Developing Prediction Rules and Evaluating Observation Patterns Using Categorical Clinical Markers

Substantial uncertainty often remains at the time that important diagnostic or therapeutic decisions must be made, despite the availability of multiple clinical indicators. Multiple in dicators may be used to define observation patterns that are associated with the presence or absence of disease. Clinical prediction rules based on groups of observation patterns have been used to quantify probabilities and reduce error rates for some medical problems, but efficient use of multiple indicators remains a major challenge in medical practice. Medical outcomes and clinical observations are frequently categorical. Two statistical techniques appropriate for generating prediction rules from categorical data are logit analysis (LA) and recursive partitioning analysis (RPA). LA and RPA were compared in evaluating observation patterns for fractures among 666 upper-extremity injuries in children, and in developing prediction rules for selective radiographic assessment. Fracture estimates and error reduc tions provided by RPA and LA were very similar. Each technique generated a set of prediction rules with a range of misclassification probabilities, and evaluated the probabilities of fracture for all observation patterns. LA used more information than RPA in observation pattern evaluations, however, and provided fracture estimates specific to each pattern. With currently available statistical software, RPA output provides better statistical guidance in generating prediction rules, whereas LA provides more statistical information of use in evaluating ob servation patterns. LA warrants attention similar to that conferred on RPA. It appears that complementary use of LA and RPA would be valuable in developing clinical guidelines. Key words: predictive models; observation patterns; prediction rules; logit analysis; recursive partitioning analysis. (Med Decis Making 1993;13:30-42)