Prediction and Variable Importance in the Prospective, Observational, Multi-Center Massive Transfusion Study (PROMMTT)

• Activity Number: 550
• Type: Contributed
• Date/Time: Wednesday, August 7, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Ivan Diaz and Anna Decker*+ and Alan Hubbard and Ivan Diaz and Mitch Cohen
• Companies: UC Berkeley and UC Berkeley and UC Berkeley and UC Berkeley and UCSF
• Keywords: prediction ; semiparametric ; causal ; cross-validation

Abstract

We present results from a semiparametric prediction and variable importance analysis of data from the PRospective, Observational, Multi-Center Massive Transfusion sTudy (PROMMTT). Our goal was to identify among the 10 sites variables that have different associations of prognostic variables and death from severe trauma. We estimated the cross-validated risk (based on misclassification), where the validation samples were defined as the sites. We defined our variable importance measures (VIM) as parameters motivated by the causal inference literature, such as the average treatment effect, defined as the change in the risk of death under interventions on blood product allocation. The cross-validated predictions of mortality at each site and the models on which the VIM's were based were based on SuperLearner, an ensemble learner that takes a library of algorithms and determines an optimal combination by minimizing the cross-validated risk. The semiparametric estimation of the VIM's used fits from SuperLearner and a bias reduction step based on targeted maximum-likelihood. The analysis identified 2 sites that appear to have different relationships of prognostic variables and death.