Multivariate estimation of reference ranges.
This paper describes a new method for estimating reference ranges of 21 frequently performed laboratory tests utilizing routine test results of selected hospital patients. Computer programs perform the selection of patient records and the statistical analysis. The patient selection criteria of (1) short hospital stay and (2) short laboratory record provide a data base rich in normal results. The statistical program excludes records with excessive variation, and in repeated data-exclusion passes through the data base progressively narrows the range of acceptability. The decision to exclude a record is based on the "pooled variation." The average value of the "pooled variation" increases as the range is narrowed. This approach insures that the process stops after a few passes, and the most consistent data remain included in the calculation of reference ranges. Statistical properties of the data sets, as well as comparison with normal ranges for healthy subjects and review of 560 clinical charts, support the idea that serviceable reference ranges are obtained. The method is suitable for monitoring the reference ranges of the laboratory at frequent intervals.