Diuretics, Beta-Blockers, and Statins and Risk of Diabetes
Diuretics, Beta-Blockers, and Statins and Risk of Diabetes
NAVIGATOR was a multinational, randomised, double blinded, placebo controlled trial examining the effects of valsartan and nateglinide on conversion to type 2 diabetes mellitus and cardiovascular outcomes in patients with impaired glucose tolerance and other cardiovascular risk factors. The study design and results have been previously published, as have the eligibility criteria (see Supplementary Appendix).
The endpoint of interest was diagnosis of new onset diabetes. We measured fasting plasma glucose every six months for the first three years of follow-up and then annually. Oral glucose tolerance tests were performed yearly. New onset diabetes was defined by a fasting plasma glucose level ≥126 mg/dL (7.0 mmol/L) or a glucose level ≥200 mg/dL (11.1 mmol/L) two hours after an oral glucose tolerance test, confirmed by an oral glucose tolerance test within 12 weeks after the increased glucose value was recorded. We separated the diabetes endpoint into 12 time windows (every six months for three years and 12 months subsequently).
We studied calcium channel blockers as a potential metabolically "neutral" control and expected that their use would be similar to that of β blockers, diuretics, and statins and, therefore, would have a similar potential for unmeasured confounding. However, calcium channel blocker use should not have any adverse or beneficial metabolic impact. As a negative control, to evaluate our methodology, we also assessed the relation between receiving a calcium channel blocker and subsequent progression to new onset diabetes. β blockers, diuretics, statins, and calcium channel blockers were prescribed to patients in the NAVIGATOR trial as part of routine clinical care, and recorded subsequent to randomisation.
Although many patients were taking cardiovascular therapies at baseline, this cohort represents a heterogeneous group with unknown, differential lengths of exposure to treatment and unknown circumstances preceding treatment initiation. In addition, patients who developed diabetes, potentially as a consequence of these drugs, would not be eligible for the study, leaving a biased sample of patients taking drugs at baseline. To avoid these biases, we evaluated treatment initiation in a population that was treatment naïve to each class of drug at baseline. Thus, of 9,306 patients enrolled in NAVIGATOR, four unique subgroups were identified for evaluating each therapy (Figure): β blockers (n=5,640), diuretics (n=6,346), statins (n=6,146), and calcium channel blockers (n=6,294). In this study, the median follow-up time for diabetes was five years from baseline, with a maximum of six years.
(Enlarge Image)
Figure 1.
Patients enrolled in NAVIGATOR trial and their use of drugs of interest at baseline. CCB=calcium channel blocker. Cohorts are not mutually exclusive but may receive other drugs
We identified potential confounders in the relation between non-randomised treatments and progression to new onset diabetes through known clinical risk factors, with additional input obtained from consensus of the clinical and analytical team. The reasons for initiating each therapy and variables known to be highly associated with new onset diabetes were listed and we mapped these to the corresponding variable in the NAVIGATOR trial ( Table 1 ). Given that the population was treatment naïve at baseline, post-baseline information on these covariates was possibly informative. Updated values for time dependent confounders were available according to the visit schedule in Table 1 .
Among the treatment naïve populations, we compared baseline characteristics between those who did and did not receive treatment. We summarised continuous characteristics using the median (25th, 75th centile) and compared them using the Student's t test; categorical variables were summarised with frequency and percentage and compared using the Pearson χ or Fisher exact test.
To evaluate the effect of the drugs on progression to diabetes, we constructed four marginal structural models, one for each drug. Unlike traditional methods, this approach can account for treatment selection bias that may depend on baseline and time dependent variables. The marginal structural model closely resembles a Cox model, with baseline covariate adjustment and treatment included as a time dependent covariate. Specifically, the treatment of interest was assigned the value 0 during periods in which treatment was withheld and 1 when treatment was used, potentially reflecting starting, stopping, and subsequent changes. The distinguishing feature of the marginal structural model is that inverse probability of treatment weighting is applied to account for time dependent variables that may be associated with initiating or discontinuing treatment. The propensity to receive treatment is recalculated according to new information at each time interval, and the weights are correspondingly updated. Under the assumption of no unmeasured confounders, baseline or time varying, the observed treatment patterns become pseudorandomised. The resemblance to the Cox model allows estimation of a hazard ratio, with the usual proportional hazards interpretation. We used the Breslow method to handle tied events times (because diabetes was measured only at six or 12 month intervals), which facilitated the use of weights. We censored the follow-up for diabetes at six years, beyond which the data were too sparse to adequately adjust for confounding.
