Higher potency statins and the risk of new diabetes: multicentre, observational study of administrative databases

Study population

We used a common analytical protocol to conduct population based cohort studies of the association between statins and diabetes in six provinces of Canada (Alberta, Manitoba, Nova Scotia, Ontario, Quebec, and Saskatchewan; source population 13.2 million in 2013), as well as in two international databases (the UK Clinical Practice Research Datalink (CPRD), population 11.6 million; and the US MarketScan database, source population about 70 million). The study populations were patients aged ≥40 years, without previously diagnosed or treated diabetes, and newly prescribed a statin (no prescription for a cholesterol lowering drug in the year before hospitalisation) between 1 January 1997 and 31 March 2011 after a hospitalisation for a major cardiovascular event (myocardial infarction, stroke, coronary artery bypass graft, or percutaneous coronary intervention). In three of the Canadian provinces (Alberta, Ontario, and Nova Scotia) data were available only for patients aged ≥65 years, and in the MarketScan database only data for patients aged <65 years were available. In four sites patient accrual began after 1997.

Data sources

All eight participating sites analysed data from their respective administrative healthcare databases, all of which have been used previously for observational research.9 10 All databases included information on the specific drug dispensed in a prescription, the quantity or days’ supply of medication dispensed (or prescribed in CPRD), and the date of dispensing. In general, physician claims and hospitalisation discharge records included date of service encounter or hospital admission, up to 25 ICD-9 or ICD-10 (international classification of diseases, ninth or 10th revision) diagnosis codes, and procedure and billing codes. Patients who were missing demographic data (sex, <1%; age, <1%) were not included in the analysis, and no imputation methods were used. We assumed that the completeness and misclassification of diagnostic coding in our datasets were similar to those in other administrative databases.

Cohorts of statin treated patients

We conducted as-treated analyses using nested case-control methodology. This type of analytical approach is well suited to studies involving time dependent exposures, and it provides good computational efficiency for analysing rare events in large databases.11 Analyses were undertaken separately at each site according to a common analytical protocol and then meta-analytic methods were used to estimate an overall rate ratio. We assembled secondary prevention cohorts, which included patients who were admitted to a hospital between 1 January 1998 (or one year after the beginning of data availability) and 31 March 2011, who received a coded diagnosis (principal or secondary) for acute myocardial infarction or stroke or a procedure for coronary artery bypass graft or percutaneous coronary intervention during their stay in hospital, who had no record of a diabetes diagnosis during their hospitalisation, and who then received a statin prescription within 90 days after being discharged, having not been prescribed a cholesterol lowering drug in the previous year (World Health Organization Anatomical Therapeutic Class C10, cerivastatin excluded). Patients were required to have a length of stay in hospital for their cohort-defining event or procedure of at least three days and no greater than 30 days. Diagnostic codes for myocardial infarction and stroke and procedural codes for coronary artery bypass graft and percutaneous coronary intervention are listed in the data supplement appendix.

Patients’ cohort entry dates were defined as the date of the first statin prescription after discharge from hospital. We excluded patients if they were <40 years old (or <66 years old in jurisdictions that only had drug data on senior citizens), had less than one year of database history, received any cholesterol lowering drug, or received any diabetes medication or a diabetes diagnosis in the previous year, including diagnoses during their index hospitalisation (drug use in hospital was not available). Our categorisation of daily statin dose was derived from a systematic review and meta-analysis of clinical trials which quantified the effects of statins on serum low density lipoprotein (LDL) cholesterol concentration.12 We defined rosuvastatin ≥10 mg, atorvastatin ≥20 mg, and simvastatin ≥40 mg as higher potency statins, and all other statins were defined as lower potency statins. That analysis showed that statins clustered around three levels of reduction of LDL cholesterol concentration: one group produced an approximate 35% reduction, one reduced concentration by about 45%, and rosuvastatin 80 mg lowered LDL concentration by 60%. This classification is similar to ones used by CNODES and others.13 14 15 We grouped rosuvastatin 80 mg with the middle group since there were relatively few such prescriptions. Therefore, statins were categorised as higher or lower in our study according to whether they would produce a theoretical <45% or ≥45% reduction in LDL cholesterol concentration.

