All-cause mortality of elderly Australian veterans using COX-2 selective or non-selective NSAIDs: a longitudinal study

Abstract

AIM

To determine hazard ratios for all-cause mortality in elderly Australian veterans taking COX-2 selective and non-selective NSAIDs.

METHODS

Patient cohorts were constructed from claims databases (1997 to 2007) for veterans and dependants with full treatment entitlement irrespective of military service. Patients were grouped by initial exposure: celecoxib, rofecoxib, meloxicam, diclofenac, non-selective NSAID. A reference group was constructed of patients receiving glaucoma/hypothyroid medications and none of the study medications. Univariate and multivariate analyses were performed using Cox proportional hazards regression models. Hazard ratios (HR) and 95% confidence intervals (CI) were estimated for each exposure group against each of the reference group. The final model was adjusted for age, gender and co-prescription as a surrogate for cardiovascular risk. Patients were censored if the gap in supply of study prescription exceeded 30 days or if another study medication was initiated. The outcome measure in all analyses was death.

RESULTS

Hazard ratios and 95% CIs, adjusted for age, gender and cardiovascular risk, for each group relative to the reference group were: celecoxib 1.39 (1.25, 1.55), diclofenac 1.44 (1.28, 1.62), meloxicam 1.49 (1.25, 1.78), rofecoxib 1.58 (1.39, 1.79), non-selective NSAIDs 1.76 (1.59, 1.94).

CONCLUSIONS

In this large cohort of Australian veterans exposed to COX-2 selective and non-selective NSAIDs, there was a significant increased mortality risk for those exposed to either COX-2-selective or non-selective NSAIDs relative to those exposed to unrelated (glaucoma/hypothyroid) medications.

WHAT IS ALREADY KNOWN ABOUT THIS SUBJECT

• Previous studies have found varying impact of exposure to COX-2 selective and non-selective NSAIDs.

WHAT THIS STUDY ADDS

• Individuals receiving a COX-2 selective NSAID had an increased risk of all-cause mortality after correction for age, sex and cardiovascular risk as measured by co-prescription.

• Despite differences in the pharmacokinetic properties of the COX-2 selective inhibitor drugs, our study lends no support to clinicians preferring any one COX-2 selective inhibitor drug, or substituting one for another on the grounds of mortality risk alone.

• The Australian Department of Veterans' Affairs data sets make it possible to conduct timely record linkage studies of all-cause mortality from use of medicines in a large and clinically relevant population.

Introduction

The rise in prescribing and subsequent voluntary withdrawal of rofecoxib has been one of the most widely written about therapeutic stories since thalidomide. There are many similarities between the life histories of the two medications. Both were introduced for a broad indication and then removed because many patients developed a particular range of serious side effects. In the case of thalidomide, only recently has it been reintroduced in a non-vulnerable non-pregnant population. The response to evidence of increased risk of myocardial infarction with rofecoxib was similarly dramatic [1, 2], though not following the initial trial [3]. Rofecoxib withdrawal led to global concerns about the continued use of other COX-2 selective NSAIDs that remained or entered the market [4–6]. There has also been debate as to the cardiotoxicity of non-selective NSAIDs [7–9] and the overall balance of gastrointestinal and cardiovascular risks [10].

Knowledge of the full range of adverse events for therapeutic products is uncertain when they are released on to the market. Quantifying risk estimates for individual medicines is initially limited to adverse event data from clinical trials, usually derived from patients meeting proscriptive selection criteria. Where multiple co-morbidities are involved, inferences are confounded and difficult to unravel. While many studies have assessed the risk of cardiac events when patients are exposed to NSAIDs, only a few have estimated the mortality risk associated with exposure to COX-2 selective and non-selective NSAIDs. A general population study of apparently healthy individuals in Denmark over a 9-year period found an increased risk of death associated with these medications [11]. A further study of the same population found an increased risk of coronary death or non-fatal MI and a trend towards an increased risk of fatal and non-fatal stroke [12]. A shorter retrospective study of US veterans who died after NSAID-associated events (gastrointestinal and cardiovascular) found advancing age, multiple co-morbidities and time spent on a selective or non-selective NSAID predicted mortality [13].

Rapid uptake by prescribers can hasten the recognition of adverse events and these events can be demonstrated in short-term studies. For example, in patients prescribed COX-2 selective NSAIDs in Australia, there was a high proportion at risk of renal and thrombotic adverse events which was identified through co-prescription when the drugs were first released onto the market [14]. This paper examines whether a longitudinal study using an administrative database can provide relevant and rapid information about all-cause mortality from relatively selective and non-selective NSAIDs in an elderly at risk population.

