I.N.D.’s reasoning that, since an autistic cohort born between 1983 and 1985 (initially diagnosed with the more restrictive DSM-III) and an autistic cohort born between 1993 and 1995 (initially diagnosed with the less restrictive DSM-IV) were both assessed with the DSM-IV and found to still be diagnosed as autistic, then the “observed increase in autism cases cannot be explained by a loosening in the criteria used to make the diagnosis”.

If that conclusion seems illogical given the procedure, that’s because it is. If it does not seem illogical to you, then I’ll borrow a well-known explanation of the issue to help you understand that it is.

To understand the fallacy of the conclusion, consider the following analogy, based on male height and graphically illustrated [below].

Suppose the criterion for “tall” was 74.5 in. and taller in the mid-1980s, but the criterion was loosened to 72 in. and taller in the mid-1990s. A diagnostic instrument based on the looser, more recent criterion of 72 in. would identify males who met the 74.5 in criterion. Although a perfectly reliable diagnostic instrument based on a looser criterion would identify 100% of the individuals who meet the looser criterion along with 100% of the individuals who meet the more restricted criterion, a highly reliable instrument might identify about 90% of each group; this is the percentage of each cohort in the California study who met the more recent autism criteria.

Most crucially, broadening the criterion will result in a dramatic increase in diagnosed cases. For instance, census data allow us to estimate that 2,778 males in McClennan County, Texas would be called tall by the more restricted 74.5-in. criterion, and 10,360 males would be called tall by the broader 72-in. criterion; if these two criteria had been applied a decade apart, a 273% increase in the number of males called tall would have emerged—without any real increase in Texans’ height. In the same way, the 273% increase from 2,778 versus 10,360 California children who received services for “autism” in 1987 versus 1998 could well be a function of broadened criteria.

What these authors argued for is no mere theoretical possibility. An empirical example where we know this happened comes from studies that have applied Kanner’s original criteria to more modern samples of autistic children. Children who qualified under Kanner’s criteria were also classified as autistic by the criteria in more recent versions of the DSM, but only about 40% of the children identified as autistic by DSM criteria met Kanner’s more stringent criteria for a diagnosis. Similarly, when autism diagnoses became somewhat more difficult to obtain with the release of the DSM-5, researchers observed that about a fifth of the children the CDC’s Autism and Developmental Disabilities Monitoring (ADDM) Network classified as autistic under DSM-IV-TR criteria did not meet the DSM-5’s criteria for diagnosis. Furthermore, the more severe a child’s DSM-IV-TR diagnosis in terms of the number of criteria they met, the more the two standards agreed. This finding also held up in Australia, where it seems that the introduction of the DSM-5 curbed the number of new diagnoses for a time.

Other details of the California study directly suggest that rising rates were driven by broader diagnostic criteria, practices, or both. Namely, rates of intellectual impairment were thirty-four percentage points lower in the more recent cohort—61% earlier and 27% later, a difference equivalent to just over 0.89 d with equivalent population variances and normality in the distribution of autism spectrum disorder (ASD) levels—and on two out of three dimensions used for autism diagnosis, the later cohort was significantly less symptomatic than the earlier cohort. These findings indicate that the California study found a diagnosis of autism had come to represent something significantly less impactful for people’s lives in a short period of time, but they nevertheless concluded that diagnosis had not broadened. How? By saying that the difference was not “clinically significant”. To return to the height example from earlier:

Comparing two cohorts of males in McClennan County diagnosed according to our more restricted (74.5-in.) versus our broader (72-in.) criterion would probably result in a statistically significant difference between the two cohorts’ average height—but the difference would be just about an inch (i.e., most likely not a clinically significant difference).

Likely the second most prominent argument for a “real” autism epidemic is that the scale of the increase in autism diagnoses is so large it must be real. This belief does not follow from its premises, and frankly, a large increase is biologically implausible whereas a diagnostic change is something we know for a fact can cause major changes in diagnosis numbers. But I digress.

One of the most common sources for the claim of a massive short-term increase is the data schools collect to report to the Department of Education for IDEA compliance. To quote the Autism Society of America: “Figures from the most recent U.S. Department of Education’s 2002 Report to Congress on IDEA reveal that the number of students with autism in America’s schools jumped an alarming 1,354% in the eight-year period from the school year 1991-92 to 2000-01.”

