Using Simulation to Compare Performance of Various Variable Selection Methods on Disease Risk Scores, Propensity Scores, and Log-Binomial Regression

• Activity Number: 564
• Type: Contributed
• Date/Time: Wednesday, August 6, 2014 : 2:00 PM to 3:50 PM
• Sponsor: Biopharmaceutical Section
• Author(s): In-Lu Amy Liu*+ and Jiaxiao Shi and Wansu Chen
• Companies: Kaiser Permanente and Kaiser Permanente and Kaiser Permanente
• Keywords: Disease risk score ; Variables Selection Methods ; Propensity score ; Log-binomial model

Abstract

In observational studies, propensity score (PS) and Disease risk score (DRS) are popular covariates balancing methods. However, very little has been reported on the issue of variable selection for DRS models. We conducted simulations to evaluate the performance of various variable selection methods (confounders only/ outcome related covariates/ exposure related covariates/ all) in a DRS and two types of modified-DRS models, a probability-based and a logit-based PS models, and a log-binomial covariate adjustment model (LB). For DRS, particularly in the scenarios with rare outcome, high exposure and high relative risk (RR), confounders-only or exposure-related-only covariates models generally performed much better than outcome-related-only or all covariates models. For PS and LB, confounders-only or outcome-related-only covariates models usually performed better than exposure-related-only or all covariates models in scenarios with rare outcome, low exposure and high RR. The differences in performance among 4 variable selections were much larger in DRS models than in PS and LB models. Users need to select appropriate variables for DRS, PS and LB models to reach optimal results.