Cyclooxygenase-2 selectivity of non-steroidal anti-inflammatory drugs and the risk of myocardial infarction and cerebrovascular accident


Dr. N. S. Abraham, Michael E. DeBakey Veterans Affairs Medical Center, 2002 Holcombe Blvd. (152), Houston, TX 77030, USA.


Aim To assess degree of cyclooxygenase-2 (COX-2) selectivity of a non-steroidal anti-inflammatory drug (NSAID) and risk of myocardial infarction (MI) or cerebrovascular accident (CVA).

Methods Prescription fill data were linked to medical records of a merged VA-Medicare dataset. NSAIDs were categorized by Cox-2 selectivity. Incidence of CVA and MI within 180 days of index prescription was assessed using Cox-proportional hazards models adjusted for gender, race, cardiovascular and pharmacological risk factors and propensity for prescription of highly COX-2 selective NSAIDs.

Results Of 384 322 patients (97.5% men and 85.4% white), 79.4% were prescribed a poorly selective, 16.4% a moderately selective and 4.2% a highly selective NSAID. There were 985 incident cases of MI and 586 cases of CVA in >145 870 person-years. Highly selective agents had the highest rate of MI (12.3 per 1000 person-years; [95% CI: 12.2–12.3]) and CVA (8.1 per 1000 person-years; [95% CI: 8.0–8.2]). Periods without NSAID exposure were associated with lowest risk. In adjusted models, highly selective COX-2 selective NSAIDs were associated with a 61% increase in CVA and a 47% increase in MI, when compared with poorly selective NSAIDs.

Conclusions The risk of MI and CVA increases with any NSAID. Highly COX-2 selective NSAIDs confer the greatest risk.


Non-steroidal anti-inflammatory drugs (NSAIDs) are commonly prescribed analgesics for the treatment of musculoskeletal, post-operative and cancer-related pain.1 They are also associated with significant morbidity and mortality, especially among the elderly.2, 3In vitro and ex vivo models suggest that NSAIDs can be classified as (i) poorly selective agents (i.e. naproxen and ibuprofen); (ii) moderately selective agents that exhibit a preferential selectivity towards the cyclooxygenase-2 (COX-2) isoform (i.e. etodolac, meloxicam, celecoxib and nabumetone); or (iii) highly COX-2 selective agents (i.e. rofecoxib, valdecoxib, parecoxib and lumiracoxib).4, 5In vivo concordance is suggested by clinical experience but has not been clearly demonstrated.6–8

The gastrointestinal literature has shown us that NSAIDs differ in toxicity1, 9–14 and this risk can be reduced, but not eliminated, with COX-2 selective NSAIDs. However, high COX-2 selectivity may increase cardiovascular events15–18 by inhibiting prostacyclin and increasing thrombotic tendencies without the beneficial effect of platelet inhibition derived from COX-1 inhibition.19 Rofecoxib was withdrawn from the market after data revealed excessive cardiac morbidity among patients with colon polyps.20 A black box warning regarding the cardiovascular safety of COX-2 selective agents was issued after reports of excess risk of myocardial infarction (MI) with valdecoxib21 similar in magnitude to that of rofecoxib (13–21/1000 patients) with the inconsistent suggestion of a potential class effect.15–17, 20–27 However, some important clinical questions remain. Is cardiovascular risk specific to certain coxibs, or due to a more general class effect? Furthermore, less is known regarding the risk of NSAIDs and cerebrovascular accident (CVA)28–30, an important morbidity to assess given that COX-2 expression is virtually absent in healthy vasculature but is abundant in atherosclerotic lesions.31

Given that the degree of COX-2 selectivity of an agent is inversely correlated with GI toxicity1, 12–14, we hypothesized that COX-2 selectivity of an NSAID would also correlate with risk of cardio- and cerebrovascular events. Our aim was to examine the biological plausibility of this hypothesis. Thus, we assessed the association between COX-2 selectivity of an NSAID and the risk of MI and CVA in a large sample of elderly patients.


Study design

We performed a retrospective cohort study among veterans ≥65 years prescribed an NSAID or a COX-2 selective NSAID at one of 176 Veterans Affairs (VA) facilities in the United States from 01/01/00–12/31/02. The Baylor College of Medicine Institutional Review Board, Houston, TX, approved the research protocol.

