Materials and Methods
Primary Care Clinical Informatics Unit
Data Source. The Primary Care Clinical Informatics Unit (PCCIU) database is an electronic primary care dataset from Scotland that captures approximately 15% of the Scottish population.[19] The PCCIU contains computerized medical records containing data from 1993 and 2011 capturing approximately 15% of the Scottish general practice population. The PCCIU contains demographics and details of patient encounters, clinical diagnosis, and prescriptions. Data access was approved by the Research Applications and Data Management Team of the University of Aberdeen, and we obtained ethics approval for this analysis from the School of Medicine, Dentistry and Biomedical Sciences Research Ethics Committee at Queen's University Belfast (reference number: 15.43)
Study Design. We conducted a nested case–control study within PCCIU. Cases had a first diagnosis of primary liver cancer, including HCC and IBDC, (based upon GP Read code: B15, excluding B153) between 1 January 1999 and 30 April 2011. Up to five controls were matched to each case on exact year of birth, sex and General Practitioners (GP) practice. The index date for the cases was defined as the date of diagnosis of primary liver cancer and cases had to be free from cancer (excluding nonmelanoma skin cancer) prior to this date. The index date for the controls was the diagnosis date of their matched case, controls were free from any cancer (apart from non–melanoma skin cancer) prior to the index date.
The start of prescription records was considered 1 January 1996 (as prescriptions prior to this date were less likely to have been electronically recorded) or the date of patient registration at a GP practice if this occurred after 1 January 1996. The shortest duration of available prescription records was determined within each matched set of a case and controls. The start of the exposure period was then set as the index date minus this duration within each matched set of a case and controls to ensure all members of the matched set had an identical length of exposure period. The end of the exposure period was 1 year prior to the index date to reduce the potential for reverse causation due to increased exposure to healthcare professionals following cancer symptoms. Cases and controls with less than 3 years of prescription records prior to their index date were excluded.
Exposure. Medication use was determined from GP prescriptions in the exposure period. For each case and control, we extracted prescriptions for PPIs[20] (including esomeprazole, lansoprazole, omeprazole, pantoprazole, rabeprazole sodium) and H2RAs[20] (including cimetidine, famotidine, nizatidine, ranitidine). A quantity of 56 tablets was assumed for the less than 0.1% of prescriptions where the quantity recorded in the PCCIU database was assumed incorrect, based upon the most common PPI prescription size. Defined daily doses (DDD) were calculated from the quantity of tablets and strength, as defined by World Health Organization.[21]
Covariates. Comorbidities were obtained from GP diagnosis codes prior to the index date, including diabetes, coronary heart disease, myocardial infarction, heart failure, peripheral vascular disease, cerebrovascular disease, cerebrovascular accident, chronic obstructive pulmonary disease, mental illness, GERD, peptic ulcer disease and liver diseases (hepatitis, cirrhosis, alcoholic fatty liver, non–alcoholic fatty liver, biliary cirrhosis). Statins and aspirin use were identified from prescription records. Lifestyle risk factors were extracted from GP records including smoking status (never smoker, previous smoker and current smoker), alcohol status (none, low [eg, moderate or light drinker], or high intake [eg, above recommended limits, chronic alcoholism]), and obesity ([BMI>30], or not obese) using the most recent record prior to the index date. Postcode of the GP practice was used to assign deprivation fifths using the Scottish Index of Multiple Deprivation.[22]
UK Biobank
Data Source. The UK Biobank contains approximately 500 000 volunteer participants aged 40–69 from England, Scotland and Wales recruited from 2006 to 2010.[23] A wide range of data was collected including lifestyle, environment, medical history and physical measures, along with biological samples. The UK Biobank is linked to cancer registry data from the Health and Social Care Information Centre (in England and Wales) and the National Health Service Central Register (in Scotland). The UK Biobank has ethical approval from the North West Multi–Centre Research Ethics Committee. All participants provided written informed consent.
Study Design. We conducted a prospective cohort study among participants in the UK Biobank. Liver cancer patients were identified using cancer registry records (based upon ICD 10 codes C22, liver and intrahepatic bile duct cancer) up to 30 September 2014. Participants with a cancer diagnosis (apart from non–melanoma skin cancer) prior to baseline or in the year after baseline were excluded (as these cancers may have been present at baseline). Consequently, cohort participants were followed from 1 year after baseline until the date of liver cancer diagnosis or censoring (on the earliest of the date of death, date of other cancer, or 30 September 2014).
Exposure. Self-reported PPI and H2RA use was first ascertained from participants using a touchscreen questionnaire at baseline, and then verified during verbal interview with a UK Biobank nurse.
Covariates. Covariates were determined from patient interview and touch screen at baseline. These included age, gender, comorbidities (GERD, peptic ulcer disease, cirrhosis, hepatitis and diabetes) and other medication use (statins and aspirin). Lifestyle risk factors including smoking (never smoker, previous smoker or current smoker) and alcohol consumption (never, <1 day per week, 1–2 days per week, 3–4 days per week or >4 days per week) were also ascertained. BMI (categorized as under or normal weight [<25], overweight [25–30], obese [>30]) was calculated from height and weight measurements recorded at baseline by trained research staff. The Townsend score based upon postcode of residence was determined as a measure of deprivation.[24]
Statistical Analysis. The characteristics of cases and control were compared using descriptive statistics (for continuous variables) or frequencies and percentages (for categorical variables).
In PCCIU, we used conditional logistic regression to estimate odd ratios (OR) and 95% confidence intervals (95% CI) for the association between PPI/H2RA use and liver cancer risk. The matched design accounted for age, sex and GP practice, and adjustments were made for comorbidities (as described), obesity, aspirin and statins use. A separate complete case analysis was conducted additionally adjusted for smoking and alcohol.
Analyses were repeated by number of prescriptions, by DDDs and by type of PPIs. Similar analyses were conducted for H2RA use. A sensitivity analysis was conducted adjusting for H2RAs and PPIs simultaneously. Additional sensitivity analyses were conducted removing prescriptions in the 2 years prior to index date (including only patients with 4 years of medical records), and in the 4 years prior to index date (including only patients with 6 years of medical records), to investigate the potential for reverse causation potentially due to gastrointestinal symptoms. A further sensitivity analysis was conducted adjusting for smoking and alcohol using multiple imputation with chained equations.[25] First, an imputation model was created using ordered logit models including age, gender, PPI, H2RA, obesity, comorbidity, statins and aspirin use, separately for cases and controls. Twenty-five imputations were conducted and results were combined using Rubin's rules.[26]
The UK Biobank cohort was analysed using Cox regression with age as the underlying time scale (individuals were considered at risk from birth and under observation from age at baseline, left truncated) to calculate hazard ratios (HR) and 95% CIs for PPI/H2RA use and liver cancer risk and by histological types (HCC based upon ICD 10 code C22.0 and IBDC code C22.1). In adjusted analyses, the model contained age, gender, deprivation, BMI, alcohol, smoking, comorbidities at baseline (GERD, peptic ulcer disease, cirrhosis, hepatitis and diabetes) and statins and aspirin use at baseline.
Sensitivity analyses were conducted adjusting for H2RAs and PPIs simultaneously and by repeating the analyses starting follow-up at 2 and 4 years after baseline (to remove cancers within 2 and 4 years, respectively, which could have influenced medication prescribing at baseline).
Aliment Pharmacol Ther. 2018;48(1):55-64. © 2018 Blackwell Publishing
