Aliment Pharmacol Ther 2011; 33: 77–88
Background The association between myocardial infarction (MI) and co-administration of proton pump inhibitors (PPIs) and clopidogrel remains controversial.
Aim To quantify the association between concomitant use of PPIs and clopidogrel and occurrence of recurrent MI.
Methods We conducted a case–control study within a cohort of acute MI patients in PHARMO Record Linkage System (1999–2008). The cases were patients readmitted for MI. PPI exposure was categorized as current (3–1 days before MI), past (30–3 days before MI), or no use (>30 days before MI). We used conditional logistic regression analyses.
Results Among 23 655 patients hospitalized following MI, we identified 1247 patients readmitted for MI. Among clopidogrel users, current PPI use was associated with an increased risk of recurrent MI (OR: 1.62, 95% CI: 1.15–2.27) when compared with no PPI use, but not when compared with past PPI use (OR: 0.95, 95% CI: 0.38–2.41). Among clopidogrel non-users, current PPI use was associated with an increased risk of recurrent MI (OR: 1.38, 95% CI: 1.18–1.61) when compared with no PPI use.
Conclusions The apparent association between recurrent MI and use of PPIs with clopidogrel depends on the design, and is affected by confounding by indication. The association is not present when (un)measured confounding is addressed by design.
Clopidogrel is an oral antiplatelet agent commonly used in addition to aspirin to reduce cardiovascular (CV) events. Clopidogrel is converted in the liver from an inactive pro-drug to its active metabolite that binds irreversible to P2RY12 adenosine diphosphate (ADP) receptors on the platelet surface, thereby preventing platelet aggregation.1
Clopidogrel conversion is catalysed by several cytochrome P450 (CYP) enzymes of which CYP2C19 is the most important. Patients with loss-of-function polymorphisms in the gene encoding for CYP2C19 have lower levels of the active metabolite and have reduced platelet inhibition during clopidogrel treatment. This results in higher rates of acute myocardial infarction (MI).2, 3 In contrast, rapid metabolizers of clopidogrel (with CYP2C19 variants leading to increased enzyme activity) have a higher rate of clopidogrel activation and more efficient platelet inhibition.4
Proton pump inhibitors (PPIs) are routinely co-administered with clopidogrel to prevent upper gastrointestinal (UGI) bleeding,5–7 which is in line with expert consensus guidelines.8 PPIs are also metabolized by CYP2C19, and can competitively bind to its catalytic site. Therefore, PPIs are potentially hindering the conversion of clopidogrel to its active metabolite. Both the European Medicines Agency (EMA) and the US Food and Drug Administration (FDA) recently advised against the concurrent use of PPIs and clopidogrel in the absence of a strong indication,9, 10 because several pharmacodynamic11–14 and clinical outcome studies15–20 suggest a detrimental interaction between clopidogrel and PPIs.
Evidence on the association between CV events and co-administration of PPIs with clopidogrel remains inconclusive. Observational studies show a small increased risk of recurrent MI in patients using PPI–clopidogrel combination therapy compared with those using clopidogrel alone. Estimated relative risks vary from 0.92 to 1.93.2, 7, 13, 15–20 These results are likely influenced by confounding by indication. Confounding by indication is introduced when more severely ill patients with a worse prognosis are more likely to receive PPIs than healthier patients. Furthermore, some published observational studies may have suffered from immortal time bias.15, 21 Both biases could distort the studied association in either direction.
Other studies failed to show an interaction between the use of PPIs and clopidogrel.2, 6, 7, 13, 19, 22, 23 Recently, results of a post hoc analysis of a randomized trial revealed no association between PPI use and the risk of the primary CV endpoint for patients treated with clopidogrel or the novel thienopyridine prasugrel (a prodrug also requiring metabolization through CYP enzymes).13 The main limitation of this randomized trial was that use of a PPI was not randomized. Preliminary analysis (prespecified sample size and follow-up time not reached) of the unpublished Clopidogrel and the Optimization of Gastrointestinal Events (COGENT) trial, a randomized double-blind trial of omeprazole 20 mg vs. placebo in patients taking dual therapy (clopidogrel and aspirin) demonstrated no significant difference in CV events between both study arms (hazard ratio = 1.02, 95% CI: 0.70–1.51).6
As a result of the potential clinical consequences for a large patient group at risk for both recurrent CV as well as UGI events, residual uncertainty about this potential drug–drug interaction should be minimized. Therefore, we conducted a nested case–control study using data from the PHARMO Record Linkage System (RLS) (1999–2008) to quantify the association between use of PPIs and recurrent MI in the absence or presence of clopidogrel while addressing the issues of both study design (avoidance of immortal time bias) and residual confounding (using past exposure to PPIs as the reference category and propensity score-based adjustments).