That might be a lot, but not noted in that message: Autism wasn’t an IDEA report category before 1991-92, and in that year, it was entirely optional for school administrators to use. Every newly-introduced reporting category starts off similarly, being underused compared to the level it will be used at, and likely underused compared to the level it should be used at.

Morton Gernsbacher has provided other examples of IDEA categories whose usage went up by a large amount for similar reasons as evidence for this proposition. As he noted, from 1991-92 to 2000-01, “traumatic brain injuries” grew by 5,059%, and by 2000-01, the category “developmental delay” had grown 663% from when it was introduced in 1997-98. These are effects that are not driven by any environmental cause or shift in genetics; there was no mass outbreak of people being smacked across the head that had coincidentally started in the year it became possible to report that sort of thing. These changes are just about increased use of a reporting category after it becomes available—a pure shift in reporting. This is common; Gernsbacher:

After the initial year, the number of children reported under the IDEA category of autism has increased by approximately 23% annually. Why the continuing annual increase? As is the case with new options in the marketplace, like cellular phones and high-speed Internet, new reporting categories in the annual child count are not capitalized upon instantaneously; they require incrementally magnified awareness and augmentation or reallocation of resources.

As an example of this, note Massachusetts’ IDEA reporting and how it’s changed over time. From 1992 to 2001, they had the country’s lowest rate, at only about 0.4-0.5 per 1,000 children. But in 2002, the rate they reported went up by 400% in a single year as a result of the choice to start actually counting the number of students with disabilities rather than attempting to calculate their number from a ratio based on the proportion of students belonging to each disability category in 1992.

Massachusetts was responsible and sensible and they did not claim that the massive one-year increase in autism reports meant autism had actually become far more common in a single year. Instead, they explained to Congress that the numbers would keep going up by large amounts for several years as “districts better understand how to submit their data at the student level” and “all districts comply completely with the new reporting methods.” This same sort of thing has happened all across the country.

Yet another prominent argument that there’s a “real” autism epidemic is that age-resolved snapshots of autism rates and constant-age tracking estimates substantially agree.

Age-resolved snapshots are cross-sectional profiles of the prevalence of autism by age. Constant-age tracking estimates are time series prevalences for successive cohorts all measured at the same age. The idea is that if the slopes are the same, then the rise must be real, because the reference age used in the constant-age tracking estimates will be one that’s a good proxy for cumulative incidence. Age-resolved snapshots are intended to better isolate birth-cohort risk because they dampen period diagnostic noise, whereas constant-age tracking gives the “how many people of a certain age need services” answer, at least in theory.

But both slopes are affected by common forces. Consider it like this:

If P_{true} is actually constant, any systematic increase in the detection function will make both the age-resolved snapshot and the constant-age tracking series rise. There are many things that will affect both slopes, and likely to similar degrees.

Earlier and more aggressive case-finding. This introduces lead-time bias and occurs because of factors like increasingly universal developmental screening, parent awareness campaigns, routine M-CHAT checks at well-child visits, and so on.

• Snapshot: In one calendar year, younger children (born more recently) already carry larger fractions of their lifetime diagnoses, so the prevalence‑by‑age line tilts upward toward the younger ages.

• Tracking: Each ADDM cycle observes eight‑year‑olds whose diagnostic “clock” started earlier, so the fraction identified by age 8 grows steadily over successive birth cohorts.

• Evidence: It’s recognized that measures “have been introduced to reduce the age at autism spectrum disorder diagnosis” not just in the U.S., but globally. Additionally, lowering the typical age at diagnosis further is frequently requested by researchers and advocates, for the sake of autistic children who might do better if they receive care at a younger age. Many researchers have recognized that this makes a contribution to autism rates. For births in California in 1990 versus 2001, for example, 12% of the rise in autism diagnosis rates was explicable by a falling age at diagnosis alone. In that same study, another 56% of the period’s rise was due to the inclusion of milder autism cases. Earlier age at diagnosis has also contributed

Broader or looser diagnostic criteria. The DSM-IV folded Asperger disorder into the spectrum and the DSM-5 kept autism as a category but relaxed specifiers for language delay, co-occurrence, and IQ, making additional borderline cases eligible for diagnosis. Furthermore, school and insurance rules for diagnosis, check-ups, and so on often mirror DSM changes within a few years, providing means for a smooth increase in diagnoses with time, without a change in true prevalence.

