Imputing Missing Values: The Effect on the Accuracy of Classification

Abstract

Data from patient records were used to classify cardiac patients as to whether they are likely or unlikely to experience a subsequent morbid event after admission to a hospital. Both a linear discriminant function and a logistic regression equation were developed using a set of nine predictor variables which were chosen on the basis of their correlations with the likelihood of a subsequent morbid event. Once the models were obtained, artificially-generated missing values were replaced with imputed values using mean substitution, regression imputation and hot-deck imputation techniques. The effect on the accuracy of the predictions using models with imputed values was determined by comparing the re-classifications using imputed data with the actual occurrence or non-occurrence of a subsequent morbid event. Mean substitution and hot-deck imputation performed slightly better than regression imputation in this application regardless of whether or not the predictor variable whose values were being imputed was categorical
or numerical.