A comparison of approaches for implementation of k-nearest neighbour imputation for missing items in cross-national, time series data sets of economic indicators

Abstract

The need for predictive accuracy in the imputation for missing data in cross‐national, time series data is discussed and the possibility of requiring unconventional approaches to imputation, namely approaches which are tailored to the specific context and applied to individual instances of missing items is also discussed.  Potential barriers to moving toward such an approach are mentioned and in particular, the demands on   resources implied by that. A taxonomy of available observations is established with the aim of being able to use it to quickly and efficiently identify potential solutions for imputing missing data. A simulation study is conducted in which the relative performance of different k‐nearest neighbor imputation implementations are related to the context in which they are set to operate with a view to providing practitioners with a‐priori understanding of which techniques are likely to perform better under any given particular set of circumstances. A multinomial model is used to begin to investigate the interaction between imputation implementations, and the role that context might play in the accuracy of their imputations.

More information

Identifiers

ORCID for Nikolaos Tzavidis: orcid.org/0000-0002-8413-8095

Catalogue record

Export record