Drug–drug interactions between antithrombotic medications and the risk of gastrointestinal bleeding

Study design

We conducted a population-based, retrospective case–control study using records in the United Kingdom (UK) General Practice Research Database from Jan. 1, 2000, through Dec. 31, 2005. This is a well-validated database of a network of more than 400 general practices in the United Kingdom6^–8 that has been widely used for pharmacoepidemiology research, including studies of gastrointestinal bleeding.9^,10 The electronic medical records in the UK General Practice Research Database include all important medical events and all prescriptions, since the general practitioner is the centre of health care in the United Kingdom. As a result, the database is a reliable source of information to study drug effects in a clinical setting.6^–8

We defined a case as any patient 18 years or older whose record in the database contained a first-ever entry of a computerized code for gastrointestinal bleeding. The date of diagnosis was taken as the index date for the case. Using incidence-density sampling, we selected up to 10 controls for every case in the database matched by practice, patient age (± 2 years) and index date. To permit a full assessment of patient comorbidity and lifestyle information, all patients had to have medical records with at least 3 years of data recorded before the index date.

We obtained ethics approval for this study from the Scientific and Ethical Advisory Group of the UK General Practice Research Database and the McGill University Health Centre Research Ethics Board.

Outcome measures

The primary exposure of interest was the co-prescription of warfarin or clopidogrel with acetylsalicylic acid or a nonacetylsalicylic-acid nonsteroidal anti-inflammatory drug. Non-acetylsalicylic-acid nonsteroidal anti-inflammatory drugs were defined according to the British National Formulary (www.bnf.org), with the vast majority of prescriptions by general practitioners being for naproxen, diclofenac and ibuprofen. (The full list of agents considered is aclofenac, dexketoprofen, diclofenac, diflunisal, etodolac, fenoprofen, ibuprofen, indomethacin, ketoprofen, ketorolac, mefenamic acid, meloxicam, nabumetone, naproxen, piroxicam, sulindac, tenoxicam and tiaprofenic acid). The cyclooxygenase-2 selective inhibitors rofecoxib and celecoxib are defined separately from older traditional nonsteroidal anti-inflammatory drugs. Current drug use was defined as a prescription in the 90 days before the index date; this definition was selected to minimize the misclassification of exposure, since it is difficult to estimate durations of warfarin prescriptions, especially because doses of warfarin may change during the course of a prescription according to changes in the patient's international normalized ratio.

Statistical analysis

We used conditional logistic regression analysis to compute odds ratios as an estimate of rate ratios of gastrointestinal bleeding associated with drug exposure. The odds ratio is a valid estimate of the rate ratio because we used incidence-density sampling to select the matched controls.11^,12 We estimated both the main effects and the statistical (multiplicative) interactions using the same statistical model. The drug–drug interaction term shows the excess risk beyond what would have been predicted from the combination of the individual effects of each drug. The adjusted effect of the drug combination is the total risk of the drug combination, including the effect of each agent individually as well as any excess risk due to the combination of the drugs.

We considered as covariates a past history (indicated by the presence in the patient's medical record of at least 1 medical code entered before the index date) of the following conditions: gastroesophageal reflux, peptic ulcer disease, a positive test result for Helicobacter pylori, hypertension, liver failure, renal failure, arthritis, diabetes, cancer (any type), chronic obstructive pulmonary disease and dementia (any type). We also considered as covariates the possible indications for warfarin use, including atrial fibrillation, pulmonary embolism, deep-vein thrombosis, congestive heart failure, myocardial infarction, angina and stroke.

We also compared demographic characteristics of the cases and controls, including age, sex, smoking status, body mass index and history of heavy alcohol use (as indicated by database medical codes). A body mass index of less than 18 kg/m² was considered to indicate underweight, of more than 30 kg/m² but less than 40 kg/m² to indicate obesity, and of 40 kg/m² or higher to indicate morbid obesity. A positive history of smoking (current or past) was grouped together as a single smoking variable given the cross-sectional nature of smoking data in the database.

All statistical analyses were adjusted for potential confounders and markers of health status, as measured by prescriptions in the 90 days before the index date or a diagnosis code for comorbid conditions entered in the database any time before the index date, as well as age, sex, smoking status and body mass index. We used indicator variables for missing demographic or lifestyle data to indicate the presence of a missing variable.