Data sources

Non-steroidal anti-inflammatory drug prescription was identified from the VA-Pharmacy Benefits Management (PBM) database, which contains dispensing elements, including facility, patient identifiers, dates of fill and total quantity of the drug dispensed. Our previously published validation study32 confirmed that the VA-PBM had 100% accuracy in identifying NSAID prescription when compared with the medical record. Patient records were identified and linked using unique identifiers to three other national VA administrative databases: the Patient Treatment File (PTF) from 1999–2003, the Out-Patient Clinic (OPC) File from 1999–2003 and the Beneficiary Identification and Records Locator Subsystem (BIRLS) Death File from 1999–2003. Each individual’s in-patient and out-patient Medicare records were also examined using the MedPAR and out-patient files from 1999–2003, to ensure accurate case ascertainment in non-VA facilities.

The diagnostic and demographic elements of the PTF have been previously validated.33, 34 It contains demographic data, one primary discharge diagnosis and up to nine secondary discharge codes, as well as endoscopic and surgical procedures for hospitalized patients. The OPC contains dates of out-patient visits and records ambulatory patient attendance at up to 15 different clinics/given day, first-listed medical or surgical diagnosis and up to nine additional diagnoses for each medical encounter. The BIRLS death file is a well-validated data source33–35, documenting date of deaths as reported to the Veterans Benefits Administration by the VA cemetery system, funeral directors, VA Medical Centers and the Social Security Administration. Additional deaths were obtained from the PTF and Medicare files.

Study population

Veterans aged 65–99 years, with a filled prescription for an NSAID, at standard musculoskeletal dose, salicylate > 325 mg/day, or coxib between 01/01/00 and 12/31/02, were eligible for study entry. Inclusion criteria included a history of prior in-patient or out-patient VA encounters in the 365 days prior to index prescription, evidence of at least a 5-day supply of NSAID, evidence of continuing use of VA facilities 60 days following index prescription as defined by any in-patient or out-patient encounter or a filled prescription for any other medication and no death.

Exposure of interest

The first NSAID or coxib prescription per patient in the study period was considered the index prescription. We included prescriptions of the following agents in our cohort: rofecoxib, valdecoxib, celecoxib, etodolac, nabumetone, meloxicam, naproxen, naproxen sodium and ibuprofen. Our classification of COX-2 selectivity was based on previously published data.5 Poorly selective agents included naproxen, naproxen sodium and ibuprofen. Moderately selective agents included celecoxib, etodolac, meloxicam and nabumetone. Rofecoxib and valdecoxib were considered highly COX-2 selective agents. Exposure was considered in a time-varying fashion, assessed on a daily basis from pharmacy fill records.

Outcomes of interest

Acute MI We identified MI with International Classification of Diseases (ICD), 9th revision, Clinical Modification of (ICD-9) code 410 (acute MI) in the first diagnostic position. Validation studies of the VA36 and Medicare37 administrative data indicate that a first diagnostic code 410 has a positive predictive value for acute MI of 92–96%. We excluded patients in whom MI had been ruled out (i.e. discharged alive within 24 h) and MI that occurred up to 90 days following cardiac procedures and cases with a modifying code indicating an event in the prior 8 weeks, as previously validated.36

CVA Cerebrovascular accident was defined37, 38 by ICD-9 codes 431, 432.9, 434–434.9 and 436 in the first diagnostic position. The positive predictive value of these codes is 91% in the VA population38 and 94% in the Medicare population37 when validated against patient charts.

Potential confounding variables and risk factors

In addition to age, gender and race, clinical co-morbidities known to be associated with cardiovascular disease were assessed in the 365 days prior to index prescription. The co-morbidities included hypertension, a history of ischemic heart disease, diabetes, congestive heart failure, peripheral arterial disease, previous MI or CVA, coronary or carotid revascularization procedures, atrial fibrillation, chronic renal failure and rheumatological diseases. We also identified potential pharmacologic risk factors, including anticoagulants, anti-platelet agents, statins, low-dose ASA use, acetaminophen and steroids 60 days prior to index prescription with overlap with NSAID prescription. We restricted ascertainment to this period to reduce the potential for information bias due to varying lengths of covariate assessment. We also assessed the potential for prescription channelling by calculating the conditional probability (i.e. the propensity score)39 for prescription of a highly or moderately COX-2 selective NSAID. The propensity score was used as a covariate to adjust the regression models as previously described.40, 41

