Comprehensive data on 3,579 patients in Baltimore nursing homes were collected by trained teams of reliable observers. The data included an implicit judgment about the level of care (LOC) deemed most appropriate for each patient. Two different approaches were used to fit the patient data to the LOC judgments: Several versions of an algorithm adapted from a utilization review scheme based on simple clinical criteria and a series of mathematical equations based on logistic regression were each tested. In the case of the latter, each equation was derived from a random half of the data and tested on the remaining half.Both approaches yielded comparable results. The best variation of the algorithm correctly identified 71 per cent of patients needing skilled care and 69 per cent of those needing unskilled care. The equations based on the logistic regressions could correctly classify 86 per cent of those needing skilled care and 63 per cent of those not needing such care. Further improvements in accuracy of prediction on one group came at the cost of less accurate identification of the other. In general, the simplest models proved the most useful.These techniques are recommended as useful for preliminary screening of nursing home patients for appropriate LOC but should not be used as a basis for final judgments. The advantages offered by this approach are the reduction of workload demands for skilled professional judgment and the availability of a very reliable preliminary screening judgment in a highly politicized atmosphere.