Patients and methods
A population-based nested case–control study (1999–2008) was conducted within a cohort of patients admitted for acute MI during the study period.
The study was conducted using data from the PHARMO RLS. This system comprises drug-dispensing records mostly from community pharmacies and hospital discharge records of more than three million inhabitants of 50 demographically defined areas in the Netherlands. For all participants, the computerized drug-dispensing histories contain data concerning the name of the dispensed drug, dispensing date, dispensed amount, prescribed dosing regimens and the legend duration of use (prescription length). All drugs are coded according to the Anatomical Therapeutic Chemical (ATC) classification.24 The hospital records include detailed information concerning discharge diagnosis, procedures, dates of hospital admission and discharge, and discharge destination (or death in the hospital before discharge). Diagnoses are consistently classified according to the International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) during the entire study period. For a detailed description of the database, we refer to a previous publication.25
The study cohort included all patients in the PHARMO RLS who were hospitalized between 1st January 1999 and 31st December 2008 for a primary diagnosis of acute MI. Cohort entry was date of discharge from hospital of an acute MI, registered as a primary discharge diagnosis code 410 according to ICD-9-CM. Secondary diagnoses were not queried because it is likely that prevalent MI events (comorbidity) are recorded as secondary diagnosis. For reasons of the dynamic study population, patients were required to have had one prescription filled at least 1 year preceding the date of cohort entry (i.e. 1 year of valid database history) to allow uniform assessment of the presence of comorbidity and confounding factors. Patients were followed from cohort entry until diagnosis of recurrent MI, last filled prescription, or the end of the study period (31st December 2008), whichever came first.
The study outcome was hospitalization for a subsequent acute MI, registered as primary diagnosis ICD-9-CM code 410. Only the first encountered recurrent acute MI during follow-up was included in the analysis. The date of rehospitalization for the MI was used as index date. To avoid misclassification of the exposure, we required a 30-day period between discharge from baseline MI and recurrent MI as outpatient time was needed to observe filling of outpatient prescriptions. In a sensitivity analysis, we used the requirement of a 90-day period between discharge from baseline MI and recurrent MI.
For each patient readmitted for acute MI, controls were randomly selected from the cohort, matched on gender, age (same year of birth), being at risk of a recurrent MI and calendar time by means of incidence density sampling.26 Only those controls were considered to be at risk of developing a recurrent MI when they had a baseline acute MI ≥30 days prior to the index date.
Definition of exposure
Drug exposure data were obtained from outpatient drug-dispensing files of community pharmacies, as recorded in the PHARMO RLS. The exposures of interest were clopidogrel and PPIs. PPIs included in the analyses were omeprazole, pantoprazole, lansoprazole, rabeprazole, and esomeprazole. Duration of use was obtained by dividing the total amount of dispensed units by the prescribed amount per day. We calculated the end of each prescription by adding the duration to the dispensing date.
Clopidogrel use was classified at the index date as either ‘no’ use (no use or use at least 30 days before index date), ‘past’ use (end of prescription between seven and 30 days before the index date), or ‘current’ use (use at index date, or end of prescription <7 days before index date). As half-time values of PPIs are short (a few hours), the current, past and no use of PPIs was defined differently: ‘no’ use (no use or use at least 30 days before index date), ‘past’ use (end of prescription between 3 and 30 days before index date) and ‘current’ use (use at index date, or end of prescription <3 days before index date). To avoid protophatic bias, only prescriptions dispensed at least 1 day before index date were considered ‘current’ use. In sensitivity analyses, the definition of past PPI use (which was the reference) was adjusted to use between 3 and 90 days or between 14 and 90 days preceding the index date. In the latter case, the definition of ‘current’ PPI switched to use at index date, or end of prescription <14 days before index date. The study is not subject to immortal time bias as the outcome under study could be assessed equally for cases and controls after the 30- or 90-day period.
