Abstract:
|
The State Inpatient Databases (SID) contain the universe of the inpatient discharge abstracts from hospitals in the U.S., thus providing a unique platform for a broad range of research in healthcare and medicine. As with any large scale data collection effort, the SID have a moderate amount of missing data in several patient-level variables. This study aims at identifying appropriate imputation methods for SID. To accomplish this aim, we compare six imputation methods (i.e., complete case analysis, mean imputation, marginal draw method, hot deck, joint multiple imputation (MI), conditional MI) through simulations. These simulations consider missing observations in continuous, binary, ordinal, and nominal variables. We report root mean square error and bias of the imputed values for continuous variables; the corrected imputed proportion for discrete variables. Simulation results indicate that the differences between these methods are marginal for continuous and binary variables. Conditional MI has the highest correctly imputed proportions for ordinal variables. Hot deck and conditional MI show similar performance and are superior to other methods for nominal variables.
|
ASA Meetings Department
732 North Washington Street, Alexandria, VA 22314
(703) 684-1221 • meetings@amstat.org
Copyright © American Statistical Association.