The GPRD was set up in 1987, then known as the VAMP (Value Added Medical Products) Research Bank 10. It consists of the electronic clinical records of patients registered with contributing general practices and aims to include complete prescribing and diagnostic information. The practices included in the GPRD are broadly representative of all practices in England and Wales in terms of geographical distribution, practice sizes and the age and gender distributions of registered patients 11. Contributing practices must meet a range of data quality criteria before they are included in the GPRD. The quality of the information in the database, including the completeness of recording of diagnoses made in medical facilities outside the practice, has been validated in a number of independent studies and has been found to be high 10. There is excellent agreement between prescribing data from the GPRD and national data from the Prescription Pricing Authority 12. General practices contributing to the GPRD originally used a software system called VAMP but during the late 1990s most practices switched to a new software system called VISION. At the start of the study period, general practitioners contributing to the GPRD used a modified version of the Oxford Medical Information Systems coding system (OXMIS) to record diagnoses 13, but in later years most practices used the READ coding system 14.

Each individual is assigned a unique identification number and all data that could identify individuals are removed from records before they are incorporated into the GPRD.

When a patient first registers with a practice contributing to the GPRD, past medical events and diagnoses judged to be clinically important by the general practitioner are recorded in the electronic record, with the date of each diagnosis where available. While a patient is registered with a contributing practice, all new diagnoses and all new drug therapies or changes in existing therapies are recorded contemporaneously. Cases were defined as people born in 1973 or later who had a first diagnosis of PDD entered into their general practice record while registered with a practice contributing to the GPRD during the study period. Cases were identified by searching the complete medical record of all people registered with participating practices during the study period for either OXMIS or READ clinical codes indicating a diagnosis of PDD: Appendix [see Additional file 1]. People with a diagnosis of a PDD made prior to their observation period, and people whose PDD diagnosis was the first entry recorded when they joined a practice, were classified as prevalent cases and were excluded from the incidence estimates.

On the basis of the recorded diagnoses patients were classified as 'autism' or 'other PDD'. Those with autistic disorders and similar presentations were grouped in the autism category and other descriptions (such as Asperger's syndrome) were grouped in the other PDD category.

Ethical approval for the study was obtained from the Scientific and Ethical Advisory Group of the GPRD and from the ethics committee of the London School of Hygiene and Tropical Medicine.

Some of the general practices contributing data to the GPRD will provide anonymised copies of hospital letters and specialist reports on individual patients. Of patients with a recorded diagnosis of PDD in the GPRD, including prevalent cases when first registered, 446 were registered with 203 general practices willing to provide this service. For 80 of these, medical records were not available as the patient was no longer registered with the general practitioner. We obtained complete case records including copies of hospital clinic letters and specialist reports for 318 (87%) of the remaining 366 people. These were reviewed by a psychologist (LH) and a random sample of 50 records were also reviewed by a child psychiatrist (EF), both of whom have long experience in the field of autism. They judged that a PDD was likely to be present in 294 children (92.5%)15. For the 211 patients (of 318) who were first diagnosed with a PDD after they entered the GPRD (and thus are included in this paper), a diagnosis of PDD was confirmed for 193 (91.5%). Thus the positive predictive value of a recorded diagnosis of PDD in the electronic record was high, but we were not able to assess the sensitivity of the ascertainment of cases, that is how often the diagnosis of a PDD may have been missed or not recorded.

Of the 211 cases reviewed and included in this paper, 5 (2.4%) were classified as 'other-PDD' on the basis of their electronic record only, whereas in the validation, 32 (15.1%) met diagnostic criteria for a PDD other than autism. Thus it is likely that a proportion of people in the 'autism' diagnostic category in this paper have a form of PDD other than autism. The inaccuracy of diagnostic descriptions of different PDDs within the GPRD is likely to reflect changes in the definition of PDD over the past two decades, in particular a broadening of PDDs other than autism 11617. Many children with Asperger's syndrome or pervasive developmental disorder not otherwise specified would only have been assigned a diagnosis of PDD from the latter half of the 1990s. In the earlier period such children may either not have received a diagnosis of PDD at all or have been diagnosed as autism. Inflation in the number of cases in later years could have occurred as other PDD diagnoses came into widespread use and some previously undiagnosed children were diagnosed. For example the OXMIS coding dictionary, used by most general practitioners contributing to the GPRD until the mid-1990s, has only two possible codes for autism and no clinical code for Asperger's syndrome, and this diagnosis could only be assigned when practices started to use the READ coding system. These changes explain the low level of diagnoses of other PDDs until the mid 1990s (table 3). During the study period there is likely to have been an increase in the diagnosis of high-functioning autism, as professionals have become more aware that autism can occur in people of normal intelligence 3. Greater ascertainment of high-functioning autism may partly explain the increased incidence of autism as well as in the other-PDD diagnoses. That there was also a marked increase in the rate of diagnosis of severe disorders, however, suggests that better detection of less severe cases alone can not explain all of the increase. Two previous studies have demonstrated falls in the rates of diagnosis of mental retardation 6 and of non-specific developmental disorders 18 during the 1990s as the rate of diagnosis of autism increased. These patterns could be partly due to improvements in the detection and diagnosis of autism.

