Algorithmic imputation techniques for missing data: performance comparisons and development perspectives

In recent years, much research has been devoted to solve the problem of missing data imputation. Although most of the novel proposals look attractive for some reason, less attention has been paid to the problem of when and why a particular method should be chosen while discarding the others. This matter is far crucial in applications, given that unsuitable solutions could heavily affect the reliability of statistical analyses. Starting from this, this work is addressed to study how well several algorithmic-type imputation methods perform in the case of quantitative data. We focus on three different logics of imputing, based respectively on the use of random forests, iterative PCA, and the forward procedure. In particular, the latter, having initially been introduced for ordinal data, has required us to develop an original adaptation so that it handles missing quantitative values.