(P-023) Handling with Vaccine Type Missing Data in a Dynamic Cohort to Assess the Link Between Time-varying Vaccination and an Autoimmune Disease

Background: Types of vaccines administrated at different number of doses and dates may be missing in health records. In Pharmacoepidemiology, handle with missing data in dynamic cohorts is required.

Objectives: To propose a method to single stochastic imputation of the type of administrated human papillomavirus vaccine (whether bi- or quadri-valent HPVv; i.e. exposure) scheduled as 2 or 3 doses in a time-varying cohort study and to compare the Hazard Ratios (HR) of thyroiditis (outcome) resulting from different strategies to deal with missing data.

Methods: A cohort study was performed using Spanish health records (BIFAP) including girls aged 9-18 years followed-up to thyroiditis in 2007-2016. Follow-up time was divided into non-exposed (reference), exposed to bi-, quadri-valent or missing type HPVv (up-to-the risk window of 2-years) and thereafter. The vaccine type for all the jabs was obtained as follows: 1) into each girl, the recorded vaccine type was carried forward or backward to the other jabs; 2) perform a single stochastic imputation based on a) determinants of both, missing data and values of 1st dose vaccine type (identified through logistic regression in a dataset with unique identifiers), b) confounders by automatic backward stepwise selection method in a dataset with multiple observations per individual (identified through risk model) and c) outcome. HR after imputation (gold standard) were compared to analysis using only girls with type recorded (i.e. complete cases; CC) and all girls, including in a missing category (MiCat) those without type.

Results: 808,201 observations for 388,411 girls were built. Type was carried out in 2.84% of 153,924 vaccinated girls remaining 35% for imputation. 15 Determinants and 21 confounders were identified. HR departed by up to 24% overestimation for bi- and 22% underestimation for quadri- valent in the MiCat, whilst 45% and 12% respectively in the CC.

Conclusions: Multiple identified determinants of vaccine type missing and values suggested missing at random (MAR) pattern but not completely at random (MCAR). Thus, CC was not trusty and additionally lost power. The proposed method is simple, but not as simple as the CC approach or Micat, which generate bias because the missing values are not completely at random. The MiCat, that does not correct for the uncertainty associated with the missing data, suggested imprecise HR in both directions.