Special Educational Needs (SEN) register count

All schools within the county of Cambridgeshire, including mainstream and special schools in both the state and the private sectors (serving the population of Cambridge City, East Cambridgeshire, South Cambridgeshire, and Fenland districts), were invited to participate in the study. These geographical areas cover a broad cross-section including urban and rural areas. Special schools in the county accept children who have moderate, severe or profound intellectual disability, emotional and behavioural difficulties, and autism-spectrum conditions. Those schools that agreed to take part were asked to provide a count of the number of children aged 5–9 years old (school years 1–4) and the number recorded on the SEN register as having an autism-spectrum condition diagnosis. Based on the available quality of information in this register, best estimate autism-spectrum condition diagnostic subgrouping using ICD–10 research criteria was carried out.

Direct school population screening

The schools that agreed to take part in the study were asked to distribute a questionnaire pack to parents of all children in the school aged 5–9 years. The total number of questionnaires requested by all schools (mainstream and special) was 11 700. It was therefore assumed that this was the total number of the schooled population that was surveyed in this study. The pack contained a diagnosis survey (online Appendix DS1), the Childhood Autism Spectrum Test (CAST)^8,22 (formerly the Childhood Asperger Screening Test; here the same acronym and the items are retained, but the title has been modified to reflect that the same instrument can be used to detect all autism-spectrum conditions and not just Asperger syndrome), as well as standardised questions about parental education and socioeconomic status.²³ The diagnosis survey enquires about various diagnoses that the child may have received, by whom these were made, along with any developmental or behavioural concerns expressed by teachers or other professionals.

The CAST was used as a screening instrument to identify those with high and borderline scores with no pre-existing diagnosis. The CAST is a 37-item parental self-completion screening questionnaire and has a scoring range from 0 to 31. For the 31 scoring items, an autism-positive response scores 1 and an autism-negative response scores 0. Some items are reverse scored so that not all `yes' responses score 1. The remaining six items are control questions on general development and are not scored. The cut-point for concerns of possible autism-spectrum condition is 15 or higher. At that cut-point, sensitivity has been shown to be 100%, specificity 97% and positive predictive value 50% using research diagnostic assessments (Autism Diagnostic Observation Schedule (ADOS)²⁴ and Autism Diagnostic Interview – Revised (ADI–R^25)).22 The CAST shows moderate to good test–retest reliability.^26,27 Scores also show heritability.²⁸ The CAST can be recommended as a screening instrument for use in epidemiological studies, but not as a general screening tool at a population level owing to the high number of false positives.²² This is a function of the low prevalence of autism-spectrum conditions in the general population and therefore the CAST has a low positive predictive value. The CAST in full has been published elsewhere.⁸

The cover letter explained that the research team was interested to hear from parents of all children aged 5–9 years, not just those with concerns about their child. The questionnaire pack was distributed to participating schools in six batches, with a batch being distributed every 2 months between February 2003 and March 2004. This was taken home for parental completion and returned directly by Freepost to the research team. Reminder letters were also distributed to participating schools, who were asked to give them out about 2 weeks later. A total of 1452 questionnaires were sent in Batch 1 and 404 (28%) were returned; 2195 were sent in Batch 2 and 610 (28%) were returned; 2686 were sent in Batch 3 and 805 (30%) were returned; 1539 were sent in Batch 4 and 592 (38%) were returned; 2119 were sent in Batch 4 and 512 (24%) were returned; and 1644 were sent in Batch 6 and 481 (29%) were returned.

Returned questionnaires were excluded from analyses if: (a) they were from parents of children from schools not invited to take part in the study (parents may have accidentally completed the questionnaire for another child attending a different school); or (b) the children were older or younger than those approached, allowing a margin of 1 year younger and older than the lowest and oldest ages expected.

All children whose parent had reported an existing autism-spectrum condition diagnosis on the diagnosis survey were excluded from the assessment phase in order to maximise available resources. The number of children whose parents reported on the diagnosis survey that their child had an autism-spectrum condition diagnosis was counted. Only these families were sent a further form asking for full information about their child's diagnosis. This included the subgroup diagnosis (if known), the clinic where the child was diagnosed, the name of the clinician who made the diagnosis and the date of the diagnosis. A child whose parents reported a clinical diagnosis of autism-spectrum condition that was consistent with their original report was considered to be a confirmed case. Only those children who had been diagnosed by a relevant health professional (e.g. paediatrician, psychiatrist, clinical psychologist) were confirmed as a case.

