Autism Classification

In 2016, the Office for Statistics Regulation reviewed the Adult Psychiatric Morbidity Survey statistical publication and confirmed the overall publication as a National Statistic (now known as an accredited official statistic). As part of this review, they specified that the autism statistics within this publication be regarded as Experimental Statistics (now known as official statistics in development), setting out the rationale for this in this letter. They outlined this status is for statistics that: “are published in order to involve users and stakeholders in their development and as a means to build in quality at an early stage.” More specifically, the letter drew out four areas for further development:

1. Narrowing of the confidence intervals that measure statistical uncertainty
2. Wider assurance of the methods from within the scientific and clinical community
3. A clearer articulation of the impact of any non-sampling errors on the estimates
4. Consideration of what existing administrative data, from clinical settings, could be used to triangulate the statistics.

I have reviewed what has been done on each of these areas since the 2016 publication and I am content that there has been sufficient progress on them to remove the experimental/in development label from the 2025 autism chapter, so they are designated as official statistics. I have also identified opportunities for further development on some of the areas, subject to resource availability. If we were to seek designation of these statistics as accredited official statistics, we would be required to invite the Office for Statistics Regulation to assess them, and a decision will be made on this as part of a review of NHS England’s statistical workplan and priority areas for assessment.

This note summarises my assessment of progress against each of the four areas and any opportunities I have identified for further development subject to resource availability.

Prevalence rates for autism published in the 2014 survey by combining the 2007 and 2014 surveys were 0.8% [95% CI = 0.5% to 1.3%]. Combining these with the latest survey has narrowed the CI range further to 0.6% to 1.2%. This narrowed range has a Relative Standard Error of 19.9. Specific age and gender breakdowns have higher relative standard errors, but the overall sample sizes are large.

Pooling data across years to increase sample size where positive cases are relatively low decreases the level of statistical uncertainty and the overall sample sizes are relatively large. The survey series has been designed with the intention that samples can be combined, especially for analyses of low prevalence disorders or subgroups. Clearly this will not pick up any small changes across years, and this is a known trade off with reporting against individual years. However, the overall pooled sample is very large for an in-depth clinical survey such as this and assessing a high proportion of the population would involve a different approach of using data recorded in clinical systems, and while this would increase sample size it would open up a new set of issues around inconsistency, which are discussed in section 4.

There has been considerable use of the autism prevalence estimates and use of the underlying instruments since the 2016 publication, providing wider assurance.

The original document describing ADOS (Autism Diagnostic Observation Schedule) has been cited 16,112 times since publication in 2000, with 9,322 of these happening since the publication of APMS 2014 in 2016 (source: Google Scholar search 15/09/2025).

Other, more recently published research assures the use of ADOS as a robust assessment tool, including:

• Assessment of Autism Spectrum Disorder - Yue Yu, Sally Ozonoff, Meghan Miller, 2024
• NICE clinical guidance CG142 (updated in 2021) recommends using ADOS to aid more complex diagnosis and assessment for adults.
• Similar instruments have been used in the Mental Health of Children and Young People Survey and have generated comparable results.

An extensive literature review and international comparison study by Talantseva et al (2023) has shown that the autism estimates produced in APMS are consistent with estimates for other developed countries, acknowledging a range of estimates. The highest prevalence was in high-income countries (0.79% [95% CI = 0.67 to 0.93]). Their meta-analysis studies ran from 2000 to 2020.

Non-sampling errors (including measurement and non-response) are any error in a data collection that is not related to sampling and can be random or systematic, and they can occur at any stage of a survey. They can be difficult to identify and quantify, and they can make the results of a study or survey less reliable.

The chapter now includes more details on these errors and links to relevant research, for example, non-response error relating to hard-to-reach subgroups. In 2023/4 we present deft tables for ethnicity and age group results by gender. The chapter also includes a caveat into the results about wide and overlapping confidence intervals.

Other non-sample errors are covered in the discussion e.g. underrepresentation of autistic adults in surveys in general, how autism manifests for men and women and examples of how this could interact with the ADOS.

The same autism sampling fractions were applied to both male and female participants, to ensure that enough women were examined.

NHS England is now capturing recorded diagnoses of autism in GP systems obtained through the Network Contract DES (NCDES) dataset. This is now summarised in published Management Information downloadable from this page. This dataset shows progressively more people receiving an autism diagnosis over the last two years. While the overall figure for the latest published month (June 2025) is just within the upper bound of the confidence interval for the overall estimated prevalence rate in APMS, the upward trajectory suggests that this will be exceeded.

There are several reasons why recorded diagnoses may differ from prevalence estimates in APMS, including potential variations in clinical practice in diagnosing people who present with possible autism. Higher figures for reported diagnosis than the underlying population estimates were found in a Swedish study and their conclusion was that administrative changes may have been important for the increase of reported diagnosis, while the underlying prevalence remained the same.

Subject to resource availability, I would recommend further analysis of recorded diagnosis data, including potential further triangulation with autism waits data from the Mental Health Services dataset.

In conclusion, I am content that there has been sufficient progress against each of the four areas that Office for Statistics Regulation identified for development when they assessed the autism chapter of the previous publication to remove the experimental statistics/official statistics in development classification. I have also identified further analysis and development that would be helpful should resource be available.

Chris Roebuck (Head of Profession for Statistics)

NHS England