Motomi Mori, PhD

Setting and Subjects

All patients presenting for vascular surgery at 5 regional Department of Veterans Affairs (VA) medical centers between January 1998 and March 2005 (3062 patients) were eligible for study entry. Patients with less than 6 months follow‐up were excluded (42 patients). The study included the remaining 3020 patients (comprising 99% of the original population). Our methods have been previously described.8 In brief, we conducted a retrospective cohort study using a regional VA administrative and relational database containing information on both the outpatient and inpatient environments. A record is generated for every contact a patient makes with the VA healthcare system, including prescription medications, laboratory values, demographic information, International Classification of Diseases, 9th Revision (ICD‐9) codes, and vital status. In addition, we used the national VA death index, the VA Beneficiary Identification and Records Locator Subsystem database, which includes Social Security Administration data, to assess vital status. A patient was considered to have a drug exposure (aspirin, ACE inhibitor, beta‐blocker, or statin) if the patient filled or renewed a prescription for the drug within 30 days before surgery. It was determined how many of these drugs were taken during this period, and in which combinations. The Institutional Review Board (IRB) at the Portland VA Medical Center approved the study with a waiver of informed consent.

Data Elements

For every patient we noted the type of vascular surgery (carotid, aortic, lower extremity bypass, or lower extremity amputation), age, sex, comorbid conditions (hypertension, cerebrovascular disease, cancer, diabetes, hyperlipidemia, chronic obstructive pulmonary disease [COPD], chronic kidney disease [CKD], coronary artery disease [CAD], or heart failure), nutritional status (serum albumin), and other medication use (also defined as filling a prescription within 30 days before surgery [insulin and clonidine]). Insulin use was documented to calculate the revised cardiac risk index (RCRI),21 and clonidine was documented to account for as a confounder.22 The RCRI was assigned to each patient. One point was given for each of the following risk factors: use of insulin, CAD, heart failure, cerebrovascular disease, CKD, and high‐risk surgery (intrathoracic, intraperitoneal, or suprainguinal vascular procedures). These variables were defined by ICD‐9 codes. CKD was defined as either an ICD‐9 code for CKD or a serum creatinine >2 mg/dL. Patients were identified by the index vascular surgery using ICD‐9 codes in the VA database, and data were extracted from both the inpatient and outpatient environments.

Statistical Analysis

Patients were included in the analysis if they either died within 6 months or were followed for at least 6 months. Data management and analyses were performed using SAS software, version 9.0. We conducted the univariate analysis of 6‐month mortality using chi‐square analysis and provided unadjusted relative risk estimates for demographic and clinical variables. Demographic variables included age, sex, year, and site of surgery. Clinical variables included preoperative use of insulin and clonidine, preoperative medical conditions, serum albumin, creatinine, RCRI score, and type of surgery.

Bias due to confounding is a problem for studies that cannot randomize subjects into treatment groups. This bias can often be reduced by adjusting for the potentially confounding variables as covariates in regression models. However, when the number of potential confounders is large, as it was in our study, and the number of events, ie, deaths, is small, the resulting regression model can be unstable and the estimates unreliable.23, 24 In such cases, it is necessary to control for confounding using another method. We chose to use propensity scoring and stratification analyses since these methods enable controlling for a large number of covariates using a single variable.

The study drugs were: aspirin, beta‐blockers, statins, and ACE inhibitors. There are 16 combinations with 120 pairwise statistical comparisons possible for these 4 drug exposures. Instead of these multiple comparisons, we chose 4 classifications of combination drug exposure to examine: all 4 drugs compared to none, 3 drugs compared to none, 2 drugs compared to none, and 1 drug compared to none. Four different propensity scores were generated since we studied 4 different drug exposure classes. For each drug exposure class, propensity analyses were performed by using logistic regression to predict the likelihood of use of the drug of interest using all potential demographic and clinical confounding variables. Each subject received a score corresponding to the probability of their having a drug exposure based on the covariates. Scores were divided into quintiles, and these quintiles were used for stratification in Cochran‐Mantel‐Haenszel analyses. Thus, we were able to test the association of patient survival to 6 months with the category of drug exposure comparisons within 30 days before surgery, while controlling for all aforementioned potential confounders. Results of the Breslow‐Day test for homogeneity indicated that no statistically significant differences existed between the results of the propensity quintiles, so the overall summary statistic was reported. All quintiles achieved a balance in the covariates. However, for the 4 study drug exposure class, there were no deaths for the first (n = 173) and second (n = 176) quintiles (corresponding to lower‐risk patients). We therefore excluded these patients from the final analysis.

Variables used in propensity scores included: age, sex, preoperative medical conditions, preoperative clonidine use, nutritional status (serum albumin), RCRI score, and year and location of surgery. To determine whether the propensity score adjustment removed imbalance among the comparisons of the combination drug classes to the no‐drug‐exposure patients, we evaluated associations between each classification of study drug exposure and predictor variables as compared to no‐drug‐exposure patients with both unadjusted chi‐square and propensity‐adjusted Cochran‐Mantel‐Haenszel analyses.

Patient Characteristics

There were 3020 patients with a median age of 67 years, and interquartile range of 59 to 75 years. Ninety‐nine percent were male, and all patients were assessed for death at 6 months after surgery (Table 1). Ten percent (304) had combination all‐4‐drug exposure, 22% (652) had 3‐drug exposure, 24% (736) had 2‐drug exposure, 26% (783) had 1‐drug exposure, and 18% (545) had no study drug exposures. Eight percent (229) of surgeries were aortic, 28% (861) were carotid, 28% (852) were lower extremity amputation, and 36% (1078) were lower extremity bypass. Twenty‐two percent (665) of patients were low risk, with a RCRI of 0, 60% (1822) were moderate risk with a RCRI of 1 to 2, and 18% (553) were high risk with a RCRI of 3. Overall the 6‐month mortality was 9.7% (294). The 6‐month mortality for carotid endarterectomy was 5.0% (43/861), for lower extremity bypass 7.6% (82/1078), for aorta repair 9.2% (21/229), and for lower extremity amputation 17.4% (148/852).

The most common single‐drug exposure was aspirin, 14% (416), followed by ACE inhibitors, 5% (163) (Table 2). The more common 2‐drug exposures included ACE inhibitors and aspirin, 7% (203), aspirin and beta‐blockers, 5% (161), and aspirin and statins, 5% (141). The common 3‐drug combinations included aspirin, beta‐blockers, and statins, 8% (229); ACE inhibitors, aspirin, and statins, 6% (167); and ACE inhibitors, aspirin, and beta‐blockers, 5% (152). ACE inhibitor exposure was common in all combinations, eg, 20.8% of the 1‐drug group had exposure to an ACE inhibitor, 40.5% in the 2‐drug group, 64.9% in the 3‐drug group, and all patients in the 4‐drug group. Overall, 39.3% of patients in the study had ACE inhibitor exposure. The gross unadjusted mortality for each drug exposure group was 10.6% for the no drug group, 11.2% for the 1‐drug group, 10.1% for the 2‐drug group, 8% for the 3‐drug group, and 7.2% for the 4‐drug group.

