Saturday, January 30, 2010

Debunking Michelle Dawson's 1 in 86 adults with autism

Recently, Michelle Dawson wrote a post called 1 in 86: the prevalence of autism among adults where she asserted, among other things, that the prevalence of autism in adults is 1 in 86.  Ms Dawson is basing this estimate on a study published in 2006 -
All 56,946 individuals comprising the targeted population cohort in this study are, as of today, the last day in the decade, 18 years of age or older. They were born between July 1, 1990 and Dec 31, 1991 and they are now all adults. 
Within this cohort, Baird et al. (2006) reported an autistic spectrum prevalence of ~116/10,000. That's 1 in 86, and all these autistics, originally assessed as such when 9 to 14yrs old, are now adults.
Well, she is correct about at least one thing - all of these children in the study were born between 1990 and 1991 and would be now all be 18 or 19 years old. However, I am not at all sure that even if autism is at a certain rate in a specific population 18 year olds that the same rate would apply to adults of all other ages.  The available data points to an increasing rate of autism and that the rate is increasing with each passing year.

But even ignoring that point for a minute, what exactly was found in the study that she is referring to?  Given the fact that the study population was 56,946, did the researchers find 662 children with a diagnosis of autism?

The short answer is no, they didn't.

What they found was 158 children who would fit the broadest possible definition of autism.  Using what would be a normal definition of autism used in studies, they only found 81 children.

So how did they get from from the 158 they identified to claiming that there were over four times that number and what exactly is the difference in the between the two groups?  The answer is complicated and has to do with the differences between actually counting the number of cases and using statistical techniques to estimate a number.  The difference also has to do with what you call autism - are you talking about what could clinical be diagnosed as autism or something who has some symptoms of autism.

Lets start with the basics, this study looked at 56,946 children born in South Thames, UK between July 1, 1990 and December 31, 1991.  When the study started, there were 255 children who had a diagnosis of autism.  If you look at what the historical rate of autism was thought to be when these children were 9 years old, it was about 1 in 250 (40 per 10,000).  The children in this group had slightly higher rate but were generally in line with the 1 in 250 number (45 per 10,000).

From there the study identified all of the children who has a special education label (or the equivalent in the UK) or had an existing diagnosis of autism.  This narrowed the population to 1,770 children.  The authors assumed that all of the children with autism would be in this sub group.  I think this is a safe assumption - any 9 year old who had any diagnosable form of autism would likely need at least some extra help at school.

I am going to stop here for a basic sanity check.  The study is asserting that there are about 662 children with autism in this population.  If you do the math you will see that they are saying that over 35% of the special ed population has autism.  Compare that to the number of existing cases (255) in the population and you will see that the rate of existing cases is about 15%.  If you consider the fact that the comparable percentage in the US for all children served under IDEA around the same time frame was about 17% you can see that the 35% is about double what would be expected.  Although to be fair, I am comparing two different countries so the comparison isn't strictly apple to apple.

Moving along, the study then used the Social Communication Questionnaire (SCQ) to help group the children.  The SCQ is a screening test that is designed to identify children with autism.  The idea is that if a child has a score above a certain level the should be further evaluated for a form of autism.  Out of the initial 1,770 there were 1,035 who returned the SCQ and consented to further evaluation.

Out of this set of 1,035, a "two-way stratified random" sample of 363 children were selected for further evaluation.  In this group of 363 there were 141 children who already had a diagnosis of autism - a little over half of the known cases.

Of the 363 only 255 were actually evaluated due to a variety of factors.  In this population of 255 there were 112 children who had an existing diagnosis of autism, which amounts to about 44% of the known population.  For each of these 255 children an ADOS and ADI-R was performed as well as assessments of IQ, language, and adaptive behaviors.  Information was collected (via the telephone interviews) from the children's teachers about their social, communicative, and other behaviors.  The children's health records were also collected.

Before I go into the specific numbers found, I wanted to talk about the "two-way stratified random" sample of 363 children. The groups (stratification) used a prior autism diagnosis and SCQ score range as the defining criteria for the groups and here lies one of the potential problems.

