Identifying predictors of cancer related quality of life using Bayesian model averaging (BMA)
*George Kypriotakis, Geriatric Research Education and Clinical Center (GRECC), Louis Stokes VAMD; CWRU School of Medicine 

Keywords: Quality of life, cancer, Bayesian model averaging

Predicting quality of life of advance stage cancer patients is a critical part in designing interventions, both behavioral and medical. We conduct Bayesian model averaging to account for the large number of potential models. Applying this methods to data from two hospitals (Louis Stokes VA Medical Center and MetroHealth Medical Center, both located in Cleveland Ohio), we identify predictors having a high posterior probability of being significant predictors of quality of life in late stage cancer. We examined the heterogeneity of the results between the two hospitals, evaluate the sensitivity of the results to the selection of both model and parameter priors,and we compared the BMA results to traditional models. Specifically we assessed the models' predictive performance by utilizing the Log Predictive Score, the Continuous Ranked Probability Score (CRPS),and by splitting the samples in testing and training sets. We found the BMA performs better in all measures of predictive accuracy.