A1071
Title: Informed presence in electronic health record data: Illustrating bias reduction approaches in longitudinal analyses
Authors: Yun Li - University of Pennsylvania (United States) [presenting]
Abstract: Electronic health record (EHR) systems capture patient information inconsistently across the lifespan, with patients generally contributing more data when they are sick than when they are healthy. This creates systematic differences in "informed presence" between captured and non-captured data, potentially biasing estimates of association. There is growing interest in analytic approaches accounting for informed presence, but practical approaches for conceptualizing, identifying, and accounting for informed presence in applied EHR-based research have limited attention. For population-level associations in longitudinal settings, a collider-bias framework is used for understanding informed presence bias, and approaches to bias reduction are demonstrated under informed presence. To illustrate, associations are investigated between steroids and cytomegalovirus viremia among pediatric solid organ transplant patients ($N=271$) in a recurrent outcomes analysis. Evidence of informed presence is identified. Incidence rate ratios dropped from 1.83 (95\% CI 1.02, 3.28) in a naive analysis to 1.37 (0.73, 2.57) when accounting for informed presence using inverse intensity weighting. Bootstrapped inverse intensity weighting and multiple outputation resulted in similar estimates of 1.37 (0.71, 2.27) and 1.40 (0.73, 2.68), respectively. When conducting longitudinal analyses with irregularly measured EHR data, it is recommended to account for outcome dependence in statistical analysis.
