gms | German Medical Science

Introduction: With unprecedented speed, a mass of vaccines has been employed to combat the COVID-19 pandemic. In Germany alone, 171.876.955 doses of different vaccines have been used until 30.03.2022 [1]. Because of small sample sizes and strict inclusion criteria, previous phase-3 studies are not sufficient to ensure the safety of applied vaccines [2]. However, routine health-care data can be employed to mitigate these problems. Despite the large number of persons enrolled in health insurances, it is not trivial to analyse vaccination-related data in Germany. Generally, an individual’s vaccination status is recorded independently from the person’s other health-care data in separate databases, such as the Digitales Impfquoten-Monitoring database [1]. Due to a lack of a persistent and database-independent person identifier, error-prone record linkage techniques have to be used to merge these databases. We aim at quantifying the impact of these record-linkage errors on the power and bias of different analysis methods when assessing COVID-19 vaccine safety based on German health-care data.

Methods: We conducted a Monte Carlo simulation study including both the Self-Controlled Case Series (SCCS) [3] method and Cox-Regression with time-dependent covariates [4] under different proportions of linkage errors. The data have been generated using a discrete-time simulation, with vaccination-related parameters retrieved from the public database operated by the Robert-Koch Institute [1]. To make the simulation as realistic as possible, we have chosen baseline and vaccine-related increase of incidence for the well documented increased risk of acute myocarditis following a mRNA vaccination [5], [6]. However, the discussed results hold for more general cases as well.

Results: Both the SCCS method and the Cox-Regression produced largely unbiased estimates with a sufficiently high power when up to 10% of the potential matches were missing. With a rising proportion of missing matches, however, the performance of the Cox-Regression deteriorated sharply, showing a substantial amount of bias and a loss in power. Specifically, the cox-regression method underestimates the true risk of undesired events following vaccination. In contrast, the SCCS method still produced unbiased results in the average, although with decreased power.

Discussion: If the information on a person's vaccination status cannot be linked to the person's health-care data, then this person necessarily has to be considered unvaccinated in the analysis. The Cox regression, based on comparing states of different individuals, consequently underestimates the true risk and most likely overestimates the baseline incidence. Since the SCCS method does not compare different individuals, but different time periods within each individual, this method is robust against random linkage. In practice, however, time-varying confounders such as, e.g., the COVID-19 infection, might complicate the mere use of SCCS designs. Due to their different susceptibilities to different types of biases, it has been recommended to use both analysis strategies [7].

Conclusion: Small amounts of random linkage errors have little impact on both bias and power of the analysis regardless of which method has been chosen. Increasing proportions of missing matches cause increasing bias resulting from Cox-regressions but not from SCCS analyses.

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.