Methods

Ethics

This study was approved by the Human Research Ethics Committees of the Australian Department of Veterans' Affairs and the University of New South Wales.

Study cohorts

The study cohorts were drawn from the administrative databases of the Commonwealth of Australia Department of Veterans' Affairs (DVA) for individuals who were continuous Gold Card Holders (GCHs) for the period of the study. Gold Card Holders are veterans, war widows and dependants with full entitlement to health care, all pharmaceutical items listed on the Repatriation or Pharmaceutical Benefits Schemes and related services for all health needs, whether or not related to military service [15].

The client database contains information on entitled individuals including: unique identifier, dates of birth and death, sex, entitlement level and the start and end date of each continuous period at a specific entitlement level for each client. The pharmaceutical claims database contains data on dispensed pharmaceutical items including: unique identifier, entitlement at time each item was dispensed, sex, date of birth, pharmaceutical item details (item code, name and strength, Anatomical Therapeutic Chemical (ATC) Classification System code, date of supply, packs supplied and number of repeats) [16]. The study period was from 1 January 1997 to 31 December 2007.

Individuals were eligible for inclusion if they had been supplied with continuous treatment of the same selective or non-selective NSAID. The criterion for initial use of a specified study medication was that the patient was not supplied with that study medication in the preceding 12 months. The date of first supply of medication was considered as the index date. The exposure period ended with any of the following events: a gap in supply of the study drug greater than 30 days, initiation of treatment with another selective or non-selective NSAID, or death. The primary study outcome was death, the date of which were reliably captured within the DVA client database. Data on cause of death were not available.

We controlled for confounders by constructing covariates from co-prescription data for pre-specified medicines as surrogates for cardiovascular risk. Risk group 1 comprised individuals co-prescribed medications for the management or prevention of cardiovascular disease (for example statins, angiotensin converting enzyme inhibitors (ACEI), angiotensin-II receptor antagonists (ATIIRA), thiazide or loop diuretics, calcium channel antagonists and β-adrenoceptor blockers). Risk group 2 comprised individuals co-prescribed antithrombotic medications (warfarin, low-dose aspirin and clopidogrel). Risk group 3 comprised individuals co-prescribed oral antidiabetic medications and/or insulin. Risk group 4 comprised individuals co-prescribed medications for rheumatoid arthritis (methotrexate, hydroxychloroquine, leflunomide, gold preparations, penicillamine and sulfasalazine). We considered these covariates separately and as a summed score as proposed by Solomon et al. [17].

Study design

Our study was designed to attribute risk to specific medications with a degree of selectivity for the COX-2 enzyme. We did this by determining hazard ratios (HR) through Cox proportional hazards regression modelling of all-cause mortality in individuals commencing treatment with a NSAID, relative to individuals supplied with an unrelated medication. Based on the literature describing the COX-2 selectivity of different NSAIDs, we pre-specified separate consideration of cohorts supplied with celecoxib, rofecoxib, meloxicam and diclofenac [7, 9, 18], with the remaining non-selective NSAIDs combined as a cohort (see Table S1). The reference group used for these analyses was composed of individuals supplied with medications for glaucoma (ATC code: S01E) or hypothyroidism (ATC code: H03AA) and not supplied with selective or non-selective NSAIDs. Glaucoma and hypothyroid medications were initially proposed by Solomon et al. [17], as they were not anticipated to have an association with cardiovascular events, but identify a group of regular users of the health care system.

Statistical analysis

All statistical analyses were undertaken with Stata version 11 (Statacorp, College Station Tx, USA). Univariate and multivariate analyses were conducted using Cox proportional hazards models to estimate mortality HRs and 95% CIs for each of the study groups against the reference group. Multivariate analyses were initially adjusted for the age and gender of the individual, with subsequent consideration of the nominated co-prescription surrogate markers for cardiovascular risk as described above. The proportional hazards assumption was assessed by examining the log–log plots of the survival curves.

Results

Baseline characteristics across cohorts

We identified a cohort of 218 670 continuous Gold Card Holders (GCHs), of whom 175 676 were new users of COX-2 selective NSAIDs or non-selective NSAIDs. These users were compared with new users of glaucoma/hypothyroid medications (n = 42 994). Baseline characteristics of the study cohorts are shown in Table 1. The non-selective NSAID group comprised users taking naproxen (22.9%), piroxicam (18.6%), ketoprofen (14.2%), indomethacin (16.8%), ibuprofen (18.7%), tiaprofenic acid (2.41%), diflunisal (1.96%), sulindac (2.37%), tenoxicam (1.95%) and mefenamic acid (0.08%).

