Online Program

Return to main conference page
Friday, February 15
Fri, Feb 15, 2:00 PM - 3:30 PM
Canal
Addressing Problematic Data

Dealing with Missing Data in a Multi-Item ICU-Mortality Scale: A Comparison of Multiple Imputation Methods (303752)

View Presentation View Presentation

*Chia-Ling Kuo, University of Connecticut Health 
Jinjian Mu, University of Connecticut 

Keywords: multiple imputation, SOFA score, composite score, R package: mice, multivariate imputation by chained equations

Missing data is particularly a problem in composite scores that require multiple items and cannot be calculated if any of the items are missing. The literature has favored multiple imputation methods and item-level imputations over scale-level imputations. Multiple imputation procedures that assume a joint model for all the variables are not feasible for large-scale data with many variables and variables in different scales. Alternatively, multivariate imputation by chained equations (MICE) can handle a huge number of variables and allows for a combination of variable types such as numeric, factor, and ordinal variables. We compare a selection of multiple imputation methods that are implemented in the mice R package. The simulations are conducted based on a real data that was initially collected to study the impact of methicillin-resistant staphylococcus aureus colonization on post-intensive care unit mortality where multiple imputation was performed to impute Sequential Organ Failure Assessment (SOFA) scores. The objective of this study is to suggest best methods for SOFA score imputation and to address whether we should exclude patients with most of SOFA items missing.