Comparing the Performance of Various Disease Risk Scores, Propensity Scores, Multivariate Logistic Regression, and Log-Binomial Regression Using Simulation

• Activity Number: 360
• Type: Contributed
• Date/Time: Tuesday, August 6, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): In-Lu Liu*+ and Jiaxiao Shi and Wansu Chen
• Companies: Kaiser Permanente and Kaiser Permanente and Kaiser Permanente
• Keywords: Propensity score ; Disease risk score ; logistic regression ; log-binomial model

Abstract

Propensity score (PS) and Disease risk score (DRS) are popular balancing covariates tools in observational studies. However, their strengths and weaknesses compared to the multiple logistic regression (MLR) and log-binomial model (LB) have not been fully examined. We conducted simulation to evaluate the performance of DRS, two DRS modifications, the probability and logit-based PS, LB and MLR. When the association among the outcome, covariates and exposure was moderate, outcome not being extremely rare (eg. 5-10%) and exposure rate being moderate (20-35%), DRS was an acceptable alternative to PS, LB or MLR. When the association between covariates and exposure or outcome was high for some covariates, most DRS, some PS methods performed poorly. DRS models with only confounders had the smallest bias and MSE compared to the models including covariates that were not related outcome or exposure. Among the three DRS, the one generated by models without exposure variables tended to perform better in bias, MSE and coverage. Users need to be aware of outcome and exposure rate, the association among the outcome, covariates and exposure, before choosing an appropriate method.