In the model we included as covariates those baseline variables that were previously identified as significantly associated with new onset diabetes in the NAVIGATOR trial. We modelled the propensity to receive treatment at each time window as a function of all time dependent covariates listed in Table 1 , updated to their value at the previous time window. The rationale was that these treatments are often used in conjunction and that known reasons for initiation of one treatment may be associated with the use of other treatments, regardless of a direct impact. Moreover, we adjusted for time dependent measures of fasting glucose and oral glucose tolerance. This adjustment allowed us to balance patients according to precursor measurements of the endpoint itself. To account for selective loss to follow-up, we similarly modelled the propensity to be censored. At each time point, we calculated stabilised weights using inverse probability of treatment and inverse probability of censoring. We fit all continuous variables using a flexible spline to account for possible non-linearity. Approximately 20% of patients were missing a variable of interest at baseline (primarily glycated haemoglobin, HbA1c). Baseline missing data were multiply imputed, and final estimates (and associated standard errors) reflect the combined analysis over five imputed datasets. We measured time dependent variables with varying frequency and additionally exhibited missing data. For time dependent variables we used a last value carried forward approach, within each imputation. We think this is reasonable as it captures the most recent available information for decision making. The measurement schedule is included in the Supplementary File.
As with inverse probability of treatment weighting, it is important to assess and minimise the impact of extreme weights, by which a few patients who are unlikely to receive treatment could exert undue influence on the results. Previous investigators recommend conducting a series of sensitivity analyses, successively truncating the weights to evaluate variation in the results. Extreme weights were rare within the six year follow-up period. We report the primary results based on minimal truncation of weights at the 0.25 and 99.75 centiles. Alternative truncation strategies did not change the results and are reported in the Supplementary File.
As with a time dependent Cox model, the primary result of the marginal structural model is a hazard ratio for treatment. To improve clinical interpretability, we sought to illustrate the discrepancy in event rates implied by the observed hazard. We used the marginal structural model to derive adjusted, five year event probabilities for two hypothetical trajectories of the time dependent covariate: no treatment throughout follow-up, or constant treatment throughout follow-up. Specifically, we used the Breslow estimator (1 minus the exponential of the negative empirical cumulative hazard estimate), fixing the level of treatment (0 and 1) and balancing other covariates by weighting (that is, fitting the marginal structural model). This estimation corresponds to the standard calculations for a Cox model, with the addition of weights. The absolute excess risk at five years was calculated as the difference between five year event probabilities, and its inverse was the number needed to harm. Confidence intervals for number needed to harm were calculated according to the Altman method.
As a point of comparison, we also fit the standard Cox proportional hazard model including the treatment of interest as a time dependent covariate, initially unadjusted and then adjusted only for baseline characteristics. This model cannot account for the changes in patients' characteristics that led to treatment decisions and therefore is likely to be biased.
Methods
NAVIGATOR was a multinational, randomised, double blinded, placebo controlled trial examining the effects of valsartan and nateglinide on conversion to type 2 diabetes mellitus and cardiovascular outcomes in patients with impaired glucose tolerance and other cardiovascular risk factors. The study design and results have been previously published, as have the eligibility criteria (see Supplementary Appendix).
Endpoint Definitions
The endpoint of interest was diagnosis of new onset diabetes. We measured fasting plasma glucose every six months for the first three years of follow-up and then annually. Oral glucose tolerance tests were performed yearly. New onset diabetes was defined by a fasting plasma glucose level ≥126 mg/dL (7.0 mmol/L) or a glucose level ≥200 mg/dL (11.1 mmol/L) two hours after an oral glucose tolerance test, confirmed by an oral glucose tolerance test within 12 weeks after the increased glucose value was recorded. We separated the diabetes endpoint into 12 time windows (every six months for three years and 12 months subsequently).
Drugs
We studied calcium channel blockers as a potential metabolically "neutral" control and expected that their use would be similar to that of β blockers, diuretics, and statins and, therefore, would have a similar potential for unmeasured confounding. However, calcium channel blocker use should not have any adverse or beneficial metabolic impact. As a negative control, to evaluate our methodology, we also assessed the relation between receiving a calcium channel blocker and subsequent progression to new onset diabetes. β blockers, diuretics, statins, and calcium channel blockers were prescribed to patients in the NAVIGATOR trial as part of routine clinical care, and recorded subsequent to randomisation.
Study Population
Although many patients were taking cardiovascular therapies at baseline, this cohort represents a heterogeneous group with unknown, differential lengths of exposure to treatment and unknown circumstances preceding treatment initiation. In addition, patients who developed diabetes, potentially as a consequence of these drugs, would not be eligible for the study, leaving a biased sample of patients taking drugs at baseline. To avoid these biases, we evaluated treatment initiation in a population that was treatment naïve to each class of drug at baseline. Thus, of 9,306 patients enrolled in NAVIGATOR, four unique subgroups were identified for evaluating each therapy (Figure): β blockers (n=5,640), diuretics (n=6,346), statins (n=6,146), and calcium channel blockers (n=6,294). In this study, the median follow-up time for diabetes was five years from baseline, with a maximum of six years.
(Enlarge Image)
Figure 1.
