Statistical models incorporating cluster-specific random effects are frequently used in hierarchical settings, such as observations clustered within patients or patients clustered within hospitals. Predicted values of these random effects are often used to “flag” extreme or outlying values, such as poorly performing hospitals or patients with rapid declines in their health. In this talk we consider whether a class of weighted predictors we previously developed can form the basis of improved flagging methods. We develop novel methods for flagging extreme values that control incorrect flagging rates, including very simple-to-use versions that we call “self-calibrated.” The new methods have higher correct flagging, while also controlling the incorrect flagging rate. We illustrate their application using data on length of stay in pediatric hospitals for children admitted for asthma diagnoses.
Speaker: Charles McCulloch, PhD, Professor of Biostatistics, UCSF
To obtain Zoom link please email: Liz Buggs