Remember how I said that the SCQ was a screening tool?  That means above a certain score you have a certain likelihood that the test will accurately identify children with autism and will not identify children who don't have autism.  All of this means that above the cut-off score the SCQ will agree with another, presumable more accurate, test, such as  the ADOS, a certain percentage of the time but below that score we really have no idea how what it says.

So the validity of the screening test is always relative to another, more accurate, test - usually the ADOS or ADI-R.  So the validity of the test isn't really relative to having autism but rather the chance of achieving a score on another test.  As a result, if you don't use the other test, or worse ignore it, then you have no real idea whats the score on the test means.  These may seem like an academic points but, as you will see below, they lead to potential problem.

To give an example, assume that we have a group of 100 children, 20 of whom score an 10 on the SCQ which is under the normal cutoff  We further test 5 of these children using the ADOS and find that 2 of them have autism.  Can we then say that 8 ( 2/ 5 * 20) children out of the 20 should have a score on the ADOS that puts them into the realm of autism?  No, we cannot because the relation of this SCQ score to the ADOS score has not been established.  If we were talking about a score above the cutoff then we would be able to make the relation.

Furthermore, if we gave the same group of children the ADOS but the, instead of giving 2 of them a label of autism, we decided to say that 4 of them did - even though that extra two did not score high enough on the ADOS to indicate autism - can we then say that 16 of the children have autism based on the SCQ scores?  Again the answer is no.

Going back to the 363 children, the stratification groups where based on prior autism diagnosis (yes, no) and range of score on the SCQ ( under 8, 8-14, 15-21, over 21 ) which gives us a total of 8 groups.  The normal cutoff of the SCQ is about 15, so I am not sure what is says when you have a score in the under 8 or 8-14 groups and you certainly can't (accurately) relate that back to other children who scored in those ranges.
Now lets consider the number of children that the study identified as having autism.  Three of the researchers involved in the study looked at all of the available information and arrived at what diagnosis, if any, that the children should have.   I think it is important to note that the researchers did not interview the children themselves but rather relied on the data that was collected by others.

Lets first take a step back, how do you know if a child has a form of autism?

Typically what is done in research is to rely on one of the "gold standard" tests - the ADOS or ADI-R.  Some of the time you will see that the children met DSM-IV criteria and usually that means that one of the researchers evaluated the child themselves and determined that the met the criteria for a form of autism.  The important part here is that either a standardized test is used OR a knowledgeable person examines the child and their history and comes to a conclusion. 

In this study, the DSM IV criteria were not used but rather a related set of criteria called the ICD-10 were used.  These criteria are similar to the DSM IV but label things differently.

Going back to the diagnosis in the subgroups, the researchers identified 53 children who met a strict (narrow) definition of autism.  This strict definition was that the ADOS and ADI-R both indicated that the child had autism.  These would be the most severe children, the ones who should be readily identifiable.  In this group 34 already had a diagnosis of autism and 19 did not.

The next group were a "consensus" group were all the three researchers agreed that it was likely that the children had autism.  This group had a total of 81 children, which included the 53 children from the strict group.  In this group there were 50 children who already had a diagnosis of autism.  For this consensus group it was not required that both of the standardized tests indicated autism and indeed, the ADOS indicated autism in only 64% of these children with a further 25% meeting the cutoff for a sub-threshold form of autism (think PDD-NOS).  The ADI-R showed better agreement with the group.

In the consensus group there were 7 who had no delay in language and no evidence of abnormal development before the age of 3.  I am not sure why these 7 would be labeled as having autism instead of a sub-threshold condition (see below).  The ICD-10 and the DSM IV both say that for a diagnosis of "autism" the symptoms must be present before the age of 3.  Under the DSM IV, these cases would likely be labeled as PDD-NOS and not autism.  I normally wouldn't think much of this, but this set makes up almost 10% of this set.  The researchers suggest that these 7 might have Asperger's but still label them as autism.

The consensus group (probably minus those 7) are likely to be what you would see included in other studies as the "autism" group since that is the group that the ADOS would have labeled as having autism.  This is also the group that the SCQ would have selected if the normal cutoff had been used.

