Predicting Unmeasured Outcomes in the Real-World Data: Bayesian and Frequentist Approaches - a Simulation Study

• Activity Number: 355 - Contributed Poster Presentations: Biopharmaceutical Section
• Type: Contributed
• Date/Time: Tuesday, July 30, 2019 : 10:30 AM to 12:20 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Wenyu Ye* and Douglas Faries and Xiang Zhang and janet ford and zbigniew kadziola and Xiaojuan Mi and Ilya Lipkovich
• Companies: Eli Lilly and Company and Eli Lilly & Company and Eli Lilly and Company and Eli Lilly and compnay and Eli Lilly and company and TechData Service Company, LLC and Eli Lilly and Company
• Keywords: prediction; Unmeasured; Real-world; Bayesian; Frequentist; Simulation

Abstract

Background: Administrative claims data are commonly used for research purposes. However,these databases have limited clinical info.Predictive models can be used to overcome this limitation. Little is known about the operating characteristics of various prediction methods.Objective:To compare Bayesian & Frequentist predictive methods when the data source is missing outcomes with or without(w/o) key predictors.Methods:Simulation(SM)data was constructed to include factors of varying correlations w/the missing outcomes & different scenario inclu&excluded unmeasured predictors. Two frequentist models logistic regression & random forest,& a Bayesian hierarchical model were implemented.Predictive performance was evaluated by the concordance (c) statistic & Hosmer-Lemeshow?2.Results:For the scenarios including all predictors(no unmeasured), the three models provided comparable performance & had higher C than those models excluding some predictors. For the scenario with unmeasured factors, Bayesian twin-regression model performed slightly better.Conclusion: SM demonstrated reasonable operating characteristics w/o missing info.Further plasmode SM are necessary to better understand scenarios.