When estimating a Markov-process model from observed micro data, it may be necessary to assume a closure state for the process#x2014;a state that can account for entities that may come into or go out of existence.When collecting data to estimate such a model, we may not be able to #x201C;see#x201D; the closure state explicitly; we must posit some value for the number of entities in it.In addition to developing maximum-likelihood estimators of the closure-state size, this paper examines the effect on point-estimate and hypothesis-test performance, as well as on prediction accuracy, of alternative methods of specifying the size of the closure state.Analytical and empirical evidence suggests that results of interest to the modeler are generally not sensitive to closure-state specification.In an advertising-media application, which truns out to support our insensitivity conclusion, the closure state holds firms that are not currently among the top advertisers.A firm entering the ranks of the top advertisers is modeled as a transition out of the closure state and into a real state for a particular electronic-media share range.A firm that is currently one of the leading advertisers but then disappears from these ranks for some reason would represent the opposite type of transition.
展开▼
