Enriching a Large-Scale Survey from a Representative Sample by Data Fusion: Models and Validation

Data Fusion is a series of operations which takes advantage of collected information. Here we present a complete, real practice of Data Fusion, focussing on all the necessary operational steps carried out. These steps define the actual key points of such a procedure: selection of the hinge variables, grafting donors and recipients, choosing the imputation model and assessing the quality of the imputed data. We present a standard methodology for calibrating the convenience of the chosen imputation model. To that end we use a validation suite of seven statistics that measure different facets of the quality of the imputed data: comparing the marginal global statistics, assessing the truthfulness of imputed values and evaluating the goodness of fit of the imputed data. To measure the adequacy of the recipient individuals in respect to the donor set, we compute the significance of the validation statistics by bootstrapping under the assumption that recipients are a random sample of the donor population. To illustrate the proposed approach, we perform a real data fusion operation on the victimization of citizens, where the collected imputation of opinion on perceived safety is used to enrich a large scale survey on citizen victimization.