In a previous article we showed that ordinary least squares with panel corrected standard errors is superior to the Parks generalized least squares approach to the estimation of time-series-cross-section models. In this article we compare our proposed method with another leading technique, Kmenta's “cross-sectionally heteroskedastic and timewise autocorrelated” model. This estimator uses generalized least squares to correct for both panel heteroskedasticity and temporally correlated errors. We argue that it is best to model dynamics via a lagged dependent variable rather than via serially correlated errors. The lagged dependent variable approach makes it easier for researchers to examine dynamics and allows for natural generalizations in a manner that the serially correlated errors approach does not. We also show that the generalized least squares correction for panel heteroskedasticity is, in general, no improvement over ordinary least squares and is, in the presence of parameter heterogeneity, inferior to it. In the conclusion we present a unified method for analyzing time-series-cross-section data.

In contrast to conventional studies on campaign finance, which focus on the aggregate effect of money on the vote, we propose a more general dynamic model based on temporally disaggregated data. The model is supported by the substantive understanding that at different stages of the campaign process candidates have different goals, and their expenditures should have different effects on the final election outcome. Using Achen's (1986) framework of quasi experiments, the model includes dynamic “assignment equations” and “outcome equations”, which address the problem of nonrandom assignment. A final vote equation is derived in which the coefficients of period-specific incumbent expenditures are constrained by an Almon polynomial. Empirical estimation provides evidence for a three-stage dynamic campaign process.

This paper uses the ideodynamic model to assess the impact of persuasive messages on candidate preference in the 1992 presidential election. The methodology is based on approaches common to the natural sciences. These approaches suggested that opinion predictions are most robust when they explicitly omit such factors as opinion poli values, incumbency, and economic performance. Instead, the only persuasive messages used to predict opinion came from 3,394 stories from the Major Paper library of the NEXIS electronic data base scored by computer for news favorable and unfavorable to George Bush, Bill Clinton, and Ross Perot. The scores were used in the equations of ideodynamics to compute media share time trends that predicted 55 public opinion polls with R2 of 0.84 and 0.79 for Clinton and Perot, respectively. These high R2 values together with the significance of the model's parameters suggested that the news media comprised the dominant influence on candidate preference. Campaign stories were also scored for coverage of different issues, and this coverage was related to good and bad news for the candidates.

Vote has been found to be among the most persistent political attitudes and behavior patterns. We specify three sources of vote persistence: persistence in (observable) factors structuring the vote, habituation through repeated vote for the same party, and stable (unobservable) personal preferences. The paper proposes an estimation procedure for distinguishing among these different factors based on one-shot election surveys, that is, without need for panel data, which are not often available. Applying the procedure to Israeli 1992 election data, the results suggest that a priori fixed preferences are relevant but have little impact on vote persistence. One's past voting record has much more impact, and continuity in political attitudes and evaluations has the most effect. We tested the suggested procedure on the American national presidential elections 1972–76 panel data, which provided further validation for our suggested method.

In recent years, more and more social scientists have begun to view the world as inherently probabilistic (Suppes 1984; Gigerenzer 1987). Without detailing the philosophical underpinnings of such a view, this subtle movement away from deterministic positivism has been fed, in large part, by a recognition of the indeterminacy of strategic interaction among individuals and the inevitability of uncertainty in social relations (Boudon 1986). Chaos theory provides an alternative viewpoint from which to view indeterminacy because the complexity we see in the real world may, in theory, be a reflection of chaotic dynamics resulting from simple deterministic structures (Huckfeldt 1990). Particularly when processes are generated by social or strategic interaction among actors, nonlinear models provide useful representations and chaotic outcomes become conceivable. In short, the existence of a complex social reality is in itself inadequate evidence of indeterminacy.

The Generalized Event Count (GEC) estimator (King 1989a) is a statistical model for event counts. Its great attraction is that it provides a general likelihood function for count data, regardless of whether the data come from a Poisson, binomial, or negative binomial distribution. In consequence, it has been used in several recent statistical studies of event counts in the social sciences.

Underlying the GEC, however, are unorthodox substantive assumptions about how the event counts have been generated (Amato, this volume). This paper gives some simple examples in which the GEC logic is clearly visible, and it shows how failures of the implicit assumptions can lead to erroneous GEC coefficient estimates and standard errors.

In this article, the mathematical and probabilistic foundations of Gary King's “generalized event count” (GEC) model for dealing with unequally dispersed event count data are explored. It is shown that the GEC model is a probability model that joins together the binomial, negative binomial, and Poisson distributions. Some aspects of the GEC's reparameterization are described and extended and it is shown how different reparameterizations lead to different interpretations of the dispersion parameter. The common mathematical and statistical structure of “unequally dispersed” event count models as models that require estimation of the “number of trials” parameter along with the “probability” component is derived. Some questions pertaining to estimation of this class of models are raised for future discussion.

King (1989) presented the “Generalized Event Count” (GEC) model as a means of dealing with event count data when the analyst is unsure whether the data are “underdispersed” or “overdispersed.” Here I establish several useful properties of the GEC model and make some practical suggestions for estimation.

We use an analogy with the normal distribution and linear regression to demonstrate the need for the Generalized Event Count (GEC) model. We then show how the GEC provides a unified framework within which to understand a diversity of distributions used to model event counts, and how to express the model in one simple equation. Finally, we address the points made by Christopher Achen, Timothy Amato, and John Londregan. Amato's and Londregan's arguments are consistent with ours and provide additional interesting information and explanations. Unfortunately, the foundation on which Achen built his paper turns out to be incorrect, rendering all his novel claims about the GEC false (or in some cases irrelevant).