Statin use in cancer survivors versus the general population: cohort study using primary care data from the UK clinical practice research datalink

We conducted a longitudinal population-based open cohort study using Clinical Practice Research Datalink (CPRD) GOLD data. CPRD GOLD is a primary care database containing the anonymised medical records for over 15 million patients from 674 general practices in the United Kingdom (UK) [20]. CPRD GOLD data, which are collated daily from participating practices that use Vision software, include information on demographics, symptoms, tests, diagnoses, and prescriptions, and are subject to data quality checks. Diagnoses in CPRD are typically recorded using National Health Service (NHS) Read codes. Approximately 7% of the UK population are included, with patients broadly representative of the UK general population in terms of age, sex and ethnicity [20]. In the UK, the General Practitioner (GP) is the first point of contact and “gatekeeper” for most non-emergency NHS care; the vast majority of the population is registered with a GP [21].

The study period was 1st January 2005 to 31st December 2013 because during this period national guidelines (Additional file 1: Table S1) remained constant in recommending statin therapy for people with predicted cardiovascular risk ≥20% over 10 years, for primary prevention. Patients who were 40 years or older during the study period and had at least 12 months of research quality CPRD follow-up prior to entering the cohort were included; follow-up began on the latest of 1/1/2005, the patient’s 40th birthday, or the date on which 12 months of follow-up had been accumulated in CPRD; follow-up ended on the earliest of 31/12/2013, the last data collection date, or the end of follow-up in CPRD. Patients with a CVD event (Additional file 2: Table S2) prior to entering the study were excluded (as our focus was on primary CVD prevention).

Read code lists to identify cancer diagnoses were developed using the methodology described in a previous study [22]. All person-time was categorised into “never-cancer”, “first year post-cancer diagnosis” and “≥ 1 year post-cancer diagnosis”. Person-time in the ≥1 year post-cancer diagnosis category made up the cancer survivor group, while “never-cancer” person-time made up the control group. Person-time in the first year post-cancer diagnosis was not analysed further. People could contribute to more than one of the comparison groups, if they developed cancer during follow-up, but could only contribute to one exposure group at any given moment of follow-up. From this overall study population, sub-cohorts were defined to address each of our objectives, as described in the sections below.

Baseline characteristics were described for the overall study population and sub-populations (defined below). We also described the proportion of patients with at least one measurement recorded in the past 5 years for each of: blood pressure (diastolic and systolic recorded on the same day), lipids (total cholesterol and/or high-density lipoprotein cholesterol (HDL) and/or cholesterol/HDL ratio), and 10 year predicted cardiovascular risk score (Framingham, QRisk, ASSIGN, Joint British Society (JBS), or unspecified; Read terms used to identify cardiovascular risk scores are included in Additional file 3: Table S3). This description of blood pressure/lipids/risk score recording was stratified by cancer survivor / control status, gender and specific points of age (namely at age 45, and in steps of 5 years up to 75); and within strata, was restricted to individuals under follow-up at the given age and with at least 5 years of prior follow-up in CPRD.

To assess uptake of statin therapy among patients with high recorded CVD risk, all patients from the overall study population with a first ever Read code or quantitative record indicating predicted 10 year cardiovascular risk of ≥20% in CPRD and no previous statin use were included in the ‘high cardiovascular risk cohort’. To exclude patients with previous high risk scores, patients were required to have at least 12 months of research quality follow-up in CPRD prior to the first high risk score record. Patients were excluded if their follow-up ended during the 31 day period after the high cardiovascular risk score, leaving inadequate eligible time to assess prescription of a statin within 1 month. For this analysis, patients were grouped as cancer survivors or controls based on their status on the date of their first high CVD risk record. The outcome was statin uptake, defined as having a first prescription for statin therapy (defined using the codes in Additional file 4: Table S4) within 31 days of the first high cardiovascular risk score. The association between cancer survivorship and statin uptake was assessed using logistic regression. First, the unadjusted association was estimated, then we adjusted for age at baseline (i.e. at first high cardiovascular risk score, in 10 year categories) and gender as a-priori confounders. The following potential confounders were considered one at a time: practice Index of Multiple Deprivation quintile (a proxy for socio-economic status) [23]; body mass index (BMI, categorised as underweight (< 18.5 kg/m²), healthy weight (18.5-25 kg/m²) and overweight/obese (> 25 kg/m²)); total cholesterol (categorised as normal (< 5 mmol), moderate-high (5-8 mmol) and high (> 8 mmol)); smoking status (never, current, ex-smoker); alcohol consumption (non-drinker, ex-drinker, high/medium/low/unknown-level current drinker); chronic kidney disease (CKD), diabetes, and chronic liver disease at baseline (both defined using Read codes); and calendar time (split into 2005–07, 2008–10, 2011–13 to account for the marked increase in statin prescribing demonstrated in previous CPRD research over the study time period [24]). These covariates were only included in the final model if the absolute change to the odds ratio (OR) was greater than 10% when the single variable was added to the unadjusted model. Only participants with complete data for the relevant confounder were included in the stepwise testing analyses (adjusted and comparative a-priori models).

We investigated whether calendar period modified the association between cancer and initiation of a statin by adding calendar period to the model as an interaction term - the stratified ORs were assessed and a likelihood ratio test was used to compare the models with and without the interaction term. There was not considered to be any a priori reason to test for any other interactions.

