An application of a complex measure to model–based imputation in business statistics

When faced with missing data in a statistical survey or administrative sources, imputation is frequently used in order to fill the gaps and reduce the major part of bias that can affect aggregated estimates as a consequence of these gaps. This paper presents research on the efficiency of model-based imputation in business statistics, where the explanatory variable is a complex measure constructed by taxonomic methods. The proposed approach involves selecting explanatory variables that fit best in terms of variation and correlation from a set of possible explanatory variables for imputed information, and then replacing them with a single complex measure (meta-feature) exploiting their whole informational potential. This meta-feature is constructed as a function of a median distance of given objects from the benchmark of development. A simulation study and empirical study were used to verify the efficiency of the proposed approach. The paper also presents five types of similar techniques: ratio imputation, regression imputation, regression imputation with iteration, predictive mean matching and the propensity score method. The second study presented in the paper involved a simulation of missing data using IT business data from the California State University in Los Angeles, USA. The results show that models with a strong dependence on functional form assumptions can be improved by using a complex measure to summarize the predictor variables rather than the variables themselves (raw or normalized).