Angiotensin Receptor Blockers and Risk of Cancer
Angiotensin Receptor Blockers and Risk of Cancer
The General Practice Research Database is a clinical database containing records from computer systems used by general practitioners to record all clinical information. The database comprises anonymised computerised medical records from general practitioners in the United Kingdom who use the Vision IT system and that have agreed at the practice level to participate; about 8% of the UK population are currently included. The UK has a publicly funded healthcare system financed through general taxation and free at the point of use to UK residents; general practitioners play a key role as they are responsible for primary healthcare and specialist referrals. Patients are affiliated to a practice, which centralises the medical information from the general practitioners, specialist referrals, and admission to hospital. Demographic information, clinically relevant lifestyle data, prescription details, clinical events, preventive care provided, specialist referrals, and hospital admissions and their major outcomes are all recorded in the database. Data collection started in 1987. Around five million patients are currently registered, with about 12 million patient records. A recent systematic review that collected together studies attempting to validate diagnoses on the database found that among such studies a median of 95% of neoplasms identified through records in the database could be confirmed with alternate data sources.
We retrieved data on all patients aged ≥18 with a first recorded prescription of an ACE inhibitor or angiotensin receptor blocker in the years 1995-2010 inclusive and dated at least six months after the patient’s registration date within General Practice Research Database. Exposure to ACE inhibitors offered an ideal comparison because the two drug classes have similar indications, reducing problems of confounding by indication, and ACE inhibitors themselves do not seem to be associated with risk of cancer. We chose the six month period to ensure that most of those included would be new users and to exclude prevalent users with unknown previous duration of treatment. ACE inhibitors were defined as all drugs classified in the British National Formulary, chapter 2.5.5.1, and angiotensin receptor blockers as all drugs classified under chapter 2.5.5.2. Patients were excluded if they had a recorded diagnosis of cancer on or before their index date (defined as the date of their first prescription of an ACE inhibitor or angiotensin receptor blocker). To prevent pre-existing cancers from affecting our results—for example, through reverse causality—we excluded the initial 12 months of person time after the first prescription for an angiotensin receptor blocker or ACE inhibitor, and individuals who developed cancer or ended follow-up during this period did not contribute to our analyses. We also excluded patients with any treatment breaks of more than 90 days during this initial qualifying period. Individuals switching from an ACE inhibitor to an angiotensin receptor blocker during follow-up left the risk set for 12 months, but could re-enter after 12 months of treatment with an angiotensin receptor blocker; this additional 12 month exclusion period after a switch was intended to prevent the estimation of a spurious association if switching were itself linked to underlying risk of cancer.
Exposure status was based on prescriptions data in the database. All prescriptions for an angiotensin receptor blocker or ACE inhibitor were retrieved and the length of each prescription was calculated on the basis of the recorded number of tablets prescribed and the daily dose; when these data were not available we assumed the median value (28 days). Clinical events are coded in the database with National Health Service (NHS) Read codes. The Read codes dictionary was first searched to identify codes indicating malignant neoplasms (excluding non-melanoma skin cancers), by using the Read code hierarchical structure, a search for keywords, and code lists used in earlier studies. We then manually reviewed and confirmed codes identified by the initial search. Subsets of codes specifically indicating neoplasms of the breast, lung, prostate, and colon were also identified during this process. In all cases, we excluded codes for borderline and suspected malignancies. Two researchers (KB and LS) developed and checked the final code lists. Finally, we identified cancers among individuals in the study by matching patients’ database records against these final code lists, with no reference to exposure status.