To assess statin persistence, patients from the overall study population who started a statin for primary prevention (i.e. with no previous CVD record) during the study time period were included in the ‘statin initiator cohort’. To exclude prevalent statin users, patients were required to have at least 12 months of research quality follow-up prior to first prescription. For this analysis, patients were grouped as cancer survivors or controls based on their status on the date of their first statin prescription. The outcome was the time from statin initiation to first cessation of therapy, calculated from prescription records. The end date of each prescription was calculated by dividing the quantity prescribed by the daily dose where this information was available. Where duration could not be calculated, the modal prescription duration (28 days) was assumed. If there was no repeat prescription within 90 days after the estimated end date of a prescription, a patient was defined as having ceased therapy at the estimated end date. A multivariate Cox model was built, with time since statin initiation as the underlying time scale and with age (time-updated), gender, BMI, smoking status and diabetes (time updated) included as a priori confounders. All regression analyses were restricted to individuals with complete data for these confounders. Other candidate confounders were assessed and added to the model using the same model-building strategy as described for the analysis of statin uptake (above).

We investigated an interaction between cancer survivorship and calendar period as uptake of statins varies over time and drivers of this may differ in cancer survivors versus the general population. We also considered heterogeneity by time since cancer diagnosis. The proportional hazards assumption was assessed by testing for an interaction between cancer survivor status and time since statin initiation (categorised as < 0.5, 0.5–1, > 1–2, > 2 years). In a sensitivity analysis, the Cox model time scale was switched to age (to control more closely for age), and time since initiation was adjusted for as a time-varying covariate. Robust standard errors were used in all analyses to account for clustering at GP practice level.

The ‘high cardiovascular risk cohort’ included 118,012 individuals (Fig. 1). Of these, 4202 individuals were cancer survivors at the time of their first high risk score and 113,035 were controls; cancer survivors were more likely than controls to be older (median age 69.1 vs 64.2 years), female (36.7% vs. 29.0%), current non-smokers (79.2% vs. 67.1%), less deprived (29.2% vs 24.0% in the lower two deprivation quintiles) and to have CKD (8.8% vs. 4.8%; Table 1, middle columns). There were also small but statistically significant differences in cholesterol and alcohol status between the groups.

Nine hundred sixty-eight of 4202 cancer survivors (23.0%) and 26,606 of 113,035 controls (23.5%) initiated a statin within a month of their first high cardiovascular risk score. There was no evidence of association between cancer survivorship and statin initiation in the initial unadjusted analysis: OR 0.97 (95% CI: 0.90–1.05, p = 0.443, Table 2). There was evidence that statin initiation was more likely among females, those with BMI ≥25 kg/m², total cholesterol > 5 mmol/L, ex-smokers, and those with diabetes or CKD, and less common in people aged over 80 years, heavy drinkers, underweight individuals, and those in later calendar periods. None of the tested covariates individually changed the association between cancer survivorship and statin initiation by more than 10%; the maximum change being 1.5%. Therefore only age and gender were included in the final logistic regression model as a priori confounders associated with both cancer status and statin initiation. The final adjusted odds ratio was 0.98 (95% CI: 0.92–1.05, P = 0.626) (Table 2). There was no evidence that the association between cancer survivorship and statin initiation differed by calendar period (p-interaction = 0.709).

Three hundred eighty thousand eight hundred fifty-five individuals were newly prescribed a statin for primary prevention (Fig. 1). Twelve thousand one hundred forty-two were cancer survivors at the time of statin initiation, 366,280 were controls and 2433 were excluded from the analysis as statins were initiated within 1 year of cancer diagnosis. Characteristics of the two groups are described in Table 1 (right hand columns); differences between the groups were similar to those observed in the high cardiovascular risk cohort. In the statin survivor and control group, 6141 (50.6%) and 190,044 (51.9%) individuals discontinued statin therapy over a median follow-up time of 1.56 and 1.65 years respectively. Following exclusion of subjects with missing a-priori confounder values, 366,771 subjects (96.9%) were included in the Cox regression models. The overall discontinuation rates were 20.7 and 19.8 per 100 person years for the cancer survivor and general population groups. The unadjusted hazard ratio (HR) for association between cancer survivorship and statin discontinuation was 1.02 (95%CI: 0.96–1.09, p = 0.50; Table 3). Associations between covariates and statin discontinuation are also shown in Table 3.

Only the a priori confounders (age, sex, BMI, smoking, diabetes status) were included in the final model as none of the other covariates changed the HR for cancer survivorship by more than 10%. The maximum change was 3.6% with the inclusion of calendar year (HR = 1.02, 0.98–1.06). The final adjusted Cox regression model provided evidence that the risk of statin discontinuation was higher in cancer survivors than controls (HR 1.07 95% CI: 1.01–1.12; p = 0.02). The same model built with age as the underlying timescale and adjusted for time since statin initiation gave similar results (HR = 1.07, 1.01–1.13, p = 0.02). There was no evidence of interaction between cancer survivorship and calendar period (p-interaction = 0.44) or of differences in the association according to time since cancer diagnosis (p = 0.70). However, the association between cancer survivorship status and statin discontinuation appeared to change with time since statin initiation (i.e. evidence of non-proportional hazards, p < 0.001): there was no evidence of an effect of cancer survivorship on statin discontinuation in the first 6 months (HR = 1.00 (0.96–1.04)) or at 6 months to 1 year of therapy (HR = 1.01 (0.94–1.08)); from the second year of therapy onwards, there was good evidence that the risk of statin discontinuation rate was higher among cancer survivors (HR = 1.09 95% CI = 1.02–1.17; p = 0.008 for the 2nd year of therapy, and HR = 1.22 95% CI =1.16–1.28; p < 0.001 for > 2 years after initiation, Fig. 3).