Subset Calibration Methods

This project will compare Model-based multiple imputation (MI), survey calibration, and inverse probability of censoring weighted targeted maximum likelihood estimator (IPCW-TMLE) approaches for handling partially observed data on confounders in the setting where the effectiveness/safety of a medication exposure is being evaluated with claims and electronic healthcare record (EHR) data. It will consider a real data example from an integrated healthcare system with data from insurance claims, pharmacy dispensations, and patient care visits. For the primary data example, the workgroup will evaluate the risk of self-harm following initiation of anti-depressant treatment. This setting is a compelling use case as there will be confounders, such as depression severity from the Patient Health Questionnaire (PHQ) and suicidal ideation severity from the Columbia Suicide Severity Rating Scale, available in the EHR only for a subset of individuals/visits. These highly prognostic variables may exhibit differential missingness patterns between different medication exposure groups, a common challenge in pharmacoepidemiological settings of interest to Sentinel. Several numerical experiments will be conducted to compare methods with real and/or simulated data that vary missing data mechanisms. Emphasis will be on comparing methods practical to implement in standard software, with consideration for methods that show promise for application across multiple sites within a distributed data network.