Every child scoring above the cut-point of 15 on the CAST, as well as a randomly selected 33% of 12–14 (borderline) scorers, were invited for a detailed assessment to establish how many undiagnosed cases of autism-spectrum condition there might be in the population. We were then able to calculate the ratio of undiagnosed:diagnosed cases. In total, 33% of borderline scorers were chosen as this was sufficient to enable estimation of undiagnosed cases in this score band without overloading the team's assessment capacity. No child scoring ⩽11 was sampled for assessment and all in this score band were assumed to not have autism. These sampling bands were chosen because our previous studies suggest that no children scoring ⩽11 points are diagnosed with an autism-spectrum condition, whereas children on the borderline of the cut-point (12–14) may show some social and communication difficulties.^8,22

Where there were missing CAST data, the midpoint score was used; that is, the maximum score (calculated by recoding missing items to one) plus the observed score (all missing data treated as if the item score would have been zero) divided by two. Although there are numerous methods to treat missing data, we decided that such an approach should be adopted to ensure maximum capture of cases from incomplete CAST questionnaires for the purposes of sampling. Five children were not invited for assessment, since the amount of missing data in their screens was considered excessive as the maximum and the observed scores were discrepant by more than 5 points.

Diagnostic assessment

Assessment consisted of a standardised interview (ADI–R²⁵) with the parent or primary caregiver and a standardised observation (ADOS²⁴) with the child. The ADI–R and ADOS were administered by members of the research team trained in the use of these instruments to reliability levels for research purposes. All assessors were unaware of CAST scores at the time of assessment. A random selection of assessments (35%) were video- and audiotaped. IQ data were not collected, which is a limitation of this study. However, developmental delay was established using the background and introductory items in the ADI–R, and communication impairment was established using items 9 and 10 in the ADI–R.

Case definition

Following face-to-face assessment, cases were defined in two ways: (a) an assessment diagnosis strictly based on meeting ADI–R and ADOS algorithm criteria; and (b) a consensus diagnosis, based on information obtained through the ADI–R and ADOS in conjunction with the clinical judgement of members of the research team. The latter is recommended because the ADI–R and ADOS alone are less reliable in detecting more subtle cases of autism-spectrum condition (the ADI–R only provides a cut-point for classic autism^25,29). A case definition for wider-spectrum conditions including Asperger syndrome and PDD other was required for maximum capture of all autism-spectrum cases. For all children who received a consensus autism-spectrum condition diagnosis, a review of their assessment data was conducted so that cases could be subdivided into the four major diagnostic subgroups using ICD–10 research criteria.

The SEN register

A total of 162 schools (142 state, 13 private mainstream schools, and 7 special schools) were approached. Of these, 96 schools (86 state and 6 private mainstream, and 4 special schools) agreed to participate in the study (Fig. 1). For simplicity, state and private mainstream schools (non-special schools) were merged for analysis and labelled as `mainstream'. Schools across Cambridgeshire were grouped geographically (between 14 and 33 schools per group) and invited across six time points (about every 2 months) between February 2003 and March 2004. School participation at each time point ranged from 46 to 71% but there were no noted differences in geographical distributions of schools that participated v. schools that refused. Overall, 79 out of the 92 mainstream schools and none of the 4 special schools reported the number of children with autism-spectrum condition from their SEN register. In total, 8824 children were attending the 79 schools. Since 11 635 questionnaires were distributed to schools, it was assumed that 2811 children attended the 13 schools that did not report the number of children with autism-spectrum condition on the SEN register. In total, 83 cases of autism-spectrum condition were reported from the 79 schools (Fig. 1), 10 of which were reported to be childhood autism. It was not possible to verify the reported diagnoses from the SEN register since legislation regarding data protection would not allow this. Further, no information was provided regarding gender distribution of the 83 cases.

View larger version:

• In this page 
• In a new window

• Download as PowerPoint Slide

Fig. 1

Special Educational Needs (SEN) register: methods and results.

