Abstract
A specification issue which has been handled differently in various empirical applications is whether or not to include alternative specific constants in models of choice behavior. Some applications have excluded constants, others have included a full set of constants, and a third class of examples uses unique constants for some alternatives, but not all. In logit models in which each individual has the same set of alterna tives, the exclusion of constants in the estimation of models when the correct model actually has alternative specific effects leads to inconsistent estimates of the coefficients of the remaining independent variables. However, the inclusion of constants when no such effects exist does not affect the consistency of the estimates of the coefficients. These results are illustrated by simple hypothetical examples and by empirical examples. When nonratio scale variables are used in logit models, the coefficients of the independent variables are not invariant under arbitrary scale shifts when alternative specific constants are excluded. Finally, the use of models to predict the response to new alternatives and the transferability of models which might or might not include alternative specific effects is discussed. The major conclusion is that the inclusion of: a full set of alternative specific constants in logit models estimated with large samples is generally preferred over the exclusion of one or more alternative specific constants.
