A0184
Title: Robustness under missing data: A comparison with special attention to inference
Authors: Carole Baum - ULiege (Belgium) [presenting]
Arnout Van Messem - ULiege (Belgium)
Holger Cevallos-Valdiviezo - ESPOL Polytechnic University (Ecuador)
Abstract: Missing value imputation is a highly studied topic. A plethora of techniques have been proposed over the years to find suitable values to replace missing data. Nowadays, imputation techniques are widely used, but a large-scale comparison of these methods - especially in terms of their robustness against outliers - seems to be missing. During a first attempt to fill this gap, we evaluate a large selection of imputation techniques involving classic and robust procedures by means of a simulation study with continuous data and different configurations of missing data and outliers. To evaluate the imputation capability and robustness of the imputation techniques, we computed the error between the original and the imputed values. However, often, the main concern is on the analysis that is performed after imputation. Therefore, in the second phase of our research, we evaluated the inferences and predictions made by different robust regression methods combined with an imputation technique in a simulation study. Both row-wise and cellwise outliers were generated, so we considered in the evaluation row-wise robust regression techniques as well as cellwise robust regression techniques. To evaluate the combined regression and imputation strategies in terms of inference capability, we measured the bias and variance of the estimated regression coefficients.
