The Long-term Risk for Myocardial Infarction or Stroke After Proton Pump Inhibitor Therapy (2008–2018)

Michael Nolde; Nayeon Ahn; Tobias Dreischulte; Ina-Maria Rückert-Eheberg; Florian Güntner; Alexander Günter; Roman Gerlach; Martin Tauscher; Ute Amann; Jakob Linseisen; Christa Meisinger; Sebastian-Edgar Baumeister

Disclosures

Aliment Pharmacol Ther. 2021;54(8):1033-1040. 

In This Article

Methods

Data Source

For this study, we analysed claims data from the Allgemeine Ortskrankenkasse (AOK) Bayern, a large regional German Statutory Health Insurance Provider. The dataset included about 6.1 million adult persons, who received health insurance cover from the AOK Bayern for at least 2 years since January 2007. Outpatient and hospital diagnoses were coded according to the German Modification of the International Statistical Classification of Diseases and Related Health Problems (ICD-10-GM), released by the German Institute of Medical Documentation and Information (DIMDI).[15] Drugs purchased over-the-counter, or administered in hospital, are not contained in the database. For data protection reasons, the data were pseudonymized. The study received approval from the Ethics Committee of the LMU Munich and the institutional review board of the AOK Bayern. It was registered at ENCePP.eu (EUPAS31559), where the study protocol, including a detailed description of the emulated target trial, was deposited. The investigators had full control over protocol development, analyses and publication. Patients or the public were not involved in the design, or conduct, or reporting, or dissemination plans of our research. This study adhered to the RECORD-PE guidelines.[16]

Study Population

The study cohort included new users of PPIs and new users of H2RAs, who started therapy between 2009 and 2018. We demanded no prior treatment with PPIs or H2RAs with at least 1year of medical history before treatment initiation recorded in the data. Cohort entry was the day of the first dispensed prescription of any of those drugs. All patients were required to be at least 18 years old and free of prevalent cardiovascular (ICD-Codes I21, I22, I23, I24.1, I25.2) or cerebrovascular disease (ICD-Codes I60, I61, I63, I64, G46) at cohort entry. A graphical depiction of our study design is shown in Figure 1.[17]

Figure 1.

Study design

Medication Exposure, Follow-up and Outcomes

New users of PPIs (ATC Code A02BC) were compared to new users of H2RAs (ATC Code A02BA), as the use of an active comparator might reduce the potential for confounding by indication, compared to a non-user control.[18] We defined exposure by identifying drug dispensing in prescription claims. Follow-up for study outcomes started the day after initiation of treatment and continued in an 'as-started' approach[19] until the occurrence of an outcome of interest, death, disenrollment or the end of the study period on 31 December 2018 (Figure 1).[17] The two study endpoints were primary MI and primary IS. Patients were considered a case of MI or IS after a hospital admission with the corresponding main discharge diagnosis (MI: I21; IS: I63, G46.5, G46.6). The validity of these claims-based diagnoses has been established.[20,21]

Covariates

We controlled for several confounders, assuming that direct causes of the exposure or outcome, excluding possible instrumental variables, would identify a sufficient set of confounding variables.[22] Accordingly, we adjusted for demographics (age, sex and nationality), calendar time of inclusion (in quarters and years), relevant comorbidities and medications. It remains unclear, whether treatment indication (e.g. gastroesophageal reflux disease [GERD]) has any direct or indirect effect on our outcomes.[23] Therefore, adjusting for treatment indication could mean adjusting for an instrumental variable, and introduce bias instead of reducing it. Despite that, we included treatment indications in the model for the propensity scores to minimize unmeasured confounding and indication bias.[24] Patient baseline characteristics were measured during the 90 days before and including the date of cohort entry. We also adjusted for the number of concurrently used drugs and the Elixhauser comorbidity score,[25] adapted to administrative data, taking both inpatient and outpatient diagnoses into account.[26] Due to intrinsic properties of the data, both were measured in the quarter preceding treatment initiation. A complete list of baseline patient characteristics and a definition of covariates are provided in Tables S1 and S2.

Statistical Analysis

We used inverse probability of treatment (IPT) weighting to adjust for confounding.[27] Propensity scores were estimated from a confounder-adjusted logistic regression model and used to calculate stabilized weights.[28] Standardized mean differences were used to assess balance in patient characteristics between treatment groups before and after weighting.[29] Raw incidence rates per 1000 person-years were computed. Overall exposure-specific survival was plotted as adjusted Kaplan-Meier estimates.[30,31] We estimated hazard ratios (HRs) with the corresponding 95% confidence intervals (CI) using weighted Cox proportional hazards models with robust standard errors. Sensitivity analyses included a comparison of PPI initiators with non-initiators, and the consideration of 97 pre-selected negative control (tracer) outcomes (NCOs)[32] to detect potential unmeasured confounding. We imposed various lag times by excluding events that occurred during the first 10, 30, 90 and 180 days after baseline.[33] The statistical software R (version 3.6.3, Foundation for Statistical Computing) was used.

processing....