While most vaccine safety research has focused on acute adverse events, a 2013 Institute of Medicine report recommended observational studies on chronic, long-term outcomes following cumulative exposure to vaccines in early childhood. The Centers of Disease Control and Prevention's Vaccine Safety Datalink (VSD) has initiated such research using electronic health record (EHR) and medical claims data from health plans nationwide; however, outcome misclassification has been cited as a major methodological challenge. Misclassification of clinical outcomes has garnered considerable attention in EHR-based research, with numerous studies validating algorithms for identifying outcomes in these data. However, there has been minimal attention given to using results from such validation studies to quantify and correct for bias of an exposure-outcome association caused by such misclassification.

In this dissertation, simulations of VSD immunization schedule studies were developed to measure the magnitude of bias caused by outcome misclassification, and to test analytic solutions for this problem. The simulations' results revealed that previous approaches, such as estimating overall outcome positive predictive values (PPVs), offer limited benefits for correcting bias (Aim 1a), while quantitative bias analysis (QBA) methods are effective (Aim 1b). For example, within a cohort of n= 257,010 children, simulations showed that Type 1 error rates of 100% occurred with outcome PPVs of 90% when differential outcome specificity was present. Since PPVs are the most commonly-reported estimate of algorithm validity, lesser-known bias analysis methods using only predictive values were evaluated (Aim 2). These results revealed that outcome PPVs should be estimated by exposure, a departure from current practices in EHR-based research. Practical guidance was developed for integration of quantitative bias analysis within immunization schedule research (Aim 3), including demonstrating these methods within an example study of immunization schedule exposure and pediatric asthma.

Bias from misclassification of clinical outcomes in EHR data is a long-recognized problem with underappreciated solutions. Failure to address outcome misclassification bias could lead to erroneous conclusions about the safety of the early childhood immunization schedule. The work presented in this dissertation challenges prevailing assumptions about the impact of EHR data misclassification, and establishes a foundation for implementing methodological solutions to address this problem.