During the 6 complete years of the study (1998‐2004) the frequency of combination exposure for all 4 study drugs increased from 3.5% to 13.4%; 3‐drug exposure also increased, 14.7% to 27.8%; 2‐drug exposure remained relatively stable, 24.5% to 22%; and single‐drug exposure declined, 24.9% to 12.7% (Figure 1). Individual study drug exposures over the 6 years of the study generally also increased with respect to the other combinations: ACE inhibitor use increased, 34.5% to 42.5%; beta‐blocker, 27.8% to 53.4%; statin, 22.6% to 52.2%. The exception was aspirin, which was relatively stable, 54.5% in 1998, and 57.2% in 2004 (Figure 2).

Figure 1

Figure 2

We also compared the use of the study drug exposures at 6 months after surgery to use within 30 days before surgery (Table 2). In the VA healthcare system aspirin is cheaper for some patients to purchase over‐the‐counter. Aspirin is likely underestimated in this dataset. The frequency of follow‐up drug exposure at 6 months was overall similar to the drug exposure within 30 days before surgery. When aspirin was 1 of the combination exposures, the frequencies declined, and when aspirin was not 1 of the exposures, the frequencies generally increased. The frequency of no‐drug exposures increased from 18.1% before surgery to 24.5% 6 months after surgery, and the frequency of all 4 drug exposures decreased from 10.1% to 6.1%, respectively.

Univariate Analysis

There were statistically significant differences in 6‐month mortality for the combination drug exposure classes compared to no‐drug exposure; P value for linear trend = 0.018 (Table 1).

Propensity‐adjusted Analysis

Patients categorized in each combination drug exposure group were significantly different in their demographic and clinical characteristics compared to the no‐drug exposure patients using unadjusted chi‐square P values (Appendix Table 1). However, after the propensity adjustments, only hyperlipidemia was statistically different for the combination 4‐drug exposure patients compared to no‐drug exposure patients (Appendix Table 1). All other demographic and clinical characteristics for the comparison of the drug exposure classes to no‐drug exposure patients had statistically nonsignificant propensity‐adjusted P values. The range of propensity score distribution was fairly comparable for each combination drug exposure group. The Breslow‐Day test for homogeneity was not significant among the quintiles for any of the drug exposure classes (Table 3; Appendix Table 2), indicating that there was not a statistically significant difference in stratum‐specific relative risks between the different quintiles. Therefore, the summary adjusted result was reported for each drug exposure group. Patients with all 4 drug exposures (with the first [n = 173] and second [n = 166] quintiles excluded due to zero deaths) compared to no‐drug exposure patients had a marginally significant association with decreased mortality, overall propensity‐adjusted relative risk (aRR) 0.52 (95% confidence interval [CI], 0.26‐1.01; P = 0.052), number needed to treat (NNT) 19; patients with the combination 3‐drug exposure had a significant association with decreased mortality, aRR 0.60 (95% CI, 0.38‐0.95; P = 0.030), NNT 38; as well as patients with combination 2‐drug exposure, aRR 0.68 (95% CI, 0.46‐0.99; P = 0.043), NNT 170 (Table 3). Patients with 1 drug exposure did not have an association with decreased mortality compared to no‐drug exposure patients, aRR 0.88 (95% CI, 0.63‐1.22; P = 0.445).

Setting and Subjects

All patients presenting for vascular surgery at 5 regional Department of Veterans Affairs (VA) medical centers between January 1998 and March 2005 (3062 patients) were eligible for study entry. Patients with less than 6 months follow‐up were excluded (42 patients). The study included the remaining 3020 patients (comprising 99% of the original population). Our methods have been previously described.8 In brief, we conducted a retrospective cohort study using a regional VA administrative and relational database containing information on both the outpatient and inpatient environments. A record is generated for every contact a patient makes with the VA healthcare system, including prescription medications, laboratory values, demographic information, International Classification of Diseases, 9th Revision (ICD‐9) codes, and vital status. In addition, we used the national VA death index, the VA Beneficiary Identification and Records Locator Subsystem database, which includes Social Security Administration data, to assess vital status. A patient was considered to have a drug exposure (aspirin, ACE inhibitor, beta‐blocker, or statin) if the patient filled or renewed a prescription for the drug within 30 days before surgery. It was determined how many of these drugs were taken during this period, and in which combinations. The Institutional Review Board (IRB) at the Portland VA Medical Center approved the study with a waiver of informed consent.

Data Elements

For every patient we noted the type of vascular surgery (carotid, aortic, lower extremity bypass, or lower extremity amputation), age, sex, comorbid conditions (hypertension, cerebrovascular disease, cancer, diabetes, hyperlipidemia, chronic obstructive pulmonary disease [COPD], chronic kidney disease [CKD], coronary artery disease [CAD], or heart failure), nutritional status (serum albumin), and other medication use (also defined as filling a prescription within 30 days before surgery [insulin and clonidine]). Insulin use was documented to calculate the revised cardiac risk index (RCRI),21 and clonidine was documented to account for as a confounder.22 The RCRI was assigned to each patient. One point was given for each of the following risk factors: use of insulin, CAD, heart failure, cerebrovascular disease, CKD, and high‐risk surgery (intrathoracic, intraperitoneal, or suprainguinal vascular procedures). These variables were defined by ICD‐9 codes. CKD was defined as either an ICD‐9 code for CKD or a serum creatinine >2 mg/dL. Patients were identified by the index vascular surgery using ICD‐9 codes in the VA database, and data were extracted from both the inpatient and outpatient environments.

Statistical Analysis

Patients were included in the analysis if they either died within 6 months or were followed for at least 6 months. Data management and analyses were performed using SAS software, version 9.0. We conducted the univariate analysis of 6‐month mortality using chi‐square analysis and provided unadjusted relative risk estimates for demographic and clinical variables. Demographic variables included age, sex, year, and site of surgery. Clinical variables included preoperative use of insulin and clonidine, preoperative medical conditions, serum albumin, creatinine, RCRI score, and type of surgery.

Bias due to confounding is a problem for studies that cannot randomize subjects into treatment groups. This bias can often be reduced by adjusting for the potentially confounding variables as covariates in regression models. However, when the number of potential confounders is large, as it was in our study, and the number of events, ie, deaths, is small, the resulting regression model can be unstable and the estimates unreliable.23, 24 In such cases, it is necessary to control for confounding using another method. We chose to use propensity scoring and stratification analyses since these methods enable controlling for a large number of covariates using a single variable.

The study drugs were: aspirin, beta‐blockers, statins, and ACE inhibitors. There are 16 combinations with 120 pairwise statistical comparisons possible for these 4 drug exposures. Instead of these multiple comparisons, we chose 4 classifications of combination drug exposure to examine: all 4 drugs compared to none, 3 drugs compared to none, 2 drugs compared to none, and 1 drug compared to none. Four different propensity scores were generated since we studied 4 different drug exposure classes. For each drug exposure class, propensity analyses were performed by using logistic regression to predict the likelihood of use of the drug of interest using all potential demographic and clinical confounding variables. Each subject received a score corresponding to the probability of their having a drug exposure based on the covariates. Scores were divided into quintiles, and these quintiles were used for stratification in Cochran‐Mantel‐Haenszel analyses. Thus, we were able to test the association of patient survival to 6 months with the category of drug exposure comparisons within 30 days before surgery, while controlling for all aforementioned potential confounders. Results of the Breslow‐Day test for homogeneity indicated that no statistically significant differences existed between the results of the propensity quintiles, so the overall summary statistic was reported. All quintiles achieved a balance in the covariates. However, for the 4 study drug exposure class, there were no deaths for the first (n = 173) and second (n = 176) quintiles (corresponding to lower‐risk patients). We therefore excluded these patients from the final analysis.

