Application of Elastic Net Logistic Regression for Propensity Score Prediction

• Activity Number: 249
• Type: Contributed
• Date/Time: Monday, August 4, 2014 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Consulting
• Author(s): Natalya Makarova*+ and Alparslan Turan and Jarrod E. Dalton
• Companies: Cleveland Clinic and Cleveland Clinic and Cleveland Clinic
• Keywords: shrinkage methods ; elastic net ; propensity score

Abstract

The elastic net is a regularized regression method that linearly combines the penalties of the lasso and ridge "shrinkage" methods. Essentially, elastic net regression is a methodology wherein the overall size of fitted model coefficients is purposely biased towards zero. The elastic net regression effectively selects among redundant and highly correlated potential predictors and produces estimates with predictive errors lower than the full model. In the consulting project we studied the association between presence of rheumatoid arthritis (RA) and postoperative risk of cardiovascular complications. Data contained 21.78 million hospital discharges in 2009-10 across the seven states. Records included age, gender, diagnosis codes with present-on-admission (POA) indicators and procedure codes. We estimated propensity scores (probabilities of having RA) from the aggregated 476 POA diagnosis-related predictors using elastic net logistic regression. Then we matched RA records with controls on age, gender, state of discharge, principal procedure, and a propensity score and run multivariable logistic regression to compare matched RA patients and controls on the outcomes.