Behavioral multistate duration models : what should they look like?

This paper discusses how specification of probabilistic models for multistate duration data generated by individual choices should be justified on a priori theoretical grounds. Preferences are assumed represented by random utilities, where utilities are viewed as random also to the agent himself. First, the paper proposes a characterization of exogenous preferences, (that is, in the special case with no state dependence effects). The main assumption asserts that when preferences are exogenous the current and future indirect utilities are uncorrelated with current and past choices, given unobservables that are perfectly known to the agent. It is demonstrated that under rather weak and general regularity conditions this characterization yields an explicit structure of the utility function as a so-called Extremal stochastic process. Furthermore, from this utility representation it follows that the choice process is a Markov Chain (in continuous- or discrete time), with a particular functional form of the transition probabilities, as explicit functions of the parameters of the utility function and choice set. Subsequently, we show how the model can be extended to allow for structural state dependence effects, and how such state dependence effects can be identified. Moreover, it is discussed how a version of Chamberlain's conditional estimation method applies in the presence of fixed effects. Finally, we discuss two examples of applications. -- Duration models ; Random utility models ; Habit persistence ; True state dependence ; Extremal process ; Markov chain