Aspirin reduces cardiovascular events in patients with pneumonia: a prior event rate ratio analysis in a large primary care database

Study design

This analysis was performed using data from the Clinical Practice Research Datalink (CPRD), a large database of coded, routine primary care health records considered representative of the wider UK population. This includes all diagnoses coded by primary care clinicians, as well as events occurring in hospital that are coded in the primary care record (e.g. via a discharge summary), but does not necessarily include all hospitalised events. We included all patients who had a coded diagnosis of pneumonia from inception of the database until January 2019. Primary care records were linked to Office for National Statistics (ONS) mortality data. If patients had multiple episodes of pneumonia, the first was taken irrespective of the time between events. The study period started 12 months prior to pneumonia diagnosis and lasted for 6 months after diagnosis.

Outcomes

For all analyses, the primary outcome was the combined outcome of myocardial infarction (MI) or ischaemic stroke. The two secondary outcomes were MI and ischaemic stroke, individually.

Definitions

Codes for pneumonia were derived from the previous literature [9] and were updated to ensure current validity. There has been extensive use of the CPRD for studying pneumonia and coding in it is generally felt to be of good quality [9–13]. Specific explicit codes for pneumonia rather than simple lower respiratory tract infection were included (appendix 1) and these were again based on the previous literature. As this is a primary care dataset, all presentations are those coded from primary care records.

Codes for outcomes and comorbidity were derived from the Manchester ClinicalCodes repository [14]. Code lists are also presented in appendix 1. For MI in particular, the code list was edited to ensure that all events represented acute events, rather than chronic myocardial ischaemia. Mortality was based upon ONS mortality data, which was derived from death certificates. Time of death was taken from ONS data rather than CPRD date, as this was more reliable [15, 16].

Patients who had outcomes on the same day as the pneumonia diagnosis were excluded. This was to ensure reliable data on which event happened first and to limit reverse confounding (pneumonia as a consequence of MIs or stroke).

Aspirin use was defined in relation to study entry (12 months prior to pneumonia diagnosis). Aspirin users were defined as patients who had received two or more prescriptions for aspirin at a daily dose of <100 mg in the 6 months prior to study entry. This ensured chronic aspirin use at study entry and this requirement was then repeated in the next two, 6-month blocks, to ensure a continued prescription of low dose aspirin over the whole 12-month period prior to pneumonia diagnosis.

Aspirin non-users were defined as patients who did not meet this criterion in all three 6-month blocks. Patients who met the criteria in some but not all of the three blocks were censored. Other drugs examined in secondary analyses were defined similarly. Age, gender and index of multiple deprivation (IMD), an area-based measure of socioeconomic deprivation, were included as covariates in the analysis, alongside medical comorbidities.

Comorbidity data were collected on smoking status, hypertension, stroke, ischaemic heart disease (IHD), diabetes and peripheral vascular disease (PVD). All comorbidities were defined by codes derived from the ClinicalCodes repository and must have occurred prior to the whole study period (i.e. 12 months prior to the pneumonia diagnosis). The absence of a comorbidity was taken to mean the absence of that condition. Smoking was defined as a categorical variable based on the most recent smoking code (current, former, never, or missing).

Statistical methodology

Our primary analysis methodology was a propensity matched PERR analysis. In this analysis, confounding was taken into account in two ways. First, similar baseline characteristics were ensured across both groups by propensity score matching (PSM) of aspirin users to non-aspirin users. Propensity matching was performed with a 2:1 matching, using covariates that were likely to be relevant for aspirin prescription (age, smoking, gender, history of IHD, stroke, or PVD). An optimal, nearest neighbour approach was utilised using the MatchIt package in the R software environment (R Foundation for Statistical Computing, Vienna, Austria; www.r-project.org). Covariate imbalance was assessed using the standardised mean difference (SMD) with a detection threshold of 0.1 [17]. Secondly, PERR analysis was used to account for unmeasured confounders. Double adjustment was carried out for any covariates that remained imbalanced after PSM [18].

PERR analysis uses a form of self-controlled design in which participants act as their own controls to reduce confounding. In a standard PERR analysis, the rate of events (in this case, cardiovascular events) prior to and after a discrete exposure are compared within the same population group, allowing for unmeasured confounders to be taken account of and a more reliable effect size to be calculated [19–21].

