The stability of various item discrimination indices when a particular item assumes membership within different sets of items was investigated. Results suggest that when item discrimination is used as a significant factor in determing which items from several forms will be selected for inclusion in a test, the actual items selected are less a function of the chance combinations of items which made up the various forms when Gulliksen's item reliability index and Findley's difference index are used as indices, than when phi coefficients, and particularly point biserial and biserial coefficients are used as estimates of item discrimination. (Author)