Analytic methods

Each day of NSAID exposure was assessed for the presence of NSAID and NSAID selectivity. An individual exposure period started at t0 and ended with termination of days’ supply or gap in fill dates of 60 days. A gap of 60 days was chosen based on the distribution of days’ supply observed in our prior published work2. NSAID selectivity was considered as a four-level class variable, including time spent on no NSAID. If more than one NSAID was prescribed on any given day, the highest selectivity group was assigned to that day.

The cohort was stratified by NSAID selectivity. Using the chi-square test for categorical variables and the anova (parametric and non-parametric) for continuous variables, we tested for differences in demographic and clinical characteristics among sub-groups. The incidence density (number of events/person-years of follow-up) of MI and CVA was calculated for the 180 days following index prescription. Incidence density ratios were calculated to compare rates occurring among patients prescribed a highly or moderately COX-2 selective NSAID or no NSAID vs. those prescribed a poorly selective NSAID.

A Cox-proportional hazards model stratified by gender and race was constructed for each outcome of interest, while adjusting for potential confounding variables. A multivariate stepwise procedure selected covariates for inclusion in the final model. Patients were censored for completion of the follow-up period without an event, last day of NSAID exposure or death. The Wald’s chi-square test was used to test for the significance of the influence of each independent variable. Hazard ratios (HR) and their 95% confidence intervals (CI) were calculated. Planned subgroup analyses included CVA risk among low-risk patients, defined as patients without a history of carotid revascularization, CVA or atrial fibrillation in the 365 days before NSAID prescription, and no evidence of anticoagulant or anti-platelet agent use at index NSAID prescription. We also performed a subgroup analysis of MI risk among low-risk patients by excluding those with a history of MI, percutaneous coronary intervention or coronary artery bypass graft (CABG) in the 365 days prior to index. Finally, to distinguish between MI and CVA risk among specific agents, we performed a subgroup analysis among patients prescribed only rofecoxib, celecoxib, etodolac, nabumetone, ibuprofen or naproxen.


We identified 384 322 veterans prescribed an NSAID of interest and meeting inclusion criteria. Most (79.4%) were prescribed a poorly selective NSAID, 16.4% were prescribed a moderately selective agent and only 4.2% were prescribed a highly selective agent at index. Overall, 97.5% of the cohort was male and 85.4% were white. Diabetes was observed in 11.1% of the cohort, hypertension in 23.0% and peripheral arterial disease in 2.6%. Baseline characteristics of the cohort, stratified by NSAID selectivity, are highlighted in Table 1.

Table 1.   Baseline characteristics of patients with NSAID prescriptions (N = 384 322), as assessed at index prescription stratified by COX-2 selectivity
CharacteristicLowMedHighP value
  1. IHD, ischemic heart disease; ASA, acetylsalicylic acid; PUD, peptic ulcer disease; MI, myocardial infarction; CVA, cerebrovascular accident (stroke); PCI/CABG, percutaneous coronary stenting/coronary arterial bypass graft; NSAID, non-steroidal anti-inflammatory drug; COX-2, cyclooxygenase-2.

 305 129 (79.4%)62 934 (16.4%)16 259 (4.2%) 
Average age (s.d.)     73.0 (5.5)    74.0 (5.6)    74.8 (5.7)<0.001
Male (vs. Female) (%)98.097.697.0<0.0001
 White (%)80.987.688.2<0.0001
 Black (%)<0.0001
 Hispanic (%)<0.0001
 Other (%)<0.0001
Atrial fibrillation (%)<0.0001
Congestive heart failure (%)<0.0001
Diabetes (%)11.210.611.5<0.0001
History of IHD (%)9.210.312.5<0.0001
Hypertension (%)22.822.723.50.07
Peripheral arterial disease (%)<0.0001
Carotid revascularization (%)<0.001
Low-dose ASA use (%)<0.0001
Chronic renal failure (%)<0.0001
Anti-platelet agent use (%)<0.0001
Anticoagulant use (%)<0.0001
Statin use (%)31.034.637.4<0.0001
Chronic NSAID (%)29.921.220.0<0.0001
High-dose NSAID (%)<0.0001
History of PUD (%)<0.0001
Rheumatologic disease (%)<0.0001
Steroid use (%)<0.0001
History of MI (%)
History of CVA (%)

