Performance evaluation of imputation based on Bayesian Networks

Abstract

The issue of missing data may arise for researchers who deal with data gathering problems. Different methods of missing data imputation have been proposed to deal with such problems. The Bayesian Network is one of the proposed methods that has been recently used in missing data imputation. In this research, to consider the effect of different kinds of missingness mechanism (ignorable and nonignorable) on the performance of imputation methods, three methods of imputation: random overall hot-deck imputation, within-class random hot-deck imputation and imputation using Bayesian Networks are compared using two indices: (1) a distance function and (2) a Modified Kullback-Leibler index. In addition, by applying Value of Information Analysis (VIA), the performance of Bayesian Networks when the missingness mechanism is nonignorable, is examined and compared with that of Bayesian Networks constructed from data perturbed by ignorable mechanisms. The comparison is made using Mean Value of Information (MVI) index. Results indicate the high quality of Bayesian Networks relative to other imputation methods.