Methods
Participants and Setting
We used data from The Health Improvement Network (THIN), a longitudinal primary care database containing anonymised medical, prescribing and lifestyle data for patients registered with participating UK general practices. At the time the dataset was generated, THIN held data on 3.9 million patients registered across 255 general practices. Information from secondary care is received by GPs and recorded in patient records. Data are recorded in THIN using Read codes, a clinical terminology system based on the International Classification of Diseases V.10 (ICD-10).
Study participants were drawn from an open cohort of 180 064 children in THIN who were born between January 1988 and September 2004 and whose primary care records had been linked to their mothers' primary care records, as previously described. Children had to have been registered at the general practice within 60 days of birth to maximise the likelihood of capturing their first fracture event. Cases and controls were a subset of children from a previous case–control study assessing risk factors for poisonings, burns and fractures. Fracture cases were children less than 5–years old who had a first fracture event in their medical record. For every case, up to 10 controls were selected at random. Controls were matched to cases on general practice and were children less than 5 years old who had not had a fracture before or on the injury date of their matched case. Children were not matched by age and sex to enable exploration of the effects of these variables.
Definition of Long-bone Fracture Cases
From the case–control population described above, we identified long-bone fracture cases using Read codes referring to fractures, as classified by ICD-10, of the femur (S72.0–S72.9, T93.1), humerus (S42.2–S42.4, S42.7), ulna and/or radius (S52.0–S52.9) and tibia and/or fibula (S82.1–82.9). Less specific Read codes such as 'broken arm' and 'greenstick fracture' were included in the definition as it was likely these codes indicated a long-bone fracture. As some Read codes do not specify an anatomical site (eg, 'fracture not otherwise specified'), we examined Read codes entered within 3 months of the first fracture Read code, to identify if the fracture had occurred in a long-bone. Three months was chosen through examining the distribution of Read codes entered onto the medical record, and to allow time for additional information from secondary care to be entered into the medical record.
Risk Factor Variables
Potential child, maternal and household risk factors for fractures were identified from existing literature. Those available in THIN included child age, sex and birth order. Maternal risk factors included age at delivery, smoking status and perinatal depression (a diagnosis of depression during pregnancy or within 6 months of delivery). Mothers were classified as having a history of alcohol misuse if they had a Read code indicating harmful or hazardous drinking documented in their medical record before the fracture event. Household risk factors included the number of children in the household (those aged 16 or under) and socioeconomic status measured using quintiles of the Townsend index of material deprivation, representing relative socioeconomic position at a national level.
Statistical Analyses
We estimated unadjusted and adjusted ORs and 95% CIs for the association of long-bone fracture with each risk factor using conditional logistic regression models. Backward elimination, as described by Collett, was used to build the multivariable models, with likelihood ratio tests (LRTs) used to assess significance and p<0.05 considered statistically significant. Child age and sex were included in all models. We included the whole study population in all multivariable models to ensure comparability. To account for missing maternal smoking and Townsend quintile data, we included a missing data category for these variables in the regression models. Potential interactions, based on theoretical plausibility, were assessed by adding interaction terms to models and testing their significance using LRTs, with p<0.01 considered significant (due to large study size). We tested for multicollinearity using the covariate correlation matrix and by calculating the variance inflation factor. Analyses were carried out in Stata V.10.1.
Statistical power was calculated using the prevalence of the rarest risk factor, recording of alcohol misuse within the same primary care population (0.48%). To obtain 80% power to detect an OR of 2.2 at the 5% significance level, using a correlation coefficient of 0.2 to allow for matching by general practice, 2019 cases with 10 matched controls per case were required. For all other risk factors of higher prevalence there was a greater level of power.
As fracture Read codes varied from those that were highly specific in defining the anatomical site of fracture to less specific codes such as 'broken arm' and 'greenstick fracture', misclassification of the outcome could have been introduced by including Read codes where there was uncertainty about the anatomical site of fracture. We therefore carried out two sensitivity analyses (Table 1), first excluding fractures where only a 'greenstick fracture' Read code was used, and second, restricting the definition to the most precise Read codes for long-bone fractures.
Ethics Statement
Approval for this study was granted in October 2009 by the THIN Scientific Review Committee.