Variables used in propensity scores included: age, sex, preoperative medical conditions, preoperative clonidine use, nutritional status (serum albumin), RCRI score, and year and location of surgery. To determine whether the propensity score adjustment removed imbalance among the comparisons of the combination drug classes to the no‐drug‐exposure patients, we evaluated associations between each classification of study drug exposure and predictor variables as compared to no‐drug‐exposure patients with both unadjusted chi‐square and propensity‐adjusted Cochran‐Mantel‐Haenszel analyses.

Patient Characteristics

There were 3020 patients with a median age of 67 years, and interquartile range of 59 to 75 years. Ninety‐nine percent were male, and all patients were assessed for death at 6 months after surgery (Table 1). Ten percent (304) had combination all‐4‐drug exposure, 22% (652) had 3‐drug exposure, 24% (736) had 2‐drug exposure, 26% (783) had 1‐drug exposure, and 18% (545) had no study drug exposures. Eight percent (229) of surgeries were aortic, 28% (861) were carotid, 28% (852) were lower extremity amputation, and 36% (1078) were lower extremity bypass. Twenty‐two percent (665) of patients were low risk, with a RCRI of 0, 60% (1822) were moderate risk with a RCRI of 1 to 2, and 18% (553) were high risk with a RCRI of 3. Overall the 6‐month mortality was 9.7% (294). The 6‐month mortality for carotid endarterectomy was 5.0% (43/861), for lower extremity bypass 7.6% (82/1078), for aorta repair 9.2% (21/229), and for lower extremity amputation 17.4% (148/852).

The most common single‐drug exposure was aspirin, 14% (416), followed by ACE inhibitors, 5% (163) (Table 2). The more common 2‐drug exposures included ACE inhibitors and aspirin, 7% (203), aspirin and beta‐blockers, 5% (161), and aspirin and statins, 5% (141). The common 3‐drug combinations included aspirin, beta‐blockers, and statins, 8% (229); ACE inhibitors, aspirin, and statins, 6% (167); and ACE inhibitors, aspirin, and beta‐blockers, 5% (152). ACE inhibitor exposure was common in all combinations, eg, 20.8% of the 1‐drug group had exposure to an ACE inhibitor, 40.5% in the 2‐drug group, 64.9% in the 3‐drug group, and all patients in the 4‐drug group. Overall, 39.3% of patients in the study had ACE inhibitor exposure. The gross unadjusted mortality for each drug exposure group was 10.6% for the no drug group, 11.2% for the 1‐drug group, 10.1% for the 2‐drug group, 8% for the 3‐drug group, and 7.2% for the 4‐drug group.

During the 6 complete years of the study (1998‐2004) the frequency of combination exposure for all 4 study drugs increased from 3.5% to 13.4%; 3‐drug exposure also increased, 14.7% to 27.8%; 2‐drug exposure remained relatively stable, 24.5% to 22%; and single‐drug exposure declined, 24.9% to 12.7% (Figure 1). Individual study drug exposures over the 6 years of the study generally also increased with respect to the other combinations: ACE inhibitor use increased, 34.5% to 42.5%; beta‐blocker, 27.8% to 53.4%; statin, 22.6% to 52.2%. The exception was aspirin, which was relatively stable, 54.5% in 1998, and 57.2% in 2004 (Figure 2).

Figure 1

Figure 2

We also compared the use of the study drug exposures at 6 months after surgery to use within 30 days before surgery (Table 2). In the VA healthcare system aspirin is cheaper for some patients to purchase over‐the‐counter. Aspirin is likely underestimated in this dataset. The frequency of follow‐up drug exposure at 6 months was overall similar to the drug exposure within 30 days before surgery. When aspirin was 1 of the combination exposures, the frequencies declined, and when aspirin was not 1 of the exposures, the frequencies generally increased. The frequency of no‐drug exposures increased from 18.1% before surgery to 24.5% 6 months after surgery, and the frequency of all 4 drug exposures decreased from 10.1% to 6.1%, respectively.

Univariate Analysis

There were statistically significant differences in 6‐month mortality for the combination drug exposure classes compared to no‐drug exposure; P value for linear trend = 0.018 (Table 1).

Propensity‐adjusted Analysis

Patients categorized in each combination drug exposure group were significantly different in their demographic and clinical characteristics compared to the no‐drug exposure patients using unadjusted chi‐square P values (Appendix Table 1). However, after the propensity adjustments, only hyperlipidemia was statistically different for the combination 4‐drug exposure patients compared to no‐drug exposure patients (Appendix Table 1). All other demographic and clinical characteristics for the comparison of the drug exposure classes to no‐drug exposure patients had statistically nonsignificant propensity‐adjusted P values. The range of propensity score distribution was fairly comparable for each combination drug exposure group. The Breslow‐Day test for homogeneity was not significant among the quintiles for any of the drug exposure classes (Table 3; Appendix Table 2), indicating that there was not a statistically significant difference in stratum‐specific relative risks between the different quintiles. Therefore, the summary adjusted result was reported for each drug exposure group. Patients with all 4 drug exposures (with the first [n = 173] and second [n = 166] quintiles excluded due to zero deaths) compared to no‐drug exposure patients had a marginally significant association with decreased mortality, overall propensity‐adjusted relative risk (aRR) 0.52 (95% confidence interval [CI], 0.26‐1.01; P = 0.052), number needed to treat (NNT) 19; patients with the combination 3‐drug exposure had a significant association with decreased mortality, aRR 0.60 (95% CI, 0.38‐0.95; P = 0.030), NNT 38; as well as patients with combination 2‐drug exposure, aRR 0.68 (95% CI, 0.46‐0.99; P = 0.043), NNT 170 (Table 3). Patients with 1 drug exposure did not have an association with decreased mortality compared to no‐drug exposure patients, aRR 0.88 (95% CI, 0.63‐1.22; P = 0.445).

Setting and Subjects

All patients presenting for vascular surgery at 5 regional Department of Veterans Affairs (VA) medical centers between January 1998 and March 2005 (3062 patients) were eligible for study entry. Patients with less than 6 months follow‐up were excluded (42 patients). The study included the remaining 3020 patients (comprising 99% of the original population). Our methods have been previously described.8 In brief, we conducted a retrospective cohort study using a regional VA administrative and relational database containing information on both the outpatient and inpatient environments. A record is generated for every contact a patient makes with the VA healthcare system, including prescription medications, laboratory values, demographic information, International Classification of Diseases, 9th Revision (ICD‐9) codes, and vital status. In addition, we used the national VA death index, the VA Beneficiary Identification and Records Locator Subsystem database, which includes Social Security Administration data, to assess vital status. A patient was considered to have a drug exposure (aspirin, ACE inhibitor, beta‐blocker, or statin) if the patient filled or renewed a prescription for the drug within 30 days before surgery. It was determined how many of these drugs were taken during this period, and in which combinations. The Institutional Review Board (IRB) at the Portland VA Medical Center approved the study with a waiver of informed consent.

