Impact of Oral Antihyperglycemic Therapy on All-Cause Mortality Among Patients With Diabetes in the Veterans Health Administration

We used existing Veterans Health Administration (VHA) data from the Diabetes Epidemiology Cohort (DEpiC), which is an ongoing study of virtually all VHA patients with diabetes since October 1996. The creation of the DEpiC has been previously described (9,10). We restricted our study to those who responded to the 1999 Large Health Survey of Veteran Enrollees (LHSVE) between April 1999 and April 2000 (11). The LHSVE selected a stratified random sample of >1,500,000 VHA patients identified in the summer of 1999, representing nearly 43% of the entire VHA enrollee population. The core survey included questions on health status measured by the Veterans RAND 36-item Health Survey, occurrence of medical and mental conditions, BMI, and many sociodemographic factors including health behaviors and social support. Vital status was ascertained from the Beneficiary Identification and Record Locator Subsystem file and data from Centers for Medicare and Medicaid Services to ensure greater accuracy.

From the DEpiC integrated database, we selected those individuals that were at least 18 years old and alive as of 31 December 2000. To minimize the variability in A1C lab assay methodology, we restricted our analyses to those patients who received care from medical facilities that used assays certified by the National Glycohemoglobin Standardization Program (12).

We used a retrospective cohort design. Drug utilization information was available from 1 October 1998 to 30 September 2000, and mortality status was obtained from calendar year 2001. Each patient had a fixed 1-year window of drug exposure, followed by a 15-month period during which drug exposure was not assessed before death or a randomly assigned study termination date. Current use of therapy was assessed during the window of drug exposure, irrespective of prior therapy. We included the criterion of surviving the 15 months after the window of drug exposure due to the concern that many events may occur in the year or so before death that can influence treatment choice and outcomes and that these influences may not represent a stable phase of care and long-term risks or benefits.

Overall, drug exposure was classified into the following cohorts: sulfonylurea monotherapy, metformin monotherapy, metformin plus sulfonylurea, TZD users, and no drug therapy. TZD users consisted of those on monotherapy (7%), those on TZD plus sulfonylurea (64%), those on TZD plus metformin (7%), and those on triple combination therapy (22%). There were few deaths in the various TZD subgroups: TZDs alone (2 deaths), TZD plus metformin (2 deaths), TZD plus sulfonylureas (33 deaths), and TZD plus metformin plus sulfonylureas (11 deaths); therefore, these subgroups were pooled (TZD users). Users of insulin, either alone or in combination, were excluded from all analyses. All doses and dosage forms were combined for each drug class. Users of sulfonylurea monotherapy made up the reference cohort for all analyses because of its large size and relevance to prior literature.

Patients within the VHA receive care from specific medical facilities; therefore, to account for possible intercorrelation between patients in the same facility, we used a mixed-logit model for the mortality outcome in which facility was treated as an independent cluster, and a random effect was added to the intercept (13,14). From the mixed-logit model, odds ratios and 95% CIs were calculated to assess the relationship between drug use and mortality.

Confounding factors were adjusted via the propensity score methodology (15,16). The propensity score represents the likelihood of receiving one of the treatments given an individual's characteristics. This method has the appeal of being able to control for many covariates in a very efficient manner. The list of covariates that were assessed during the window of drug exposure and included in the propensity score can be found in the online appendix (available at http://dx.doi.org/10.2337/dc06-2272). To ease the complexity of using the propensity score methodology for an exposure variable that is not dichotomous, we created five separate propensity scores, using logistic regression, for each of the treatment groups relative to the reference group. The c-statistic was used to test the performance of each propensity score model. The c-statistic is equivalent to the area under the receiver operating curve and ranges from 0.5 to 1, with 1 indicating perfect prediction and 0.5 indicating a chance prediction. The generated propensity scores were then included as a covariate in the respective mixed-logit models.

Several variables from the 1999 LHSVE had relatively high rates of missing observations (up to 8.4%), which were assumed to be at random (17). Incomplete lab profiles also were common but may be due to the fact that a particular test was not done rather than an actual missing observation. Regardless, up to 25.2% did not have complete lab profiles. To preserve the sample size and enhance generalizability to the rest of the VHA population, we used an unavailable category for several covariates, including several continuous covariates that were converted to nominal variables for this purpose.

