Informative missingness in electronic health record systems: the curse of knowing

Table 1 Measures of predictive performance under different scenarios of missing data

Scenario–method	Mean prediction error (SD)	RMSPE	C-statistic	Brier score	Calibration-in-the-large
Reference
No missing values	− 0.009 (0.244)	0.244	0.663	0.209	0.016
Scenario 1
Zero imputation	− 0.004 (0.217)	0.217	0.699	0.197	0.019
Mean imputation	− 0.004 (0.217)	0.217	0.699	0.197	0.023
CCA	− 0.005 (0.244)	0.244	0.618	0.239	0.017
Multiple imputation	− 0.005 (0.269)	0.269	0.622	0.216	0.021
Scenario 2
Zero imputation	0.104 (0.245)	0.266	0.663	0.220	− 0.467
Mean imputation	0.104 (0.245)	0.266	0.663	0.220	− 0.467
CCA	0.104 (0.245)	0.266	0.663	0.220	− 0.467
Multiple imputation	− 0.042 (0.246)	0.249	0.663	0.211	0.199
Scenario 3
Zero imputation	− 0.024 (0.292)	0.293	0.541	0.234	0.119
Mean imputation	− 0.040 (0.299)	0.302	0.541	0.239	0.210
CCA	− 0.104 (0.245)	0.266	0.662	0.220	− 0.461
Multiple imputation	− 0.043 (0.264)	0.268	0.663	0.212	0.207
Scenario 4
Zero imputation	− 0.151 (0.278)	0.316	0.500	0.248	0.782

Abbreviations: CCA complete case analysis, SD standard deviation, RMPSE root mean squared prediction error. See main text for a description of the scenarios and details about the methods

ISSN: 2397-7523