Data Elements

For every patient we noted the type of vascular surgery (carotid, aortic, lower extremity bypass, or lower extremity amputation), age, sex, comorbid conditions (hypertension, cerebrovascular disease, cancer, diabetes, hyperlipidemia, chronic obstructive pulmonary disease [COPD], chronic kidney disease [CKD], coronary artery disease [CAD], or heart failure), nutritional status (serum albumin), and other medication use (also defined as filling a prescription within 30 days before surgery [insulin and clonidine]). Insulin use was documented to calculate the revised cardiac risk index (RCRI),21 and clonidine was documented to account for as a confounder.22 The RCRI was assigned to each patient. One point was given for each of the following risk factors: use of insulin, CAD, heart failure, cerebrovascular disease, CKD, and high‐risk surgery (intrathoracic, intraperitoneal, or suprainguinal vascular procedures). These variables were defined by ICD‐9 codes. CKD was defined as either an ICD‐9 code for CKD or a serum creatinine >2 mg/dL. Patients were identified by the index vascular surgery using ICD‐9 codes in the VA database, and data were extracted from both the inpatient and outpatient environments.

Statistical Analysis

Patients were included in the analysis if they either died within 6 months or were followed for at least 6 months. Data management and analyses were performed using SAS software, version 9.0. We conducted the univariate analysis of 6‐month mortality using chi‐square analysis and provided unadjusted relative risk estimates for demographic and clinical variables. Demographic variables included age, sex, year, and site of surgery. Clinical variables included preoperative use of insulin and clonidine, preoperative medical conditions, serum albumin, creatinine, RCRI score, and type of surgery.

Bias due to confounding is a problem for studies that cannot randomize subjects into treatment groups. This bias can often be reduced by adjusting for the potentially confounding variables as covariates in regression models. However, when the number of potential confounders is large, as it was in our study, and the number of events, ie, deaths, is small, the resulting regression model can be unstable and the estimates unreliable.23, 24 In such cases, it is necessary to control for confounding using another method. We chose to use propensity scoring and stratification analyses since these methods enable controlling for a large number of covariates using a single variable.

The study drugs were: aspirin, beta‐blockers, statins, and ACE inhibitors. There are 16 combinations with 120 pairwise statistical comparisons possible for these 4 drug exposures. Instead of these multiple comparisons, we chose 4 classifications of combination drug exposure to examine: all 4 drugs compared to none, 3 drugs compared to none, 2 drugs compared to none, and 1 drug compared to none. Four different propensity scores were generated since we studied 4 different drug exposure classes. For each drug exposure class, propensity analyses were performed by using logistic regression to predict the likelihood of use of the drug of interest using all potential demographic and clinical confounding variables. Each subject received a score corresponding to the probability of their having a drug exposure based on the covariates. Scores were divided into quintiles, and these quintiles were used for stratification in Cochran‐Mantel‐Haenszel analyses. Thus, we were able to test the association of patient survival to 6 months with the category of drug exposure comparisons within 30 days before surgery, while controlling for all aforementioned potential confounders. Results of the Breslow‐Day test for homogeneity indicated that no statistically significant differences existed between the results of the propensity quintiles, so the overall summary statistic was reported. All quintiles achieved a balance in the covariates. However, for the 4 study drug exposure class, there were no deaths for the first (n = 173) and second (n = 176) quintiles (corresponding to lower‐risk patients). We therefore excluded these patients from the final analysis.

Variables used in propensity scores included: age, sex, preoperative medical conditions, preoperative clonidine use, nutritional status (serum albumin), RCRI score, and year and location of surgery. To determine whether the propensity score adjustment removed imbalance among the comparisons of the combination drug classes to the no‐drug‐exposure patients, we evaluated associations between each classification of study drug exposure and predictor variables as compared to no‐drug‐exposure patients with both unadjusted chi‐square and propensity‐adjusted Cochran‐Mantel‐Haenszel analyses.

Patient Characteristics

There were 3020 patients with a median age of 67 years, and interquartile range of 59 to 75 years. Ninety‐nine percent were male, and all patients were assessed for death at 6 months after surgery (Table 1). Ten percent (304) had combination all‐4‐drug exposure, 22% (652) had 3‐drug exposure, 24% (736) had 2‐drug exposure, 26% (783) had 1‐drug exposure, and 18% (545) had no study drug exposures. Eight percent (229) of surgeries were aortic, 28% (861) were carotid, 28% (852) were lower extremity amputation, and 36% (1078) were lower extremity bypass. Twenty‐two percent (665) of patients were low risk, with a RCRI of 0, 60% (1822) were moderate risk with a RCRI of 1 to 2, and 18% (553) were high risk with a RCRI of 3. Overall the 6‐month mortality was 9.7% (294). The 6‐month mortality for carotid endarterectomy was 5.0% (43/861), for lower extremity bypass 7.6% (82/1078), for aorta repair 9.2% (21/229), and for lower extremity amputation 17.4% (148/852).

The most common single‐drug exposure was aspirin, 14% (416), followed by ACE inhibitors, 5% (163) (Table 2). The more common 2‐drug exposures included ACE inhibitors and aspirin, 7% (203), aspirin and beta‐blockers, 5% (161), and aspirin and statins, 5% (141). The common 3‐drug combinations included aspirin, beta‐blockers, and statins, 8% (229); ACE inhibitors, aspirin, and statins, 6% (167); and ACE inhibitors, aspirin, and beta‐blockers, 5% (152). ACE inhibitor exposure was common in all combinations, eg, 20.8% of the 1‐drug group had exposure to an ACE inhibitor, 40.5% in the 2‐drug group, 64.9% in the 3‐drug group, and all patients in the 4‐drug group. Overall, 39.3% of patients in the study had ACE inhibitor exposure. The gross unadjusted mortality for each drug exposure group was 10.6% for the no drug group, 11.2% for the 1‐drug group, 10.1% for the 2‐drug group, 8% for the 3‐drug group, and 7.2% for the 4‐drug group.

During the 6 complete years of the study (1998‐2004) the frequency of combination exposure for all 4 study drugs increased from 3.5% to 13.4%; 3‐drug exposure also increased, 14.7% to 27.8%; 2‐drug exposure remained relatively stable, 24.5% to 22%; and single‐drug exposure declined, 24.9% to 12.7% (Figure 1). Individual study drug exposures over the 6 years of the study generally also increased with respect to the other combinations: ACE inhibitor use increased, 34.5% to 42.5%; beta‐blocker, 27.8% to 53.4%; statin, 22.6% to 52.2%. The exception was aspirin, which was relatively stable, 54.5% in 1998, and 57.2% in 2004 (Figure 2).

Figure 1

Figure 2

We also compared the use of the study drug exposures at 6 months after surgery to use within 30 days before surgery (Table 2). In the VA healthcare system aspirin is cheaper for some patients to purchase over‐the‐counter. Aspirin is likely underestimated in this dataset. The frequency of follow‐up drug exposure at 6 months was overall similar to the drug exposure within 30 days before surgery. When aspirin was 1 of the combination exposures, the frequencies declined, and when aspirin was not 1 of the exposures, the frequencies generally increased. The frequency of no‐drug exposures increased from 18.1% before surgery to 24.5% 6 months after surgery, and the frequency of all 4 drug exposures decreased from 10.1% to 6.1%, respectively.

Univariate Analysis

There were statistically significant differences in 6‐month mortality for the combination drug exposure classes compared to no‐drug exposure; P value for linear trend = 0.018 (Table 1).

