When categorical measures are taken on independent groups, the statistical procedure of choice is the familiar Pearson chi-square test. The Pearson chi-square should be used only when the categorical data are independent, meaning that each subject is counted once and only once, and each subject's score is independent of each other subject's score. There are two commonly used experimental designs that produce nonindependent categorical data: designs in which the subjects in the different groups have been matched in some way, and designs in which the subjects are measured more than once, for example, a prepost design in which the subject is measured before and after a treatment. Unfortunately, with the latter, a somewhat bewildering number of statistical procedures are available, depending upon the type of categorical data and the number of repeated measures involved. If the data were continuous rather than categorical, one would ordinarily use a repeated measures analysis of variance. If the data were such that the magnitude of change from pretest to posttest could be ranked, the Wilcoxon signedrank test would be appropriate. If only the direction of change from pre to post could be measured, then the sign test or McNemar's test for change could be used. Generalizations of these tests to more complex situations include: Friedman's test, which applies to three or more trials, and for which the subject's data across trials must be ordered or ranked; and Cochran's test, for which the data across repeated measures are either Os or Is, indicating that an event or state either did or did not occur. A third generalization of the sign test or McNemar's test is the Stuart-Maxwell test, the topic of the present paper. The Stuart-Maxwell Test. The Stuart-Maxwell test (Stuart, 1955; Maxwell, 1970) is appropriate when there are either two measures on the same subjects or two matched groups of subjects and the data are categorical. If the data have two categories only, the Stuart-Maxwell test is identical to McNemar's test. When there are three data categories (see, e.g., Table 1), the Stuart-Maxwell test can be solved by a direct algebraic expression (Heiss & Everitt, 1971). When there are more than three data categories, the procedure requires matrix inversion, which is the reason the program described herein was written. The data for the Stuart-Maxwell test are usually written in the form of a contingency table. For example (see Table 1), we ask 78 subjects for which candidate they intend to vote before some important event (such as a con-