Building predictive medical models on incomplete data

While working on the National Cardiovascular Network (NCN) Outcomes Management Report our group was confronted with a high percentage of missing data, despite the large size of our registry. One of our goals was to find a way to compare the results achieved at different sites. Excluding cases with missing data significantly decreased the number of cases and, in some instances, all the data from a particular center was eliminated, thereby removing them from comparison. To avoid such a scenario, we utilized multiple imputation. The obtained results and methods used are subject of this article.