Abstract:
|
EM variable selection (EMVS), a deterministic methodology utilizing the EM-algorithm to perform Bayesian variable selection for normal linear models, reduces computational burden compared to MCMC based methods by estimating the posterior mode instead of sampling the posterior distribution. This method iteratively estimates a covariate's inclusion through an associated missing variable in the E-step and then maximizes the unknown parameters for inference in the M-step. Oftentimes in public health research, there is a need for analyzing binary outcomes of heterogeneous populations, such as disease onset or behavioral trends. This paper develops EMVS for a logistic regression model to select the most probable subset of covariates. We explore this methodology's performance under different priors through simulation, motivated by data collected to investigate the relationship between genetic and psychosocial risk factors and smoking initiation in a cohort of Mexican American youth. We apply the Newton-Raphson method for optimization in the M-step. Logistic EMVS's performance in this work substantiates its usefulness for exploratory variable selection.
|
ASA Meetings Department
732 North Washington Street, Alexandria, VA 22314
(703) 684-1221 • meetings@amstat.org
Copyright © American Statistical Association.