PRM41 - Data Imputation for Missing Values in a Claim-Based Administrative Database: Comparison of Imputatation Approaches

Abstract

A federal-founded project aims to evaluate an intensive care management (ICM) model among Medicaid-insured youth with severe mental health disorders. Because many discharge dates were missing due to administrative issues, researchers faced technical challenges to examine long-term benefits after youth were discharged from the ICM. The objectives of this study were to compare two data imputation methods and to provide a framework to evaluate the performance of imputed data. Data were from one Mid-Atlantic state Medicaid claims linked with the ICM records. First, youth with completed discharge dates were identified and randomly assigned into two datasets: training and test set. Within the training set, we conducted regression-based single imputation to predict the length of stay (LOS) in ICM. Second, discharge dates in the test set were treated as missing. Regression coefficients from the training set and multiple imputation were applied to the test set, respectively. Third, we compared the LOS distribution and cross-validation metric- root mean squared error (RMSE)- to determine which imputation method performs better. Finally, we applied the better imputation method to youth with real discharge dates. Of the 709 youth who enrolled in the ICM, 126 (21.6%) had missing discharge dates. Compared with single imputation, multiple imputation performed better based on the similar distribution of LOS in ICM and smaller RMSE between actual and imputed values (RMSE =257.6 for single imputation; RMSE= 208.7 for multiple imputation). We applied the multiple imputation method to impute the LOS in ICM for youth with real missing dates. The study findings suggested multiple imputation performed better than single imputation if the missing observations cannot be simply ignored. Future work is needed to investigate the proposed framework, develop methods for data imputation diagnostics, and provide more guidance on handling missing data in observational studies.