Response time of an information system can be improved by reducing the number of buckets accessed when retrieving a document set. One approach is to cluster the document base in such a way to ensure greater probability that identifying records will be physically contiguously located. This paper considers a clustering algorithm that controls the file space density through the use of a user-specified threshold value for which we will show how this influences file density and retrieval performance. The method is particularly well-suited to the reorganization of traditional multi-attribute files. Statistical models are constructed to predict the file performance parameters. Our results reveal those models, before and after clustering, are reasonably accurate.