Follow-up time began at the end of the 12 month qualifying period after the first prescription of either an ACE inhibitor or angiotensin receptor blocker and ended at the earliest of an incident diagnosis of cancer or death with cancer listed as a cause, death from causes other than cancer, transfer out of the database practice network, or last data collection date for the practice. For the analyses of specific cancers (lung, breast, prostate, colon), follow-up was censored at the first occurrence of any other type of cancer. We examined associations between risk of cancer and use of an angiotensin receptor blocker in two ways. For the primary analysis, we used ever exposed to angiotensin receptor blockers versus never exposed to angiotensin receptor blocker (that is, treatment with an ACE inhibitor only) as the main exposure variable, and, after calculating crude incidence ratios, we included the variable in adjusted Cox regression models for each outcome. The variable ever exposed to angiotensin receptor blocker was treated as time updated so that, for example, people starting treatment with an ACE inhibitor and switching to an angiotensin receptor blocker would contribute initial person time as never exposed to an angiotensin receptor blocker, but person time after the switch as ever exposed to an angiotensin receptor blocker (Figure 1). Once exposed to angiotensin receptor blockers, however, patients remained in the ever exposed group even if they subsequently stopped treatment; we used this strategy because any causal effect of angiotensin receptor blocker use on cancer risk would likely manifest as a diagnosis only some time after harmful exposure. In a secondary analysis, we re-ran our adjusted Cox models using a time updated variable capturing cumulative duration of angiotensin receptor blocker exposure (classified as 12-24, 25-36, 37-48, 48-60, >60 months) in place of the variable ever exposed to an angiotensin receptor blocker. Cumulative duration of angiotensin receptor blocker exposure did not increase during treatment interruptions or discontinuations of treatments (Figure 1). In all cases, our Cox models were adjusted for the following potential confounders, evaluated at the index date: age group (18-54, 55-64, 65-74, ≥75, based on approximate quartiles), sex, body mass index (BMI; underweight, normal, overweight/obese), smoking status (non-smoker, current smoker, ex-smoker), alcohol status (non-drinker, ex-drinker, or current drinker—further classified as light, moderate, heavy, unknown), diabetes status (categorised as no evidence of diabetes, diabetes without metformin or insulin use, diabetes with metformin but no insulin use, and diabetes with any insulin use), hypertension (based on a recording of diagnosed hypertension or blood pressure >140/90 mm Hg), heart failure, statin use, index of multiple deprivation score (a measure of socioeconomic status), and calendar year. All of these variables were included because they could plausibly be associated with both treatment and cancer risk. First order interactions between these variables were tested in a preliminary modelling stage, with a cut off of P<0.001 to reflect the number of tests involved (55); this led to the identification of several interactions that were significantly associated with the outcome (between age and sex, BMI and smoking, smoking and index of multiple deprivation score, and heart failure and calendar year). Inclusion of these interactions, however, had little effect on the hazard ratio associated with the exposure of interest—that is, there was no suggestion that omission would lead to confounding—therefore the interactions were not included in subsequent modelling, in the interests of parsimony and ease of interpretation of the final model. For the analyses of risk of breast and prostate cancer, the study population was restricted to women and men, respectively, and sex was omitted from the confounder model. The primary analysis excluded the 10% of individuals with missing data on smoking, alcohol, or BMI; these patients were included in later sensitivity analyses with missing category and multiple imputation methods to deal with the missing covariates.
(Enlarge Image)
Figure 1.
Assignation of time varying variable for ever exposed to angiotensin receptor blocker (ARB) and duration of angiotensin receptor blocker use during follow-up under four example scenarios
To estimate the effect of angiotensin receptor blocker exposure on absolute cancer rates, we included in a Poisson model variables for which our previous Cox model suggested evidence of an association with cancer. The Poisson model was used to predict event rates by exposure to angiotensin receptor blocker in the most recent calendar period (2005 onwards) among those at highest and lowest baseline risk (men aged ≥75 who were current smokers and had a history of heart failure, and women aged 18-54 who had never smoked and had no history of heart failure, respectively). Finally, in an exploratory analyses we examined the role of specific angiotensin receptor blockers (losartan, candesartan, irbesartan, valsartan, telmisartan, other). For this analysis, patients were assumed to be exposed to only a single type of angiotensin receptor blocker during follow-up, taken as the first one prescribed.
For cancer outcomes that seemed to be associated with use of angiotensin receptor blockers, we assessed (post hoc) associations with use in the initial 12 months of exposure to provide further information on whether the observed associations were likely to represent a causal relation: associations in the first 12 months would argue against this as a causal effect would be unlikely to operate on such a short timescale.
Several planned sensitivity analyses were then carried out. Firstly, we extended our initial 12 month exclusion period after the first prescription for an angiotensin receptor blocker or ACE inhibitor to two and then three years. Secondly, in case of any effect of ACE inhibitor use on cancer risk, we used those prescribed thiazides and related diuretics (as defined by chapter 2.2.1 of the British National Formulary) as an alternative comparator group. Thirdly, we repeated the analysis including the 10% of patients with missing data on smoking, alcohol status, or BMI. We used two methods for handling the missing data: firstly, we introduced a separate missing data category for each of the incomplete variables, and secondly, we used multiple imputation by chained equations, with the imputation model including all covariates from the main outcome model. Fourthly, we repeated the primary analysis using attained age, rather than time since start of treatment, as the principal time scale, to provide the finest possible control for age. Fifthly, we checked the proportional hazards assumption by testing for a zero slope in the scaled Schoenfeld residuals over time and, as necessary, introduced interactions with analysis time for any covariates with evidence of non-proportional hazards. Finally, we carried out a post hoc sensitivity analysis, restricting the study population to patients without diabetes because of an imbalance in prevalence of diabetes between treatment groups.
