An Autism Challenge - Cremieux Recueil

Cremieux Recueil

SubscribeSign in

An Autism Challenge<br>0% of the rise in autism diagnoses is due to autism becoming more common

Cremieux<br>Oct 15, 2025<br>∙ Paid

62

13

Share

Here’s my challenge to anyone who thinks the explanation for the so-called “autism epidemic” is anything but diagnostic drift: prove me wrong.<br>To do this, you’ll need to show that in a proper decomposition of the rise, the liability for autism in the general population has gone up, that there’s more autism than there used to be. This means going beyond mere diagnosis numbers. Do something like this:

Background

There are two reasons why the diagnosed rate of autism can increase:<br>Liability Shift: More kids are born autistic or more people become autistic.

Diagnostic Drift: The criteria used to diagnose become broader and/or more broadly applied.

There’s a lot that can go into these categories. You could see liability increasing due to things like increased assortative mating and advancing parental age, or, in theory, due to heretofore unprecedented things like environmental exposures. You could see diagnostic drift due to changes in incentives to diagnose, awareness of conditions, diagnostic substitution, and so on. But ultimately, these are the two categories.<br>With the right data in hand, we can assess how much of the rise in autism diagnoses can be attributed to either of these broad categories. We can do this if we know about symptom scores and the proportions getting diagnosed in the population. Diagnosis is a threshold on latent liability for autism:<br>\(L_t\sim N(\mu_{L,t},1)\)

An individual is diagnosed if:<br>\(L_t > \tau_t\)

Putting diagnosis on the probit scale gives a simple identity, where pₜ is the diagnosed proportion:<br>\(z_t\equiv\Phi^{-1}(1-p_t)=\tau_t-\mu_{L,t}\)

This identity says observed diagnosis equals a threshold minus the liability mean. Conceptually, we recover a cohort-level liability index based on symptom scores, place diagnoses on the probit (tail) scale, and, because<br>\(z_t=\tau_t-\mu_{L,t},\)

changes in zₜ must come from either μ_{L,t} moving—a liability shift—or changes in τₜ—diagnostic drift. We then construct a liability-only counterfactual that holds the threshold for diagnosis fixed over time and ask the question: How much of the observed rise in pₜ would occur if only μ_{L,t} changed? The rest is diagnostic drift.<br>This brings us to the desired quantity:<br>\(\text{Liability Share} = \frac{\Delta p^{\text{(Liability-Only)}}}{\Delta p^{\text{(Observed)}}}, \text{Drift Share} = 1 - \text{Liability Share}\)

Setup

We have exactly the required data courtesy of Lundström et al.’s 2015 BMJ study of the prevalence of autism in Sweden between 1993 and 2002.<br>The first piece of data comes from Table 1. This table has the numbers of people in the population-representative Child and Adolescent Twin Study in Sweden (CATSS), alongside the numbers of them who have passed certain thresholds on the Autism-Tics, ADHD and other Comorbidities (A-TAC) inventory and numbers of people in the National Population Register (NPR) who had obtained an autism diagnosis.

Lundström et al. 2015, Table 1<br>The second piece of data comes from Table 2. This table has the mean autism score and the prevalences for the cutoffs and registered diagnoses.

Lundström et al. 2015, Table 2<br>I’ve charted exactly this data elsewhere in an article where I explained the so-called “autism epidemic”, but in case you didn’t see that, here it is.

We have two tail counts from the CATSS dataset, the score mean and 95% CIs, and registered diagnostic prevalences from the NPR, we just need to handle the ascertainment issue where the sample’s last assessment year is 2009, so some of the kids included in it are too young to have been diagnosed at the same ages as those in earlier years. As the authors noted:<br>Parents of twins born during 1993-95 were interviewed when the twins were 12 years old, and those born during 2000-02 only had 7-9 years of follow-up in the national patient register. As a consequence there was a seeming decrease in the annual prevalence of autism spectrum disorder for those born in 2000-02. However, when only including those who had been registered with a diagnosis in the national patient register before the age of 7 years , and thus having had exactly the same length of follow-up, a monotonic increase from 0.07% (1993) to 0.43% (2002, P[Emphasis mine.]

Several strategies all work fine enough in theory. We could leverage the age-7 diagnoses, scale diagnoses up linearly to age 12, cut off the analysis in 1997 for full age comparability or use 1999 for near-diagnostic comparability, or rescale based on autism data for the whole country, or we could compare across different cuts such as 1993-1996, 1996-1999, and 1999-2002. Each approach provides comparable results.<br>With our required data in hand, let the score take integer values<br>\(s\in \{ 0,...,K \}\)

From Tables 1 and 2, compute...