At baseline, we considered the following factors as potential covariates: percutaneous transluminal coronary artery angiography (PTCA) within 30 days of baseline MI, prior hospitalization for cardiogenic shock, congestive heart failure, cancer, cerebrovascular disease, acute renal failure, chronic renal failure, cardiac dysrhythmia (including atrial fibrillation), or UGI ulcer. The covariate diabetes mellitus was a combination of prior hospitalization for diabetes mellitus and antidiabetic drug use. As a proxy for (the degree of) comorbidity, we aggregated all prescriptions per ATC-code in the 1-year preceding cohort entry and we refer to this ‘score’ as the number of different prescriptions. In addition, we considered use of commonly prescribed CV medications (anti-hypertensive drugs, low-dose acetylsalicylic acid, β-adrenergic antagonists, digoxin, and lipid lowering drugs) or NSAIDs at baseline. We defined comorbidity and medication at baseline rather than at recurrent MI to avoid adjustment on intermediates in the pathway from exposure to outcome. Medications inducing or inhibiting cytochrome P450 2C19 or cytochrome P450 3A4 (directly affecting levels of clopidogrel and PPIs),27 as well as standard secondary prevention medications such as lipid lowering drugs and anti-hypertensive agents were assessed at the index date. Furthermore, days of follow-up (number of days between date of discharge of baseline MI and index date) and length of hospital stay for the baseline MI were also considered as covariates.
Descriptive analyses were performed to describe the cohort with regard to baseline characteristics. Prescription rates for PPIs and clopidogrel for the cohort were calculated per calendar month to assess prescription time trends. Survival rates were estimated using the Kaplan–Meier method and evaluated the probability of recurrent MI free survival during follow-up.
Baseline characteristics were compared between cases with a recurrent MI and controls without recurrent MI in patients unexposed to PPIs, and between all PPI users and PPI non-users in patients without recurrent MI using univariate conditional logistic regression analyses. We applied conditional logistic regression analysis to estimate matched and adjusted odds ratios (OR) and 95% confidence intervals (95% CI) for the association between recurrent MI and exposure to PPI. The analyses were performed in two subgroups: (i) current clopidogrel users and (ii) clopidogrel non-users to examine whether there was a true drug–drug interaction between clopidogrel and PPIs. The reference group included patients with no prescription of PPIs in the previous month (non-PPI users). A second analysis that more effectively deals with confounding by indication was performed applying past PPI users as the reference category. In the adjusted model, we included all factors that changed the unadjusted odds ratio with more than 10%. To test the robustness of our finding, we examined the effect of extension of the past PPI exposure window on the observed associations as a sensitivity analysis. Finally, we estimated the OR for current use of omeprazole and esomeprazole, using current pantoprazole as the reference group.
As the use of PPIs was not randomly assigned, we developed a propensity score (summary exposure risk score). The propensity score represents the likelihood of exposure to a PPI for each individual patient.28 The propensity score for use of PPI was developed using logistic regression modelling. The log odds of the probability that a patient received a PPI at any point of time during follow-up was modelled as a function of all previously described baseline covariates and was included in our dataset as a separate confounder. The PPI propensity score was only included in the adjusted analysis when it changed the odds ratio with more than 10%, unless we adjusted for individual covariates, to avoid double adjustment for the same covariate.
Statistical significance was assumed for two-sided P-values < 0.05. All statistical analyses were performed using spss software version 15.0 (SPSS Inc, Chicago, IL, USA).