• Result: Every cohort diagnosed under the newer rules—whether you view them at a specific age in tracking data or at any age in a snapshot—captures additional mild cases that would have been below threshold earlier.

• Evidence: The CDC’s ADDM data shows that the proportion of autism cases without intellectual disability has been climbing for decades, even including recent waves. Between 2000 and 2016, 84.6% of the increase in the ADDM was due to non-profound autism, with the remainder due to the profound variety. Within both categories, diagnoses became less symptomatic; thus, autism in general has become less symptomatic. This indicates criteria creep rather than an etiologic surge, as rising autism diagnoses with falling severity means some part of the diagnosis numbers going up reflects autism becoming a less alarming thing to be diagnosed with.

Improved surveillance infrastructure. Electronic health records, state autism registries, and linkage with early‑intervention databases systematically capture cases that would have been missed two decades ago.

• Thus, each new surveillance cycle benefits from better data feeds—boosting the constant-age estimate.

• Within a single year, the platforms tend to have the most complete data on recent births (who have been in the record-linked systems since birth), nudging the snapshot upward for younger ages.

Diagnostic substitution and category migration. Schools and clinicians may relabel children who would once have received “speech delay”, “learning disability”, or “intellectual disability” as autism because the label carries better services or clearer guidance. Parents have even been noted to push for their child carrying a diagnosis like “intellectually disabled” to be diagnosed with autism instead, because it provides additional funds and services.

The upward tracking slope comes from each successive cohort being subjected to earlier, broader, and arguably more complete ascertainment by a given age. The upward tilt in the snapshot comes from those same forces acting within the single surveillance year, giving younger children a head start on accumulating diagnoses.

Acknowledging all this, the argument from similarity in slopes becomes fallacious and presumptuous. Additional evidence for this point comes from the way the results are presented. The slopes are allowed to differ radically—regularly by amounts like 20 to 40%!—and they’re still declared to be so similar as to support a “real” autism epidemic, when the difference indicates variability that’s so great it should cast doubt on any such inference.

A novel argument that autism has “really” increased is to be incredulous that researchers might have missed cases by modern criteria when they went out and sought cases with old criteria. This sounds absurd, but it was a view voiced in the recent autism press conference at the HHS. I’ll explain.

Between 1984 and 1985, researchers in North Dakota contacted “all relevant health and service providers” in the state and asked them “to provide names and records of all patients who had autistic symptoms”. The identified patients were between the ages of two and eighteen, and after getting in touch, they were provided with a comprehensive evaluation. During that evaluation, it was observed that twenty-one met DSM-III criteria for autism, two met the criteria for childhood onset pervasive developmental disorder (COPDD), and thirty-six were diagnosed with atypical pervasive developmental disorder. Thus, the prevalence for North Dakota was estimated at 1.16 per 10,000 for autism, 0.11 per 10,000 for COPDD, and 1.99 per 10,000 for APDD, for an overall pervasive developmental disorder rate of 3.26 per 10,000.

The way this study was understood by RFK Jr. was extremely differently from its reality. His remarks were as follows:

In 1987, it was another exhaustive study, a peer-reviewed study in North Dakota. Set out to county every child in the state with a pervasive developmental disorder including autism. That study meticulously combed through every record, every diagnosis, and even conducted in-person assessments of the entire population of 180,000 children under 18. The autism rate they found was 3.3 per 10,000. That’s in line with the 1 in 10,000 that was found in Wisconsin 17 years earlier.

For context, today the last number of 1 in 36 is 83 times higher. In 1987, out of every one million kids, 330 were diagnosed with autism. Today there are 27,777 for every million.

If you accept the epidemic denier’s narrative, you have to believe that researchers in North Dakota missed 98.8% of the children with autism. Thousands of profoundly disabled children were somehow invisible to doctors, teachers, parents, and even their own study. The same researchers who followed the original cohort for 12 years double-checked their number. They went back in 2000 and found that they had missed exactly one child.

