ABSTRACT
Gibbons and Chakraborti (1992, p. 135) gave seven possible treatments of ties for rank tests in general. A quick review of nonparametric texts indicates that the most commonly recommended procedure seems to be ignoring the tied observations; see for example Lehmann (1975, p. 123 and p. 144), Siegel and Castellan (1988) and Daniel (1990). Ignoring ties is an approach used in StatXact (1995), arguably a most comprehensive nonparametric statistics package. Consider a simple example: suppose that two chess players play nine matches, of which five are won by player A, none by B, and there are four draws. In testing if both players are equally matched, the draws are treated as ties, and clearly have positive probability. Moreover they are meaningful in comparing the players.
