Multiple imputation of the Glasgow Coma Score

Abstract

Background: To investigate whether multiple imputation (MI) of missing Glasgow Coma Scale (GCS) values generates more accurate GCS frequency distributions and less biased parameter estimates in logistic regression models predicting mortality than the standard procedure of excluding observations with missing GCS values.

Methods: The study population consisted of 5,065 patients with complete GCS information from the trauma registry of a Level 1 trauma center. Missing GCS values were imposed on the data set, and the performance of MI (extrapolating missing GCS from a data prediction model) and of deleting all data observations with missing GCS (list-wise deletion) were evaluated. GCS and Trauma and Injury Severity Score (TRISS) frequency distributions and parameter estimates were compared with true values from the original data set.

Results: GCS and TRISS frequency values generated by MI were much more accurate than those generated by list-wise deletion. GCS and TRISS parameter estimates generated by MI all had acceptable bias and coverage rates when compared with true values. List-wise deletion provided biased parameter estimates for the GCS, the Revised Trauma Score, and the Injury Severity Score.

Conclusion: MI is a valid solution to the problem of missing GCS data in trauma research. It allows the conservation of precious data observations and leads to unbiased estimates in consequent analyses. Analyses, which exclude observations with missing GCS data, provide biased results.