What’s Missing? Effects of Missing Data and Imputation on Predictive Performance

• In Event: Conflict Modeling, Prediction and Forecasting

Abstract

Missing data is a ubiquitous problem in science in general, and social science in particular. It is well-known that not dealing properly with the missing data problem may cause inferences to be biased. Less attention has been given to the implication of missing data and imputation techniques on the predictive performance of different models. This paper explores these issues through generating missing data in a dataset used for forecasting the number of fatalities from political violence in all countries during 1989-2021. The results show that the choice of imputation technique should be informed by a number of factors such as the suspected missingness mechanism, the choice of forecasting model, whether missingness is intermittent or case-wise, and the proportion of missing values. Intermittently missing data can safely be imputed using simple and computationally inexpensive methods such as last observation carried forward/backward while case-wise missing data are best replaced using multiple imputation techniques. The paper also highlights broader implications of the missing data problem in predictive studies in political science and the use of missing data as data in order to improve performance of predictive models.

Author

• David Randahl, Department of Peace and Conflict Research, Uppsala University