A recent report by the Information Centre of the UK’s National Health Service says that the autism prevalence for adults is about 1 in 100.
The report, Autism Spectrum Disorders in adults living in households throughout England, states:
Using this recommended threshold score on the ADOS, 1.0%of the adult population had ASD. The rate was higher in men (1.8%) than women (0.2%), which fits with the profile found in childhood population studies.
This isn’t very new news, Anthony Cox and Kev Leitch on this blog have already discussed this.
I would suggest that the first reaction from pretty much everyone (including myself) was wrong.
What should the take-away message be from this?
Simple–there are a lot of unidentified adult autistics who are likely getting little, no, or inappropriate support.
From the report:
“Perhaps most important of all is the finding that adults with ASD are socially disadvantaged, less well educationally qualified, less able intellectually and possibly under-supported by services. Much of this could be alleviated with greater involvement of existing established social, educational, welfare and health care services.
Another point that really sticks out to me is:
There was no indication of any increased use of treatment or services for mental or emotional problems among people with ASD.
We don’t know if this is because they don’t need more services or they just aren’t getting them.
Two more statistics are also worth pointing out: 4.5% of unmarried male adults are identified as ASD. 8% of male adults in social housing are identified as ASD.
That’s huge. Imagine walking through a specific housing complex and 1 in 12 men you see are autistic.
Now that I am done lecturing everyone (including myself) about what I think the important message is from this report, let’s take a look at the report itself. Specifically, let’s consider the complaints that are being levied against it.
The first thing that struck me is that this is a report, not a published study. It is not, to my knowledge, peer reviewed in the same sense as a journal article. It is in the same class as the MMWR reports that the CDC puts out (a government report) that are used by almost everyone to discuss the prevalence of autism. So, if you use 1 in 150 or 1 in 166 prevalence numbers, don’t complain about this UK report being non-peer reviewed.
People are complaining about how the study was conducted. Here is the basic process:
A. Phase one AQ-20 self-completion screen
B. Selection of cases for phase two assessment
C. Phase two ADOS assessment of a subset of cases
D. Weighting to adjust for selection probabilities and non-response.
This was not worded well, since many people assume that the phase one screen only used the AQ-20. As I will discuss below, this is not the case.
The initial screen started by identifying addresses that were residences, and selecting some by random selection. They selected 13,171 possible households for phase-1. Of these, 57% agreed to respond.
9% of sampled addresses were ineligible because they contained no private households, while 4%were addresses of unknown eligibility (see Section 3.2.5). This left an estimated base of 13,171 known eligible or probable eligible households for the phase one interview. The proportion of selected adults who agreed to take part in an initial interview is shown in Figure 3B. At the phase one interview, 57% of those eligible agreed to take part in an interview. This included 50 partial interviews where the respondent completed the service use and CIS-R modules, but did not reach the end of the interview.
These interviews were not just the AQ-20 screen, as noted below:
The phase one interviews were carried out by NatCen interviewers. These included structured assessments and screening instruments for mental disorders, as well as questions about other topics, such as general health, service use, risk factors and demographics. These interviews lasted about 90 minutes on average.
90 minutes per interview, with 7,353 full interviews works out to 5.3 man-years of labor.
This was no small effort.
Many complaints I have seen concentrate on the AQ-20 test. The full AQ test, or Autism Quotient, is a product of the Cambridge Autism Research Centre, and has been studied already. It is usually a 50 question test, but it was shortened to 20 questions (hence AQ20) and adapted for the NHS survey. They shortened the AQ to save time.
Many people have been confused that this short questionnaire was the method used to make the autism diagnoses. People pull questions out and question whether they could be used to diagnose autism. The AQ20 wasn’t used to make the diagnoses. It was only a part of the phase-one, pre-screen, part of the study. The diagnoses were made using the ADOS.
They used the information from phase 1 to select a smaller subgroup for the more intensive phase-2 part of the study. Amongst this smaller group, the researchers chose people they thought were more likely to have psychosis, Asperger syndrome or personality disorder.
7461 respondents provided a productive phase one interview. Of these 58 were proxy respondents and therefore not eligible for the phase two interview (see Figure 3E). A probability of selection was calculated for each respondent based on their answers to the phase one screening questions on psychosis, Asperger syndrome, and personality disorder: as outlined in Section 3.2.6. 5,329 respondents had a probability of selection of greater than zero: 4050 of these also agreed to be recontacted for a phase two interview (76%). After the application of the highest of the four disorder specific sampling fractions, 849 respondents were selected for a phase two interview. Phase two interviews were conducted with 630 of these (74%).
618 of the 630 people who underwent the phase-2 interviews were given the ADOS test. A score of 10 on the ADOS was chosen to indicate a diagnosis of an ASD. Of these, 19 scored above 10 on the test and were diagnosed as having an ASD.