Methods
General Practice Research Database
The General Practice Research Database is a clinical database containing records from computer systems used by general practitioners to record all clinical information. The database comprises anonymised computerised medical records from general practitioners in the United Kingdom who use the Vision IT system and that have agreed at the practice level to participate; about 8% of the UK population are currently included. The UK has a publicly funded healthcare system financed through general taxation and free at the point of use to UK residents; general practitioners play a key role as they are responsible for primary healthcare and specialist referrals. Patients are affiliated to a practice, which centralises the medical information from the general practitioners, specialist referrals, and admission to hospital. Demographic information, clinically relevant lifestyle data, prescription details, clinical events, preventive care provided, specialist referrals, and hospital admissions and their major outcomes are all recorded in the database. Data collection started in 1987. Around five million patients are currently registered, with about 12 million patient records. A recent systematic review that collected together studies attempting to validate diagnoses on the database found that among such studies a median of 95% of neoplasms identified through records in the database could be confirmed with alternate data sources.
Study Participants
We retrieved data on all patients aged ≥18 with a first recorded prescription of an ACE inhibitor or angiotensin receptor blocker in the years 1995-2010 inclusive and dated at least six months after the patient’s registration date within General Practice Research Database. Exposure to ACE inhibitors offered an ideal comparison because the two drug classes have similar indications, reducing problems of confounding by indication, and ACE inhibitors themselves do not seem to be associated with risk of cancer. We chose the six month period to ensure that most of those included would be new users and to exclude prevalent users with unknown previous duration of treatment. ACE inhibitors were defined as all drugs classified in the British National Formulary, chapter 2.5.5.1, and angiotensin receptor blockers as all drugs classified under chapter 2.5.5.2. Patients were excluded if they had a recorded diagnosis of cancer on or before their index date (defined as the date of their first prescription of an ACE inhibitor or angiotensin receptor blocker). To prevent pre-existing cancers from affecting our results—for example, through reverse causality—we excluded the initial 12 months of person time after the first prescription for an angiotensin receptor blocker or ACE inhibitor, and individuals who developed cancer or ended follow-up during this period did not contribute to our analyses. We also excluded patients with any treatment breaks of more than 90 days during this initial qualifying period. Individuals switching from an ACE inhibitor to an angiotensin receptor blocker during follow-up left the risk set for 12 months, but could re-enter after 12 months of treatment with an angiotensin receptor blocker; this additional 12 month exclusion period after a switch was intended to prevent the estimation of a spurious association if switching were itself linked to underlying risk of cancer.
Drug Exposure and Outcomes Data
Exposure status was based on prescriptions data in the database. All prescriptions for an angiotensin receptor blocker or ACE inhibitor were retrieved and the length of each prescription was calculated on the basis of the recorded number of tablets prescribed and the daily dose; when these data were not available we assumed the median value (28 days). Clinical events are coded in the database with National Health Service (NHS) Read codes. The Read codes dictionary was first searched to identify codes indicating malignant neoplasms (excluding non-melanoma skin cancers), by using the Read code hierarchical structure, a search for keywords, and code lists used in earlier studies. We then manually reviewed and confirmed codes identified by the initial search. Subsets of codes specifically indicating neoplasms of the breast, lung, prostate, and colon were also identified during this process. In all cases, we excluded codes for borderline and suspected malignancies. Two researchers (KB and LS) developed and checked the final code lists. Finally, we identified cancers among individuals in the study by matching patients’ database records against these final code lists, with no reference to exposure status.
