Risk of thrombocytopenia and thromboembolism after covid-19 vaccination and SARS-CoV-2 positive testing: self-controlled case series study

Study design and period

We undertook a self-controlled case series from 1 December 2020 to 24 April 2021 (the latest date for which outcome data were available) to examine the associations between ChAdOx1 nCoV-19 or BNT162b2 mRNA vaccines and thrombotic events during the ongoing covid-19 vaccination programme in England. We also investigated the association between a SARS-CoV-2 positive test and the thrombotic events of interest among the same vaccinated population.

The self-controlled case series was originally developed to assess risks of adverse events to vaccination.16 The case series determines the relative incidence of the outcome of interest for exposed time periods (eg, after vaccination or SARS-CoV-2 infection) compared with unexposed baseline periods in people with the outcome of interest (see supplementary fig 1). Inference is within people and therefore this implicitly controls for all covariates that remain constant during the study period. We selected patients with each outcome during the study period and determined dates when they had the vaccine or tested positive for SARS-CoV-2. Separate analyses were carried out for each outcome of interest.

Data sources

We used the National Immunisation Management System register of covid-19 vaccination to identify vaccine exposure, which includes vaccine type, date, and doses for all people vaccinated in England. We linked individual level data to national data for mortality, hospital admissions, and SARS-CoV-2 infection. Linkages are conducted to protect patient anonymity because the hashed ID used to link patients cannot be reversed. Vaccine and SARS-CoV-2 positive test data were available until 24 April 2021. Hospital admission data were obtained from two sources: Hospital Episode Statistics (up to 31 March 2021), which includes any admission, and Secondary Users Service (up to 24 April 2021), which includes only admissions with an outcome (death or discharge).

We used the QResearch database of 12 million patients linked to the above dataset to calculate background incidence rates for each outcome before the pandemic (2015-19). QResearch includes demographic, clinical, and drug data, and is used for clinical1718 and drug safety research.1920 QResearch is one of the largest and most representative primary care research databases nationally,21 covering approximately 20% of the population of England.

Inclusion criteria

We included all people aged ≥16 years who had first doses of the ChAdOx1 nCoV-19 or BNT162b2 mRNA vaccines and any outcome of interest. We excluded people who received the mRNA-1273 (Moderna) vaccine because of very small numbers. We used the first event in the study period and excluded patients from the analysis of each outcome if they had a hospital admission for the same event in the two years before the start of the study period.

Exposure

Our main exposures were the first dose of the ChAdOx1 nCoV-19 or the BNT162b2 mRNA vaccine. We censored people on the earliest of the following: date of their second dose, date of death, or study end date (24 April 2021). Our secondary exposure was a positive SARS-CoV-2 test result, using the first date of a positive test (not the date of reporting) as the exposure date. We defined the exposure risk intervals as the following prespecified time periods: 0, 1-7, 8-14, 15-21, and 22-28 days after each exposure date (vaccination date or date of positive SARS-CoV-2 test). Multiple risk intervals were used to differentiate between acute and non-acute phases after vaccination. The baseline period (without exposure) was defined as any time between 1 December 2020 and 29 days before the exposure date, and from 29 days after the exposure date until 24 April 2021 or the censored date if earlier. A pre-risk interval of 1-28 days before each exposure date was included to deal with possible bias that might arise if the occurrence of the outcome temporarily influenced the likelihood of exposure.22 Histograms of the data by interval between vaccination and event day are shown in supplementary figure 2a. The graphs show a drop of hospital admissions or deaths for each outcome about 28 days before vaccination, hence the choice of the pre-risk period. Hospital admissions for the event of interest are probably the trigger for the covid-19 test. Such events might be caused by SARS-CoV-2 infection, but the reverse causality involved in their detection induces bias. To reduce the bias, which could overestimate or underestimate the effect of infection, we decided to allocate day 0 to a risk period on its own.23

Outcomes

Our three composite primary outcomes were hospital admission or death associated with thrombocytopenia, venous thromboembolism, and arterial thromboembolism. Our prespecified secondary outcomes were subsets of the primary outcomes: CVST, ischaemic stroke, myocardial infarction, and other rare arterial thrombotic events. Hospital inpatient admissions from 1 December 2020 to 24 April 2021 with an ICD-10 (International Classification of Diseases version 10) diagnosis code in their first 13 diagnoses fields that indicated the outcome of interest were identified in the Hospital Episode Statistics and Secondary Users Service databases. Supplementary table 1 gives a list of ICD-10 codes and groupings for each outcome. We used the earliest date of hospital admission or date of death for the relevant event as the event date. Cause of death in the UK is assessed by the medical practitioner who attended the patient during their last illness, and if not seen in the preceding 14 days, the cause is determined by a coroner based on assessment of medical evidence.