Propensity‐adjusted Analysis

Patients categorized in each combination drug exposure group were significantly different in their demographic and clinical characteristics compared to the no‐drug exposure patients using unadjusted chi‐square P values (Appendix Table 1). However, after the propensity adjustments, only hyperlipidemia was statistically different for the combination 4‐drug exposure patients compared to no‐drug exposure patients (Appendix Table 1). All other demographic and clinical characteristics for the comparison of the drug exposure classes to no‐drug exposure patients had statistically nonsignificant propensity‐adjusted P values. The range of propensity score distribution was fairly comparable for each combination drug exposure group. The Breslow‐Day test for homogeneity was not significant among the quintiles for any of the drug exposure classes (Table 3; Appendix Table 2), indicating that there was not a statistically significant difference in stratum‐specific relative risks between the different quintiles. Therefore, the summary adjusted result was reported for each drug exposure group. Patients with all 4 drug exposures (with the first [n = 173] and second [n = 166] quintiles excluded due to zero deaths) compared to no‐drug exposure patients had a marginally significant association with decreased mortality, overall propensity‐adjusted relative risk (aRR) 0.52 (95% confidence interval [CI], 0.26‐1.01; P = 0.052), number needed to treat (NNT) 19; patients with the combination 3‐drug exposure had a significant association with decreased mortality, aRR 0.60 (95% CI, 0.38‐0.95; P = 0.030), NNT 38; as well as patients with combination 2‐drug exposure, aRR 0.68 (95% CI, 0.46‐0.99; P = 0.043), NNT 170 (Table 3). Patients with 1 drug exposure did not have an association with decreased mortality compared to no‐drug exposure patients, aRR 0.88 (95% CI, 0.63‐1.22; P = 0.445).

Setting and Subjects

We conducted a retrospective cohort study using a regional Department of Veterans Affairs (VA) administrative and relational database, the Consumer Health Information and Performance Sets (CHIPs), which automatically extracts data from electronic medical records of all facilities in the Veterans Integrated Services Network 20, which encompasses Alaska, Washington, Oregon, and Idaho. CHIPs contains information on both outpatient and inpatient environments, and a record is generated for every contact a patient makes with the VA health care system, which includes picking up prescription medications, laboratory values, demographic information, International Classification of Diseases, 9th Revision (ICD‐9), codes, and vital status. In addition, we used the Beneficiary Identification and Records Locator Subsystem database, which is the national VA death index and includes Social Security Administration data that has been shown to be 90%95% complete for assessing vital status.20

Data for all patients who had vascular surgery at 5 VA medical centers in the region from January 1998 to March 2005 was ascertained. If a patient had a second operation within 2 years of the first, the patient was censored at the date of the second operation. A patient was defined as taking a statin or beta‐blocker if a prescription for either of these medications had been picked up within 30 days before or after surgery. The IRB at the Portland VA Medical Center approved the study with a waiver of informed consent.

Data Elements

For every patient we noted the type of vascular surgery (carotid, aortic, lower extremity bypass, or lower extremity amputation), age, sex, comorbid conditions (hypertension, cerebrovascular disease, cancer, diabetes, hyperlipidemia, chronic obstructive pulmonary disease [COPD], chronic kidney disease [CKD], coronary artery disease [CAD], heart failure), tobacco use, ethnicity, nutritional status (serum albumin), and medication use, defined as filling a prescription within 30 days before surgery (insulin, aspirin, angiotensin‐converting enzyme [ACE] inhibitor, and clonidine). Each patient was assigned a revised cardiac risk index (RCRI) score.21 For each the risk factors: use of insulin, CAD, heart failure, cerebrovascular disease, CKD, and high‐risk surgery (intrathoracic, intraperitoneal, or suprainguinal vascular procedures) 1 point was assigned. These variables were defined according to ICD‐9 codes. CKD was defined as either an ICD‐9 code for CKD or a serum creatinine > 2 mg/dL. Patients were identified by the index vascular surgery using ICD‐9 codes in the CHIPs database, and both inpatient and outpatient data were extracted.

Statistical Analysis

All patients were censored at the point of last contact up to 5 years after surgery to focus on more clinically relevant long‐term outcomes possibly associated with vascular surgery. We conducted 3 separate analyses: (1) statin exposure regardless of beta‐blocker exposure; (2) beta‐blocker exposure regardless of statin exposure, and; (3) combined exposure to statins and beta‐blockers.

Propensity score methods were used to adjust for imbalance in the baseline characteristics between statin users and nonusers, beta‐blocker users and nonusers, and combination statin and beta‐blocker users and nonusers.22, 23 The range of the propensity score distribution was similar in drug users and nonusers in the individual analyses. There was sufficient overlap between the 2 groups in each stratum. To derive propensity scores for the individual drug analyses, statin use and beta‐blocker use were modeled independently with the demographic and clinical variables using stepwise logistic regression with a relaxed entry criterion of = 0.20. Only 1 variable (hyperlipidemia) remained significantly different between statin users and nonusers, and it was included in the subsequent analyses as a potential confounder. The variable albumin had 511 missing values. To keep this variable in the propensity scores, the missing values were replaced by the predicted values of albumin from the multiple linear regression model that included the other demographic variables. The propensity scores were grouped into quintiles and used as a stratification variable in the subsequent analyses. To confirm that the propensity score method reduced the imbalances, the demographic and clinical characteristics of statin and beta‐blocker users and nonusers and combination users and nonusers were compared using Cochran‐Mantel‐Haenzel tests with the respective propensity score as a stratification variable.

For the combined use of both study drugs, we performed univariate analysis with adjustment only for RCRI (as this was a powerful predictor of mortality in our dataset; Table 1) as well as a propensity score analysis in an exploratory manner. There have been limited applications of propensity score methods to multiple treatment groups. Similar to that in the study by Huang et al.,24 we developed a multinomial baseline response logit model to obtain 3 separate propensity scores (statin only vs. none, beta‐blocker only vs. none, and both vs. none). Because of the limited sample size, the data were stratified according to the median split of each propensity score. Each score had similar ranges for each treatment group. All but 5 variables (CAD, hypertension, hyperlipidemia, ACE inhibitor use, and type of surgery) were balanced after accounting for strata. These 5 variables were then included in the final stratified Cox regression model as potential confounders.

To comment on patient‐specific risk by stratification with the RCRI, we used a fixed time point of the 2‐year mortality estimated from the Cox regression model to analyze use of study drugs singly or in combination compared with use of neither.

Chi‐square tests were used to categorize and compare demographic and clinical characteristics of statin users and nonusers, of beta‐blocker users and nonusers, and combination users and nonusers. Survival curves were estimated using the Kaplan‐Meier method and compared using the log‐rank test. Stratified or unstratified Cox regression was used to estimate the hazard ratios of statins and beta‐blockers, with or without adjustment for the propensity score. All analyses were performed using SAS (Statistical Analysis System) software, version 9.1.

Patient Characteristics

The study included 3062 patients whose median age was 67 (interquartile range, 5974; Table 1). Ninety‐nine percent of the study patients were men. Overall, ambulatory use of statins and beta‐blockers was found in 44% and 53% of patients, respectively, and combination use occurred in 30%. Sixty‐one percent of patients had an RCRI of 1 or greater; among them 71% were statin users (Table 2), 68% were beta‐blocker users (Table 3), and 75% were combination users (Table 4). Sixty‐four percent of surgeries were either lower extremity bypass or amputation, 29% were carotid, and 8% aortic. Median follow‐up for all patients was 2.7 years (interquartile range, 1.24.6). Of the whole study cohort, 53% and 62% filled a prescription for a statin or beta‐blocker within 1 year of surgery, respectively, and 58% and 67% filled a prescription within 2 years of surgery, respectively. Overall mortality at 30 days was 3%, at 1 year 14%, and at 2 years 22%.

