The extensive use of counts of events data to model foreign policy behavior has universally assumed that the events reflect an underlying continuum of foreign policy orientation. This research strongly breaks with this tradition by assuming, as the original collectors of events data assumed in their justification for the data collection projects that events data represents the basic 'stuff' of politics, and not merely a discrete event manifestation of some hidden process or nebulous 'posture'. Thus modeling foreign policy with events data will prove most fruitful for tests of theory if events are modelled and predicted as events - not as aggregate counts over some fixed temporal domain. In order to do this, a set of baseline measures for goodness-of-fit for events data models will be offered. In addition a means of translating an aggregate data model will be suggested to assess the relative contribution of these two means of analyzing events data.  Once these baseline measures are accepted, theoretical propositions concerning foreign policy behavior may be integrated as a set of rules. This rule-based approach might most properly be labeled an expert system. Should this experimental approach to events data analysis prove an acceptable paradigm for testing foreign policy models, the integration of theoretical statements from such diverse perspectives and levels of analysis as game theory, systemic influences, political economy, national attributes, learning, adaptation and reaction models, psychological models of leadership, and even contextual interpretation from traditional area studies or diplomatic history may be integrated under a single framework. If so, this will produce an exceptionally powerful framework for testing foreign policy models across a wide spectrum of the discipline.

Discrete Event Analysis Table of Contents

• Introduction
• A brief aside on the nature of events data
• Some basic Models of event-interaction processes
• Baseline measures for discrete event data analysis
• A description of the object-oriented model of discrete event prediction
• The basic model algorithm
• Translation of and comparison with an aggregate model
• Expectations for the Model

Over the last two decades hundreds of research articles and books have been published using foreign policy events data as either the behavior to be explained or as an exogenous influence on other behaviors.  After basic demographics, events data are the most commonly used data in foreign policy analysis.  Invariably, this data is used as counts or averages of events aggregated into regular temporal units (e.g., monthly foreign policy conflict/cooperation scores).  In these analyses foreign policy events such as threats, promises, visits of state, wars, and trade agreements all get pooled together to produce a monthly (or annual) picture of the overall status of foreign policy activity between pairs of nations, or dyads. Much of the debate over the proper use of events data has centered around the proper way to collect, aggregate, or scale the various events data counts into a usable form. Indeed, in one of the more prominent discussions of the scaling of events data, Goldstein summarizes both the pattern and the problem in events data research, "These time-series studies must rely on one or another method to convert event-by-event data into aggregate time-series representing a country's behavior toward another over time" (1992: emphasis added). This traditional method of using aggregate international events data is so prevailing as to be not only the norm, but the dominant paradigm for events data research.  This project seeks to break this mold.

The extensive use of aggregate counts of foreign policy events to model foreign policy behavior has universally assumed that the events reflect an underlying continuum of foreign policy orientation. This analysis breaks strongly with this tradition by assuming, as did the original collectors of these events data sets, that events data represents the basic 'stuff' of politics, and not merely a discrete event manifesting from some hidden process or nebulous 'posture'. It is believed that modeling foreign policy with events data will prove most fruitful for tests of theory if events are modeled and predicted as events - not as aggregate counts over some fixed time domain.

A simple, if possibly naive, reason can be offered for the prominence of aggregate measures. As the proverbial cliché goes, "If the only tool you have is a hammer, you tend to see every problem as a nail." The staple tool of the empirical political scientist (and much of the rest of the world) is regression analysis, and regression does not deal well with irregularly spaced time intervals. So, if your data do not conform to your needs, rework them till they do - hence the monthly event count or average. This aggregation of events data into annual or monthly time-series scores is posited on the very reasonable view that this gives a summary measure of the overall foreign policy attitude of the actor. The event count is used therefore to tap the underlying continuous variable that we cannot directly operationalize (King, 1988).  Hence studies of foreign policy cooperation and conflict 'posture' abound. The absence of techniques which will examine raw events data is striking, once one backs away from the standard practice. The early days of events data research saw summary counts used, but the desire to move beyond description ushered in the use of conventional time-series designs due to a complete dearth of alternate methods.

