The Impact of Modeling Strategy and Informative Censoring on Estimated Readmission Risk

• Activity Number: 308
• Type: Contributed
• Date/Time: Tuesday, August 11, 2015 : 8:30 AM to 10:20 AM
• Sponsor: Health Policy Statistics Section
• Author(s): Brittney Bailey* and Abigail Shoben and Randi Foraker
• Companies: The Ohio State University and The Ohio State University and The Ohio State University
• Keywords: 30-day readmission ; Cox proportional hazards regression ; heart failure ; simulation ; logistic regression

Abstract

Penalties for exceeding a risk-standardized threshold in 30-day readmissions have prompted hospitals to investigate predictors of readmission in conditions such as heart failure. However, different modeling approaches have produced discrepant results, in magnitude and/or direction, even when the same predictors are included. Accordingly, our goal was to compare modeling approaches in order to provide guidance to investigators and policy-makers in interpreting previous results and conducting future studies. We fit two logistic models with different exclusion criteria and one Cox proportional hazards model to simulated data, varying non-administrative censoring scenarios, readmission times, and administrative censoring times in order to evaluate their respective impacts. The Cox model produced stable estimates of the hazard ratios that appropriately reflected the differences in readmission between levels of exposure, whereas the logistic models produced highly variable estimates of the odds ratios driven by type and degree of censoring as well as by choice of exclusion criteria. To avoid the latter issues, we recommend using Cox regression in future studies of readmission.