Statistical Match of Samples Using Additional Information

Abstract

Statistical matching methods combine records from two samples with different but overlapping sets of variables in order to construct a new sample that comprises all variables. From theoretical reasons the two samples do not contain information about all statistical interactions present in U. Therefore every classical statistical matching method implicitly presupposes the conditional independence of specific variables. However often this assumption is violated and this causes distortions of the resulting sample. In this paper information from additional samples is used to replace the conditional independence assumption. The EM-algorithm is used to estimate population parameters from different sources. Using these estimates several criteria for the combination of records are constructed.