Univariate Survival Analysis

Univariate Cox regression analysis revealed a strong effect of the composite RCRI, which was predictive of mortality in a linear fashion over the course of the study compared with an RCRI of 0 (Table 1). Univariate analysis showed significant associations with decreased mortality for statins (hazard ratio [HR] = 0.66 [95% CI 0.580.75], P < .0001) and beta‐blockers (HR = 0.74 [95% CI 0.660.84], P = .0001); see Table 1. Of note, compared with that in 1998, mortality did not change for all the years for which data were complete. In addition, compared with taking neither study drug, taking a statin only, a beta‐blocker only, or both was associated with decreased mortality (P = .0002, P = .0079, and P < .0001, respectively; Fig. 1).

Figure 1

Propensity Score Analysis for Single Study Drug

There were significant differences in demographic and clinical characteristics between statin‐users versus statin nonusers, and between beta‐blocker users versus beta‐blocker nonusers. These differences became insignificant after the propensity score adjustment, with the exception of hyperlipidemia for statins, P = .02, which was added to the model as a confounder (Table 2). The distribution of the propensity scores was similar for study drug users and nonusers within each stratum. The association with decreased mortality remained significant after adjusting for propensity score (for statins, HR = 0.78 [95% CI 0.670.92, P = .0050], number needed to treat [NNT] = 22; for beta‐blockers HR = 0.84 [95% CI 0.730.96, P = .0103], NNT = 30; Fig. 2).

Figure 2

Combination Study Drugs and Revised Cardiac Risk Index: Univariate Analysis

We wanted our results to closely model those of combination use of the study drugs by patients in a clinical situation. Therefore, we first examined the effects of ambulatory statins alone, beta‐blockers alone, and a combination of statins and beta‐blockers by univariate analysis. Grouping patients by study drug use has not commonly been reported in the literature. We also examined the statistical interaction between the study drugs and the RCRI. The main‐effects model adequately explained all‐cause mortality, and the statistical interaction between the study drugs and the RCRI was not significant.

The final univariate Cox regression model, which compared use of a statin alone, a beta‐blocker alone, and a statin and beta‐blocker in combination with using neither study drug, demonstrated that the combination of statins and beta‐blockers had an HR over the whole study period of 0.43 (95% CI 0.360.51, P < .0001), statins alone had an HR of 0.59 (95% CI 0.480.72, P < .0001), and beta‐blockers alone had an HR of 0.71 (95% CI 0.610.83, P < .0001).

To clarify the effects of the study drugs on patients at different levels of risk, we stratified patients by the RCRI and evaluated the effects of the study drugs on mortality at 2 years, comparing the results to a referent of taking no study drugs. The use of both a statin and a beta‐blocker consistently produced a relative risk reduction (RRR) of approximately 50% with an NNT of 410, with highly statistically significant results for patients at all levels of risk (Table 5). As patient risk level increased, the NNT decreased, consistent with higher‐risk patients benefiting most from combination therapy with statins and beta‐blockers.

In addition, the range of outcomes can be clearly seen for both patient‐specific risk level and study drug use. For example, overall mortality at 2 years for all patients was 22%. For the study drugs, mortality ranged from 16% for patient who used both a statin and a beta‐blocker to 27% for those patients who used neither study drug. The use of the RCRI showed that the healthiest patients who were taking both a statin and a beta‐blocker did the best, with a 2‐year mortality of 9%, compared with the sickest patients who were taking neither study drug, whose 2‐year mortality was 59%. Use of both study drugs by the sickest patients was associated with a reduction in 2‐year mortality to 32% (P < .0001; Table 5).

Propensity Score Analysis of Use of Combination Study Drugs

Because there was very limited literature to guide us in the use of propensity score analysis of multiple treatment groups, we performed these analyses in an exploratory manner. There were significant differences between combination statin and beta‐blocker users and nonusers. These differences became insignificant after adjusting for propensity score, except for the 5 variables previously mentioned, which were added to the model as potential confounders (Table 4). The propensity‐adjusted Cox regression model comparing use of each study drug alone and in combination with taking neither over the whole study period still showed an association with decreased mortality. The combination of statins and beta‐blockers had an HR of 0.56 (95% CI 0.420.74), P < .0001; statins alone had an HR of 0.79 (95% CI 0.620.99), P = .0472; and beta‐blockers alone had an HR of 0.80 (95% CI 0.670.94), P = .0183.

Combination Study Drugs and Revised Cardiac Risk Index: Propensity Analysis

We performed the stratified Cox regression adjusted for the propensity scores for each level of RCRI and estimated 2‐year mortality. The use of both a statin and a beta‐blocker compared with using none was still consistently statistically significant, with an RRR of approximately 36% and an NNT of 820 for all levels of patient risk (Table 6). Possibly because of the reduced number of patients in each RCRI category, neither single‐agent study drug compared with none showed a statistically significant decrease in mortality at any level of patient‐specific risk (Table 6). Again, higher‐risk patients benefited most from combination therapy.

Study Drug Timing: Subcohort Analysis

A subcohort analysis was performed to clarify the timing of the study drugs. Of the patients taking statins, 69 of 1346 (5.1%) took the drug before surgery only, 119 of 1346 (8.8%) took the drug after surgery only, and 1158 of 1346 (86%) took the drug both before and after surgery. Of the patients taking beta‐blockers, 54 of 1617 (3.3%) took the drug before surgery only, 397 of 1617 (24.6%) took the drug after surgery only, and 1166 of 1617 (72.1%) took the drug both before and after surgery. The use of statins and beta‐blockers had a correlation of 0.29 (contingency coefficient).

Setting and Subjects

We conducted a retrospective cohort study using a regional Department of Veterans Affairs (VA) administrative and relational database, the Consumer Health Information and Performance Sets (CHIPs), which automatically extracts data from electronic medical records of all facilities in the Veterans Integrated Services Network 20, which encompasses Alaska, Washington, Oregon, and Idaho. CHIPs contains information on both outpatient and inpatient environments, and a record is generated for every contact a patient makes with the VA health care system, which includes picking up prescription medications, laboratory values, demographic information, International Classification of Diseases, 9th Revision (ICD‐9), codes, and vital status. In addition, we used the Beneficiary Identification and Records Locator Subsystem database, which is the national VA death index and includes Social Security Administration data that has been shown to be 90%95% complete for assessing vital status.20

Data for all patients who had vascular surgery at 5 VA medical centers in the region from January 1998 to March 2005 was ascertained. If a patient had a second operation within 2 years of the first, the patient was censored at the date of the second operation. A patient was defined as taking a statin or beta‐blocker if a prescription for either of these medications had been picked up within 30 days before or after surgery. The IRB at the Portland VA Medical Center approved the study with a waiver of informed consent.