So doctors and therapists in the past were not stupid. They weren’t missing all these cases. The epidemic is real.

Practically every detail RFK Jr. provided during this section was incorrect or a misunderstanding. My suspicion is that someone on his team wrote these remarks for him and he believed what he was told. That person must have either been ignorant or severely biased, because they cannot have written this given what the study actually was. RFK Jr. should fire them for misleading him into spouting misinformation in an official statement of the HHS.

For one, the study was not as exhaustive as RFK Jr. was misled into believing it was. The researchers would have liked to detect every child in the state who had the conditions they were screening for, but they relied on recorded cases in the few years after the introduction of the DSM-III, before widespread recognition of autism. Moreover, the authors only contacted providers and asked them to refer patients to them for interviews, and there was no verification of the quality of those reports. Without the force of law behind them, there is no reason to suspect each provider would have provided all of their cases. Given the time the study took place, they certainly did not have adequate records, either, for financial and diagnostic reasons I’ve already gone into above. Contrary to what RFK Jr. was told, the researchers did not personally go out to do interviews with every child in the state either.

For two, the original study did not estimate the autism rate as 3.3 per 10,000. That’s what they estimated the overall pervasive developmental disorder rate at. This may seem to favor RFK Jr.’s point even more, but why this is relevant is unclear since the practices and funding for diagnosis and awareness of autism then differ radically from today.

For three, since we know their original estimate was an under-estimate, it’s not clear what value is supposed to be provided by mentioning the follow-ups. Well, maybe. The follow-ups do provide some value, but not for reasons stated by RFK Jr. In the follow-up, the researchers got in touch with just fifty-two of the original fifty-nine subjects. Ten declined to participate, leaving data for forty-two to work with. The researchers who did this follow-up did not, as RFK Jr. stated, find that they missed a child, because they did not do any novel discovery. The only time ‘just one child’ came up in the study, it was that one child had died between ascertainment and follow-up. The other forty-one whose data were revisited were found to generally show lower severity compared to their earlier assessment, with a 20% reduction using the criteria of the DSM-III-R, and a 23% reduction using the DSM-IV. Also: “Global Assessment of Functioning improved 19%, and the average number of comorbidities decreased 45%. Thirty-seven percent of patients improved in all four measures, whereas only 5% improved in only one measure.”

What value this data might provide to RFK Jr. is that it might reinforce his case about how dramatic the incline in autism identification has been, because it naïvely seems likely that profound cases have increased even more dramatically given this baseline. But the reality is that severe cases were seriously underidentified, so it cannot really do that. We know this data is being misunderstood for many reasons.

• North Dakota still had asylums, and as the autism researcher and UCLA professor Edward Ritvo noted in his newsletter “No Epidemic of Autism”, these were important at the time. North Dakota had only just begun shutting down its asylums as it entered the 1980s, and the common practice then was to “warehouse” autistic people who had been “improperly diagnosed and housed.” This accounts for many of the most severe cases being missing from the sample and much of the rest of the country in this period.

• Co-diagnosis of autism and other disorders was not common until many years after the period 1984-85 in which the children were ascertained. If the children had a more severe disorder or—in many cases—even a less severe one that excludes an autism diagnosis, they would likely have been diagnosed with that alone rather than with infantile autism.

• The age-range for the sample prohibits credible diagnosis for most cases. To be referred into the sample, the diagnosis had to have been logged before the researchers contacted the provider in 1984-85, and the DSM-III had only just come out in 1980. Furthermore, the DSM-III criteria used required onset before thirty months of age, and the style of retrospective diagnosis required was and remains rare and impractical. Due to the diagnostic criteria in use in the study, most ages of child who could have theoretically been in the sample had to be seriously underdiagnosed.

• Dakota is a state that is very White relative to the rest of the country, and the whole country used to be considerably more White than it is now. Recently, the increase in autism diagnoses has been greater for non-White persons. This suggests underdiagnosis of non-Whites is something that used to be quite prominent, but more relevantly, it speaks to the need to demographically adjust future estimates to make them comparable to ones from the past. Even more importantly, it speaks to issues with the CDC ADDM’s—RFK Jr.’s primary reference—identification of the profoundly autistic. Non-Whites tend to have lower IQs than Whites, and a profound diagnosis can occur on the basis of an IQ score equal to or below 50 alone. The increase in non-White diagnoses that likely stems from better service provisioning and reduced rates of neglect for diagnosis and other psychiatric services in their communities has thus elevated the rate of profound diagnoses without adaptive behavior deficits—profound diagnoses in name alone! The same thing has happened across class lines, as autism has gone from a wealthy family diagnosis to one that’s becoming more common in poor families.

