Variable Importance Estimation in a Kaiser Permanente Database

• Activity Number: 237
• Type: Contributed
• Date/Time: Monday, August 2, 2010 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Sherri Rose*+ and Bruce Fireman and Mark J. Van der Laan
• Companies: University of California, Berkeley and Kaiser Permanente and University of California, Berkeley
• Keywords: variable importance ; targeted maximum likelihood estimation ; super learner ; effect estimation ; methods

Abstract

Kaiser Permanente is located in Northern California and provided medical services to approximately 350,000 persons over the age of 65 in the year 2003. We were interested in establishing which of 184 disease and diagnosis variables (medical flags) from 2003 had the greatest impact on death in 2004. Variable importance measures were estimated for the top medical flags using targeted maximum likelihood estimation (TMLE). TMLE is a flexible procedure that allows the use of machine learning (such as super learner) to estimate both causal and variable importance parameters. This method provided a ranked list of medical flags in terms of their contributions to determining the outcome of death, adjusted for the remaining medical flags as well as age and gender. Variable importance analysis for 2004, 2005, and 2006 was also performed.