Political researchers are often confronted with unordered categorical variables, such as the vote-choice of a particular voter in a multiparty election. In such situations, re-searchers must choose an appropriate empirical model to analyze this data. The two most commonly used models are the multinomial logit (MNL) model and the multinomial probit (MNP) model. MNL is simpler, but also makes the often erroneous independence of irrelevant alternatives (IIA) assumption. MNP is computationally intensive, but does not assume IIA, and for this reason many researchers have assumed that MNP is a better model. Little evidence exists, however, which shows that MNP will provide more accurate results than MNL. In this paper, I conduct computer simulations and show that MNL nearly always provides more accurate results than MNP, even when the IIA assumption is severely violated. The results suggest that researchers in the field should reconsider use of MNP as the most reliable empirical model.