Abstract
Traditional missing data techniques of imputation of the MAR (missing at random) schemes focus on prediction of the missing value based on other observed values. In the case of continuous missing data the imputation of missing values often focuses on regression models.
In the case of categorical data, usual techniques are then focused on classification techniques which sets the missing value to the ‘most likely’ category. This however leads to overrepresentation of the categories which are in general observed more often and hence can lead to biased results in many tasks especially in the case of presence of dominant categories.
We present original methodology of imputation of missing values which results in the most likely structure (distribution) of the missing data conditional on the observed values. The methodology is based on the assumption that the categorical variable containing the missing values has multinomial distribution.
Values of the parameters of this distribution are than estimated using the multinomial logistic regression.