The study cohort was the object of a previous report on NSAIDs and acute MI, where the methods are described in detail.7
In brief, the original study population, identified using the computerized databases of the universal health care programs of the province of Quebec, Canada, consisted of a random sample of 125 000 residents of the province, 66 years of age or older, who were dispensed an NSAID between Jan. 1, 1999, and June 30, 2002, and had been enrolled in the health plan for a period of at least one year. The date of the first such prescription was taken as cohort entry. To identify people starting NSAID therapy, we excluded those who had received such an agent in the year preceding cohort entry (n
= 1193). In addition, we excluded those who had received ASA but no other NSAID (n
= 1552) or had received prescriptions from 2 or more NSAID categories on the day of cohort entry (n
= 153). The latter criterion made it possible to define mutually exclusive exposure groups. To study the risk of a first event, people with a hospital discharge diagnosis of MI (International Classification of Diseases, Ninth Revision [ICD-9] codes 410 and 412, all diagnostic fields) any time before cohort entry were also excluded (n
= 8168). An additional 7 cohort members were excluded because their follow-up was less than 0 days (possible data errors). The remaining 113 927 people were followed until the earliest of the date of a first study end point, termination of health coverage (i.e., death or emigration), death, or Dec. 31, 2002 (end of study). The beneficiary file, prescription drugs, physician services, hospital admissions and vital statistics databases were linked, using a unique encrypted identifier, to provide individual-level information on health insurance coverage, demographic characteristics, medical and drug history, and vital status. This information was available for a minimum of 5 years preceding cohort entry as well as the duration of follow-up.
The study end point was a first hospital admission with a discharge diagnosis of acute MI (ICD-9 code 410), nonfatal or fatal, occurring any time after cohort entry. The date of admission was taken as the event date (referred to as the index date), and the MI was classified as fatal if the person died within 30 days of this date. The MI was considered a valid study end point if the hospital stay was 3 days or greater unless the person had been transferred to or from another institution or had undergone percutaneous coronary angioplasty.8
All NSAIDs available during the study period, including over-the-counter ASA and ibuprofen, were covered by Quebec's drug benefit program and available without prescribing restrictions. However, only agents obtained by prescription were captured by this database.
Since the results of our previous study demonstrated that the risk of MI was limited to those currently exposed to rofecoxib,7
the present analyses target these people (n
= 239). Current users of celecoxib were also targeted (n
= 287) because data published to date for this COX-2 have been conflicting. People for whom the duration of the last prescription dispensed in the year preceding the index date overlapped with this date were classified as current users; otherwise, people were considered past users. Those who did not receive at least one NSAID prescription in the year preceding the index date were considered non-users during this period. Current users of rofecoxib and celecoxib were categorized according to the following patterns of use: first-time use
(the current prescription represented first exposure to that agent) and prevalent use
(at least one additional prescription for the same agent had been received during follow-up). First-time users were further divided into naive users
(no other NSAIDs had been dispensed during follow-up) or switchers
(at least one other NSAID from a different category had been dispensed in addition to that defining current use). Prevalent users were further divided into users of monotherapy
(only prescriptions for the same agent had been dispensed during follow-up) or switchers
(at least one other NSAID from a different category had been dispensed in addition to 2 or more prescriptions for the COX-2 inhibitor that defined current use).
Two measures of cumulative exposure were used to assess the effect of duration of use in the year preceding the index date: total number of prescriptions dispensed for the same agent, and proportion of days exposed to the same agent, calculated as the sum of the number of days supplied with each prescription divided by 365 days. In this way, we were able to assess the effect of rofecoxib and celecoxib even among intermittent users. For the purposes of the analysis, the number of prescriptions dispensed was divided a priori into 4 categories: 1, 2–4, 5–8, and more than 8 prescriptions, whereas the proportion of days exposed was divided into quartiles.
The time since treatment discontinuation was calculated as the number of days between the date of the last available dose (based on the number of days supplied for the last prescription dispensed) and the index date. For the purposes of assessing the risk of MI after treatment cessation, this variable was divided a priori into 3 categories: 0–7 days, 8–30 days, and more than 30 days.
To study the effect of exposure to rofecoxib and celecoxib in the period preceding the event (i.e., the etiologically relevant time window) while simultaneously controlling for the potentially confounding effect of calendar time, we used a time-matched, nested case–control analysis of the cohort.9,10
This approach has been shown to provide unbiased estimates of the rate ratios that would be obtained from a traditional time-to-event analysis of the full cohort, with little or no loss in precision but significant gains in computational efficiency, particularly when analyzing time-varying exposures within large cohorts.11–13
This is not surprising, given that the primary difference between these 2 analytical techniques is the number of non-cases or control subjects that is retained in the analysis. With the standard time-to-event analysis, all people remaining at risk of the event (i.e., non-cases or control subjects) on the date of the case-event (i.e, the failure time that defines the “risk-set”) are retained in the analysis, whereas with the nested case–control approach only a random sample of the tens of thousands of controls are retained. The significant gains in computational efficiency are obtained from analyzing a subset of the non-cases and from analyzing exposure only in the etiologically relevant time period, thereby avoiding the need to program and evaluate a person's exposure status on a daily basis.
The index (event) date of each case was used to define the risk sets from which people who were still at risk of the event (i.e., non-cases or control subjects) were chosen. For each study end point (case), we randomly selected 20 control subjects matched on month and year of cohort entry and age (within one year) and assigned them the case's index date. Thus, age and follow-up time were identical for cases and control subjects in each risk set. We compared the risk of a first MI associated with various patterns and duration of use of rofecoxib and celecoxib as well as time since treatment discontinuation of users to that of non-users in the year preceding the index date (reference group). Rate ratios for these associations were estimated using conditional logistic regression to account for individual-level matching.14,15
These measures of association are equivalent to the hazard ratios that would be estimated from the corresponding Cox proportional hazards regression.16
Since our previous study results demonstrated no gain in precision with the use of a parsimonious model, all rate ratios were fully adjusted for the potentially confounding effects of conventional determinants of MI, other comorbid conditions, measures of health care utilization, and indices of comorbidity.17–19
With the exception of the latter 2, which were assessed in the year preceding the index date, all other covariates were assessed at baseline (year preceding cohort entry). We identified prognostic factors and comorbid conditions using both hospital admissions and drugs for specific corresponding treatments.
We tested the robustness of our definition of cumulative exposure using intermittent users by repeating the analysis using continuous users only.
This study was approved by the ethics boards of the province of Quebec (Commission d'accès à l'information du Québec) and the Royal Victoria Hospital, McGill University.