Patients prescribed a highly selective agent at index were more likely to have a history of cardiovascular events (ischemic heart disease, MI, CVA and carotid revascularization). Patients prescribed a highly selective NSAID at index were also more likely to be prescribed an anti-platelet agent (i.e. ticlopidine and clopidogrel), anticoagulant or steroid, and have a history of rheumatologic disease or peptic ulcer disease, congestive heart failure, atrial fibrillation or chronic renal failure. In contrast, patients with chronic NSAID prescriptions (days’ supply ≥ 90 days) or low-dose ASA use were prescribed a low selectivity agent at index. A greater proportion of these patients also had a history of prior CABG or percutaneous coronary stenting.

Outcomes of interest

MI There were 985 incident cases of MI in 145 870 person-years of observation. Table 2 shows the rates of MI in three subgroups of interest and among patients who were not on an NSAID at the time of their event. Seventy-four events occurred among patients on highly selective NSAIDs in 6037 person-years of follow-up or 12.3 per 1000 person-years of exposure. The incidence density of MI among patients on a low or moderately selective agent was similar, 7.4 per 1000 person-years and 7.6 per 1000 person-years, respectively. The lowest incidence density was observed among patients who were not on any NSAID at the time of their event (3.9 per 1000 person-years).

Table 2.   Incidence density of MI (n = 985) among elderly patients prescribed an NSAID from 01/01/02–12/31/02 (N = 384 322)
COX-2 SelectivityEvents (N)Follow-Up (PY)Incidence Density (95% CI)/1000 PYIncidence Density Ratio† (95% CI)
  1. † Incidence density ratio given the reference category is NSAID (poor selectivity).

  2. PY, person-years; CI, confidence interval; NSAID, non-steroidal anti-inflammatory drug; COX-2, cyclooxygenase-2.

No NSAID14837 6703.93 (3.91–3.95)0.53 (0.19–0.84)
Poor60481 2317.44 (7.42–7.45)Reference
Moderate15920 9337.60 (7.56–7.63)1.02 (1.01–1.04)
High74603712.23 (12.17–12.35)1.65 (1.30–2.38)

Univariate analysis demonstrated that prescription of highly selective NSAIDs was associated with a 56% increase in MI events when compared with prescription of low selectivity agents. The absence of exposure to NSAIDs was most protective (HR 0.65; 95% CI: 0.54–79). A number of known cardiovascular risk factors were also strongly associated with MI, including a past history of MI (HR 4.8; 95% CI: 3.5–6.7); congestive heart failure (HR 2.6; 95% CI: 1.9–3.5); a history of ischemic heart disease (HR 2.2; 95% CI: 1.9–2.6); peripheral arterial disease (HR 2.1; 95% CI: 1.5–2.8); diabetes (HR 1.7; 95% CI: 1.4–2.0); and the use of low-dose ASA (HR 1.6; 95% CI: 1.3–1.9), anticoagulants (HR 1.7; 95% CI: 1.3–2.3) or anti-platelet agents (HR 1.8; 95% CI: 1.3–2.4). Multivariate analysis confirmed the significant association of highly selective agents, when adjusted for prior history of ischemic heart disease or MI, congestive heart failure, peripheral arterial disease, diabetes and the use of low-dose ASA, anticoagulants or anti-platelet agents. The absence of NSAID exposure was again shown to be protective (Figure 1).

Figure 1.

 Multivariate analysis of NSAID selectivity and MI among average and low-risk patients. * Where reference category is NSAID of poor selectivity. NSAID, non-steroidal anti-inflammatory drug; and MI, myocardial infarction. Analysis of low-risk patients excludes patients with a history of MI or revascularization procedures, and those with concomitant anticoagulant or anti-platelet agent use.