In the late 1980s and early 1990s autism was only likely to be diagnosed by specialist child psychiatrists who increased in number by 40% between 1988 and 2001 (personal communication: Royal College of Psychiatrists). Through the 1990s developmental and community paediatricians began to diagnose PDDs, and this, combined with increased awareness of autism and PDDs among the general public, may have contributed to increased ascertainment of the disorders 19. The marked geographical variation in rates of diagnosis and in the ratio of diagnoses of autism to other PDDs (table 4) may reflect differences in service provision and parental awareness in different regions.

It is likely that children with PDDs other than autism will generally be first diagnosed at later ages than children with autism. Greater ascertainment of other PDDs during the latter part of the study period (table 3) could have led to the observed increase in incidence being largely restricted to older age groups. However, the increased rates were observed for all age groups (table 2).

Because of low numbers of cases in some years, indirect standardisation was used to calculate SIRs. When comparing SIRs, marked differences in the age distribution of the populations being compared can lead to a biased comparison 20. However, the differences in age distributions were not great in our study and the patterns seen in the crude rates and the age standardised rates did not differ materially, suggesting the comparison of SIRs was valid.

The accuracy of the denominator data may have changed during the study period. When patients move geographical area they may delay registering with a new general practice until they have a specific reason to visit them. Thus our estimates of person-years at risk may be too low for those moving into a practice and too high for those moving out. Person-years at risk may also be inflated as patients who emigrate may omit to inform their general practitioner and, furthermore, administrative delays and errors may result in individuals being registered with more than one general practitioner. For example, in 1997 in England 51 million people were registered with a general practitioner 21 whereas the total population size was only around 49 million people22. It has also been suggested that in recent years the period of time for which patients who have moved out of an area remain registered with a general practice in that area has shortened as health authorities and general practices have streamlined procedures 23. This may mean that the inflation of denominators could have been greater in the early years compared to later years. However, these factors could explain only a very small part of the increased rates observed. An additional issue is that the person-years used as denominators in our analysis did not exclude the period following diagnosis. However, given the relative rarity of the disease, this will have produced only a very small inflation in the denominators.

A previous study based on the GPRD assessed time trends in the diagnosis of autism from 1988 to 1999 2425. The study did not examine other PDDs, was restricted to children aged less than 13 years and included a total of 305 cases. The previous paper was based on a sub-set of the GPRD data held by the Boston Collaborative Drug Surveillance Program. This is reflected in the person-years of observation contributing to the two studies – just under 3.1 million person years in the previous study compared with over 14 million person years in our paper. The two studies were undertaken in substantially different populations. However, the extent of the rise in incidence in the two studies was similar: from around 0.3/10,000 person-years in 1989/1990 to around 2.0/10,000 person-years by 1999 24. In addition, the increased risks of autism observed in successive birth cohorts were similar to those observed in our study 25. A study in the West Midlands area of the United Kingdom based on diagnoses made at child development centres found cumulative incidence rates of autism for children between the ages of one to four years of 2.22/10,000 person-years in 1991–3 (based on 20 cases), rising to 4.75/10,000 person-years in 1994–6 (based on 42 cases) 26. The corresponding cumulative rates from our study were 1.00/10,000 person-years in 1991–3 rising to 1.97/10,000 person-years for 1994–6. Restricting our data to the West Midlands region we observed rates of recorded diagnosis of autism of 1.57/10,000 person-years in 1991–3 and 1.56/10,000 person-years in 1994–6, based on 14 and 16 cases respectively. It is unclear why the rates observed in the West Midlands study were different from the rates we observed. However, these area and time period specific rates are based on small numbers of cases. The validation of cases in the GPRD can not necessarily be assumed to apply to all geographical areas. A recent study in North London estimated the prevalence of autism by year of birth and concluded there was a rise from 1979 to 1992 after which there was a plateau between 1992 and 1996 27. These authors excluded children with Asperger's. The review of clinical records of children in our study indicated that some children who would have been diagnosed as having autism in the early 1990s actually had Asperger's and would probably not have been classified as 'autism' in later years. This shift in diagnostic labelling from autism to Asperger's may explain, at least in part, the plateau in incidence described by Lingam et al. and would be compatible with the reduction in average age at diagnosis observed in that study.