What makes for a good missing value imputation? New algorithmic approaches in PCA and random forests-based techniques

The increasing availability of data often characterized by missing values has paved the way for the development of new powerful algorithmic imputation methods for handling missing data, among which two recent proposals seem most promising: Stekhoven and Bühlmann's method (missForest), a nonparametric technique based on a random forest, and Josse, Pagès, and Husson's imputation method (missMDA) based on the EM-PCA algorithm. In this work, a new PCA-based procedure is developed by drawing on the forward-imputation approach introduced by Ferrari, Annoni, Barbiero, and Manzi in the context of ordinal data (ForImp). Comparisons with the two methods above are then considered