Statistical analysis

We described the characteristics of each cohort (patients who had been vaccinated with the outcomes of interest) in terms of age, sex, and ethnicity. The self-controlled case series models were fitted using a conditional Poisson regression model with an offset for the length of the risk period. Incidence rate ratios, the relative rate of hospital admissions or deaths due to each outcome of interest in risk periods relative to baseline periods, and their 95% confidence intervals were estimated using each model. Exposure terms for both vaccines and for infection with SARS-CoV-2 were included in the same model. To account for temporal changes in background rates, we divided the study period into weekly blocks starting on 1 December 2020 and adjusted for these changes as discrete covariates in the analysis. We used Wald tests to compare risks associated with ChAdOx1 nCoV-19 and BNT162b2 mRNA vaccines. We investigated whether the associations between vaccine exposures and outcomes are sex or age dependent by running the analyses in separate subgroups by sex and age group (younger or older than 50 years).

Sensitivity analyses

We conducted six sensitivity analyses: (1) excluding those who died from the outcome; (2) restricting analysis to the period after vaccination or after a SARS-CoV-2 positive test, without censoring at death; (3) censoring at 12 weeks after vaccination; (4) censoring on 10 March 2021; (5) restricting the study period until 31 March 2021; and (6) restricting the analysis to patients who had their vaccination after 1 January 2021. The first two analyses tested the assumption that the occurrence of an outcome event did not influence the probability of subsequent exposures, such as through death.22 The third analysis tested the sensitivity to our approach of censoring patients at the time of the second dose (to avoid contamination of baseline time). The additional censoring at 12 weeks after the first dose of vaccine was used because this is the recommended time for a second dose in the UK. Concerns over CVST and blood clots were raised first in Europe around 10 March 2021, and so we censored on this date for the fourth sensitivity analysis; this allowed us to include only time unaffected by any notoriety bias (because of media attention, thrombotic events might have been more likely to be reported on death certificates or hospital admission records if healthcare professionals were aware of recent ChAdOx1 nCoV-19 vaccination). The fifth analysis tested the completeness of the data, and the sixth analysis was done to have comparable time periods and priority groups for the two vaccines.

Assessing the self-controlled case series assumptions

To further assess the assumptions of the self-controlled case series and our modelling choices, we visually examined the data. We plotted a histogram of the number of occurrences of an event by time before or since vaccination for each outcome to assess the possibility that a hospital admission for that event might affect subsequent vaccination (supplementary fig 2a). We plotted a histogram of the time from event to actual end of observation in patients who were censored and uncensored to assess if event dependent observation periods could be a problem for the analysis (supplementary fig 2b).

Event dependent exposures—supplementary figure 2a shows the number of occurrences of an event by time before or after vaccination. We observed a decrease in the 28 days immediately before vaccination, indicating that occurrence of an event might have reduced the likelihood of vaccination. This pattern is similar for most of the outcomes and for both vaccines. Therefore, we have added the pre-risk period of 28 days.

Event dependent observation periods—supplementary figure 2b shows the frequency of days from event to actual end of observation in censored and uncensored patients. A spike close to zero is apparent in the censored data histogram for most of the outcomes, excluding CVST. This finding indicates the presence of event dependent observation periods (censoring on death date due to outcome), which we tested further with the first and second sensitivity analyses (excluding those who died from the outcome; restricting analysis to the period after vaccination, without censoring at death). These additional analyses agreed with the main analysis, suggesting that there should be little concern about these outcomes.

Absolute measure of risk

In self-controlled case series analysis, results are presented in relative terms—the ratio of the incidence in the exposure risk periods relative to the incidence in control periods. We supplemented these results with measures of effect of each exposure in absolute terms using a method developed to estimate the number of exposures needed to produce one excess adverse outcome and the excess number of events per 10 million exposed for each outcome.24 These measures were computed for a period of 8-28 days after vaccination to remove the healthy vaccinee effect seen at 0-7 days after vaccination. To make numbers comparable, the same measures were computed for the same period after a SARS-CoV-2 positive test.

Negative or positive controls

We examined the associations of exposures with coeliac disease as a negative control outcome25 because it was not thought to be associated with exposure to vaccination or SARS-CoV-2 infection; and with anaphylaxis as a positive control outcome because it could occur shortly after either vaccine.26

Patient and public involvement

This project is supported by a patient and public involvement advisory panel which we thank for its continued support and guidance. The input of the panel has helped us identify priority questions for consideration. PPIE (patient and public involvement and engagement) advisers were supportive of the vital importance of reporting on thrombosis risks associated with vaccination against covid-19 and covid-19 itself. We have reviewed the findings of this study with our PPIE advisers. A lay summary has been developed with patient and public involvement input and feedback, including an infographic.