Implicit Linear Inequality Edits Generation and Error Localization

The Census Bureau’s SPEER editing system applies the Fellegi-Holt model to economic establishment surveys under ratio edits and a limited form of balancing. If the complete set of explicit and implicit edits is available, then Fellegi-Holt methods have the advantage that they determine the minimal number of fields to change so that a record satisfies all edits in one pass through the data. For most surveys, implicit edits are not generated because the generation requires days-to-months of computation. In some situations, when implicit edits are not available, Fellegi-Holt systems use integer programming methods to solve the error localization problem directly and slowly. With only a small subset of the needed implied edits, the current version of SPEER (Draper and Winkler, 1997) applies ad hoc heuristics that finds error-localization solutions that are not optimal for as much as five percent of the edit-failing records. In this paper we present recent modifications to the SPEER editing system that maintain its exceptional speed and do a better job of error localization. The new SPEER has an auxiliary program for generating implicit linear inequality edits using the Fourier-Motzkin elimination method. Using this methodology we generate a large subset of the implied edits prior to SPEER error localization. We present the theory, computational algorithms, and results from evaluating the feasibility of this approach