• Diagnostic substitution has resulted in increases in the number of profound cases on the basis of intellectual disability and shifting people with serious non-autism disorders into being classified as autistic. It is not that profound cases were missed per se, it is that definitions have indisputably radically changed.

• People have started to lie to receive diagnoses. People seek out diagnoses, even as adults, in order to acquire greater levels of service for their children, and to receive additional welfare benefits for themselves. As the popularity of autism has increased via social media, more people have begun to self-diagnose and to seek out professional diagnoses as well.

• The 1-in-36 and 1-in-31 figures do not refer to severe cases, they refer to all cases, increasingly many of which are non-profound. Talking about this change as if it represents an increase in profound autism alone that could not have possibly been missed, or as if symptom levels within the profound and non-profound categories haven’t fallen, or even talking about it as if it’s something that needs to be diagnosed for those affected to live a normal life is profoundly in error.

We can be certain that autism rates have gone up for artefactual reasons—diagnosis, changing awareness and incentives, etc. rather than real increases in the number of people with autism—by exploiting policy changes. For example, above, I mentioned the Massachusetts saw autism reports increase 400% in one year due to a change in school reporting. Similarly, when states implement a policy of rewarding school districts for diagnoses, diagnoses immediately rise by about 25%. The underlying prevalence doesn’t shift, the incentives to diagnose do, simple as that.

Something similar happens when autism insurance mandates are put into effect. When insurers are legally compelled to provide treatments for children with autism, the number of people diagnosed with autism increases right away.

One study looked at mandates that affected children covered by UnitedHealthcare, Aetna, and Humana between 2008 and 2012. States that implemented the mandates initially had a diagnosis prevalence of 1.8 per 1,000, but after a year of the mandates, that crept up by 0.17, then by 0.27 in the second year, and by 0.29 in the third. The effect of a mandate in this period was a roughly 16% increase in the number diagnosed over trend.

This result was entirely predicted by the insurers and is a major reason why they attempted to prevent the mandates from being enacted. Supporting the arguments of insurers about run-away cost growth, a related finding is that when these mandates are placed on them, families shift from public to privately funded insurance and people shift towards using considerably more services, but the evidence on spending increases is not unambiguous, and the amount of the increase varies considerably by initial spending levels. Age caps on mandates also help to limit spending.

Meta-analytically, the meaning of autism has drifted in terms of the things it correlates with. As noted above, we know this is true for demographic correlates like race and class, and it’s becoming ever more true for sex as well. But some correlates are more biologically proximal, and those too have drifted. For example, emotion recognition and theory of mind deficits? Down in more recent cohorts. Cognitive flexibility and planning problems? Diminished. P3b amplitudes and brain volumes? Not as exaggerated in autism any longer!

All of this means that autism is becoming like what we might’ve once known autism to be. Where individuals with autism used to be more distinct from those without, they’re increasingly “normal”. Compare this to the situation with schizophrenia, which has a far more (keyword) pathognomonic presentation: time goes on and schizophrenia’s correlates are maintained.

An under-discussed source of autism diagnoses is that providers are self-interested. Autism is a business, no matter how bad that sounds to say. The place providers do their diagnosing indexes characteristics of the providers and of the quality of the diagnosis, the cases, etc. as well. With that said, we know that, in some times and places, a large part of the increase in diagnoses can be chalked up to increased use of outpatient diagnosis. In this Danish example, about 40% of the increase in reported autism prevalence could be chalked up to the inclusion of outpatient contacts alone. A further third could be attributed to diagnostic change alone, and together, those factors explained about 60% of the increase.

One of the more influential settings affecting autism diagnoses is probably going to be proximity to autistic people. People learn socially, and many of them have learned about their own autism from the experience of encountering their autistic friends or seeing autism mentioned online. Many parents have learned that their kids might have autism from seeing their kids’ friends. A brilliant study using data from California provided perhaps the best evidence for this sort of effect.