Statistical Methods
Follow-up time began at the end of the 12 month qualifying period after the first prescription of either an ACE inhibitor or angiotensin receptor blocker and ended at the earliest of an incident diagnosis of cancer or death with cancer listed as a cause, death from causes other than cancer, transfer out of the database practice network, or last data collection date for the practice. For the analyses of specific cancers (lung, breast, prostate, colon), follow-up was censored at the first occurrence of any other type of cancer. We examined associations between risk of cancer and use of an angiotensin receptor blocker in two ways. For the primary analysis, we used ever exposed to angiotensin receptor blockers versus never exposed to angiotensin receptor blocker (that is, treatment with an ACE inhibitor only) as the main exposure variable, and, after calculating crude incidence ratios, we included the variable in adjusted Cox regression models for each outcome. The variable ever exposed to angiotensin receptor blocker was treated as time updated so that, for example, people starting treatment with an ACE inhibitor and switching to an angiotensin receptor blocker would contribute initial person time as never exposed to an angiotensin receptor blocker, but person time after the switch as ever exposed to an angiotensin receptor blocker (Figure 1). Once exposed to angiotensin receptor blockers, however, patients remained in the ever exposed group even if they subsequently stopped treatment; we used this strategy because any causal effect of angiotensin receptor blocker use on cancer risk would likely manifest as a diagnosis only some time after harmful exposure. In a secondary analysis, we re-ran our adjusted Cox models using a time updated variable capturing cumulative duration of angiotensin receptor blocker exposure (classified as 12-24, 25-36, 37-48, 48-60, >60 months) in place of the variable ever exposed to an angiotensin receptor blocker. Cumulative duration of angiotensin receptor blocker exposure did not increase during treatment interruptions or discontinuations of treatments (Figure 1). In all cases, our Cox models were adjusted for the following potential confounders, evaluated at the index date: age group (18-54, 55-64, 65-74, ≥75, based on approximate quartiles), sex, body mass index (BMI; underweight, normal, overweight/obese), smoking status (non-smoker, current smoker, ex-smoker), alcohol status (non-drinker, ex-drinker, or current drinker—further classified as light, moderate, heavy, unknown), diabetes status (categorised as no evidence of diabetes, diabetes without metformin or insulin use, diabetes with metformin but no insulin use, and diabetes with any insulin use), hypertension (based on a recording of diagnosed hypertension or blood pressure >140/90 mm Hg), heart failure, statin use, index of multiple deprivation score (a measure of socioeconomic status), and calendar year. All of these variables were included because they could plausibly be associated with both treatment and cancer risk. First order interactions between these variables were tested in a preliminary modelling stage, with a cut off of P<0.001 to reflect the number of tests involved (55); this led to the identification of several interactions that were significantly associated with the outcome (between age and sex, BMI and smoking, smoking and index of multiple deprivation score, and heart failure and calendar year). Inclusion of these interactions, however, had little effect on the hazard ratio associated with the exposure of interest—that is, there was no suggestion that omission would lead to confounding—therefore the interactions were not included in subsequent modelling, in the interests of parsimony and ease of interpretation of the final model. For the analyses of risk of breast and prostate cancer, the study population was restricted to women and men, respectively, and sex was omitted from the confounder model. The primary analysis excluded the 10% of individuals with missing data on smoking, alcohol, or BMI; these patients were included in later sensitivity analyses with missing category and multiple imputation methods to deal with the missing covariates.
(Enlarge Image)
Figure 1.
Assignation of time varying variable for ever exposed to angiotensin receptor blocker (ARB) and duration of angiotensin receptor blocker use during follow-up under four example scenarios
To estimate the effect of angiotensin receptor blocker exposure on absolute cancer rates, we included in a Poisson model variables for which our previous Cox model suggested evidence of an association with cancer. The Poisson model was used to predict event rates by exposure to angiotensin receptor blocker in the most recent calendar period (2005 onwards) among those at highest and lowest baseline risk (men aged ≥75 who were current smokers and had a history of heart failure, and women aged 18-54 who had never smoked and had no history of heart failure, respectively). Finally, in an exploratory analyses we examined the role of specific angiotensin receptor blockers (losartan, candesartan, irbesartan, valsartan, telmisartan, other). For this analysis, patients were assumed to be exposed to only a single type of angiotensin receptor blocker during follow-up, taken as the first one prescribed.
Pos Hoc Analyses, Sensitivity Analyses, and Model Checking
For cancer outcomes that seemed to be associated with use of angiotensin receptor blockers, we assessed (post hoc) associations with use in the initial 12 months of exposure to provide further information on whether the observed associations were likely to represent a causal relation: associations in the first 12 months would argue against this as a causal effect would be unlikely to operate on such a short timescale.
Several planned sensitivity analyses were then carried out. Firstly, we extended our initial 12 month exclusion period after the first prescription for an angiotensin receptor blocker or ACE inhibitor to two and then three years. Secondly, in case of any effect of ACE inhibitor use on cancer risk, we used those prescribed thiazides and related diuretics (as defined by chapter 2.2.1 of the British National Formulary) as an alternative comparator group. Thirdly, we repeated the analysis including the 10% of patients with missing data on smoking, alcohol status, or BMI. We used two methods for handling the missing data: firstly, we introduced a separate missing data category for each of the incomplete variables, and secondly, we used multiple imputation by chained equations, with the imputation model including all covariates from the main outcome model. Fourthly, we repeated the primary analysis using attained age, rather than time since start of treatment, as the principal time scale, to provide the finest possible control for age. Fifthly, we checked the proportional hazards assumption by testing for a zero slope in the scaled Schoenfeld residuals over time and, as necessary, introduced interactions with analysis time for any covariates with evidence of non-proportional hazards. Finally, we carried out a post hoc sensitivity analysis, restricting the study population to patients without diabetes because of an imbalance in prevalence of diabetes between treatment groups.
