The General Practice Research Database is a computerised database containing anonymised patient records for about 6 million people in the United Kingdom registered with a National Health Service primary care physician (general practitioner).3
Every prescription issued by the general practitioner, all consultations with the general practitioner, test results and diagnoses from primary and secondary care, referrals to outpatient clinics, hospital admissions, and deaths are coded by the general practitioner and entered into the database, as are basic demographic data and certain lifestyle data. General Practice Research Database prescription data have been shown to be virtually complete, and the data on incidence of cancer (based on hospital records) are around 95% valid and complete.4 5
Individual patients are recorded as entering the General Practice Research Database when they are registered with a participating general practice and leave the database when they move to a non-participating general practice, leave the NHS (for example, through emigration), or die. The database thus consists of longitudinal medical records in which patients’ length of follow-up is the time between entering and leaving the database.
We did a nested case-control study of gastrointestinal cancer in the General Practice Research Database. We defined cases as men and women aged at least 40 years with a diagnosis of incident invasive cancer of the oesophagus (ICD-10 code C15), stomach (C16), or colorectum (C18-20) recorded between 1995 and 2005 and with at least 12 months of follow-up within the General Practice Research Database before the date of diagnosis. For each case, we selected five controls with no record of gastrointestinal cancer before the index date (defined as the date of diagnosis of the case) matched on age at index date (to within 2 years), sex, participating general practice, and observation period in the database. The observation period for this study was, for both cases and their matched controls, the period between the date of entry of the case into the General Practice Research Database and the date of diagnosis (that is, patients were eligible as controls only if their follow-up time in the database included the observation period of their matched case, and for the analyses we set the observation period of the controls to match that of the cases exactly). We defined patients as exposed to bisphosphonates if they had a record within the observation period of at least one prescription for any oral bisphosphonate preparation that is licensed in the UK for use in osteoporosis (British National Formulary section 6.6.2). We excluded patients with prescriptions for bisphosphonates licensed to treat Paget’s disease or bone metastases. We estimated duration of use of bisphosphonates as the time between the first prescription and last prescription within the observation period.
We used conditional logistic regression to calculate relative risks and 95% confidence intervals for oesophageal, stomach, and colorectal cancer in relation to prescription of oral bisphosphonates. We used the Stata computing package (release 10.1) for all analyses. We adjusted the main analyses for smoking status (latest record before index date: never, past, current, missing), alcohol intake (latest record before index date: non-drinker, drinker, missing), and body mass index (latest record at least two years before index date: <25, 25-30, ≥30, missing). We assigned missing values for the confounding variables to a separate category.
We did three sensitivity analyses: defining bisphosphonate exposure as two or more prescriptions; restricting analyses to patients with complete data on smoking, alcohol, and body mass index (complete case analysis); and restricting data on bisphosphonate prescription, smoking, and alcohol use to that recorded more than one year before the index date. The results of sensitivity analyses, including the complete case analysis, suggested that use of a multiple imputation method for dealing with missing data was not necessary in this dataset and may in any case not be appropriate, given the association between bisphosphonate use and the extent of missing data on potential confounding variables.
We also did analyses of the risk of cancer associated with prescription of bisphosphonates within groups defined by various factors that may be related to prescription of bisphosphonates and to risk of gastrointestinal cancer. In addition to age at diagnosis, sex, smoking status, alcohol drinking, and body mass index, these included diagnosis of osteoporosis or osteopenia within the observation period (yes or no); diagnosis of fracture (any site) recorded before the first bisphosphonate prescription (yes or no: analysis restricted to those with at least 12 months’ observation before the first bisphosphonate prescription); diagnosis of upper gastrointestinal disease (including oesophagitis, gastro-oesophageal reflux disease, hiatus hernia, oesophageal ulcers, Barrett’s oesophagus, gastritis, duodenitis, peptic ulcers, and dyspepsia) recorded before the first bisphosphonate prescription (yes or no: analysis restricted to those with at least 12 months’ observation before the first bisphosphonate prescription); and prescription of non-steroidal anti-inflammatory drugs (British National Formulary section 10.1.1, including aspirin), corticosteroids (sections 6.3.2 and 10.1.2), or acid suppressant drugs (including H2 receptor agonists, section 1.3.1, and proton pump inhibitors, section 1.3.5) (yes or no; either at any time during the observation period or before the first bisphosphonate prescription).
We calculated estimates of the absolute risk of oesophageal cancer in bisphosphonate users from the relative risks obtained here, applied to incidences of oesophageal cancer typical for men and women aged 60-79 years in Europe and North America.6