To estimate prevalence using data from the SEN register, exact binomial methods were used to calculate 95% confidence intervals. For childhood autism, this gives a prevalence estimate of 11 per 10 000 (95% CI 5–21). For all autism-spectrum conditions, this gives a prevalence estimate of 94 per 10 000 (95% CI 75–116) or 1 in 106.

Diagnosis survey and CAST population and response

In total, 11 635 questionnaire packets were distributed to participating mainstream schools, one to each child in school years 1–4. Of these, 3370 were returned; 28 questionnaires were excluded because they were outside the age range approached or from a non-participating school, leaving 3342 diagnosis surveys and CAST questionnaires from mainstream schools for analysis (response: 29%).

In addition, 65 questionnaires were distributed to participating special schools, as this was the number of children who were aged 5–9 years at the time of the survey, and 31 were returned (48%). All but one of the special schools had fewer than 100 pupils on the roll at the time of the survey, and the age range of pupils in special schools is wider than for mainstream schools. For example, one school accepts children and young adults from age 2 to age 19.

Thus, altogether there were 3373 questionnaires available for analysis (response: 29%). Of these, 1683 (50%) were for boys and 1672 (50%) were for girls, and for 18 (1%), gender was not stated by the parent/guardian. Table 1provides the age distribution and demographic details of the screened population. Table 2 indicates that the population is biased towards Social Class I (around 50% coming from this social class). Regarding the number of parents educated to degree level or higher (leaving education at age 21 or later), 956 of the 3373 responders (29% of the current sample) fit this category. This is similar to the 2001 census data for Cambridge, Cambridge East and Cambridge South, which indicates that 32% of the total population from those areas was educated to degree level or higher.³⁰

Reported cases of autism-spectrum conditions

Overall, 41 cases of autism-spectrum conditions were reported by parents using the diagnosis survey: 37 from mainstream schools and 4 from special schools (Fig. 2). All cases from special schools were assumed confirmed since entry to a special school would require extensive assessment of individual special needs. Of the 37 cases reported from mainstream schools, 24 (65%) parents replied to the request for further information providing details of the reported diagnosis. Two children had not actually received diagnoses. Of the remaining 22 cases, 19 (16 boys, 3 girls) included information consistent with having a diagnosis of an autism-spectrum condition (the 3 other parent-reported cases were too vague to assign an ICD–10 classification) (Fig. 2). The ratio of confirmed:reported cases from the diagnosis survey was therefore 19:24.

View larger version:

• In this page 
• In a new window

• Download as PowerPoint Slide

Fig. 2

Diagnosis survey: methods and results.

All reported diagnoses (from the diagnosis survey) were adjusted by the ratio of confirmed:reported. `Confirmed' refers to those parents who not only reported that their child had a diagnosis of an autism-spectrum condition but who also replied to the request for further information, and where this further information was consistent with their initial report. This provided a weight to apply to this group, giving an estimate of known cases within the screened population. Using the weight of 19/24 as an index of diagnostic validation, we can adjust for possible slight overreporting by the 13 parents who reported that their child (9 boys, 5 girls) had a diagnosis of an autism-spectrum condition from the diagnosis survey but who did not reply to the request for further details (assuming the rate of confirmation of diagnosis is the same in the non-responders). On this basis, the overall parent-reported prevalence estimate is 13×(19/24) + 19 (confirmed cases) + 4 (special school cases: 3 boys, 1 girl). This equates to 33.3 known cases in the screened population.

To estimate prevalence, confidence intervals were calculated using 1000 non-parametric bootstrap samples,³¹ reporting percentile confidence intervals (95% CI 2.5–97.5). This allows the confidence intervals to be asymmetrical about the prevalence estimate and includes the uncertainty of the weighted estimate. This gives a prevalence estimate of 99 per 10 000 (95% CI 65–141) or 1 in 101. For boys and girls separately, the prevalence estimate is 153 per 10 000 (95% CI 94–217) or 1 in 66 and 42 per 10 000 (95% CI 9–79) or 1 in 208 respectively.