Table 1.

Baseline characteristics for cohorts

Characteristic	Celecoxib	Rofecoxib	Meloxicam	Diclofenac	Non-selective NSAIDs	Glaucoma/Hypothyroid medications (reference group)	All individuals
n	43 257	22 558	9651	33 951	66 259	42 994	218 670
Age, mean (SD) (years)	77 (9.1)	77 (9.5)	77 (10.3)	75 (9.3)	75 (8.8)	79 (6.5)	77 (8.8)
Age, median (IQR)	79 (75–82)	79 (75–82)	80 (74–84)	77 (73–81)	77 (73–80)	79 (76–83)	78 (74–82)
Male, n (%)	25 006 (58.1%)	12 043 (53.4%)	4654 (48.2%)	22 263 (65.6)	45 527 (68.1%)	22 379 (52.1%)	131 447 (60.1%)
Year identified, n
1997	–	–	–	3 547	9 398	7 564	20 509
1998	–	–	–	8 116	20 936	9 026	38 078
1999	–	–	–	5 858	10 246	4 167	20 271
2000	14 605	–	–	3 168	5 288	3 394	26 455
2001	12 218	8 038	–	3 173	5 597	7 017	36 043
2002	7 180	7 597	1625	3 354	5 278	4 555	29 589
2003	3 785	4 680	1804	2 220	2 961	2 271	17 721
2004	2 687	2 243	1639	1 547	2 102	1 658	11 876
2005	1 049	–	1775	1 461	1 987	1 443	7 715
2006	1 005	–	1601	978	1 441	1 154	6 179
2007	728	–	1207	529	1 025	745	4 234
Risks by co-prescription
Cardiovascular disease	30 517 (70.6)	15 814 (70.1)	7306 (75.7)	21 666 (63.8)	42 459 (64.1)	26 898 (62.6)	144 660 (66.2)
At risk of thrombosis	13 877 (32.1)	7 452 (33.0)	3898 (40.1)	6 931 (20.4)	12 196 (18.4)	10 097 (23.5)	54 451 (24.9)
Rheumatoid arthritis	788 (1.8)	377 (1.7)	146 (1.5)	339 (0.99)	737 (1.1)	426 (0.99)	2 813 (1.3)
Diabetes	3 414 (7.9)	1 704 (7.6)	798 (8.3)	2 197 (6.5)	4 485 (6.8)	3 241 (7.5)	15 839 (7.2)
Median (IQR) duration of follow-up (days)	60 (60–60)	60 (60–60)	60 (60–60)	60 (60–74)	60 (60–133)	60 (60–172)	60 (60–87)

The majority (60%) of the cohort were male and the mean age of individuals in the cohort was 77 (SD = 8.8) years. Approximately 66% and 25% of the cohort were supplied a cardiovascular or antithrombotic medication, respectively; 7.2% were supplied medications for diabetes and 1.3% were supplied medications specific for treatment of rheumatoid arthritis. Entry of individuals into the study was dated from the commencement of supply of the NSAID or reference medication. Individuals supplied rofecoxib entered the cohort over a period of 4 years, reflecting availability on the Australian market. Median duration of follow-up for all individuals in the cohort was 60 (IQR = 60–87) days and maximum duration of follow-up was 9.1 years. Approximately 85 807 patient years of follow-up were represented in the cohort.

The numbers of deaths within all cohorts, person-years of follow-up, unadjusted mortality incidence rates per 100 person-years and median time to death are shown in Table 2.

Table 2.

Unadjusted mortality incidence and median time to death in years for cohorts

Characteristic	Celecoxib	Rofecoxib	Meloxicam	Diclofenac	Non-selective NSAIDs	Glaucoma/Hypothyroid medications (reference group)	All individuals
Number in each cohort	43 257	22 558	9,651	33 951	66 259	42 994	218 670
Number of deaths	729	424	159	507	1217	588	3624
Person-years	13 495.2	7008.7	1971	14367.3	30500.4	18464	85806.58
Mortality incidence/100 PYFU (95% CI)	5.40 (5.02, 5.81)	6.05 (5.50, 6.65)	8.07 (6.91, 9.42)	3.53 (3.23, 3.85)	3.99 (3.77, 4.22)	3.18 (2.94, 3.45)	4.22 (4.09, 4.36)
Median days to death (IQR)	47 (25–72)	42 (22–67)	40 (20–64)	44 (19–69)	45 (22–69)	48 (26–69)	45 (23–69)

PYFU, person-years of follow-up.

