Imputation strategies in the trauma registration
Trauma databases often contain relatively high proportions of missing physiologic values. Multiple imputation (MI) could be a possible adequate solution for the missing values. This study aimed to demonstrate the influence of more simplified imputation models on standardized W statistic (Ws) (number of excess survivors per hundred patients that would be achieved if the study center treated identically the same case mix as the reference population).METHODS
Data from three trauma care networks in the Netherlands were used to investigate local differences in missing data. Five different imputation models (MI 1 to 5) were created based on literature and expert opinion. A sixth database was created using maximal single imputation and a seventh database with only complete case analysis (CCA). The Ws values were calculated for the three regions separately.RESULTS
A total of 8,853, 24,487, and 8,599 observations were examined in region 1, region 2, and region 3, respectively. The Ws in region 1 ranged from −0.48 (95% confidence interval [CI], −1.71 to 0.80) for CCA to 0.53 (95% CI, −0.19 to 1.26) for MI 4 and a range of 0.40 (95% CI, −0.91 to 0.10) for CCA to −0.32 (−0.69, 0.04) for MI 1 and MI 4 was found in region 2. The Ws for region 3 ranged from −0.19 (−0.83 to 0.45) in all MI data sets to −0.12 (−0.76 to 0.52) in the CCA data set. Although there were no significant differences between the Ws of the imputation data sets and the CCA analysis, large differences were found in the region with the most missing values.CONCLUSION
Different imputation strategies did influence Ws values. Supplementary variables showed no additional value for the imputation process and a more simplified imputation model could be used to adequately impute missing data.LEVEL OF EVIDENCE
Prognostic, level II.