These authors had access to California’s Birth Statistical Master Files and they used that data alongside the CDDS’ autism diagnosis data to track thousands of people who had autism and their siblings longitudinally, even through residential moves. With this data in hand, it becomes possible to test whether living physically near someone diagnosed with autism increases the odds of a child being themselves diagnosed. It does. The probability of being diagnosed significantly increases the closer a family is to another child diagnosed with autism.

After demonstrating this, the authors showed that this finding was not driven by underlying time trends, whereby, due to the rising prevalence of autism, “proximity to the nearest child with autism has increased over time simply because there are more children with autism.” They then ruled out population density by refitting the model with a median split, finding that the effect held up. This is pretty critical, since it suggests toxicant-based explanations have to affect both urban and rural areas. Their fixed-effects model also rules out non-time-varying toxicant sources.

The proximity effect on diagnosis also seems to be qualified by the type of autism. Namely, it leads to a larger increase in autism of the moderate-to-high functioning variety than of the low-functioning variety. The source of the referral for diagnosis also seems to be correlated based on proximity. Meaning, if the nearest child diagnosed with autism got referred by, say, a physician or hospital, then it’s more likely that the child now being referred was referred by the same source. With this in mind, the authors’ found suggestive evidence that the culprit—or savior?—could be school districts, as the effect only tended to show up when children were close but shared the same school district. Unfortunately, the samples that were close but in different districts were small, so that finding has to be viewed tentatively. But what doesn’t is the effect of moving.

Leveraging this data, we can compute a causally-informed estimate of the population-attributable fraction (PAF) of autism cases due to proximity effects. In this study, 40% of people were exposed in the sense of having an autistic child living within 500 meters of their home, and the effect was considerable, leading us to a PAF of 16%. In other words, 16% of autism cases can be attributed to living near other children with autism, for whatever set of reasons, from sharing a school district on down to living near the same hospitals or even an effect of cultural-geographic clustering. Additionally, bringing the quality of Medi-Cal coverage up—or lowering it down?—to the level of other insurance plans in this period would have increased diagnoses by another 13% due to the underdiagnosis effects of coverage for the poor or the overdiagnosis effects of other plans. The jury is out on how to interpret that effect.

Everything I’ve talked about so far is quite beside the point. A single piece of evidence indicates that there is no real epidemic of autism. As remarked in a review in a 2020 Nature Reviews Disease Primers article:

No significant evidence is available supporting that autism is rarer in older people, which provides further evidence against the suggestion that autism is increasing in prevalence over time.

The most major misconception in attempts to diagnose the causes and reality of the “autism epidemic” is this: that it’s much more common among kids than adults.

If you’ve read this far, you’ll no doubt have realized that the difference in diagnosis rates between children and adults stems from the former being exposed to a far greater level of novel identification criteria and practices, because autism is supposed to be diagnosed early. Active case-finding expeditions virtually all concern children and children alone, with adults given short shrift and, at best, being classified as autistic based on their rare opportunities to obtain a diagnosis earlier in life, or through self-selection.

When researchers go out of their way to screen adults based on ICD-10/DSM-IV criteria, they observe an autism rate of just over 1%. When the same methods are used by researchers who go out of their way to screen children, they observe an autism rate of just over 1%. If you were to use more or less expansive criteria, you would almost-certainly still find the same rates among the young and the old, albeit with a bias for very old ages due to survival deficits associated with autism and difficulties reliably reaching old people. In truth, the concordance in rates across ages is a lucky coincidence of the expansion of autism, because it is extremely unlikely that many autistic individuals would survive to old age were they classified under Kanner’s criteria as opposed to the more modern, looser definitions in use today.

Some of the most ‘slam-dunk’ evidence comes from places like Sweden, Denmark, Finland, and Norway, which have population registers. These registers provide them with lots of information about people that could be used for understanding if there’s been diagnostic runaway without a similar increase in autism symptom levels. As it turns out, using Swedish data, we can see that there was no significant time-trend for measured symptom scores (p = 0.85), but there was a highly-significant and highly-positive time trend for autism diagnoses (p < 0.001).