Data Elements

For every patient we noted the type of vascular surgery (carotid, aortic, lower extremity bypass, or lower extremity amputation), age, sex, comorbid conditions (hypertension, cerebrovascular disease, cancer, diabetes, hyperlipidemia, chronic obstructive pulmonary disease [COPD], chronic kidney disease [CKD], coronary artery disease [CAD], heart failure), tobacco use, ethnicity, nutritional status (serum albumin), and medication use, defined as filling a prescription within 30 days before surgery (insulin, aspirin, angiotensin‐converting enzyme [ACE] inhibitor, and clonidine). Each patient was assigned a revised cardiac risk index (RCRI) score.21 For each the risk factors: use of insulin, CAD, heart failure, cerebrovascular disease, CKD, and high‐risk surgery (intrathoracic, intraperitoneal, or suprainguinal vascular procedures) 1 point was assigned. These variables were defined according to ICD‐9 codes. CKD was defined as either an ICD‐9 code for CKD or a serum creatinine > 2 mg/dL. Patients were identified by the index vascular surgery using ICD‐9 codes in the CHIPs database, and both inpatient and outpatient data were extracted.

Statistical Analysis

All patients were censored at the point of last contact up to 5 years after surgery to focus on more clinically relevant long‐term outcomes possibly associated with vascular surgery. We conducted 3 separate analyses: (1) statin exposure regardless of beta‐blocker exposure; (2) beta‐blocker exposure regardless of statin exposure, and; (3) combined exposure to statins and beta‐blockers.

Propensity score methods were used to adjust for imbalance in the baseline characteristics between statin users and nonusers, beta‐blocker users and nonusers, and combination statin and beta‐blocker users and nonusers.22, 23 The range of the propensity score distribution was similar in drug users and nonusers in the individual analyses. There was sufficient overlap between the 2 groups in each stratum. To derive propensity scores for the individual drug analyses, statin use and beta‐blocker use were modeled independently with the demographic and clinical variables using stepwise logistic regression with a relaxed entry criterion of = 0.20. Only 1 variable (hyperlipidemia) remained significantly different between statin users and nonusers, and it was included in the subsequent analyses as a potential confounder. The variable albumin had 511 missing values. To keep this variable in the propensity scores, the missing values were replaced by the predicted values of albumin from the multiple linear regression model that included the other demographic variables. The propensity scores were grouped into quintiles and used as a stratification variable in the subsequent analyses. To confirm that the propensity score method reduced the imbalances, the demographic and clinical characteristics of statin and beta‐blocker users and nonusers and combination users and nonusers were compared using Cochran‐Mantel‐Haenzel tests with the respective propensity score as a stratification variable.

For the combined use of both study drugs, we performed univariate analysis with adjustment only for RCRI (as this was a powerful predictor of mortality in our dataset; Table 1) as well as a propensity score analysis in an exploratory manner. There have been limited applications of propensity score methods to multiple treatment groups. Similar to that in the study by Huang et al.,24 we developed a multinomial baseline response logit model to obtain 3 separate propensity scores (statin only vs. none, beta‐blocker only vs. none, and both vs. none). Because of the limited sample size, the data were stratified according to the median split of each propensity score. Each score had similar ranges for each treatment group. All but 5 variables (CAD, hypertension, hyperlipidemia, ACE inhibitor use, and type of surgery) were balanced after accounting for strata. These 5 variables were then included in the final stratified Cox regression model as potential confounders.

To comment on patient‐specific risk by stratification with the RCRI, we used a fixed time point of the 2‐year mortality estimated from the Cox regression model to analyze use of study drugs singly or in combination compared with use of neither.

Chi‐square tests were used to categorize and compare demographic and clinical characteristics of statin users and nonusers, of beta‐blocker users and nonusers, and combination users and nonusers. Survival curves were estimated using the Kaplan‐Meier method and compared using the log‐rank test. Stratified or unstratified Cox regression was used to estimate the hazard ratios of statins and beta‐blockers, with or without adjustment for the propensity score. All analyses were performed using SAS (Statistical Analysis System) software, version 9.1.

Patient Characteristics

The study included 3062 patients whose median age was 67 (interquartile range, 5974; Table 1). Ninety‐nine percent of the study patients were men. Overall, ambulatory use of statins and beta‐blockers was found in 44% and 53% of patients, respectively, and combination use occurred in 30%. Sixty‐one percent of patients had an RCRI of 1 or greater; among them 71% were statin users (Table 2), 68% were beta‐blocker users (Table 3), and 75% were combination users (Table 4). Sixty‐four percent of surgeries were either lower extremity bypass or amputation, 29% were carotid, and 8% aortic. Median follow‐up for all patients was 2.7 years (interquartile range, 1.24.6). Of the whole study cohort, 53% and 62% filled a prescription for a statin or beta‐blocker within 1 year of surgery, respectively, and 58% and 67% filled a prescription within 2 years of surgery, respectively. Overall mortality at 30 days was 3%, at 1 year 14%, and at 2 years 22%.

Univariate Survival Analysis

Univariate Cox regression analysis revealed a strong effect of the composite RCRI, which was predictive of mortality in a linear fashion over the course of the study compared with an RCRI of 0 (Table 1). Univariate analysis showed significant associations with decreased mortality for statins (hazard ratio [HR] = 0.66 [95% CI 0.580.75], P < .0001) and beta‐blockers (HR = 0.74 [95% CI 0.660.84], P = .0001); see Table 1. Of note, compared with that in 1998, mortality did not change for all the years for which data were complete. In addition, compared with taking neither study drug, taking a statin only, a beta‐blocker only, or both was associated with decreased mortality (P = .0002, P = .0079, and P < .0001, respectively; Fig. 1).

Figure 1

Propensity Score Analysis for Single Study Drug

There were significant differences in demographic and clinical characteristics between statin‐users versus statin nonusers, and between beta‐blocker users versus beta‐blocker nonusers. These differences became insignificant after the propensity score adjustment, with the exception of hyperlipidemia for statins, P = .02, which was added to the model as a confounder (Table 2). The distribution of the propensity scores was similar for study drug users and nonusers within each stratum. The association with decreased mortality remained significant after adjusting for propensity score (for statins, HR = 0.78 [95% CI 0.670.92, P = .0050], number needed to treat [NNT] = 22; for beta‐blockers HR = 0.84 [95% CI 0.730.96, P = .0103], NNT = 30; Fig. 2).

Figure 2

Combination Study Drugs and Revised Cardiac Risk Index: Univariate Analysis

We wanted our results to closely model those of combination use of the study drugs by patients in a clinical situation. Therefore, we first examined the effects of ambulatory statins alone, beta‐blockers alone, and a combination of statins and beta‐blockers by univariate analysis. Grouping patients by study drug use has not commonly been reported in the literature. We also examined the statistical interaction between the study drugs and the RCRI. The main‐effects model adequately explained all‐cause mortality, and the statistical interaction between the study drugs and the RCRI was not significant.

The final univariate Cox regression model, which compared use of a statin alone, a beta‐blocker alone, and a statin and beta‐blocker in combination with using neither study drug, demonstrated that the combination of statins and beta‐blockers had an HR over the whole study period of 0.43 (95% CI 0.360.51, P < .0001), statins alone had an HR of 0.59 (95% CI 0.480.72, P < .0001), and beta‐blockers alone had an HR of 0.71 (95% CI 0.610.83, P < .0001).

To clarify the effects of the study drugs on patients at different levels of risk, we stratified patients by the RCRI and evaluated the effects of the study drugs on mortality at 2 years, comparing the results to a referent of taking no study drugs. The use of both a statin and a beta‐blocker consistently produced a relative risk reduction (RRR) of approximately 50% with an NNT of 410, with highly statistically significant results for patients at all levels of risk (Table 5). As patient risk level increased, the NNT decreased, consistent with higher‐risk patients benefiting most from combination therapy with statins and beta‐blockers.

