Abstract
This paper discusses the problem of variance specification in models for event count data. Event counts are dependent variables that can take on only nonnegative integer values, such as the number of wars or coups d'etat in a year. I discuss several generalizations of the Poisson regression model, presented in King (1988), to allow for substantively interesting stochastic processes that do not fit into the Poisson framework. Individual models that cope with, and help analyze, heterogeneity, contagion. and negative contagion are each shown to lead to specific statistical models for event count data. In addition. I derive a new generalized event count (GEC) model that enables researchers to extract significant amounts of new information from existing data by estimating features of these unobserved substantive processes. Applications of this model to congressional challenges of presidential vetoes and superpower conflict demonstrate the dramatic advantages of this approach.
Highlights
Event counts are dependent variables that take on nonnegative integer values for each of n observations
The number of visible uses of military force initiated by the United States in each six-month interval (Stoll, 1984), the number of presidential vetoes per year (Rohde and Simon, 1985), the frequency of formal and informal military alliances (Russett, 1971; McGowan and Rood, 1975), and the annual number of presidential appointments to the Supreme Court (King, 1987) are examples of time series counts
Examples of cross-sectional event count studies include the number of coups d'etat in each black African state (Johnson, Slater, and McGowan, 1984)and the number of political activities engaged in and reported by Soviet emigres (Di Franceisco and Gitelman, 1984).l
Summary
Event counts are dependent variables that take on nonnegative integer values for each of n observations. The Poisson regression model is used in many disciplines in lieu of these models, but it makes two key assumptions about the way unobserved processes generate event counts that are implausible in many applications Providing estimates of these unobserved processes, instead of assuming them, can lead to important insights about empirical data. If these assumptions do not hold, but the Poisson model is applied anyway, parameter estimates will be inefficient and standard errors inconsistent, a situation analogous to heteroscedasticity in least squares models. This estimator is consistent in the presence of unknown forms and levels of under-, over-, or Poisson dispersion. These models can be considered special cases of the event count models discussed below
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