Among low-risk patients (N = 367 749) there were 845 incident events. Exposure to a highly selective NSAID was again associated with the highest incidence of MI, 11.8 per 1000 person-years of exposure (95% CI: 11.6–11.8 per 1000 person-years) and was most predictive of MI (HR 1.6; 95% CI: 1.2–2.1) when adjusted for age, atrial fibrillation, congestive heart failure, a history of ischemic heart disease, peripheral arterial disease, diabetes, hypertension and low-dose ASA (Figure 1). The absence of NSAID exposure was protective (HR 0.7; 95% CI: 0.6–0.8), with an incidence of 3.8 per 1000 person-years (3.8–3.9 per 1000 person-years).

CVA A total of 586 incident events were observed in 145 960 person-years of follow-up. The incidence of CVA among highly, moderately or poorly selective agents was 8.1 (95% CI: 8.0–8.2), 5.4 (95% CI: 5.3–5.4) and 4.2 (95% CI: 4.2–4.3) per 1000 person-years of exposure, respectively. The lowest incidence of CVA was observed among patients with no exposure to NSAIDs (2.2 per 1000 person-years; 95% CI: 2.2–2.3 per 1000 person-years). Multivariate analysis confirmed the significant association of highly selective agents (HR 1.6; 95% CI: 1.2–2.2) when adjusted for the presence of atrial fibrillation (HR 1.5; 95% CI: 0.9–2.3), a history of ischemic heart disease (HR 1.3; 95% CI: 0.99–1.6) or history of CVA (HR 6.2; 95% CI: 4.6–8.4), diabetes (HR 1.5; 95% CI: 1.2–1.9) and the use of anticoagulants (HR 1.4; 95% CI: 0.9–1.9) or anti-platelets (HR 2.3; 95% CI: 1.7–3.0) (Figure 2).

Figure 2.

 Multivariate analysis of NSAID selectivity and CVA among average and low-risk patients. * Where reference category is NSAID of poor selectivity. NSAID, non-steroidal anti-inflammatory drug; and CVA, cerebrovascular accident. Low-risk cohort excludes patients with a history of CVA, carotid revascularization, atrial fibrillation and excludes patients on anticoagulants or anti-platelet agents.

Among low-risk patients, those without a history of atrial fibrillation, CVA, carotid revascularization or evidence of anticoagulant or anti-platelet agent use at the time of index prescription, 447 events occurred in 134 042 person-years of follow-up. The incidence of CVA was once again greatest among patients exposed to highly selective agents (5.8 per 1000 person-years; 95% CI: 5.7–5.8 per 1000 person-years), lower among patients who were on poor (3.7 per 1000 person-years; 95% CI: 3.7–3.8 per 1000 person-years) or moderately selective agents (3.8 per 1000 person-years; 95% CI: 3.8–3.9 per 1000 person-years) and lowest among patients who were not exposed to NSAIDs at the time of their event (1.9 per 1000 person-years; 95% CI: 1.9–2.0 per 1000 person-years). Multivariate analysis revealed no exposure to NSAID as protective for CVA, when adjusted for NSAID selectivity, congestive heart failure, history of ischemic heart disease, peripheral arterial disease and diabetes (Figure 2).

A post hoc analysis examined the influence of NSAID selectivity on type of CVA. The majority of CVA events were ischemic in nature (86%; n = 507), the remaining 14% (n = 79) were coded in the administrative data as haemorrhagic. In a competing risk analysis using Cox regression42, we could not conclude that the degree of NSAID selectivity was significantly different among the two types of stroke (data not shown).

Comparison of individual drugs When the most commonly prescribed NSAIDs in the cohort were examined for differences in event rates, the incidence density of both MI and CVA revealed an increase in event rate among patients prescribed NSAIDs with greater degree of COX-2 selectivity when compared with exposure to naproxen (Table 3). In multivariate models, rofecoxib was associated with the greatest risk for MI among both average (HR 2.5; 95% CI: 1.8–3.4) and low-risk (HR 2.5; 95% CI: 1.8–3.5) individuals. Rofecoxib was also associated with the greatest risk of CVA among average-risk individuals (HR 3.0; 95% CI: 2.0–4.4), as well as in low-risk individuals (HR 2.6: 95% CI: 1.3–4.2). For both MI and CVA, the moderately and poorly selective agents were similar in risk when average- and low-risk individuals were examined (Figures 3 and 4) and compared with periods of no NSAID exposure.