The initial study population comprised 27 513 patients who were admitted to the hospital with an acute MI during follow-up. After exclusion of patients of whom not one year of valid data before baseline MI was available, the final cohort comprised 23 655 patients. The mean age was 64.7 [±standard deviation (s.d.) 13.2] years and 15 897 (67.2%) were men. The median follow-up time after discharge was 42.6 months [interquartile range (IQR) 16.8–71.7 months]. Comorbidity at baseline was substantial: 4.4% of patients had been hospitalized for cancer, 3.3% for cardiac dysrhythmia, and 2.8% for cerebrovascular disease. Prevalence of diabetes mellitus was 21.6%. The median number of distinct prescribed drugs in the one year before cohort entry was 5 (IQR 2–9 prescriptions) (Table 1; column 2). At baseline, 10.5% of the patients used PPIs and 315 patients (1.3%) used clopidogrel.
|Cases with recurrent MI|
|Controls without recurrent MI|
|In patients unexposed to PPIs||In controls without recurrent MI|
|Number||23 655 (100)||616 (100)||126 817 (100)||69 313 (100)||126 817 (100)|
|Mean age (s.d.)||64.7 (13.2)||66.1 (13.2)||66.3 (10.4)||68.1 (10.3)||66.3 (10.4)|
|Median follow-up time in months (IQR)||42.6 (16.8–71.7)||11.4 (4.1–29.4)*||29.5 (13.1–51.6)*||34.5 (16.0–58.4)*||29.5 (13.1–51.6)*|
|Median number of different prescriptions (IQR)†||5 (2–9)||5 (2–8)*||3 (1–6)*||6 (3–10)*||3 (1–6)*|
|Median length of stay baseline MI in days (IQR)||7 (4–10)||7 (5–9)||7 (5–10)||7 (5–10)*||7 (5–10)*|
|Male gender||15 897 (67.2)||436 (70.8)||72 020 (85.1)||55 377 (79.9)||72 020 (85.1)|
|PTCA during or within 30 days of baseline MI||7889 (33.4)||126 (20.5)||23 808 (28.1)||21 851 (31.5)*||23 808 (28.1)*|
|Presence of diabetes mellitus||5100 (21.6)||125 (20.3)*||13 774 (16.3)*||14 022 (20.2)*||13 774 (16.3)*|
|Hospitalizations before baseline MI|
|Cardiogenic shock||12 (0.1)||1 (0.2)||31 (0.0)||37 (0.1)||31 (0.0)|
|Congestive heart failure||551 (2.3)||14 (2.3)*||858 (1.0)*||961 (1.4)*||858 (1.0)*|
|Cancer||1045 (4.4)||13 (2.1)||2380 (2.8)||2710 (3.9)*||2380 (2.8)*|
|Cerebrovascular disease||651 (2.8)||11 (1.8)||1378 (1.6)||1836 (2.6)*||1378 (1.6)*|
|Acute renal failure||32 (0.1)||0 (0.0)||49 (0.1)||84 (0.1)*||49 (0.1)*|
|Chronic renal failure||210 (0.9)||4 (0.6)||449 (0.5)||667 (1.0)*||449 (0.5)*|
|Cardiac dysrhythmia||775 (3.3)||18 (2.9)||1898 (2.2)||1960 (2.8)*||1898 (2.2)*|
|UGI ulcer and gastritis||153 (0.6)||1 (0.2)||101 (0.1)||630 (0.9)*||101 (0.1)*|
|Medication use at the time of baseline MI|
|Angiotensin-converting-enzyme inhibitor||3141 (13.3)||84 (13.6)*||9122 (10.8)*||8601 (12.4)*||9122 (10.8)*|
|Angiotensin-receptor antagonist||1787 (7.6)||38 (6.2)*||4200 (5.0)*||5034 (7.3)*||4200 (5.0)*|
|Acetylsalicylic acid||5364 (22.7)||153 (24.8)*||16 181 (19.1)*||17 120 (24.7)*||16 181 (19.1)*|
|β-Adrenergic antagonist||5531 (23.4)||148 (24.0)*||16 915 (20.0)*||16 891 (24.4)*||16 915 (20.0)*|
|Calcium-channel antagonist||3150 (13.3)||92 (14.9)*||9257 (10.9)*||10 031 (14.5)*||9257 (10.9)*|
|Digoxin||592 (2.5)||14 (2.3)||1204 (1.4)||1311 (1.9)*||1204 (1.4)*|
|Spironolactone||336 (1.4)||5 (0.8)||455 (0.5)||559 (0.8)*||455 (0.5)*|
|Lipid lowering drugs||4144 (17.5)||113 (18.3)*||13 034 (15.4)*||13 307 (19.2)*||13 034 (15.4)*|
|Thiazide diuretic||806 (3.4)||16 (2.6)||2080 (2.5)||1958 (2.8)||2080 (2.5)|
