National Trauma Data Bank® (NTDB)

NTDB Imputation Symposium
Notes
Wednesday, March 14, 6:30-9:00 pm
Grand Hyatt Denver, Mt. Harvard

Chair: David Clark

1 ) Welcome and Introductions

2) S. Goble’s presentation “What should we do about missing data?”
Highlighted the three different natures of missing data and the importance of sensitivity analysis. The following missing data methods were discussed: rational substitution, available case analysis, single imputation, hot deck, and multiple imputation.

3) Experiences with imputation by Dr. Mann.
There are 2 sources of missing data in NTDB; missing cases (never makes it into NTDB) and missing data). Highlighted that it is Important to understand the purpose of the study: estimation or hypothesis testing, this will guide which kind of method to use for imputation.  Introduced one additional method of imputation: FIML (Full Information Maximum Likelihood)

4) Dr. Nathan’s presentation: “Implication of missing data on the interpretation of measures of hospital performance” An applied approached was presented based on adjusted ranking of hospitals. The adjusted ranking was first done with all available cases and the change in ranking using the following imputation methods were summarized: 1) add a category for the missing data, 2) mean, 3) regression, 4) hot deck. Approx 91% of the hospitals changed the rank > = 3 when imputation was done. 

5) Dr. Clark Open discussion. Handed out:  “Multiple Imputation: a primer”
J. Schafer, Department of statistics, The Pennsylvania Statue University, University Park,  Pennsylvania, USA.  Discussion on the circular element to using an outcome variable to create the imputed value and then predict the outcome using that imputed value. Dr. Clark presented how other data bases are handling missing data.  The consensus was that NTDB are not ready to include imputed values in the data base, but we should continue to try to educated about missing data. In addition, some facilities are missing so much data so >20% so for those it might not make sense to impute data. Dr. Levhy thought of the imputation more as part of the analysis plan and not the responsibility of the data base.

Action items:
User manual will be updated with guidance for handling missing data
Guidance for handling missing data will be posted on website

NTDB

Trauma Programs