Data source
We conducted a retrospective cohort study using data from the Clinical Practice Research Datalink (CPRD), a large database containing primary care medical records of 6.9% of the United Kingdom (UK) population, representative of the wider population in terms of age and sex distribution [17]. Practices included in CPRD receive training on recording clinical information, with data from a practice being used only when it has reached a certain standard of quality (up-to-standard; UTS). Patient information is recorded using a coded thesaurus of clinical terms (Read codes).
Study population
Separate exposed populations were defined for PMR and GCA, which were not mutually exclusive. Patients with PMR and/or GCA, aged 40 and over, were identified based on the presence of one or more relevant Read codes (Additional file 1: Table S1) documented in the patient’s electronic medical record between 1990 and 2004. Each patient was assigned an index date corresponding to the date of their first recorded disease diagnosis. Those diagnosed with PMR or GCA before the study period or within 3 months of their registration with a practice were considered prevalent cases and excluded [18].
Two control populations were separately defined for PMR and GCA. Controls did not have a diagnosis of any inflammatory conditions (PMR, GCA, gout, ankylosing spondylitis, inflammatory bowel disease, rheumatoid arthritis, systemic lupus erythematous or psoriasis) recorded in their entire electronic medical record up until the end of study. For each exposed patient, four controls were randomly selected matched on age, sex and general practice. The non-exposed patient’s index date was defined as their matched exposed patient’s index date.
Follow-up
The study start date was defined as the date of a patient’s registration with their practice, the date their practice was defined as UTS, the patient’s index date or on January 1, 1990, whichever came latest. The study end date was defined as the earliest of the date of the patient’s death, the date the patient transferred out of the practice, the date of last data collection from that practice, the date of first fracture, or August 31, 2015. Those with less than 12 months UTS data prior to index date and less than 3 years UTS follow-up after the index date were excluded.
Outcome definition
The event of interest was time from index date until first fracture. As we were concerned with fragility fractures, fractures at four sites encompassing the definition for ‘major osteoporotic fracture’ [19] were selected (vertebrae, humerus, wrist and hip); we also included general codes for fragility fracture where site was unspecified. First fracture was identified using Read codes (Additional file 1: Table S1), which have been previously validated in CPRD [20]. In order to ensure this was the first fracture, patients with a Read code for fracture prior to their index date were excluded.
Covariates
We extracted information on patient demographics (age, sex) at their index date, lifestyle-related characteristics (body mass index (BMI), smoking status and alcohol consumption) using the measurement nearest to their index date (ever prior to index and up to 1 year after), comorbidities (summarised using the Charlson comorbidity index [21]) and prescription of medications (glucocorticoid, methotrexate, bisphosphonates and proton pump inhibitors (PPIs)) prior to the outcome for both the exposed and non-exposed. BMI was categorised according to the World Health Organization classification as underweight (<18.5 kg/m2), normal (18.5–25 kg/m2), overweight (25–29.9 kg/m2) or obese (≥ 30 kg/m2). Those with missing information on BMI, smoking and alcohol use were included as a separate category. Information on falls was collected during the study period. Dose and duration of each glucocorticoid prescription was derived from available information using the algorithm described in Additional file 2, which was subsequently used to calculate the average daily dose. Doses of oral glucocorticoid were converted into a prednisolone-equivalent dosage. Where a condition/prescription is present in the record, we assume that the patient received this diagnosis/prescription at the time it was recorded. Where there is no record of a condition/prescription, it is assumed that the patient did not have this condition/prescription.
Statistical analysis
All data summaries and analyses were performed separately for the PMR and GCA cohorts. Demographics and lifestyle-related characteristics were summarised using frequencies and percentages. Incidence rates were expressed as the number of first fractures per 10,000 person years. To assess the association between exposure and time to osteoporotic fracture, Cox proportional hazard models were used to obtain estimates of hazard ratios (HRs) with 95% confidence intervals (CIs), based on robust standard errors to account for matching. Unadjusted estimates were obtained followed by adjustment for age, sex, BMI, alcohol, smoking, Charlson comorbidity index and PPI use. To avoid over-adjustment, models were only adjusted for confounding factors (associated with both the outcome and exposure) which affected estimates by > 10%. Proportionality of hazards assumption was tested throughout using Schoenfeld residuals. Subgroup analyses by sex and age group were performed. The timing of fracture reporting in relation to PMR and GCA diagnosis was then assessed by comparing the absolute rate of fracture at yearly intervals up to 5 years after PMR/GCA diagnosis among the exposed and non-exposed in terms of incidence rate ratios using a Poisson regression model. Analyses were stratified by fracture site. The effect of the use of methotrexate and the cumulative dose of glucocorticoid on the incidence of fracture was evaluated by estimating hazard ratios within those with PMR/GCA, excluding the controls without PMR/GCA from this analysis. Any methotrexate use was compared to none and each quintile of average daily glucocorticoid dose was compared to lowest quintile. The analyses were adjusted as above, with the addition of bisphosphonate use.
Sensitivity analyses were performed, defining the exposure status as a Read code for PMR or GCA, plus two or more prescriptions for glucocorticoids during the study period; hence, patients who did not have two or more glucocorticoid prescriptions were excluded, along with their matched controls.
Patients with GCA may have symptoms of both PMR and GCA, whereas patients with PMR would not be expected to have symptoms of GCA; therefore, additional sensitivity analysis was performed estimating the risk of fracture in the group of patients with PMR diagnosis, excluding those with GCA codes. Finally, analyses considering complete cases only (no missing category for smoking, alcohol use and BMI) were performed as sensitivity analyses and compared to the main results.
All analyses were performed using Stata/MP 14.2 (Stata Corporation, TX, USA).