What is proposed breaks completely with the traditional use of events data, and indeed, is best described as experimental. Only limited attempts to examine events data in alternative ways have emerged (Dixon, 1988; Hudson, 1991). These few studies have approached events data analysis as questions of pattern recognition (Schrodt, 1991), Hidden Markov Models (Schrodt, 1997), rule-based expert systems (Job and Johnson, 1991) and as historical time for the context for events( Duffy, 1991). The incorporation of artificial intelligence in this literature is fledgling, but it is gaining in recognition and prominence, especially in the area of machine coding of events data (Gerner, et al. 1993; Schrodt, 1994). Yet no studies have emerged which attempt to define a means of evaluation of events data models on their performance of event prediction. This project seeks to redress this problem.

This set of research pages starts from the premise that in order to test theory with actual events, a set of baseline measures for goodness-of-fit for events data models need to be offered and evaluated. Given appropriate baseline measures upon which to assess model accuracy, theoretical propositions concerning foreign policy behavior may then be integrated as a set of rules. Such rule-based approaches might most properly be labeled computational models or expert systems. Should this experimental approach to events data analysis prove an acceptable paradigm for testing foreign policy models, the integration of theoretical statements from such diverse perspectives as game theory, systemic influences, political economy, national attributes, learning, adaptation and reaction models, psychological models of leadership, and even contextual interpretation from traditional area studies or diplomatic history may be integrated under a single framework. If so, this will produce an exceptionally powerful framework for testing foreign policy models across a wide spectrum of research. Lastly, the development of a somewhat esoteric expert system generally places such analyses beyond the means of most conventional scholars. Thus, an additional aspect of this research will be the development of an interactive object-oriented approach to model development coupled with its availability on the World Wide Web.  It is hoped that a set of generic templates for the introduction of new rules will allow any scholar with access to the Web the ability to formulate and test alternative models.

Objectives

Several features that make this an especially interesting approach to modeling events data.

1. The research design develops a fundamentally different approach to theory development and theory testing in an area central to foreign policy analysis - an approach that allows for the data to be modeled without the loss of information that occurs in aggregation.
2. The method of analysis will be an interactive computer simulation which builds upon an "object-oriented" paradigm to make the modeling tools usable by a broader audience.
3. The analysis will utilize the JAVA language and HTML scripting to provide access to the method and design for scholars of any level of computer skill.  This design will thus promote interactive research via the World Wide Web.

1. The development of a series of benchmarks (such as temporal domain, spontaneous events, number of rules, etc.) for testing the goodness-of-fit for a model on raw events data.
2. The integration of several diverse models of foreign policy into a discrete event model formulation and comparison of model performance to the benchmarks.
3. The conversion of one or more traditional time-series model of foreign policy behavior into a discrete event analog, and comparison of goodness of fit between a direct translation of a time-series model and a discrete event formulation of the same basic model.
4. The articulation of the rule-based system as set of interactive application modules so that the model may be used interactively over the World Wide Web and thus open to scholars without training in programming.

To set the stage for this discussion, we can start by noting that all events data have the following attribute: they are either "spontaneous" (i.e., exogenous to the foreign policy interaction process), or they are "endogenous" reactions, responses, or interactions. They also have a temporal component. That is they can be located in time, and therefore, when event interactions occur, the temporal dimension becomes a measurable and theoretical component. Measurable error in modeling a discrete event process can manifest itself in a number of different manners. Predicted events thus may be the wrong type, or predicted at the wrong time. Events may be predicted which never occur, and models may fail to predict many that do. Models may predict single events only to have an actor use several events as a response.  All of these sources of variation serve to provide empirically measurable benchmarks for testing theory with events data. Lastly, theories should be parsimonious to be considered more general. Hence, this project will establish a number of useful benchmarks for the comparison and testing of theories.

All of these benchmarks will be initially be established by using the COPDAB (Conflict and Peace Data Bank) events data set. The COPDAB data set is appealing given its scaled nature. This will facilitate model development in the early stages.

In order to set the stage for discussing a discrete events data model, a brief aside on the nature of events data will prove useful.
 
Back to Main Events Data ResearchWeb Table of Contents
Back to Discrete Event Analysis Table of Contents