|Other diuretic, excluding thiazide||2713 (11.5)||63 (10.2)||5299 (6.3)||6094 (8.8)*||5299 (6.3)*|
|Nonsteroidal anti-inflammatory drug||1720 (7.3)||30 (4.9)||3910 (4.6)||6384 (9.2)*||3910 (4.6)*|
|Medication use at index date‡|
|Angiotensin-converting-enzyme inhibitor||213 (34.6)||29 328 (34.6)||24 817 (35.8)||29 328 (34.6)|
|Angiotensin-receptor antagonist||50 (8.1)||9301 (11.0)||9535 (13.8)*||9301 (11.0)*|
|Acetylsalicylic acid||408 (66.2)||56 621 (66.9)||45 511 (65.7)||56 621 (66.9)|
|β-Adrenergic antagonist||377 (61.2)||52 200 (61.7)||43 575 (62.9)||52 200 (61.7)|
|Calcium-channel antagonist||116 (18.8)||14 252 (16.8)||13 727 (19.8)*||14 252 (16.8)*|
|Lipid-lowering drugs||356 (57.8)||54 475 (64.4)||45 776 (66.0)*||54 475 (64.4)*|
|Thiazide diuretic||24 (3.9)||2690 (3.2)||2672 (3.9)||2690 (3.2)|
|Other diuretic, excluding thiazide||130 (21.1)*||12 606 (14.9)*||14 510 (20.9)*||12 606 (14.9)*|
|Medication use at recurrent MI|
|Cytochrome P450 2C19 inhibitor(s)||7 (1.1)||679 (0.8)||676 (1.0)*||679 (0.8)*|
|Cytochrome P450 2C19 inducer(s)||13 (2.1)||1043 (1.2)||2973 (4.3)*||1043 (1.2)*|
|Cytochrome P450 3A4 inhibitor(s)||56 (9.1)||5725 (6.8)||5854 (8.4)*||5725 (6.8)*|
|Cytochrome P450 3A4 inducer(s)||8 (1.3)||823 (1.0)||724 (1.0)||823 (1.0)|
Figure 1 shows that prescription rates of both PPIs and clopidogrel changed over time between 1st January 1999 and 31st December 2008, illustrating the increased use of the study drugs over a decade. Within the study cohort, 1247 patients were readmitted to the hospital with an acute MI. The risk of a recurrent MI at least 30 days after discharge was 2.8% (95% CI: 2.6–3.0%) in the first year and 10.3% (95% CI: 9.1–11.5%) in 10 years.
Of the cases with a recurrent MI, 1224 (98.2%) could be matched to at least one control without a recurrent MI. Characteristics of cases and controls (in the unexposed group to PPIs) are depicted in Table 1 (columns 3 and 4). As expected, cases were more likely to have been hospitalized for comorbidities associated with an increased risk of recurrent MI, such as congestive heart failure and diabetes mellitus. At baseline, cases used several types of CV drugs more frequently, for instance angiotensin-converting-enzyme (ACE) inhibitors and lipid-lowering drugs. To understand whether selective prescribing (channelling) of PPIs to persons at increased risk of recurrent MI occurred, we assessed the presence of risk factors in PPI users and PPI non-users (in patients who did not develop a recurrent MI) (Table 1; columns 5 and 6). Patients using PPIs were older, were more frequently men, had been hospitalized more frequently before cohort entry for several comorbid conditions, and more frequently used CV medications at baseline.
Among current clopidogrel users, a significant association was observed between current PPI use and recurrent MI when compared with PPI non-use (OR: 1.62, 95% CI: 1.15–2.27). When applying past PPI use as the reference category rather than PPI non-use, we found no association between recurrent MI and current use of PPI (OR: 0.95, 95% CI: 0.38–2.41) (Table 2). Repeated analyses with variations in the definition of past PPI use (3–90 days or 14–90 days) did not affect the study findings appreciably; OR: 1.15, 95% CI: 0.51–2.61 and OR: 1.34, 95% CI: 0.50–3.62, respectively.