Several sensitivity analyses were performed to evaluate the robustness of the design and analysis. It is possible that A1C is in the causal pathway between drug treatment and mortality. To evaluate the effect of A1C, we performed the analysis without adjusting for A1C. Based on data from the LHSVE, the VHA was not the sole source of medical care for 44.8% of patients in this cohort; therefore, we also conducted an analysis that included only those who used the VHA as the sole source of medical care. We also used an alternative design in which the window of drug exposure was fixed as of 1 October 1998 through 30 September 2000, mortality was assessed in calendar year 2001, and the period between exposure and mortality was variable for each individual depending on the date of death or the random study termination date. An analysis excluding those with any missing information also was conducted.

We identified 39,721 patients with diabetes who responded to the LHSVE and received care from 125 different medical facilities across the nation. Of these, 19,053 (48.0%) were in the reference (sulfonylurea monotherapy) cohort, 2,988 (7.5%) were metformin monotherapy users, 13,820 (34.8%) were users of metformin in combination with sulfonylureas, 675 (1.7%) were TZD users, and 3,185 (8.0%) were not on any antihyperglycemic therapy.

A selection of patient characteristics is presented in Table 1. Several variables indicate a possible selection bias (confounding by indication), suggesting that the more aggressive treatment cohorts (oral combination therapy with metformin and TZD users) had more advanced diabetes. Mean A1C was 6.6% for those not on drug therapy, ∼7% for both the sulfonylurea and metformin monotherapy cohorts, and ∼8% for the metformin plus sulfonylurea combination and TZD users. In general, relative to the sulfonylurea monotherapy cohort, metformin plus sulfonylurea combination users and TZD users had more advanced disease, as indicated by several diabetes severity indicators, including diabetes duration, A1C, proportion of patients receiving care from a diabetes specialist, and the mean number of diabetes-related physician visits. On the other hand, sulfonylurea monotherapy users were the oldest (aged 68.2 years) and had higher Charlson scores, higher LDL cholesterol, higher creatinine, slightly worse health status, and a higher percentage of patients with a history of myocardial infarction and congestive heart failure relative to metformin use either alone or in combination with sulfonylureas.

The c-statistics from the propensity score models were 0.69 for metformin users, 0.75 for metformin plus sulfonylurea users, 0.85 for TZD users, and 0.71 for nonusers. c-statistics between 0.7 and 0.8 are generally considered as acceptable and between 0.8 and 0.9 as excellent (18). Because several of the propensity score model c-statistics were <0.8, residual confounding caused by some of the strongest predictors of drug use is possible. Therefore, these variables were identified (age, diabetes duration, A1C, creatinine, diabetes-related physician visits, and utilization of lipid-lowering and hypertension medications) and included in the subsequent multivariate analyses along with the propensity score to control for any residual confounding imposed by these variables.

The highest unadjusted mortality rate was for the TZD user cohort (7.1%), and the lowest was among metformin monotherapy users (2.7%). The unadjusted logit models suggested that metformin monotherapy users, metformin plus sulfonylureas users, and those not on drug therapy had a reduced risk of all-cause mortality, whereas TZD users experienced an increased risk of mortality. However, from the multivariate logit models, the risk of all-cause death was attenuated for all groups, and none remained statistically significant. (Table 2)

Table 3 contains the results of the various sensitivity analyses. After testing models without A1C, excluding those who may receive care outside the VHA, and evaluating an alternative study design, we found the original results were robust. In the alternate design, the odds ratio for the TZD users changed direction, but was not significant, and was within the 95% CI of the original analysis. When patients with incomplete data were excluded, sample sizes decreased substantially; there were 9,360 sulfonylurea monotherapy users, 1,607 metformin monotherapy users, 7,497 metformin plus sulfonylurea users, 454 TZD users, and 1,596 who were not on any antihyperglycemic therapy. Despite the changes in sample size, the results are consistent with the original findings.

In this diabetic population of veterans, we found no significant drug effect on all-cause mortality. We had the unique ability to control for many variables that are not typically available in database studies, including standardized lab values and several important patient-reported covariates and well-researched measures of comorbid conditions. Despite this, there may still have been some residual unmeasured confounding for which we were unable to account.