Diagnoses are simply not a workable criteria for understanding prevalence without explicit and credible attempts to go out and find everyone who fits different diagnoses. The autism prevalence for Swedish kids based on symptom score measurements? A consistent 0.95% that didn’t significantly and systematically go up or down! But the diagnoses nevertheless just kept going, more than doubling between 1993 and 2002. But this doubling still wasn’t up to snuff, because the diagnosis numbers never actually met the known community prevalence of autism, and, as you now know, that prevalence drifts based on how strictly or loosely you define autism.

So remember this:

There is not, and has never been, any credible evidence for a “real” epidemic of autism or for any of its proposed environmental causes. Answering the question of if there is a “real” autism epidemic has provided a complete answer to the question of why there might be.

People have proposed tons of explanations for rising autism rates. These have almost all been careless attempts, since they have reasoned from the existence of a “real” epidemic, and thus about something to be explained in terms of increasing exposures. People have tried to argue for all sorts of out-there causes that imply we’re mismanaging our health systems, giving out drugs too willy-nilly, getting infected en masse and just generally being imbeciles about toxins in ways that are counter-indicated by real data.

I’ve previously mentioned now-falsified examples of proposed causes like acetaminophen use during pregnancy, epidural use at birth, and infection during pregnancy. More recently, people have proposed ultrasound use during pregnancy as something to fear (it’s not supported), and much more popularly, lots of people believe the rise in diagnoses—or, per them, the real rise in autism—is driven by vaccination.

The theory that autism is driven by vaccines makes no biological sense, fails to comport with the well-known extreme heritability of autism, and predicts nonsensical things given the expansion in diagnosis rates over time. For example, if vaccination has driven the expansions in rates, then it should have highly nonlinear effects: when vaccines are added to the childhood vaccination schedule, the overwhelming majority of people get them immediately, so why does the autism diagnosis rate keep climbing steadily? Without an accompanying rise in autism rates in the year vaccines are introduced, or in way that’s consistent with actual vaccine uptake, there’s no statistical plausibility to the autism-vaccine link. That there simply is no statistical link in large cohorts is therefore unsurprising. That there is definitely no causal link, as discerned with sibling studies, is doubly unsurprising.

But despite its origins in fraud, and its repeated, decisive refutations, the vaccine-autism link refuses to die in a large part of the public’s imagination. Some part of this is because people want it to be true. A not-insignificant number of parents even feel that there’s a lot of truth in it, because they’ve fallen for the old fallacy that, because something happened after something else, they must’ve isolated the cause.

In many cases, this happens because parents are looking for an explanation for why they just started to notice their child was autistic, and they’re happy to chalk it up to vaccines and well visits. People go so far in looking for support that they flat-out misread studies on the alleged harms of vaccination and they fall into theory rabbit holes, where the link is supported because they believe in some mechanism behind it. Mercury content, thimerosal, and vaccination in general have all had mechanisms proposed to explain their association with autism, and they’ve all failed, probably because there’s no association between vaccination and autism in the first place.

Plainly, if you want to avoid falling into the same traps, think causally and eschew consideration of “evidence” presented by people who cannot reason in terms of standard evidence like trials, large cohort studies, the rapidly expanding number of causal cross-sectional studies (i.e., DiDs, RDDs, IVs, etc.), and realistic animal evidence. When a hypothesis is supported by conspiracy theorizing and cross-sectional associations that don’t replicate or can be seen to be due to confounding, then drop it and the person proposing them. The person pushing the idea likely wants to convince you more than they want to be right, and you should let them know you know that by ignoring them. Demand good evidence, or you might find yourself fooled into thinking there’s been a “real” autism epidemic and it’s down to someone’s cause célèbre.

Though no one has ever provided quality evidence for an environmental cause of autism, there may still be at least two things that could be causing slightly—but not massively—rising rates of autism. These are

1. Older parents. The age at first marriage and the age at first childbirth are increasing, and this seems to have a causal effect on children’s risk of autism. Could this be environmental, driven by the womb? Maybe, but since the paternal age effect is larger, it seems like the age-related accumulation of mutations in sperm will be more likely to take the blame.

2. Our increasingly awesome healthcare systems. People who would never have survived childbirth or early childhood in previous generations are surviving nowadays at historically unprecedented rates. The examples range from those with major congenital defects to people with minor, but still risk-disposing issues, like autism.