Cox proportional hazards modelling

Univariate analysis showed that male gender, increasing age and the supply of co-prescribed cardiovascular or antithrombotic medications, antidiabetic medications or medications specific for the management of rheumatoid arthritis resulted in a significantly increased mortality risk. In the univariate and in both multivariate models (Table 3), all individuals initiating use of a COX-2 selective NSAID or a non-selective NSAID had a significantly higher mortality risk when compared with those initiated treatment with a glaucoma/hypothyroid medication. After adjusting for age, gender and cardiovascular risk factors assessed by co-prescription, individuals supplied with a non-selective NSAID had the highest increased mortality risk, and among the COX-2 selective inhibitor drugs, the risks from highest to lowest were for rofecoxib, meloxicam, diclofenac and celecoxib, respectively.

Table 3.

Hazard ratios and 95% CI for all-cause mortality amongst users of COX-2 selective NSAIDs and non-selective NSAIDs in univariate and two different multivariate models, compared with users of glaucoma/hypothyroid medications

	Univariate	Multivariate model 1*	Multivariate model 2*
Exposure group
Glaucoma/hypothyroid	1.00 (ref)	1.00 (ref)	1.00 (ref)
Celecoxib	1.40 (1.26, 1.56)	1.52 (1.36, 1.70)	1.39 (1.25, 1.55)
Rofecoxib	1.58 (1.40, 1.80)	1.74 (1.53, 1.97)	1.58 (1.39, 1.79)
Meloxicam	1.69 (1.42, 2.02)	1.74 (1.46, 2.07)	1.49 (1.25, 1.78)
Diclofenac	1.16 (1.03, 1.30)	1.42 (1.26, 1.60)	1.44 (1.28, 1.62)
Non-selective NSAIDs	1.36 (1.23, 1.50)	1.71 (1.54, 1.88)	1.76 (1.59, 1.94)
Covariates
Age (per 5-year increase)	1.68 (1.64, 1.73)	1.75 (1.70, 1.80)	1.71 (1.66, 1.76)
Male (vs. female)	1.68 (1.56, 1.81)	1.86 (1.73, 2.00)	1.78 (1.65, 1.91)
Risks through co-prescription
Cardiovascular	1.64 (1.52, 1.77)		1.20 (1.11, 1.30)
Thrombotic	2.37 (2.22, 2.54)		1.77 (1.65, 1.90)
Rheumatoid arthritis	1.61 (1.28, 2.02)		1.59 (1.26, 2.00)
Diabetes	1.65 (1.49, 1.83)		1.49 (1.34, 1.66)

^*Multivariate model 1 is adjusted for gender and age; multivariate model 2 is adjusted for gender, age and cardiovascular risks by co-prescription.

Discussion

In this large cohort of Australian veterans exposed to COX-2 selective and non-selective NSAIDs, there was a significant increased mortality risk for those exposed to NSAIDs relative to those exposed to unrelated (glaucoma/hypothyroid) medications. Our study provides data over the life history of what purported to be a new class of drugs and as such provides a unique contribution from pharmacoepidemiology. The history of prescribing celecoxib and rofecoxib can be traced from early adoption, where the proposition was that the newer drugs were ‘safer NSAIDs’, to disenchantment arising from previously unrecognized serious adverse effect profiles. While the effect size is small and perhaps, therefore reassuring, the high prevalence of use lends public health significance and indicates the value of monitoring with linked databases in the post-marketing phase of the drug life cycle. We were not able to include cause of death in our analysis. However, even though our cohort members were elderly, our study endpoint of all-cause mortality was a valid outcome measure in itself, reflecting the net result of benefits and harms from exposure to the medication(s).

While the relatively selective NSAIDs and the non-selective NSAID group were associated with an increased mortality risk in our study, the lowest relative risk was with celecoxib. The literature about NSAIDs is often contradictory, particularly about celecoxib, where the dose may be important. Increased risk of myocardial infarction and death with celecoxib and other non-selective NSAIDs has been found in large, well-designed epidemiological studies [11, 19, 20] as well as under clinical trial conditions [6]. It may also be relevant that, in contrast to some other studies, our cohort population was elderly with multiple co-morbidities assessed by co-prescription surrogates: 70% had at least one cardiovascular risk factor, 23% had two and 3.03% had three or more risk factors.