Diabetes endpoint

We defined our diabetes endpoint as the first occurrence of a hospitalisation with a principal or secondary diagnosis for diabetes (ICD-9 diagnostic code 250; ICD-10 codes E10, E11, E12, E13, E14) or a prescription for insulin or an oral antidiabetic medication. These were most likely to be incident cases, as most patients admitted for a cardiovascular event would have been screened for diabetes, and anyone with a diabetes diagnosis at that hospital admission was excluded. Using hospital discharge data to capture diabetes was previously evaluated to have 90% sensitivity and 92% specificity in the Canadian province of Ontario.16

High dimensional propensity scores

We estimated high dimensional propensity scores for all patients at each site. The high dimensional propensity score algorithm is available as downloadable SAS software files from the Brigham and Women’s Hospital.17 As described in detail elsewhere,18 the algorithm prioritises thousands of drug, diagnostic, procedure, and demographic variables according to their potential to cause bias in the estimate of an exposure-outcome association (such as rate ratio). Typically, the 200-500 variables most likely to cause bias are included in a propensity score model that is estimated using logistic regression. In our analysis, we used high dimensional propensity scores to estimate the predicted probability (propensity score) of exposure to higher potency statins versus lower potency statins, conditional on all of the included covariates.

In addition to the 500 covariates empirically selected for the propensity score models, we also included the following pre-specified covariates: year of cohort entry and, from the year preceding cohort entry, binary indicator variables for whether the patient was hospitalised, had a laboratory test, received more than four distinct prescription drugs dispensed, received a loop diuretic, had more than four physician visits, or had a diagnosis for hypertensive disease, hypercholesterolemia, peripheral vascular disease, or congestive heart failure. After estimating propensity scores, we trimmed from the analysis the patients with the smallest 5% and largest 5% of propensity scores.

As-treated analysis

A follow-up end date was defined for each patient as the earliest occurrence of a diabetes endpoint, date of death, date of emigration, date of loss of continuous health or drug plan enrolment, 24 months after start of statin treatment, a dispensing for cerivastatin, or 31 March 2011. All cases of the diabetes endpoint that occurred during follow-up were identified, and the date of the event was defined as the index date. Ten controls were randomly selected for each case from the risk set of patients who remained at risk of diabetes, after matching on sex, age (within two years), and a requirement to have entered the cohort within 90 days of the cohort entry date of the case.

We used conditional logistic regression to estimate matched odds ratios between patients treated with higher potency statins and those treated with lower potency statins for up to two years of statin exposure. Adjusted matched odds ratios were estimated by including tenth of propensity score (nine binary indicator variables) and type of event for which the patient entered the cohort (three binary indicator variables for stroke, coronary artery bypass graft, and percutaneous coronary intervention; myocardial infarction was the reference category). Because of our method for sampling controls, odds ratios from our conditional logistic regressions were equivalent to rate ratios. In addition, we examined three mutually exclusive durations of current exposure within the two year exposure timeframe (≤120 days, 121-365 days, and 366-730 days). Rate ratios of diabetes were compared between higher and lower potency statin-treated patients within each duration category. Patients who received both a higher and lower potency statin in the same current exposure category were categorised as higher potency patients in that particular category.

Given that the mechanism by which statins might cause diabetes is unknown, we conducted a sensitivity analysis in which patients meeting our diabetes definition within the first 90 days of statin treatment were censored (90 days and 180 days at the Ontario site); in effect assuming that early events could not be the result of treatment. Thus, events included in the sensitivity analysis in the current cumulative exposure categories were events that occurred at least 90 days after exposure to statins. Patients who had events in the first 90 days were censored from further follow-up and could not serve as controls at a later time, as was the case with any patient who had an event. This sensitivity analysis was the same as the main analysis in all other respects.

Meta-analysis of site-specific results

The data analysts at each site were kept blind to the results of other sites until after the meta-analysis was completed. For the meta-analysis, we pooled the results from each site using fixed effects and random effects models. Inverse variance weighted rate ratios and 95% confidence intervals were calculated to estimate the overall effect across sites for each pre-specified duration of statin exposure.