NON PARAMETRIC BAYESIAN NETWORKS FOR MEASUREMENT ERROR DETECTION

In this paper the problem of detection and correction of errors in the Banca d’Italia Survey on Household Income and Wealth is considered. A novel error correction procedure is proposed. It is based on non parametric Bayesian networks that are graphical models expressing dependence structure via bivariate copulas associated to the edges of the graph. The network structure and the misreport probability are estimated using a validation sample. The correction procedure is applied to bond amounts.