Table 3.   Incidence of MI and CVA among elderly patients prescribed specific NSAIDs
Incidence Density (95% CI)/1000 PYIncidence Density Ratio† (95% CI)Incidence Density (95% CI)/1000 PYIncidence Density Ratio† (95% CI)
  1. † Incidence density ratio given the reference category is naproxen exposure; ‡ Average median daily dose for each NSAID was rofecoxib 25 mg/day, celecoxib 200 mg/day, etodolac 800 mg/day, nabumetone 1000 mg/day, ibuprofen 1800 mg/day and naproxen 1000 mg/day.

  2. MI, myocardial infarction; CVA, cerebrovascular accident; NSAID, non-steroidal anti-inflammatory drug; PY, person-years; CI, confidence interval.

Rofecoxib (6022 PY)12.29 (12.20–12.38)1.76 (1.34–2.68)8.13 (8.06–8.20)1.80 (1.29–3.15)
Celecoxib (10 339 PY)7.25 (7.20–7.31)1.04 (1.02–1.08)4.54 (4.50–4.58)1.00 (1.00–1.01)
Etodolac (8994 PY)8.01 (7.95–8.06)1.15 (1.06–1.31)6.45 (6.39–6.50)1.42 (1.15–2.11)
Nabumetone (1521 PY)7.89 (7.75–8.03)1.13 (1.05–1.27)3.94 (3.84–4.04)0.87 (0.70–0.96)
Ibuprofen (46 456 PY)7.77 (7.75–7.80)1.12 (1.05–1.24)4.00 (3.98–4.02)0.88 (0.73–0.97)
Naproxen (34 437 PY)6.97 (6.94–7.00)Reference4.53 (4.50–4.55)Reference
No NSAID (28 708 PY)3.41 (3.39–3.44)0.49 (0.12–0.85)2.09 (2.07–2.11)0.46 (0.07–0.90)
Figure 3.

 Multivariate analysis examining specific-NSAIDS and risk of MI among average and low-risk patients. * Where reference category is periods of no NSAID exposure. NSAID, non-steroidal anti-inflammatory drug; and MI, myocardial infarction. Analysis of low-risk patients excludes patients with a history of MI or revascularization procedures, and those with concomitant anticoagulant or anti-platelet agent use.

Figure 4.

 Multivariate analysis examining the association between specific-NSAIDS and risk of CVA among average and low-risk patients. * Where reference category is periods of no NSAID exposure. NSAID, non-steroidal anti-inflammatory drug; and CVA, cerebrovascular accident. Low-risk cohort excludes patients with a history of CVA, carotid revascularization or atrial fibrillation and excludes patients on anticoagulants or anti-platelet agents.


This study examined whether the degree of COX-2 selectivity of an NSAID is associated with an increased risk of MI and CVA. We compared the risk conferred by poorly, moderately and highly COX-2 selective agents and found that highly COX-2 selective agents conferred the greatest risk. However, moderately selective agents did not appear to be associated with an increased risk compared with poorly selective agents. Furthermore, among this cohort, the risk of MI or CVA was greatest during periods when the patient had exposure to an NSAID of any kind, when compared with periods without exposure to an NSAID.

Our study confirms the results of prior case-control studies in select populations, a large Health Maintenance Organization (HMO) in California22, five counties in Pennsylvania24, Medicare beneficiaries16 and Tennessee Medicaid15 and offers additional evidence of a differential effect on cardiovascular endpoints among NSAIDs with varying capacities to inhibit the COX-2 enzyme as suggested by Solomon et al.43 in a smaller longitudinal study of Pennsylvania Medicare Beneficiaries. Our results extend the existing literature by providing estimates in a substantially larger national cohort in which each patient’s exposure to NSAIDs and the degree of COX-2 selectivity of that NSAID is measured on a daily basis, thus permitting comparisons between agents on the entire spectrum of COX-2 selectivity with greater precision and without regional biases that may influence prescribing habits. In this fashion, we are able to demonstrate that all NSAIDs increase the risk of cardiovascular endpoints when compared with periods of time when no NSAID exposure is observed in the prescription fill data. We were also able to distinguish that among coxibs, the observed cardiovascular morbidity does not represent a ‘class effect’ but appears to be a function of highly selective COX-2 inhibitors. Our estimates of risk are in keeping with the estimates from the Vioxx Gastrointestinal Outcomes Research (VIGOR) trial (relative risk [RR] 1.4; 95% CI: 1.4–4.0)44 and the Adenomatous Polyp Prevention on Vioxx (APPROVe) trial (RR 1.9; 95% CI: 1.2–3.1)45. Rofecoxib, the most highly COX-2 selective drug in our cohort, was also responsible for greater cardiothrombotic events when individual drugs were compared.