In addition, the range of outcomes can be clearly seen for both patient‐specific risk level and study drug use. For example, overall mortality at 2 years for all patients was 22%. For the study drugs, mortality ranged from 16% for patient who used both a statin and a beta‐blocker to 27% for those patients who used neither study drug. The use of the RCRI showed that the healthiest patients who were taking both a statin and a beta‐blocker did the best, with a 2‐year mortality of 9%, compared with the sickest patients who were taking neither study drug, whose 2‐year mortality was 59%. Use of both study drugs by the sickest patients was associated with a reduction in 2‐year mortality to 32% (P < .0001; Table 5).

Propensity Score Analysis of Use of Combination Study Drugs

Because there was very limited literature to guide us in the use of propensity score analysis of multiple treatment groups, we performed these analyses in an exploratory manner. There were significant differences between combination statin and beta‐blocker users and nonusers. These differences became insignificant after adjusting for propensity score, except for the 5 variables previously mentioned, which were added to the model as potential confounders (Table 4). The propensity‐adjusted Cox regression model comparing use of each study drug alone and in combination with taking neither over the whole study period still showed an association with decreased mortality. The combination of statins and beta‐blockers had an HR of 0.56 (95% CI 0.420.74), P < .0001; statins alone had an HR of 0.79 (95% CI 0.620.99), P = .0472; and beta‐blockers alone had an HR of 0.80 (95% CI 0.670.94), P = .0183.

Combination Study Drugs and Revised Cardiac Risk Index: Propensity Analysis

We performed the stratified Cox regression adjusted for the propensity scores for each level of RCRI and estimated 2‐year mortality. The use of both a statin and a beta‐blocker compared with using none was still consistently statistically significant, with an RRR of approximately 36% and an NNT of 820 for all levels of patient risk (Table 6). Possibly because of the reduced number of patients in each RCRI category, neither single‐agent study drug compared with none showed a statistically significant decrease in mortality at any level of patient‐specific risk (Table 6). Again, higher‐risk patients benefited most from combination therapy.

Study Drug Timing: Subcohort Analysis

A subcohort analysis was performed to clarify the timing of the study drugs. Of the patients taking statins, 69 of 1346 (5.1%) took the drug before surgery only, 119 of 1346 (8.8%) took the drug after surgery only, and 1158 of 1346 (86%) took the drug both before and after surgery. Of the patients taking beta‐blockers, 54 of 1617 (3.3%) took the drug before surgery only, 397 of 1617 (24.6%) took the drug after surgery only, and 1166 of 1617 (72.1%) took the drug both before and after surgery. The use of statins and beta‐blockers had a correlation of 0.29 (contingency coefficient).

Setting and Subjects

We conducted a retrospective cohort study using a regional Department of Veterans Affairs (VA) administrative and relational database, the Consumer Health Information and Performance Sets (CHIPs), which automatically extracts data from electronic medical records of all facilities in the Veterans Integrated Services Network 20, which encompasses Alaska, Washington, Oregon, and Idaho. CHIPs contains information on both outpatient and inpatient environments, and a record is generated for every contact a patient makes with the VA health care system, which includes picking up prescription medications, laboratory values, demographic information, International Classification of Diseases, 9th Revision (ICD‐9), codes, and vital status. In addition, we used the Beneficiary Identification and Records Locator Subsystem database, which is the national VA death index and includes Social Security Administration data that has been shown to be 90%95% complete for assessing vital status.20

Data for all patients who had vascular surgery at 5 VA medical centers in the region from January 1998 to March 2005 was ascertained. If a patient had a second operation within 2 years of the first, the patient was censored at the date of the second operation. A patient was defined as taking a statin or beta‐blocker if a prescription for either of these medications had been picked up within 30 days before or after surgery. The IRB at the Portland VA Medical Center approved the study with a waiver of informed consent.

Data Elements

For every patient we noted the type of vascular surgery (carotid, aortic, lower extremity bypass, or lower extremity amputation), age, sex, comorbid conditions (hypertension, cerebrovascular disease, cancer, diabetes, hyperlipidemia, chronic obstructive pulmonary disease [COPD], chronic kidney disease [CKD], coronary artery disease [CAD], heart failure), tobacco use, ethnicity, nutritional status (serum albumin), and medication use, defined as filling a prescription within 30 days before surgery (insulin, aspirin, angiotensin‐converting enzyme [ACE] inhibitor, and clonidine). Each patient was assigned a revised cardiac risk index (RCRI) score.21 For each the risk factors: use of insulin, CAD, heart failure, cerebrovascular disease, CKD, and high‐risk surgery (intrathoracic, intraperitoneal, or suprainguinal vascular procedures) 1 point was assigned. These variables were defined according to ICD‐9 codes. CKD was defined as either an ICD‐9 code for CKD or a serum creatinine > 2 mg/dL. Patients were identified by the index vascular surgery using ICD‐9 codes in the CHIPs database, and both inpatient and outpatient data were extracted.

Statistical Analysis

All patients were censored at the point of last contact up to 5 years after surgery to focus on more clinically relevant long‐term outcomes possibly associated with vascular surgery. We conducted 3 separate analyses: (1) statin exposure regardless of beta‐blocker exposure; (2) beta‐blocker exposure regardless of statin exposure, and; (3) combined exposure to statins and beta‐blockers.

Propensity score methods were used to adjust for imbalance in the baseline characteristics between statin users and nonusers, beta‐blocker users and nonusers, and combination statin and beta‐blocker users and nonusers.22, 23 The range of the propensity score distribution was similar in drug users and nonusers in the individual analyses. There was sufficient overlap between the 2 groups in each stratum. To derive propensity scores for the individual drug analyses, statin use and beta‐blocker use were modeled independently with the demographic and clinical variables using stepwise logistic regression with a relaxed entry criterion of = 0.20. Only 1 variable (hyperlipidemia) remained significantly different between statin users and nonusers, and it was included in the subsequent analyses as a potential confounder. The variable albumin had 511 missing values. To keep this variable in the propensity scores, the missing values were replaced by the predicted values of albumin from the multiple linear regression model that included the other demographic variables. The propensity scores were grouped into quintiles and used as a stratification variable in the subsequent analyses. To confirm that the propensity score method reduced the imbalances, the demographic and clinical characteristics of statin and beta‐blocker users and nonusers and combination users and nonusers were compared using Cochran‐Mantel‐Haenzel tests with the respective propensity score as a stratification variable.

For the combined use of both study drugs, we performed univariate analysis with adjustment only for RCRI (as this was a powerful predictor of mortality in our dataset; Table 1) as well as a propensity score analysis in an exploratory manner. There have been limited applications of propensity score methods to multiple treatment groups. Similar to that in the study by Huang et al.,24 we developed a multinomial baseline response logit model to obtain 3 separate propensity scores (statin only vs. none, beta‐blocker only vs. none, and both vs. none). Because of the limited sample size, the data were stratified according to the median split of each propensity score. Each score had similar ranges for each treatment group. All but 5 variables (CAD, hypertension, hyperlipidemia, ACE inhibitor use, and type of surgery) were balanced after accounting for strata. These 5 variables were then included in the final stratified Cox regression model as potential confounders.