N = 1224
N = 153 967
|ORmatched (95% CI)||ORadjusted (95% CI)|
|Current clopidogrel/no PPI||90||11 147||1 (ref)||1 (ref)|
|Current clopidogrel/current PPI||78||4715||1.89 (1.37–2.63)||1.62 (1.15–2.27)‡|
|Current clopidogrel/past PPI||6||436||1 (ref)||1 (ref)|
|Current clopidogrel/current PPI||78||4715||1.15 (0.46–2.86)||0.95 (0.38–2.41)§|
Among clopidogrel non-users, current use of PPI was associated with recurrent MI when compared with PPI non-use (OR: 1.38, 95% CI: 1.18–1.61). When applying past PPI users as the reference category, a weak association between recurrent MI and current use of PPI was found (OR: 1.22, 95% CI: 0.79–1.88), although the latter’s confidence interval crosses the null, which was likely due to insufficient power (Table 3).
N = 1224
N = 153 967
|ORmatched (95% CI)||ORadjusted (95% CI)|
|No clopidogrel/no PPI||766||110 002||1 (ref)||1 (ref)|
|No clopidogrel/current PPI||237||23 335||1.56 (1.34–1.81)||1.38 (1.18–1.61)‡|
|No clopidogrel/past PPI||25||2941||1 (ref)||1 (ref)|
|No clopidogrel/current PPI||237||23 335||1.22 (0.79–1.88)||1.22 (0.79–1.88)§|
The analysis by type of PPI among current clopidogrel users showed that pantoprazole was most frequently used, followed by omeprazole and esomeprazole. Lansoprazole and rabeprazole were rarely used. Using current pantoprazole as the reference category, current use of omeprazole and of esomeprazole were not associated with an increased risk of recurrent MI (respectively, OR: 1.07, 95% CI: 0.58–1.99 and OR: 0.83, 95% CI: 0.40–1.69, respectively) (Table 4).
N = 1224
N = 153 967
|ORmatched (95% CI)||ORadjusted† (95% CI)|
|Current clopidogrel + pantoprazole||36||2271||1 (ref)||1 (ref)|
|Current clopidogrel + omeprazole||26||1339||1.03 (0.57–1.84)||1.07 (0.58–1.99)|
|Current clopidogrel + esomeprazole||13||827||0.96 (0.48–1.92)||0.83 (0.40–1.69)|
|Current clopidogrel + lanso- or rabeprazole||2||186||–||–|
This population-based cohort study showed that the association between clopidogrel–PPI co-therapy and risk of recurrent MI is highly affected by confounding by indication,29 which may explain the contrasting results in the literature. To illustrate the problem of confounding, we compared current PPI users with none or past PPI users in the absence of clopidogrel. To deal with the problem of confounding by indication, we compared current use of clopidogrel plus current use of PPI not only with current clopidogrel without PPIs, but also with current clopidogrel use plus past use of PPIs, which reduces confounding by indication. When current PPI use was compared with past PPI use, the association disappeared, suggesting that the observed association between current PPI use and recurrent MI when PPI non-use was the reference, may have been the result of residual confounding.
Our findings are in line both with other observational studies showing an increased risk among PPI users when compared with PPI non-users,15–20 and those that did not show such an association if better control of confounding was applied.6, 13, 19, 30 In observational studies or post hoc analyses in randomized controlled trials, the use of PPI is by definition not randomly assigned, most likely leading to confounding by indication. It is arguable whether adequate adjustment for this harmful type of confounding is possible as much of the confounding will be subtle and unmeasurable. Design choices and progressive adjustment techniques can be used to avoid residual confounding, as we have shown. By careful selection of the study design, one can match cases and controls on important covariates, including calendar time. We illustrated the increased use of the study drugs over a decade (Figure 1), and thereby the importance of matching on index date (thus calendar time) as potential confounder. In addition, one can use a so-called ‘active comparator’ as reference, such as past users of PPIs instead of non-users. For adjustment, high-dimensional propensity scores,19 summary disease risk scores (probability of short-term mortality),16 or conventional propensity scores13, 20 can be applied. Table 5 provides an overview of techniques used in other observational studies on this topic to deal with confounding by indication.
