Una nuova procedura di imputazione di dati mancanti basata sugli alberi di decisione

A procedure based on the use of Sequential Regression and Classification Trees for the imputation of missing data is proposed. We dealt with the case of non-monotone patterns of missing data with mixed measurement level of the variables. The aim of the analysis is to obtain a completed data matrix with optimal characteristics (with respect to means, variances and correlations of the variables) which is often the main demand for a statistical office. Moreover we want to obtain a measure of the additional variability due to the presence of missing values. A simulation case with qualitative and quantitative data is analyzed and the results compared with other procedures. In particular, the performance of Multiple Imputation, using IVEWARE, were compared with our proposal using a large simulation with artificial data and with the EU-SILC cross-sectional data. Our non-parametric method showed to be very competitive on these simulations.