Using directed acyclic graphs to ajust for confounding in influenza vaccine effectiveness studies

The annual estimation of the Influenza Vaccine Effectiveness (IVE) requires the implementation of observational studies that are prone to potential bias. Significance testing or change-in-point estimate approaches have been commonly used to selectect covariates for models adjustment in IVE studies. In EuroEVA, the test-negative study Portuguese component of the IMOVE network, for the past 8 seasons the potential confounders were identified if they changed crude OR estimate. The set of variables changed every season and some model adjustment issues were raised. This study aimed evaluating the use of causal directed acyclic graph (DAG) approach for IVE studies to allow adjustment with a more consistent and stable set of variables. We used data from EuroEVA study from seasons 2011-12 to 2015-16. The set of collected variables included sociodemographic (sex, education, cohabitants), health status (immunocompromising, non-immunocompromising conditions), health behaviour (smoking), severity (hospitalizations), previous vaccination, functional status (help to bath), health seeking behaviour (nr of GP consultations), target group for vaccination (health professional or caregiver), time in the influenza season. Three steps were done to build the theoretical model of causal relation between influenza vaccination and medically attended influenza: 1) literature review; 2) development of DAG using DAGitty software; 3) adjustment of logistic regression model using minimum sufficient set of covariates to adjust for confounding (testing for conditional independencies). DAGitty proposed three different models, where the variables age, both health conditions, being health professional and time in the season were common to all models. Other additional variables included were smoking and education (model 1), nr of GP visits (model 2) and sex (model 3). IVE were 58% (95%CI:41%-70%), 57% (95%CI: 40%-70%) and 58% (95%CI:40%-70%) for model 1, 2 and 3 respectively. The minimum variable set to adjust for confounding included six variables. Both IVE adjusted estimates and confidence intervals were similar between the three proposed models. DAG resulted in a good framework to study analysis and leaded to more stable set of variables that allows inter season comparison. However, it should be noted that two important variable were not included (previous infection and vaccination). Also, DAG needs to be updated if changes occur in the vaccination programme or individuals behaviour towards the vaccine and infection.The project 'I-MOVE+ Integrated Monitoring of Vaccines Effects in Europe: a platform to measure and compare effectiveness and impact of influenza and pneumococcal vaccines and vaccination strategies in the elderly' received funding from the European Union´s Horizon 2020 research and innovation programme under grant agreement Nº 634446N/