Patients enrolled in NAVIGATOR trial and their use of drugs of interest at baseline. CCB=calcium channel blocker. Cohorts are not mutually exclusive but may receive other drugs
Covariate Follow-up
We identified potential confounders in the relation between non-randomised treatments and progression to new onset diabetes through known clinical risk factors, with additional input obtained from consensus of the clinical and analytical team. The reasons for initiating each therapy and variables known to be highly associated with new onset diabetes were listed and we mapped these to the corresponding variable in the NAVIGATOR trial ( Table 1 ). Given that the population was treatment naïve at baseline, post-baseline information on these covariates was possibly informative. Updated values for time dependent confounders were available according to the visit schedule in Table 1 .
Statistical Analysis
Among the treatment naïve populations, we compared baseline characteristics between those who did and did not receive treatment. We summarised continuous characteristics using the median (25th, 75th centile) and compared them using the Student's t test; categorical variables were summarised with frequency and percentage and compared using the Pearson χ or Fisher exact test.
To evaluate the effect of the drugs on progression to diabetes, we constructed four marginal structural models, one for each drug. Unlike traditional methods, this approach can account for treatment selection bias that may depend on baseline and time dependent variables. The marginal structural model closely resembles a Cox model, with baseline covariate adjustment and treatment included as a time dependent covariate. Specifically, the treatment of interest was assigned the value 0 during periods in which treatment was withheld and 1 when treatment was used, potentially reflecting starting, stopping, and subsequent changes. The distinguishing feature of the marginal structural model is that inverse probability of treatment weighting is applied to account for time dependent variables that may be associated with initiating or discontinuing treatment. The propensity to receive treatment is recalculated according to new information at each time interval, and the weights are correspondingly updated. Under the assumption of no unmeasured confounders, baseline or time varying, the observed treatment patterns become pseudorandomised. The resemblance to the Cox model allows estimation of a hazard ratio, with the usual proportional hazards interpretation. We used the Breslow method to handle tied events times (because diabetes was measured only at six or 12 month intervals), which facilitated the use of weights. We censored the follow-up for diabetes at six years, beyond which the data were too sparse to adequately adjust for confounding.
In the model we included as covariates those baseline variables that were previously identified as significantly associated with new onset diabetes in the NAVIGATOR trial. We modelled the propensity to receive treatment at each time window as a function of all time dependent covariates listed in Table 1 , updated to their value at the previous time window. The rationale was that these treatments are often used in conjunction and that known reasons for initiation of one treatment may be associated with the use of other treatments, regardless of a direct impact. Moreover, we adjusted for time dependent measures of fasting glucose and oral glucose tolerance. This adjustment allowed us to balance patients according to precursor measurements of the endpoint itself. To account for selective loss to follow-up, we similarly modelled the propensity to be censored. At each time point, we calculated stabilised weights using inverse probability of treatment and inverse probability of censoring. We fit all continuous variables using a flexible spline to account for possible non-linearity. Approximately 20% of patients were missing a variable of interest at baseline (primarily glycated haemoglobin, HbA1c). Baseline missing data were multiply imputed, and final estimates (and associated standard errors) reflect the combined analysis over five imputed datasets. We measured time dependent variables with varying frequency and additionally exhibited missing data. For time dependent variables we used a last value carried forward approach, within each imputation. We think this is reasonable as it captures the most recent available information for decision making. The measurement schedule is included in the Supplementary File.
As with inverse probability of treatment weighting, it is important to assess and minimise the impact of extreme weights, by which a few patients who are unlikely to receive treatment could exert undue influence on the results. Previous investigators recommend conducting a series of sensitivity analyses, successively truncating the weights to evaluate variation in the results. Extreme weights were rare within the six year follow-up period. We report the primary results based on minimal truncation of weights at the 0.25 and 99.75 centiles. Alternative truncation strategies did not change the results and are reported in the Supplementary File.
As with a time dependent Cox model, the primary result of the marginal structural model is a hazard ratio for treatment. To improve clinical interpretability, we sought to illustrate the discrepancy in event rates implied by the observed hazard. We used the marginal structural model to derive adjusted, five year event probabilities for two hypothetical trajectories of the time dependent covariate: no treatment throughout follow-up, or constant treatment throughout follow-up. Specifically, we used the Breslow estimator (1 minus the exponential of the negative empirical cumulative hazard estimate), fixing the level of treatment (0 and 1) and balancing other covariates by weighting (that is, fitting the marginal structural model). This estimation corresponds to the standard calculations for a Cox model, with the addition of weights. The absolute excess risk at five years was calculated as the difference between five year event probabilities, and its inverse was the number needed to harm. Confidence intervals for number needed to harm were calculated according to the Altman method.
As a point of comparison, we also fit the standard Cox proportional hazard model including the treatment of interest as a time dependent covariate, initially unadjusted and then adjusted only for baseline characteristics. This model cannot account for the changes in patients' characteristics that led to treatment decisions and therefore is likely to be biased.
Source...