|Authors||Study type||Study population||Outcome||Methods to deal with confounding|
|Simon et al.2||Prospective cohort||Patients with acute MI||Death, MI, stroke||Multivariable adjustments|
Propensity analysis for CYP2C19 loss-of-function-alleles which was used to match five controls for each patient with two variant alleles
|Bhatt et al.6||Double-blind randomized trial||Patients with ACS undergoing coronary stent placement||MI, stroke, CABG, PCI, CV death||Randomization|
|Ray et al.7||Retrospective cohort study||Patients hospitalized for MI, coronary artery revascularization or unstable AP on clopidogrel||Serious CV disease [(non)fatal MI, stroke, other CV death]||Multivariable adjustments|
Propensity score for the use of PPI was converted in deciles and included as variable in model
|O’Donoghue et al.13||Retrospective cohort within RCT||Patients with ACS undergoing PCI||CV death, MI, stroke||Multivariable adjustments|
Propensity score for the use of PPI and strata matched on this score
|Ho et al.15||Retrospective cohort study + case–control study||Patients with ACS on clopidogrel||All-cause mortality, rehospitalization for ACS||Multivariable adjustments|
Matched on duration of follow-up
|Juurlink et al.16||Nested case–control||Patients with acute MI on clopidogrel||Recurrent MI, death||Multivariable adjustments|
Matched on age, PCI, date of hospital discharge, predicted probability of short-term mortality
|Pezalla et al.17||Retrospective cohort study||Patients adherent to clopidogrel||Acute MI||Subgroup analysis; restriction to patients who all had diagnosis of ischaemic heart disease, congestive heart failure, hypertension, hyperlipidemia and diabetes before the start of clopidogrel therapy|
|Stanek et al.18||Retrospective cohort study||Patients adherent to clopidogrel following coronary stenting||CV events (MI, unstable AP, TIA/stroke, coronary vascularization, CV death)||Multivariable adjustments|
|Rassen et al.19||Retrospective cohort study||Patients undergoing PCI or hospitalized for ACS on clopidogrel||MI, revascularization, all-cause mortality||Multivariable adjustments|
High-dimensional propensity score and strata matched on this score (1:1)
|Stockl et al.20||Retrospective cohort study||Patients with MI or coronary stent on clopidogrel||Acute MI||Propensity score and strata matched on this score (1:1)|
|Collet et al.22||Prospective cohort||Survivors of MI (<45 years)||CV death, nonfatal MI, urgent revascularisation||Multivariable adjustments|
|Dunn et al.23||Retrospective cohort within RCT||Patients undergoing PCI||Death, MI, stroke (1 year)||No information on adjustment to covariates|
Our study further confirms the relevance of confounding, illustrated by the fact that PPI users had more comorbidity and used more co-medication than PPI non-users at baseline and also by the results of the analysis in clopidogrel non-users. The increased risk for recurrent MI with current PPI use (OR 1.38, 95% CI: 1.18–1.61) compared with no PPI use in the absence of clopidogrel therapy provides further support for this. Another explanation would be that PPIs have a harmful effect irrespective of clopidogrel status. We are aware of two studies that conducted similar analysis in clopidogrel non-users that also showed significantly elevated risks with PPI use (adjusted OR 1.29 and 1.55) compared with PPI non-use.23, 31 In two other observational studies, however, no significant increase in CV events was shown in patients prescribed PPIs without clopidogrel.15, 32
We addressed the confounding issue by all possible design and adjustments measures, but still the comparison against non-use of PPI seemed confounded. Applying past PPI use as active comparator changed the association from increased to no effect and was the most powerful approach to deal with channelling of PPIs, more than adjustment for the propensity score. The propensity score did not have large explanatory power in this study due to the fact that relatively little clinical information was available to construct the score.
