If one wishes to assess the degree of association be- tween the variables in an ordered two-way contingency table, there are various measures available. These meas- ures have several advantages and disadvantages in terms of interpretation, ease of computation, and appropriate- ness for the data under discussion. One attractive meas- ure of association is Goodman and Kruskal's (1954, 1959, 1963, 1972) gamma, defined as (C - D)I(C + D), where C and D are the numbers of concordant and discordant pairs, respectively, in the table. We will denote this meas- ure by G; it is an estimate of a population parameter y. For hypothesis testing and other inferential purposes it is desirable to know as much as possible about the distributional properties of G. Goodman and Kruskal (1959) showed that G has an asymptotically normal dis- tribution, with a calculable asymptotic variance. Rosen- thal (1966) did a Monte Carlo study of the distribution for small samples. Since the exact formula for the asymptotic variance is awkward to handle, Goodman and Kruskal gave an upper bound, which is much easier to calculate but which seems in practice to be a substantial overes- timate of the exact formula. However, computer pro- grams exist for finding variance estimates and related significance tests based on the exact formula for asymptotic variance (see Berry, Mielke, and Jacobsen 1977). Brown and Benedetti (1977) also worked on modifications of the variance formula. Not enough has been known about the distribution of G for small and moderate sample sizes; and in view of the uses being made of G in applied research and its appearance in elementary texts, further examination of these distributional questions is appropriate