If one took these data to calculate a crude prevalence, a value of 19/618 or about 3% (1 in 32) would be obtained. No one is saying this is an accurate estimate, but it is worth noting for this reason: quite obviously, the phase-1 screen was successful in finding a larger number of autistics than
These data were weighted to estimate a prevalence for the overall study group, including non-responders. They came up with a prevalence of 1% (1.8% for males 0.2% for females).
Here are some of the complaints I have read about this report:
1) They define adult with ages as young as 16.
this is supposed to be a sign that they are trying to fudge the data by including people who are really part of the so-called “epidemic”.
Well, the UK Census defines adults as people age 16 and over. If including 16 and 17 year olds were a problem, one would expect the younger age category to have a much higher prevalence. It doesn’t.
age group 16-44: prevalence 1.1%
age group 45-74: prevalence 0.9%
age group 75+: prevalence 0.8%
(from table 2B)
Another complaint is that there were no ethnic or racial minorities in the group of 19 identified autistics. This is a good check for internal consistency, but it isn’t a valuable check in this case.
The UK has about 90% white population. They tested 618 individuals with the ADOS. If the selection were random in ethnicity, they would have 61 minority participants. A 1% autism rate would lead us to expect 0.6 ethnic minority autistics.
People have complained that the study only shows adults in residential housing. I.e. they didn’t check institutions.
The report is very clear about this (it is even in the title). They note:
The sample for APMS 2007 was designed to be representative of the population living in private households (that is, people not living in communal establishments) in England. People living in institutions are more likely than those living in private households to have ASD, however this group was not covered in the survey reported on here and this should be borne in mind when considering the survey’s account. At the time of the 2001 Census, 2% of the English population aged 16 years or over were resident in a communal establishment.
So, yes, by leaving out those in institutions they didn’t measure the prevalence in the total UK population. That is a limitation of this report. The prevalence in the institutions is very likely to be higher than in residential settings. I.e. they would have found more autistics had they look in institutions and they would have found a higher prevalence overall.
One complaint is that they only identified 19 adult autistics in phase two. This is definitely worth considering as that puts some big error bars on the results. This becomes especially when they try to break the results down by age, gender, or other category.
It has been proposed that one can’t extrapolate from the 19.
But, just for fun, let’s make the assumption that the 19 adult autistics ID’d in phase 2 group (of about 618) are all there would be in the entire group studied (7353).
That would give a prevalence of 25 per 10,000.
This is much higher than people have been claiming the prevalence should be amongst adults. Many people claim the prevalence rate for adults should be about 1 in 10,0000 or 3.3 in 10,000.
To put it simply, those claiming there is an epidemic of autism are off by at least a factor of 7.
Or, to put it another way,
It isn’t a question of whether there are factors such as widening of the criteria or diagnositic substitution/accretion that have caused some of the rise in the autism “rates”. The question is what factors have been in play and how big of an effect did they have.
Another criticism I have seen is that the male:female ratio is quite high, 18:1, as opposed to the 3:1 or 4:1 found in most studies. They found 19 autistics. A 3:1 male:female ratio would lead us to expect 5 females. I’ll let the statisticians tell me if this is significant, but they did find fewer women than I would expect. The report notes repeatedly that the small number of women limits the analysis.
I have been told in the comments on this blog that the study was done cheaply. I wouldn’t be surprised. However, it does represent a significant effort. I am impressed that anyone undertook to do a prevalence study on adults.
This isn’t a definitive or end-all study. Far from it. But it is a reasonable study and a very good start. I hope this is the beginning of a much greater effort to gather more information on adult autistics. It is pretty frightening to think that a very large segment of the autistic population could be undiagnosed and possibly receiving inappropriate supports.
Left Brain Right Brain 
Only one in a hundred? (sigh) I used to think I was one in a million!
The math is straightforward in retrospect, and completely sound, BTW. It did take me a while to figure it out, though.
It’s straightforward primarily because the probabilities of selection are stratified. That is, there’s a finite set of probabilities that are assigned to phase 1 participants. So you can think of it this way: There are groups of autistics. There’s the group that was assigned a probability of 1.0, there’s the one that got 0.65, the one with 0.25, and so forth.
Try solving this, and you’ll completely understand how it works:
Suppose that among the 19 autistics identified with the ADOS, 5 had been assigned a probability of 1.0 and 14 had been assigned a probability of 0.25.
Q: How many autistics from the original group (phase 1) were assigned a probability of 1.0? (That’s trivial.) How many roughly were assigned a probability of 0.25? How many autistics were there in total, approximately?
You miss one of the main points in the study:
“The prevalence of ASD was only 0.4%in those with a V-IQ above 100. People with the
lowest V-IQ score (70-85) were much more likely to have ASD (2.7%of all adults, 4.3%of
men) than those in higher scoring V-IQ groups”.
The extrapolated data shows that 2.7 of all adults were of borderline intelligence with only .4 unaffected by lower than normal IQ scores.
The data shows no evidence that a large segment of adults in England are unrecognized ‘Asperger’ types, ppeople with an ASD and normal or above IQ.