Of the 41 children whose parents reported a previous diagnosis of an autism-spectrum condition, all (100%) from the special school sample (n = 4) scored above the cut-point of 15 on the CAST (range 18–28, using the midpoint score). In total, 30/37 (81%) children who were reported by parents as having an autism-spectrum condition diagnosis from the mainstream school sample scored above the cut-point of 15 on the CAST and 4 (10.8%) children scored 12–14. Two (5.4%) children scored ⩽11. Of these, one parent replied to the request for further information and the child was confirmed as not having autism-spectrum condition. In the other case, the parent replied to the request for further information, which was consistent with an autism-spectrum condition diagnosis and was thus treated as a confirmed case. All children whose parents initially reported that their child had a diagnosis were excluded (before the request for further information was made) from the assessment phase owing to resource limitations.

Sampling

From the screened population (excluding those with an autism-spectrum condition diagnosis reported on the diagnosis survey), 90 children scored ⩾ 15 on the CAST and were invited for assessment. Of these, 52 (58%) attended. Overall, 122 children scored 12–14 on the CAST and 33% were invited for assessment. Of these, 25 (63%) attended. Thus, 77 research assessments were undertaken. Of these, 8 were of children from special schools and 69 from mainstream schools. There were no significant differences between those who accepted and those who refused the invitation for assessment in terms of CAST score, gender, age of child or concerns expressed by professionals (allP>0.2). Ten (eight boys, three girls) children who scored ⩾15 and one (boy) who scored 12–14 were given a research diagnosis of an autism-spectrum condition. Of these 11 children, 4 met algorithm criteria on both the ADOS and the ADI–R (including the child who scored 14 on the CAST) and 5 met criteria on either the ADI–R or the ADOS, but not on both. The other two children did not meet algorithm cut-points on either instrument but were showing clear difficulties in all three areas of the diagnostic `triad' (social interaction difficulties, communication difficulties, and/or repetitive and stereo-typed behaviours and interests) and were given a consensus autism-spectrum condition research diagnosis. Thus, 11 out of 77 assessments resulted in a consensus diagnosis of autism-spectrum condition, of which 4 resulted in a consensus diagnosis of childhood autism. This suggests that diagnosis was cautious and conservative. Table 3 provides full assessment scores for the 11 identified cases of autism-spectrum condition. Figure 3 shows the study flow from the CAST screening and assessment phase.

View larger version:

• In this page 
• In a new window

• Download as PowerPoint Slide

Fig. 3

Childhood Autism Spectrum Test (CAST) screening: methods and results.

Inverse probability weighting methods were used to adjust the prevalence estimates for the known non-response to the invitation for assessment within each sampling score band, and for the 33% random sampling for face-to-face assessment among the borderline scorers from the CAST screening. As described earlier, assessment invitations were weighted towards the high scorers. Low scorers (score ⩽11) were assumed to have no new cases since no cases of autism-spectrum conditions were identified in previous studies of children scoring ⩽ 11.^8,22 Using the weightings of 90/52 (90 children who scored ⩾15 on the CAST and 52 children in this score band who participated in the assessment) and 122/25 (122 children who scored 12–14 on the CAST and 25 children in this score band who attended for a research assessment), the overall directly observed prevalence estimate for all autism-spectrum conditions is (10×(90/52)) + (1×(122/25)), which corresponds to 22.2 new (undiagnosed) cases from the screened population.

Calculating the ratio of known:unknown cases of autism-spectrum conditions

This ratio is calculated by adding the 33.3 known cases of autism-spectrum conditions from the diagnosis survey data and the 22.2 new research cases from the CAST screening and assessment, which amounts to 55.5 total cases. Dividing this by 33.3 (known cases) gives 1.67 as a multiplier that can be applied to prevalence estimates generated from studies that screen the primary population for known cases of autism-spectrum conditions (95% CI 1.24–2.32). For boys, the ratio is calculated by adding 19.2 (weighted number of male new research cases from the CAST) and 25.7 (weighted number of male known cases from the diagnosis survey), amounting to 44.9 total cases of autism-spectrum conditions in boys. This can be divided by the known male cases (25.7) to give 1.75 (95% CI 1.3–2.6) as the multiplier that can be applied for boys. For girls, the calculation is 3.0 (weighted new female research cases) added to 7.0 (weighted known cases from the CAST) divided by 7.0. This gives a multiplier of 1.43 (95% CI 1.1–3.3). Overall, for boys and girls together, for every three known cases, there are at least two undiagnosed cases of autism-spectrum conditions in the primary school population, or a ratio of 3:2 (known:unknown). There is no evidence of a difference in the undetected estimate of cases for boys and girls, despite a clear difference in the overall ratio of boys:girls with autism-spectrum conditions.