Our study used a time to event approach to attribute mortality risk during the period of drug exposure only, and spanned a period from 1997 to 2007. Our median follow-up for cohort members was 60 days, which reflects the stringent censoring rules which isolate individuals supplied with each study medication only and attribute risk to the use of each individual medication (no switching, and gap in supply of 30 days). Because the majority of our cohort had a short period of exposure, the majority of our events also occurred within a short time after drug prescription. Fosbol et al. noted this increase in risk even after short exposure to diclofenac, celecoxib and rofecoxib in their population-based study in Denmark [11]. Indeed, other investigators have reported an increase in cardiovascular outcomes after short-term use of both selective or non-selective NSAIDs, even at low doses [9, 12, 21, 22]. In contrast to these studies and our own, Mangoni et al. reported a reduced risk of incident myocardial infarction, heart failure and all-cause mortality in elderly subjects, but they used a wider definition of NSAID exposure and a case control design assessing risk using a conditional logistic regression approach [23].

In unadjusted analyses, meloxicam was associated with the highest increase in risk of mortality, but this was reduced in the multivariate model where we adjusted for all covariates. Meloxicam was a late entrant to the Australian market, establishing itself at a time when prescribers and the general public were aware of the cardiac risks associated with rofecoxib. Barozzi et al. [24] have demonstrated the switch to meloxicam after the withdrawal of rofecoxib. By 2008, meloxicam had become the most widely subsidized NSAID on the Australian Pharmaceutical Benefits Scheme [25]. In the case of meloxicam, there may be insufficient longitudinal data in these cohorts for a clear pattern to emerge, and therefore continuing monitoring will be essential. Although the excess mortality risk in our cohorts for all of the selective and non-selective NSAID group was small, this increase in attributable risk assumes greater importance with increasing numbers of patients being treated with any individual drug.

There are a number of limitations to our study. The majority of people in our study also received prescriptions for paracetamol, so the contribution and influence of paracetamol on mortality rates could not be determined. As all cohort members in this study had complete coverage of all medications, we believe that most were only taking the prescribed NSAID they were dispensed, and there was only a small chance that some cohort members purchased over-the-counter NSAIDs. Significant differences exist in the pharmacological and pharmacokinetic properties of the NSAIDs. For example, meloxicam, like rofecoxib, has a long half-life. There are marked differences in response to selective inhibitors of COX-2 within and between individuals [26]. Such variability confounds interpretation of observational studies and might be considered a limitation of this study. Because we considered the non-selective NSAIDs as a group and we did not consider dose in our analyses, we were unable to detect differences in the all-cause mortality rates for individual non-selective agents, nor could we assess the effect of dispensed dose of any agent on all-cause mortality. Despite these limitations, the database used in this study allowed for a long period for data extraction covering celecoxib and rofecoxib prescribing from initial launch through strategic marketing initiatives to the uncovering of major adverse events and therapeutic disenchantment. Our data follow-up period was comparable with another large and comprehensive published study, which had a follow-up period of 9 years ending in 2005 [11]. However, we were able to follow-up outcomes longer, that is, for a full 3-year period following the withdrawal of rofecoxib. Abraham et al. [13] had a shorter study period (1 January 2000 to 31 December 2002) and enumerated all-cause mortality in the 365 days after an index NSAID-associated event.

Elderly patients often require multiple medications in the face of age-related changes in pharmacokinetics and pharmacodynamics [27]. The mean age of patients in our cohort was 77 years and more than half of the individuals within each cohort who received NSAID treatment were also supplied with medications related to increased cardiovascular risk. Thus, our study highlighted the importance of adjusting for confounders before determining predictors of all-cause mortality. We have shown this can be done using administrative data sets provided clinicians experienced in the field ensure relevance to clinical practice.

While there may be individuals whose arthritic pain responds better to one NSAID than another, our study lends no support to clinicians preferring to prescribe any one COX-2 selective NSAID, or substituting one for another on the grounds of mortality risk. The attributable risk of mortality in our elderly population with multiple co-morbidities and taking any NSAID is not so large as to justify the clinician too readily moving patients with pain, generally from arthritis, to opioids such as oxycodone and buprenorphine patches.

Competing Interests

There are no competing interests to declare.

Supporting Information

Additional Supporting Information may be found in the online version of this article:

Table S1 Study medications and associated ATC codes

Please note: Wiley-Blackwell are not responsible for the content or functionality of any supporting materials supplied by the authors. Any queries (other than missing material) should be directed to the corresponding author for the article.