The interpretations of the data is problematic given the inability of ASD researchers to tease out people of less than normal IQ who may have social-communicative problems related to ID but who have enough isolated symptoms to qualify for an ASD diagnosis (as do a high percentage of adult stroke patients who share social-communicative problems that are observed in ID).
Actually RAJ, according to the study, all autistics located are high functioning (all have VIQ of 70 or higher.)
That part of the study is pretty much nonsense, of course. The VIQ estimation method they use is very rudimentary to say the least. It’s based on being able to properly pronounce a list of words. In some cases I believe they also had to listen to pronunciation. Is it surprising autistics would do relatively poorly in this test?
Think about this part:
Why is it OK to avoid estimating the VIQ of non-English speakers using this method? Is it because it would be unfair?
Additionally, I doubt that method is able to estimate IQs below 70 or above 130.
I’m not sure what the practical value of that part of the study is, except to say “you see, autistics are inferior.” It could’ve easily been left out, and readers of the paper would’ve been spared a few wasted neurons.
Joseph, you said –
“Suppose that among the 19 autistics identified with the ADOS, 5 had been assigned a probability of 1.0 and 14 had been assigned a probability of 0.25.”
This is not a valid assumption. We do not know that, in general, there is a relation between and ADOS score and the probabilities arrived at using the results of the phase 2 screen. All we know is that these are the relations that were found in a subset. These relations cannot be used to extrapolate back the original group.
That would be like saying I have a group of 100 people and I want to find all the number of people with red hair. I test these people with an indirect measure that I just made up that I think will guess their hair color. This test is indirect because I am not actually looking at their hair color, just things that I believe to be related to red hair. After giving this test I use the results to select the most promising 10 candidates and then actually look at their hair color. I find that 2 of the 10 have red hair and also had a certain score on my indirect test. Can I then assume that everyone who had that score in the original group had red hair?
The answer is no, you can’t. The reason is you need two pieces of information to make this assumption – how many redheads you found in the subgroup and how accurate your test is. You can get one result from your test – you cannot get both.
You can’t say that I found two in the subgroup but there were 8 in total because this score is related to having red hair and that you know this score is related to having red hair because there were 8 people in the total group with red hair that had this score. This is circular logic.
Am I not making a lot of sense? Is the methodology of the paper still beyond comprehension?
Let’s use that. Suppose there are 50 red-heads and 50 non-red-heads, but this is unknown to us. To ‘screen’ we ask them if they are white. If someone is white, we assign them a probability of 0.4, and if they are not white, we assign them a probability of 0.2. (And yes, we make up these probabilities pretty much as we please.)
Among the red-heads, let’s say 80% are white. Among the non-red-heads, 50% are white. But again, as researchers we are completely unaware of this at this point.
After the screening, we’ll have a group roughly like this: 33 people total, 18 of whom are red-heads, and 15 are non-red-heads. Notice that it’s not 50/50. (If the arithmetic to get these figures is not straightforward, let me know.) Additionally, among the red-heads identified, we check back and see that 16 had been assigned a probability of 0.4, and 2 had been assigned a probability of 0.2.
What’s in the last paragraph is all we know. Presumably we have tool that determines hair color very precisely among the 33 people screened. And we know the probabilities assigned to each person in phase 1.
Since 16 identified red-heads had been assigned a probability of 0.4, we can extrapolate back and determine that about 40 red-heads from stage 1 had been assigned a probability of 0.4 (that’s 16 / 0.4).
Similarly, since 2 had been assigned a probability of 0.2, we can extrapolate back that 10 red-heads from stage 1 had been assigned a probability of 0.2.
In total, we conclude that just about 40 + 10 (50) red-heads must have been in the original group, even though we don’t really know who they are. This is the right answer.
Joseph, the problem with your example is still the assumption that the probabilities that you are using are accurate and represent a real relationship.
The problem with the survey is that the relationship between the autism diagnosis and the screening test have not been validated or tested – and without that the extrapolated results are very weak.
This survey could have been so much stronger if the researchers had identified the 53 or so additional people who they thought would have had autism and given them an ADOS. In that way the results could have been validated and we would have a much stronger, believable result.
No one has made that assumption, but I’m just repeating myself at this point. It’s desirable that there’s a relationship, and you can reasonably expect that there’s some relationship.
An interesting study that proves very little. The sample size is way too small for one. One female and 18 males shows the high variance in this study without doing much statistical analysis.
I did not read the study, but it sounds like they extrapolated using assumptions to get to the 1%. There is no way that they could assert no difference across age groups with only 19 cases.
What this study does do is call into question the rise in ASD over time, but everyone knows that only the most severe cases were diagnosed 10+ years ago. Any HFA or AS would not have been counted. Many were labeled mentally retarded, ADD or ADHD.
No study of any size can assert a negative. The best you can do is say “it can’t be distinguished from ‘no difference’ statistically” which is what occurred in this case.
There was a slight downward trend with age, as discussed in the prior post.
16-45: 1.1%
45-75: 0.9%
75+: 0.8%