Combining estimates and sensitivity analysis

Data from the SEN register and the diagnosis survey estimated the prevalence of known cases of autism-spectrum conditions, whereas data from the CAST screening and assessment estimated the number of unknown cases. It is not surprising that over 60% of children in the assessment phase were given a consensus diagnosis of an autism-spectrum condition (rather than childhood autism) since cases of lower functioning (childhood autism) are more likely to have been detected by school age. Of those who received a consensus diagnosis of childhood autism, three-quarters were described as `high-functioning' and met ICD–10 research criteria for this category because these children had a delay in language (so could not be classified as having Asperger syndrome) with no reported cognitive impairment.

By adjusting the estimate derived from the SEN register (that is, based on whole classes and not a 29% response) for the additional number of cases that would be expected if all the children were directly observed, the prevalence estimate is 157 per 10 000 or 1 in 64 (95% CI 99–246).

Study weights were adjusted to reflect the fact that special schools and mainstream schools had a slightly different response. The bias introduced by the 29% response among parents cannot be directly estimated. Two assumptions that are expected to be lower and higher than our point estimate are: (a) there were no cases in the non-responders; and (b) data are missing at random, such that the non-responding population would have the same distribution on the CAST as the responding population. Both are unlikely and produce very different values. An intermediate position is to assume that there is a `worried parent' effect; in this case we assumed that there is a twofold increase in response in parents whose child scored on the CAST in the high or medium range (worried parents). Therefore, three sensitivity estimates from the CAST screening and assessment are provided.

Adjusting for the school non-response made no change to the prevalence estimates. The extreme sensitivity analyses, as expected, provide widely divergent estimates. If we assume there were no cases in the non-responders, the estimate is 47 per 10 000 (95% CI 32–63). If we assume non-responders had the same distribution of scores on the CAST as the responders, the estimate is 165 per 10 000 (95% CI 111–218). Assuming an increased response from parents of children with higher scores on the CAST (the `worried parents' assumption) reduces the estimate, as expected, to 113 per 10 000 (95% CI 76–149).

The different methods provide a range of estimates which, including sensitivity analyses, range from 47 to 165 per 10 000 children aged 5–9 years old. The two estimates that appear least biased are: (a) the estimate based on the multiplier from the CAST screening and assessment applied to the SEN register estimate (the ratio of known:unknown cases, 1.67×94); and (b) the estimate that makes a reasonable assumption about a possible response bias from postulated worried parents. These provide the estimates of 157 and 113 per 10 000 respectively.

The data converge on the conclusion that, despite differences in methodology, the prevalence of known cases of autism-spectrum conditions is slightly above 1%. The prevalence estimates generated are shown in Fig. 4.

View larger version:

• In this page 
• In a new window

• Download as PowerPoint Slide

Fig. 4

Summary of prevalence estimates from Special Educational Needs (SEN) register and diagnosis survey, and calculation of the multiplier from Childhood Autism Spectrum Test (CAST) screening.

Implications

Epidemiological research involving screening the school-age population for developmental conditions is not without difficulties. Our experience from this study and previous studies⁴⁷ has indicated response to surveys to be a major limitation. Although our final estimates may be built upon a selection of possible assumptions about the non-responding population, our most conservative estimate (assuming no cases of autism-spectrum conditions in the non-responders) is still nearly 12 times higher than the estimate in 1978.⁹Independent replication of this study is warranted, in particular to ascertain diagnostic verification of parent-reported and SEN-register autism-spectrum condition diagnoses. Results from this study should be treated with caution because of the low response and the finding that the socioeconomic distribution of the population was not representative of the UK population. Notwithstanding this, as autism-spectrum conditions are being increasingly recognised, these studies suggest that appropriate services should plan to meet the needs of between 1–2% of the primary school-aged population.