In the current analysis, the standard PERR methodology has been adapted to address the question of interest. Usually exposed and unexposed groups are defined based on incident exposure to a treatment at an index date (time origin). We define the index date as the date of pneumonia diagnosis. The incident exposure is defined to be experiencing the pneumonia diagnosis when a current aspirin user. The unexposed group are selected from non-aspirin users who also experience the pneumonia diagnosis. This allows calculation of an adjusted effect of aspirin use on post-pneumonia events.

For this analysis, patients were split into the two exposure groups (aspirin users and non-aspirin users) depending on their prescription records. For both groups a prior period was identified, starting 12 months prior to the pneumonia date and lasting for 6 months, as was a posterior period (the 6 months directly after diagnosis of the pneumonia) (figure 1).

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 1

Schematic of the prior event rate ratio (PERR) analysis.

Cardiovascular events occurring during these periods were recorded and used to estimate a PERR ratio of hazards for prior events to posterior events. The hazards for the prior and study periods were calculated using a Cox proportional hazards model, with group membership as the independent variable, adjusted for relevant covariates. Bootstrapping was performed to generate confidence intervals (CIs). The resulting outcome then represents the estimated effect size of aspirin use on post-pneumonia events after accounting for the effect of time-invariant confounding. Any patients who died in the posterior period were right-censored.

To avoid the possibility of patients switching from being non-aspirin users to aspirin users, which would violate the assumptions of the PERR methodology, patients were required to be either on or off aspirin throughout the whole reference period (12 months prior, until diagnosis of pneumonia). Patients who started aspirin between the prior period and diagnosis of pneumonia were censored.

The prior period deliberately ended 6 months prior to the pneumonia diagnosis. This was to ensure there was complete washout between periods. This was a pragmatic choice, aimed at reducing any concern of bias that pneumonia events could be a consequence of MI hospitalisation episodes.

Missing data for smoking status was coded as such, whereas other missing data was removed (complete case analysis) as this was rare (43 patients). We estimated a hazard ratio of 0.8, a prevalence of routine aspirin usage of 33% and incidence of events at 6 months of 3%, with a crude sample size of around 17 112 participants required, which was easily achieved before PSM.

Data were extracted from the CPRD and processed using Stata software (StataCorp, College Station, TX, USA) with the CPRDUTIL package; however, all analyses were performed using R software versions 3.6.0 and 4.0.0 using the tidyverse, ggplot, matchit, broom, survival and rio packages.

Sensitivity and secondary analyses

For the main analysis, it was possible for aspirin prescription not to overlap the pneumonia diagnosis index date (the precise timing of prescriptions can be subject to uncertainty in the CPRD). To determine whether findings were robust to the timing of aspirin use, a sensitivity analysis was performed restricting aspirin patients to those in receipt of the drug on the index date.

As participants might die in the posterior period (and thus not have coded events), there is a potential for a survival bias where one group has fewer posterior events as it has higher mortality and this should be accounted for by right-censoring. Death represents a competing risk and our chosen approach is to use a proportional cause-specific hazards model, as recommended for aetiological research in the presence of competing risks [22].

As a test for robustness we examined two alternative negative control outcomes: fracture (of any kind) and constipation, to ensure that any effect identified was specific to MI/stroke and to identify evidence of survival bias.

To look for possible evidence of residual confounding after PERR analysis, the analysis was replicated with four other drug classes, three of which were negative control exposures. First, paracetamol and levothyroxine were chosen, as these have no known cardiovascular effect and are not strongly correlated with aspirin prescription. Secondly, nonsteroidal anti-inflammatory drugs (NSAIDs, extracted by all NSAID drug codes in the CPRD) were compared as another potential drug class with an anti-inflammatory effect, albeit one known to have potential increased cardiovascular risk and which has been associated with increased events after pneumonia. Finally, proton pump inhibitors were compared as a commonly prescribed agent that is often co-prescribed with aspirin.

To compare with the PERR analysis, standard Cox regression (an analysis that accounts for measured confounders only) was undertaken using the same outcome variables. This analysis was performed on the same population used in the PERR analysis, equivalent to simply performing the analysis on the posterior period only, without adjusting for the prior period events. These models were then tested for the proportional hazards assumption (using the Therneau and Grambsch approach), to look for evidence of a time varying effect which might have impacted the PERR analysis [23].

As coding in CPRD could be a consequence of administrative activity (i.e. recording information from a hospital discharge summary or clinic letter) rather than direct clinical care provided by the GP, we undertook a subsequent sensitivity analysis only including consultations that were definitively performed by the primary care clinician (limited to only consultation types that were in the GP practice, at a home-visit, or over the telephone). This was achieved using the constype variable in CPRD.