Statin Use and Chronic Lymphocytic Leukemia Incidence: A Nested Case–Control Study in Manitoba, Canada

Data sources

Manitoba Health (MH) is the publicly funded health insurance agency providing comprehensive health insurance, including coverage for hospital and outpatient physician services, to the province's 1.3 million residents. Coverage is universal, with no eligibility distinction based on age or income, and participation rates are very high (>99%; ref. 19). Insured services include hospital, physician, and preventive services including vaccinations. MH maintains several centralized, administrative electronic databases that are linkable using a unique personal health identification number (PHIN). The completeness and accuracy of MH administrative databases are well established (20, 21). These databases have been used extensively in studies of cancer epidemiology and postmarketing evaluation of various vaccines and drugs.

CancerCare Manitoba maintains one of the oldest population-based cancer registries in the world [Manitoba Cancer Registry (MCR), in operation since 1956]. Reporting of cancer cases is mandated by provincial regulations and required for payment of physicians' service claims. The MCR is regularly audited by the North American Association of Central Cancer Registries. The quality of cancer registration has been consistently very high (21). Most cases are pathologically-confirmed (94% for cases registered between 2003 and 2007) and less than 2% of registrations originate from death certificates (21).

The MH Population Registry (MHPR) tracks addresses and insurance status (including end of coverage due to death or emigration) for all insured persons. Since 1971, the Hospital Abstracts database recorded virtually all services provided by hospitals in the province, including admissions and day surgeries (20). The data collected comprise demographic as well as diagnosis and treatment information including primary diagnosis and service or procedure codes, coded using the International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) before April 2004, and the ICD-10-CA (Canadian adaptation of the ICD-10) and the Canadian Classification of Health Interventions (CCI) afterwards. The Medical Services database, also in operation since 1971, collects similar information on services provided by physicians in offices, hospitals, and outpatient departments across the province (20).

The Drug Program Information Network (DPIN), in operation since 1995, records all prescription drugs dispensed to Manitoba residents, including most personal care home residents (22). The DPIN database captures data from pharmacy claims for formulary drugs dispensed to all Manitobans even those without prescription drug coverage. However, prescriptions dispensed to Registered First Nations, who receive their prescription benefits from the federal government, tend to be underreported by up to 20% (22). Because information is submitted electronically at the “point-of-sale,” the accuracy of the recorded prescription information is excellent (22).

This study was approved by the University of Manitoba Research Ethics Board and by MH's Health Information Privacy Committee.

Source cohort

We used the MHPR to assemble a cohort comprised of all adults ages 40 years or older (statins are rarely prescribed for younger people) who were registered with MH between 1999 and 2014. Cohort members were followed from the latest of the study start date (January 1, 1999), their 40th birthday or the date of immigration to Manitoba until the study end date (December 31, 2014), or the date of diagnosis of CLL, death, or emigration, whichever occurred first.

Definition of cases and controls

To be eligible for inclusion in the nested case–control analysis, a participant must have been: (i) free of cancer (except nonmelanoma skin cancer) before the index date, defined as the date of diagnosis for a case or the matching date for a control, and (ii) be covered by MH for a minimum of 5 years prior to the index date (to ensure that all participants had a reasonable opportunity to fill statin prescriptions before the index date).

Cases, identified by linking with the MCR, were members of the source cohort diagnosed with a pathologically-confirmed CLL (ICD-O-3 histology codes 9670 and 9823; ref. 23). We also included small lymphocytic lymphoma (SLL) in the cases group, because SLL has similar clinical and pathologic characteristics to CLL (24). Using incidence density sampling (25), we matched each case to up to 5 controls on birth date (±1 year), region of residence (1 of 5 health regions with populations ≥40 between 25,000 and 375,000), and length of coverage with insurance before the index date (to ensure a comparable length of residence in the province).

Measurement of prescription drug use

For each participant, detailed histories of dispensed statins and other drug classes were obtained from the DPIN for the period between January 1, 1995, or the coverage initiation date if it was later, and the index date. The length of these histories was ≥10 years in 59% of participants (median 11.4 years, interquartile range 7.7–15.2 years). The WHO Anatomic Therapeutic Chemical (ATC) classification was used to classify drugs, for example, statins were defined as all drugs in the Manitoba drug formulary with ATC codes C10AA (Supplementary Tables S1 and S2). Statins were further classified, a priori, according to their potency and lipophilicity, into hydrophilic (pravastatin and rosuvastatin), high-potency (simvastatin and atorvastatin), and low-potency (lovastatin and fluvastatin) lipophilic statins.

Exposure to individual and grouped statins (henceforth the index class) was characterized in 2 ways: (i) As a binary (“ever-use”) variable indicating whether a participant ever filled a prescription of any drug in the index class at any point before the index date, and (ii) as continuous and ordinal variable representations of the average annual dose of the index class calculated by dividing the total dispensed quantity of the class by its overall duration of use (measured from the dispensing date of the first prescription that included a drug in that class). All drug use in the year immediately prior to the index date was excluded to avoid protopathic bias (26). Because different drugs in the same class may have different pharmacologic potency, the total dispensed quantity for each drug was expressed as a proportion of the WHO's defined daily dose (DDD) for that drug before summing up all these proportions as the total dispensed quantity of the class (see Supplementary Table S1 for list of DDDs). The DDD is “the assumed average maintenance dose per day for a drug used for its main indication in adults” (27).

Covariates

We identified ischemic and other chronic cardiovascular diseases and diabetes (indications or proxies for indications for statin use) from the hospital abstracts and medical services databases, using previously validated algorithms (Supplementary Table S3). We also measured the number of physician visits within the 5-year period before the index date as a proxy for propensity to access healthcare services. Neighborhood average household income quintiles, which are correlated with self-reported household incomes, were determined based on place of residence and 2011 Canadian census data (28, 29).

Statistical analysis

We used conditional (to account for individual matching) logistic regression (CLR) models to calculate ORs and 95% CIs of the association between statin use and CLL diagnosis. Selection of covariates for the final models was driven by a causal diagram analysis (30) and an empirical search for confounders: a variable was considered a confounder if its inclusion in adjusted models resulted in a >10% change in OR estimates of the main exposure. The final model was adjusted for history of chronic cardiovascular disease, ever-use of non-statin lipid-lowering drugs, aspirin, non-aspirin NSAIDs, household income, and number of physician visits in the 5-year period before the index date. To test for effect modification, we stratified our analysis by each confounder and used likelihood-ratio tests to assess the statistical significance of the interaction terms. When appropriate, models also included mutual adjustment for statin groups and individual statins.

Given the long exposure histories in this cohort, participants had highly variable statin use histories. To reduce the effect of this heterogeneity, and to assess the presence of an “induction period” for statin effects (the time interval between an exposure exerting its causal effects and disease initiation or prevention; ref. 31), we repeated all analyses after dividing the exposure history into 3 successive periods: 2 to 5, 6 to 10, and 11+ years before the index date. A separate exposure index was computed for each period by limiting exposure measurements to prescriptions dispensed during that period (32). As before, CLR models were used to estimates ORs associated with drug use in each period with mutual adjustment for exposure in other periods. We used SAS 9.4 (SAS Institute) and Stata 14 (Stata Corp. LLC) for all analyses.