We keep young people alive today better than we ever have, and that means that people who have historically been at-risk of an early death are being kept alive well beyond where they would have been historically. This might not mean much for the people who receive mild autism diagnoses, but it can mean a lot for those who receive profound diagnoses, since they are at the greatest risk of an early demise.

I think that, though this increases the societal burden of the incapable to some degree, it is still to be lauded, because their lives still have value. The families broken up by having to raise these kids may be worse off, and there may be some calculus through which our medical advances are bad for society, but it’ll take a lot to move me from the position of valuing the lives of even the profoundly disabled.

Here is where I voice my pleasurable agreement with RFK Jr.’s recent speech about autism. Near the end, he noted that the cost of medical care for autism are set to balloon over $1 trillion annually by 2035. This estimate seems credible and I take no issue with it. In fact, if a pro-diagnosis radical works their position into the next iteration of the DSM or the AMA succeeds in destroying prior authorization, the costs will probably be far greater.

This is why I have tried to avoid or qualify “under-” and “overdiagnosis” in addition to avoiding calling diagnostic change “good”. Change in diagnostic criteria and clinical practice can be “good” with respect to equating rates across age and demographic groups when diagnosis disparities are driven by bias, and it can be good for identifying people who genuinely need care. But we have to face this trio of facts:

1. We diagnose way too many people with autism that’s so mild it doesn’t need treatment.

2. So much spending on care for autistic people is wasteful. Someone with a mild case of Asperger’s does not need to hit a mandated spending cap for their condition, and someone with mental retardation who convinced a clinician to give them a more favorable diagnosis should not be receiving treatment intended for autistics and found to be useless for the merely mentally retarded.

3. We have pathologized normal-range behavior.

The first two points speak for themselves. The latter is one that must be reiterated.

So many aspects of normal child—and particularly male—behavior have been turned into issues that demand psychiatric attention, and this has been done intentionally. I am actually awestruck by the fact that RFK Jr. has not fired the head of the Child Psychiatry Branch of the National Institutes of Mental Health, one Judith Rapoport, who was once quoted as saying:

I’ll call a kid a zebra if it will get him the educational services I think he needs.

To most people, admitting that you’ll lie to get someone diagnosed with a condition because you see a need should be a cause for alarm. Would you trust the psychiatrist telling you they’ll lie to get you services? Maybe, because you want those services, but would you like to pay for the same treatment for a million other kids? And if the next year, none of the kids deserve that treatment but they’re getting diagnosed anyway, do you want to foot the bill?

That the diagnostic situation has now evolved to the point where too many people are receiving services they do not need is now so well-evidenced that debate is hard to fathom and cutting the fat is more than warranted. I wish RFK Jr. all the luck in the world in these efforts, and if he wants a succinct, well-supported report on how to do the cuts without cutting care for the kids who really need it, he has my number.

We have a moral duty to get this right. I know many of my readers departed from me a few paragraphs ago when I said I value the lives of the severely handicapped, and I expect more will stop reading right now, but what I’m going to write is heartfelt:

I do not like knowing that autistic people are suffering for no reason.

I recently noted an example of this over on Twitter. That example was chelation treatment. Plenty of people followed bad science and poor reasoning—likely in many cases desperately, and understandably—to the point where apparently about 7% of autistic people had been given a harmful, often painful, and—for autistics—difficult to endure treatment to remove heavy metals from their body in the hopes that it would treat their condition. This did not work. All it did was cost a lot of money, a lot of time, a lot of relationships between parents and autistic children, and it hurt people who often wouldn’t know anything about what was going on.

If bad science leads people to avoiding epidurals, a lot of mothers are going to needlessly suffer through birth. If bad science leads people to avoiding acetaminophen, a lot of people are going to live their lives with more pain than they would have otherwise. If bad science leads people to avoiding vaccines, a lot of kids are going to suffer, a lot of kids are going to die, and a lot of parents are going to wonder why they ever believed it was better to have a dead child than an autistic one.

There is no “autism epidemic” and there never was. There are no responsible toxins to remove from the food supply or chemicals to purge from the water or the air. If we cannot get this right, I doubt we’ll be able to do anything else any better.