Evaluation of Score Functions to Aid in the 2002 Census of Agriculture Review Process

The National Agricultural Statistics Service (NASS) has expressed concern about the cost effectiveness of survey editing. This is particularly true since 1997 when NASS assumed the responsibility for the census of agriculture for which hand-editing of all questionnaires is not feasible. In the 1997 Census of Agriculture, all reported data were keyed and then edited by computer. There was no attempt to sort the records according to importance of review. Records flagged for review during the computer edit were reviewed in the order in which they were keyed. Thus time was spent reviewing records which had little impact on any county estimates. Some important records which should have been reviewed more thoroughly probably were not, due to time constraints. Survey estimates can be improved by priority sorting the records so that those records that are more likely to contain errors which will have a significant effect on the county estimates are reviewed first. This priority sorting of records for review is accomplished by first applying a score function and then sorting the records. This paper presents an evaluation of three score functions thought to be most suitable for data collected by NASS. The ultimate goal is to reduce the labor-intensive manual review of data without damaging data quality. The results indicate that using a score function presents an improvement over the previous way of reviewing census records. The paper provides detailed information on each score function, the methodology followed to evaluate each function, the recommended score function to use, and some thoughts for further research.