The nested logit model has become an important tool for the empirical analysis of discrete outcomes. There is some confusion about its specification of the outcome probabilities. Two major variants show up in the literature. This paper compares both and finds that one of them (called random utility maximization nested logit, RUMNL) is preferable in most situations. Since the command nlogit of Stata 7.0 implements the other variant (called non-normalized nested logit, NNNL), an implementation of RUMNL called nlogitrum is introduced. Numerous examples support and illustrate the differences between both specifications.