Previous research on the relationship between antihyperglycemic drug use and mortality has yielded mixed results. The UKPDS showed that among overweight patients, metformin reduced the risk of diabetes-related mortality and all-cause mortality versus conventional therapy. However, metformin also was found to increase diabetes-related deaths and all-cause mortality when added to the maximum dose of sulfonylurea (2). Several observational studies also have concluded that metformin, when added to sulfonylurea monotherapy, increased the risk of all-cause mortality (4–6); however, these studies may have been plagued by confounding by indication (19). Our findings, which controlled for many potential confounding variables, found no significant increase in mortality associated with metformin used alone or in combination with sulfonylureas. Recently, studies published by Johnson and colleagues (7,20) concluded that metformin alone and in combination was associated with a reduced risk of mortality. Our sample size and duration of follow-up may have limited our ability to detect any potential beneficial effect of drug therapy on all-cause mortality.

Several additional limitations should be recognized. The VHA databases contain limited information on the use of non-VHA services; however, the DEpiC database used for this analysis does contain health care utilization information extracted from linked Medicare files through September 2000. Given the generous pharmacy benefit in the VHA, the lower economic status of veterans compared with the general U.S. population, and the lack of a Medicare pharmacy benefit during this study period, patients who required prescription pharmacotherapy probably filled their prescriptions at VHA pharmacies. Based on data from the LHSVE, the VHA was the sole source of medical care for 55.2% of patients in this cohort. When we limit our primary results to just those that used the VHA as the sole source of medical care, the direction and magnitude of the odds ratios do not change appreciably.

The accuracy of our findings are only as good as the data that are used to create them, which can be a limitation with some sources of administrative databases. However, the agreement between information in the VHA administrative databases and written medical records has been reported to have adequate reliability for demographics (average κ = 0.92), inpatient diagnoses (diabetes principal diagnosis κ = 0.80, diabetes secondary diagnosis κ = 0.82) (21), and several diabetes quality-of-care measures (A1C <9.5%, κ = 0.93; LDL <130, κ = 0.83) (22). Also, the vital status ascertained through the Beneficiary Identification and Record Locator Subsystem has been shown to be up to 94.5% accurate (23) and when combined with CMS data, as was done for this study, is even higher.

A limitation of the dataset was the duration of retrospective data, hence our need to assess prevalent drug exposure during the window of drug exposure only. We did evaluate the agreement between the mutually exclusive drug cohorts defined by utilization from October 1998 to September 1999 and from October 1999 to September 2000 and found that 88, 71, 71, 91, and 72% of sulfonylurea monotherapy users, metformin monotherapy users, metformin plus sulfonylurea users, insulin users, and those on no drug therapy, respectively, were classified the same in both years. TZD use was less consistent across years (31%), likely due to the availability of products during the time of this analysis. We also assessed the relationship between various levels of exposure of each drug class. When we limited the analysis to only those that used at least 90 or 180 days of drugs in the 12-month window of drug exposure, the conclusions did not change, suggesting that our assumption that any indication of drug use for this prevalent cohort was a good indicator of chronic use.

It should be noted that troglitazone was introduced in January 1997, but due to its association with severe liver toxicity it was withdrawn from the U.S. market on 21 March 2000. Most troglitazone users were switched to one of the other TZDs when it was removed from the market. These characteristics of the TZD market may have influenced prescribing practices around TZDs but should not have impacted our multivariate results. Importantly, we found no excess mortality risk attributable to TZD use.

In conclusion, we found no significant drug effect on all-cause mortality for all oral treatment cohorts relative to sulfonylurea oral monotherapy. Future work should assess whether long-term exposure to oral antihyperglycemic medications reduces all-cause or cause-specific mortality.

This research was supported partially by a grant from the Veteran Affairs Clinical Science Research and Development (to L.M.P.).

The authors thank the Veterans Health Administration Office of Quality and Performance for providing access to the LHSVE data.

The findings and opinions reported here are those of the authors and do not necessarily represent the views of the Veterans Health Administration or any other organizations.