Random Forests for Offcial Statistics Imputation: towards a more efficient methodology

Abstract

Model-based imputation is a widespread practice for handling missing data within official statistics. However, the current model-based methods, such as regression imputation, present drawbacks concerning extensive procedures behind con?guration. They pose challenges relating to efficiency, especially in the context of official statistics production where a great volume of data is processed within limited timespans and resources. Therefore, there is a demand for a direction of methodology which can be implemented efficiently with a reduced burden of con?guration, as well as uniformly across various data circumstances. The present study explored into this direction, by implementing Random Forests which models variable relations automatically. With the focus on minimizing the con?guration procedures to handle speci?c data circumstances, we compared imputation performances of several Random Forests-based methods along with the current regression-based methods. Our results supported the implementation of Random Forests. In particular, Random Forests-median imputation was found as a method with great prospect, which can serve as a reference point for future methodological improvement in efficiency of model-based imputation.