Our study is internally valid as we used a population-based design, so selection bias was unlikely as all cases and controls came from the same source population. The same pertains to information bias because data were gathered prospectively without knowledge of the hypothesis studied. The method to control for confounding was discussed previously. The database is proven valid for research as the diagnoses are labelled with ICD-9-CM codes in each hospital by official coding personnel from the national registry of hospitalization discharge records and linked to pharmacies with complete information on outpatient dispensing. In the Netherlands, patients usually go to one pharmacy because of billing purposes (invoices being directly sent to insurance without the need to prepay by patient) and medication surveillance. The exposure drugs, clopidogrel and PPIs, are dispensed through regular pharmacies. With regard to external validity, the individuals captured in the database are representative for the entire population of the Netherlands in terms of age, gender, socio-economic background, morbidity and mortality, drug use, and geographic distribution.
This study adds to the existing literature because we revealed the difficulties to deal with confounding by indication in observational studies and we adjusted for confounding in the most optimal way by applying past PPI use as the reference category. Furthermore, few other studies could account for the use of low-dose aspirin as important risk factor for recurrent MI.19, 20 We avoided immortal time bias by assigning exposure to the category it belongs to and not using follow-up time to define exposure. Immortal time bias alludes to the fact that the outcome under study cannot be assessed when follow-up time is used to define the exposure status. We used a time varying exposure assessment, whereas some other cohort studies applied a fixed exposure status,13, 20 which leads to misclassification of exposure at the index date (e.g. patients were defined at baseline as PPI non-users irrespective of changes in PPI use during follow-up). In addition, we used an European database, whereas many published studies so far were conducted in North-America. None of the PPIs were available over the counter during the study period, and we have illustrated that there was no difference in the risk of recurrent MI between esomeprazole, omeprazole, and pantoprazole.
There are several important limitations to this analysis. First, we were unable to study actual drug utilization and adherence to dispensed drugs, because only dispensing data were available in the database instead of more reliable proxies for drug use. Second, using the past PPI group as reference, we may compare current PPI users to past PPI users who had not been fully adherent to the prescribed PPI and therefore still had medication available and were actually also current PPI users at the time of the recurrent MI. To address this point, we conducted several sensitivity analyses varying the time window to define past PPI use and this did not affect the estimates. In addition, although the internal validity may be jeopardized by noncompliance of the patient, this is unlikely to have affected the comparison between active compounds (e.g. omeprazole/esomeprazole compared with pantoprazole). Third, as most persons continue PPIs once started, applying past PPI use as reference reduces study power substantially (resulting in wide confidence intervals). Finally, residual confounding remained due to lack of data on important cardiac risk factors, such as smoking status, lipoproteins, or type of coronary stents, which was likely providing an overestimation of the effect measure and this was illustrated when applying PPI non-use as the reference.
In conclusion, this study provides a unique angle on the association between clopidogrel–PPI co-therapy and risk of recurrent MI by using several techniques to deal with confounding by indication and other methodological issues. Among clopidogrel users, current use of PPIs was associated with an increased risk of recurrent MI when compared with PPI non-use, but not when compared with past PPI use, which shows the magnitude of the confounding by indication that is present. In clopidogrel non-users, current PPI use was also associated with an increased risk of recurrent MI when compared with PPI non-use. This study thus demonstrates that previously reported associations between PPI–clopidogrel co-therapy and risks of recurrent MI are influenced by confounding. It further demonstrates that such confounding can be (at least partly) circumvented by careful selection of study design and conventional adjustments techniques. Further work should concentrate on well-conducted prospective randomized trials to dissolve the confusion around this controversial drug–drug interaction.
Declaration of personal interests: V. Valkhoff, E. van Soest, and G. ‘t Jong do not have any conflicts to declare. As employee of Erasmus MC, M. Sturkenboom has been involved as project leader and in analyses contracted by various pharmaceutical companies. She received unconditional research grants from AstraZeneca, Pfizer, Merck, Johnson & Johnson, Amgen, Roche, GSK, Boehringer, Yamanouchi and Altana, none of which are related to the subject of this study. M. Sturkenboom has been consultant to Pfizer, Servier and Lundbeck on issues not related to this manuscript. E. Kuipers has served as a speaker and advisory board member for AstraZeneca. Declaration of funding interests: This study was fully funded by both Departments of Gastroenterology & Hepatology, and Medical Informatics of the Erasmus MC, Rotterdam, The Netherlands.