Estimation of Influenza Vaccine Effectiveness using Secondary Data: A Cohort Study and Propensity Score-Matched Analysis of Claims Data from Baden-Wuerttemberg

Our objective was to estimate and replicate influenza vaccine effectiveness (VE) for the 2014/2015 influenza season (IS) based on routine data from a German health insurance claims dataset. In addition, we investigated associated methodological aspects. From the AOK Baden-Wurttemberg, claims data of 2.64 million insured persons residing in Baden-Wuerttemberg and aged 15 years and older were available for analysis. Based on influenza vaccine-specific reimbursement codes claimed in the vaccination period of 2014, participants were classified as either vaccinated or unvaccinated. Baseline covariates that could confound the association between vaccination and influenza infection were considered for all participants. Covariates included age, sex, place of residence, and covariates indicative of health status and health-service utilization. The primary outcome was defined as influenza hospitalization during the IS in winter and spring of 2015. Secondary outcomes included pneumonia hospitalizations, and all-cause mortality among others. Propensity score matching (PSM) was used to build a comparable set of vaccinated and unvaccinated participants. A bias analysis was conducted by estimating VE pre- and post-IS, periods in which vaccination is not thought to be effective, because influenza is not circulating in the population. A subset of 839,706 participants could successfully be matched with a 1:1 ratio. The estimated influenza VE (based on influenza hospitalization) was 27% [95% confidence interval (CI): 17%; 36%], which compares well with the estimate of the RKI for the same season (27% [95% CI: -1%; 47%]). Bias analysis revealed that result could be partially accounted for by residual confounding yielding a potential overestimation of the true underlying effect. Secondary outcomes for pneumonia hospitalizations and mortality showed similar results though likely prone to a greater extent of residual confounding. It can be concluded that (1) secondary data from German health insurances can be used to derive plausible influenza VE estimates, and (2) PSM is a useful and transparent method to derive those estimates. In addition, (3) residual confounding is a relevant problem in observational studies on VE and (4) bias analysis in pre- and post-season periods are an essential complement for interpretation of results.