The use of partitioning methods for classification is discussed. A brief outline of the method of recursive partitioning is given and a note is made of some of its potential drawbacks. An alternative approach is outlined in which a particular class of dichotomous partitions is considered. The class incorporates prior knowledge concerning the nature of an association. Strategies to choose a partition from this class are suggested. The method is illustrated by an analysis of data from patients with gastrointestinal cancer and benign disease. A partition to discriminate between cancer and benign disease is obtained using symptoms, age and tumour marker data.