Data came from the health improvement network,5
a longitudinal dataset from primary care practices in the United Kingdom. The health improvement network contains information collected in computerised primary care records from 308 practices throughout the UK. Details of demographics, primary care diagnoses, and treatments prescribed are prospectively recorded in individual patient’s records. Details of referrals, secondary care diagnoses, and deaths are also captured because of the structure of the UK healthcare system. Within this system, the population are registered with one general practitioner and remain on that general practitioner’s list while being treated in secondary care. Medical events are automatically coded at entry by use of the READ coding system. Each primary care practice participating in the database has a unique practice identification number recorded in the dataset. The quality of the database is monitored, and medical diagnoses in the database have high validity.6
The study cohort included children who had a date of birth recorded in the health improvement network database since the inception of computerised data collection and continuous enrolment until they received a diagnosis of ADHD or until their 10th birthday. The study period was 1988 to 2003. We excluded children with a diagnosis of ADHD before their second birthday. We considered children to have ADHD if the READ code for ADHD, overactive child, attention deficit disorder, hyperkinetic disorder, or hyperkinetic conduct disorder was coded any time after the child turned 2 years of age. We extracted data from the records of children for as long as they remained registered in the database; however, we noted injuries, diagnoses, and prescriptions as occurring before or after the child’s 10th birthday.
We divided the study cohort into three groups. We defined the first group, children with early head injury, as children whose record contained any READ code for head injury except for “minor head injury” and “nursing advice for head injury” before the child’s second birthday. READ codes are sufficiently specific to differentiate abrasions and lacerations to the head and face from head injury itself. We extracted information on the source of the referral for patients with early head injury and dichotomised the data according to whether or not the child was referred by a physician to a higher level of care or admitted to hospital. The second group, children with early burn or scald injury, comprised all children who had a READ code for burn or scald recorded before their second birthday. Children from the cohort without head injury or burn injury before age 2 years made up the third group (comparison group).
Data extracted for all children included practice identification number, date of registration in the database, duration of registration, date of birth, sex, codes and dates of head injuries (up to three injuries after 2 years of age) and burn or scald injury, and any code indicating prematurity (<37 weeks gestational age) or child abuse. A head injury had to occur at least two months after a previous one to be considered a new injury, to ensure that the visit was not a follow-up visit. For each child, we extracted the date of the first diagnostic code for ADHD and the date of the first prescription for drugs commonly used for ADHD, including methylphenidate, dexamfetamine, and atomoxetine. We used the Townsend deprivation index, on the basis of patients’ home postcodes, to measure socioeconomic status. The Townsend index ranks levels of deprivation throughout the UK by fifths (1=least deprived; 5=most deprived).7 8
We used frequencies and percentages for categorical variables and medians with interquartile ranges for non-normally distributed continuous variables. We used the Wilcoxon-Mann-Whitney and the Kruskal-Wallis tests to compare continuous variables with skewed distributions. We considered P<0.05 to be statistically significant. We calculated the relative risk, with 95% confidence intervals, of developing ADHD for each injury group compared with the comparison group. We used χ2 tests to explore associations between predictor variables and the outcome. We did multivariate modelling, using the generalised estimating equation method to account for clustering by practice, to examine the association of injury type with the odds of developing ADHD after adjustment for Townsend index, sex, and prematurity. We used Poisson regression to calculate the unadjusted rate of diagnosis of ADHD.