Assessing treatment effects in observational data with missing or mismeasured confounders: A comparative study of practical doubly-robust and traditional missing data methods

Music City Center 

For safety and rare outcome studies in pharmacoepidemiology, multiple, large databases are often merged to improve statistical power and create a more generalizable cohort. Medical claims data have become a mainstay in evaluating the safety and effectiveness of medications post-approval, but confounders derived from administrative data can be prone to measurement error. Electronic health records (EHR) data or data abstracted from chart review have more granular patient data than do medical claims, but the gold standard exposure data may only be available on a subset. I will discuss two practical-to-implement doubly-robust estimators for this setting, one relying on a type of survey calibration and another utilizing targeted maximum likelihood estimation (TMLE), and compare their performance with that of more traditional missing data methods in a detailed numerical study. Numerical work includes plasmode simulation studies that emulate the complex data structure of a real large electronic health records cohort in order to compare anti-depressant therapies in a setting where a key confounder is prone to missingness.

doubly-robust methods

missing data

electronic health records

targeted maximum likelihood estimation

generalized raking

survey calibration