Our results also add to the NSAID literature by being the first to quantify the risk of NSAID-related CVA in the Department of VA, the largest national equal access healthcare system in the United States. There is a paucity of literature regarding the risk of cerebrovascular events.28–30 Reicin et al.46 published a secondary analysis that examined differences in the Anti-Platelet Trialists’ Collaboration composite endpoint that included MI and CVA, as part of the rofecoxib osteoarthritis development program. These results included a blind extension phase during which patients from the placebo or rofecoxib 5 mg were switched after 6 weeks to either diclofenac or one of two doses of rofecoxib (12.5 or 25 mg). They noted no statistically significant difference in investigator-reported cardiothrombotic events when rofecoxib was compared with placebo (RR 1.1; 95% CI: 0.3–3.2). However, only 18 events were reported with a total of 672 person-years of follow-up, which might have limited the ability to demonstrate a clinically significant difference. Konstam et al. performed a pooled analysis of 23 rofecoxib trials involving 28 000 patients and over 14 000 person-years of follow-up.47 Again, no statistical significance was shown in the Anti-Platelet Trialists’ Collaboration composite outcome of cardiothrombotic events, however, total event rate (n = 52) was low (non-fatal cardiac events: 29; non-fatal CVA: 23) and the person-years of exposure to the NSAID insufficient to demonstrate a clinically significant effect. In a pooled analysis of 15 phase II to IV controlled arthritis trials using celecoxib48, the risk of cardiothrombotic events was again non-significant 1.1 (95% CI: 0.7–1.6) for all patients and 0.9 (95% CI: 0.5–1.6) for the subgroup not taking aspirin (∼90%). However, the number of events (30 non-fatal CVA and 77 MI) was again small.

The size of our cohort, the frequency of events and the ability to assess daily risk given exposure to poorly, moderately or highly selective agents or no exposure to NSAIDs is an important feature of this study that helps answer some remaining questions regarding cardiothrombotic risk of all classes of NSAIDs. Our cohort was not limited to arthritis patients alone; rather it included any patient who was dispensed an NSAID of any class at a therapeutic dose for at least 5 days’ supply. Thus, our results are generalizable to a broader group of elderly patients who may have been prescribed an NSAID for other indications. Furthermore, we were able to adjust for important confounders and risk factors such as the use of anti-platelet agents, anticoagulants, cardiovascular risk factors or the propensity for channelling to highly COX-2 selective agents.

Patients who received the highly selective NSAIDs were also the patients with the highest underlying risk for cardiovascular events. For example, patients prescribed highly selective COX-2 agents were older and had a more frequent history of atrial fibrillation, ischemic heart disease and prior MI, peripheral arterial disease, chronic renal failure and rheumatological disease, as well as prior carotid and coronary revascularization procedures. These high-risk patients were those with the greatest risk of NSAID-related upper GI toxicity. After controlling the differences in baseline characteristics and the probability of receiving a highly selective COX-2 agent given their underlying GI bleeding risk factors (i.e. propensity score), patients prescribed the highly selective COX-2 agents still exhibited the greatest risk for cardiovascular events. However, we cannot completely exclude the possibility of residual confounding because of the retrospective nature of the study. In particular, we were unable to completely assess several potential confounding variables including body mass index, tobacco use and socioeconomic status. We do not believe that these variables were differentially distributed among patients prescribed a highly selective NSAID vs. those prescribed a non-selective NSAID as previously shown49 from analysis of the Medicare Current Beneficiary Survey.

There may have been some misclassification bias introduced by our inability to capture over-the-counter NSAID prescriptions, including low-dose ASA. Thus, our cohort may be subject to both selection and misclassification effects. In addition, our cohort was primarily male, Caucasian and elderly, thus limiting the generalizability of our findings. However, our results are still informative, as they demonstrate that highly selective COX-2 agents are no more dangerous in the elderly than in the younger cohort enrolled in the APPROVe and VIGOR trials.