The last group was the "other asd" group and was made up of 77 children, 18 of which had a prior diagnosis.  Six were in this group because of late onset (although how this 6 differ from the 7 above I don't know), 61 met the ICD-10 criteria for atypical autism (PDD-NOS on the DSM IV) because the severity wasn't there, 7 had an incomplete history (lack of early medical records), and 3 met criteria for other ICD-10 disorders.  I don't know why these last three were included in the "other ASD" group because, according to the data provided, they did't have autism.

That brings the total number of children with any possible (however questionable) diagnosable form of autism to 158, of which less than half (68) had an existing diagnosis.  Or to state that another way, 44 of the children who initially had a label of autism had the label revoked and 90 were added.

Now here comes the funny part.  From these numbers the researchers estimated the prevalence of autism in the entire population and came up with the following (all per 10,000) : 24.8 have "strict" autism, 38.9 have "consensus" autism, and 77.2 have some other form for a grand total of  116.1 for all forms of autism.

So, how did the researchers get from the 158 number to a figure that was over four times that amount (662)?  Simple, they used their groupings and, after making some adjustments, extrapolated back to the original population of special education children.  They then assumed that all of the autism cases were in that subset and arrived at what they thought were the totals for the whole population.

Unfortunately, the study does not detail the exact methods that they used to do this extrapolation nor does it give the breakdown of how many of each of the asd groups (narrow, consensus, other asd) fall into each stratification groups and without those figures it is difficult to duplicate their exact numbers.  If you remember, half of the stratification groups would not have a good path back to the population since it really isn't know how a score of below the cutoff would be related to a diagnosis of autism and we can't assume that there is no relation.  Furthermore, I don't know what an SCQ score would mean once you have ignored the ADOS or ADI-R score - which was done for a number of the children.

But here is the rub - no matter how they extrapolated from this subset of 255 back to the entire population, it is at best only a guess.  The underlying principle of doing this sort of analysis is that you have a known (and verified) path from your subset back to the original population.  You can't pick an subset using some non-arbitrary method and then expect to accurately relate it back to the whole.

To give you an example, assume I have a set of 100 men and women.  I select all of the people who have dark hair and find I have 10 men and 15 women.  Can I then say that, based on hair color, there are 40 (10 / (10+15) * 100)  men in the initial set of 100?  The answer is no - you would have to know what percentage of men had dark hair to start out with.
If the researchers they had selected a completely random set of individuals from the initial population of 1,035, they would have then been able to relate it back to the whole as a simple percent because there would have been no relation between having autism and being in the selected group.  Or if they had used the SCQ as a screening test (as it was designed) and then tested all of those who scored over the cut off they would have found all of the "likely" cases of autism - remember, counting actual cases is much more accurate than estimating.

As a result, we have to treat their extrapolation from the subset back to the original population as questionable.  When you have a questionable result, you look to other sources to see if other studies have arrived at the similar results.  Well, if you look at other research done on the same age group / birth year, you will see that the prevalence numbers from this study are significantly higher than those from other studies - somewhere from about 1.5 to 4 times higher than other estimates.

If you further consider other research that was done in a subset of the population by the same researchers, you will see that they estimated, just 2 (age) years before, that the prevalence of all cases of autism in this was half (57.9 per 10,000) of what they are suggesting now.

Remember the percentages of special ed statistics I listed above?  They also suggest that the proposed numbers are twice what they should be.

On the other hand, one of the goals of the researchers was to find cases of autism that had been missed by other sources and that goal was accomplished.  They identified 90 new cases that had been missed and, at the same time, they rejected the diagnosis of 46 children.  If you put those two numbers together, you are left with a number of cases that is 40% higher than the previously number.  If you extrapolated that back to the original 255 cases of autism, you would expect there to be a total of 360 cases of autism.  While that number is higher it certainly is much less than the proposed 662.

As a result, when you consider all of the above reasons, it is unlikely that the prevalence of autism in this population is  in 86.

So is Michelle Dawson's 1 in 86 adults overall accurate?  I don't think so.

1 comment:

  1. What? Soft "science" from Michelle Dawson that just happens to support her ideological views?

    Say it aint so!