Variable Importance and Prediction Methods for Longitudinal Problems with Missing Variables

• Activity Number: 338
• Type: Topic Contributed
• Date/Time: Tuesday, August 6, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Ivan Diaz*+ and Alan Hubbard and Anna Decker and Mitch Cohen
• Companies: UC Berkeley and UC Berkeley and UC Berkeley and UCSF
• Keywords: Variable importance ; Prediction ; TMLE ; Causal inference ; Longitudinal ; Missing variables

Abstract

We present methods for prediction and variable importance (VIM) for data sets containing both continuous and binary exposures, measured longitudinally, and subject to missingness. We demonstrate the use of these methods with an example about prognosis of medical outcomes of severe trauma patients. Current medical practice involves rules of thumb and scoring methods that only use a few variables and ignore the dynamic and high-dimensional nature of trauma recovery. Well-principled prediction and VIM methods can thus provide a tool to make adequate care decisions informed by the possibly high-dimensional patient's physiological and clinical history. The VIM parameters that we use are based on methods for causal inference, and have a causal interpretation in terms of the expected outcome under a clinical intervention at a given time point. The targeted maximum likelihood estimators presented are consistent and efficient, under certain assumptions. The prediction method used is an ensemble learner that takes a user-given library of prediction algorithms and outputs a linear combination of them that is optimal in the sense that it is asymptotically equivalent to the oracle selector.