Other strengths of our study include the more complete case ascertainment of events. Veterans who are 65 years and older are eligible to receive care at the VA or in the private sector under Medicare.50 Co-utilization of VA and Medicare services has been well documented among dually eligible veterans.50, 51 By merging the VA administrative data with Medicare files for all dually eligible patients, we can ensure more complete case ascertainment of outcomes that occur at non-VA facilities. However, the potential for misclassification of the out-of-hospital cardiovascular event (acute non-fatal MI or CVA) is possible, given all patients may not present at a VA or Medicare facility for treatment. In addition, we may underestimate the risk of cardiovascular events that manifest as sudden death outside the hospital.22 These events would not be captured by the merged VA-Medicare dataset.

Given our substantial sample size, we were able to confirm the exploratory analysis of Solomon et al.43 suggesting the risk associated with highly selective agents was independent of baseline cardiovascular risk. The examination of the risk of MI and CVA in lower-risk cohorts demonstrated very similar results. In addition, because the need for anticoagulants and non-aspirin anti-platelet agents may be a marker of a higher-risk patient group, patients on these agents at the index date were also excluded in some analyses. Similarly, the lower-risk cohort for the examination of MI excluded patients with established high-risk ischemic heart disease (i.e. patients with a prior history of MI or coronary revascularization procedures or patients on anticoagulants and non-aspirin anti-platelet agents). The examination of these lower-risk subsets allowed us to conclude that the magnitude of increased risk of cardiovascular events with highly selective COX-2 agents was not really different based on the baseline risk factors in this elderly cohort.

Differences in chemical structure may explain why celecoxib appeared less hazardous than rofecoxib in our study. It is possible that cardiovascular risk is attenuated by the reduced capacity of moderately selective agents to inhibit the COX-2 enzyme due to differences in chemical structure. For example, a sulfonamide such as celecoxib differs with regard to bioavailability, half-life and hepatic metabolism when compared with a methlysulfone derivative such as rofecoxib.19, 52 Methylsufone derivatives are more potent inhibitors of COX-2 have longer half-lives and are more selective than sulfonamides in vitro.53

The increased risk in the highly selective COX-2 group is consistent with the expected thrombotic potential of these agents through highly selective COX-2-mediated inhibition of prostacyclin, a potent inhibitor of platelet aggregation, leading to an unopposed prothrombotic state because of the unopposed activity of COX-1-mediated thromboxane A2.54, 55 In addition, the possibility of a contribution of the greater increase in blood pressure, which has been noted with the highly selective COX-2 inhibitors compared with other less selective agents49, 56 in mediating an increase in CVA and MI cardiovascular risk, remains to be fully explored. Finally, recent reports suggest that the prothrombotic effect of rofecoxib may be related to increased oxidation of human low-density lipoprotein (LDL)57, an important step in the development of atheroscelerotic disease. Moderately (i.e. celecoxib and meloxicam) and poorly COX-2 selective NSAIDs (i.e. ibuprofen and naproxen) fail to demonstrate similar LDL oxidation.57 The differential effect of lipid oxidation is not yet clearly established and requires further investigation.


Among elderly veterans, all NSAIDs increase the risk of MI and CVA. However, the relative COX-2 selectivity of NSAIDs is important in determining the cardio- and cerebrovascular risk. All NSAIDs are not equal in the risks they present to patients. Highly selective COX-2 NSAIDs confer the greatest risk of MI and CVA among both high-risk and low-risk populations. Therefore, the use of highly COX-2 selective agents should be used with caution.


Declaration of personal and funding interests: Dr Abraham is supported by an American Society of Gastrointestinal Endoscopy Career Development Award, a Merit Review Award from the Department of Veterans Affairs (VA) (IIR 115–05), an Investigator Initiated Grant from TAP Pharmaceuticals and an Outcomes Research Award from the American Gastroenterological Association Foundation for Digestive Health and Nutrition. Dr El-Serag is supported by a VA Health Services Research Advanced Research Career Development Award (00-013-2). Dr Deswal is the recipient of an Advanced Clinical Research Career Development Award from the VA Cooperative Studies Program (712A).