Abstract:

Data fusion is today becoming a more and more widely used technique : when confronted for instance with the problem of dealing with media/product type of data, it is clear that "marrying together information from different surveys" [1] is both cheaper and less demanding for the respondents to each of the two surveys. Therefore, a lot of efforts have been spent by statisticians to develop various fusion techniques, allowing to "transfer information from a donor to a recipient" in some way, these procedures being mainly adapted from "the well established set of statistical procedures for dealing with missing data" [2]. However, almost all authors start saying that "the most desirable way (...) would be to analyse a single source data base" [1], This is the reason why it is first necessary to review the real value of the statistical fusion process by itself : the criteria presently used by practitioners may be seen more as empirical evaluation rules than as TRUE statistical tests. A complete evaluation of the fusion process has to be based on a clear understanding of what exactly is expected from the "synthetic file" obtained after applying a matching procedure between two initial different files.