Chapter One

International Event Data

International conduct, expressed in terms of event data, is the chief dependent variable
of international relations research. ... [This] starting point is provided as readily by
the ordering principle of classical diplomatic history as by the basic concepts of
general system analysis. Thus, we may assert that the prime intellectual task in the
study of international relations is to account for actions and responses of states in
international politics by relating these to the purposes of statecraft or, alternatively,
we can say that the problem is to account for the relations among components of the
international system by analyzing the characteristics of the various components of that
system by tracing recurring processes within these components. [Both definitions]
carry about the same information and involve nearly the same range of choices of
inquiry and analysis.
Charles McClelland 1970:6
Much of political behavior can be characterized by interactions between political actors,
whether individuals or institutions. Politics primarily consists of who did what to whom, and
most of our information about politics comes in the form of narratives about actions, reactions,
and activities. Patterns of political interactions are commonly both used to characterize existing
situationsfor example in distinguishing peaceful from conflictual periods in the relationship of
two statesand are also used as indicators of possible future situations, as when relations
between two states are characterized as deteriorating or improving.
The importance of interactions in political behavior has been recognized since the beginnings
of the efforts to systematically study international political behavior using scientific
approaches. This research has primarily taken the form of the analysis of event datanominal or
ordinal codes recording the interactions among international actors as reported in the open
presswhich break down complex political activities into a sequence of basic building blocks
(e.g., comments, visits, rewards, protests, demands, threats, and military engagements). Event
data sets were a major focus of quantitative international relations research in the 1960s and
Schrodt and Gerner
Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

Page 1-2

1970swork that generated Edward Azars (1982) Conflict and Peace Data Bank (COPDAB)
and Charles McClellands (1976) World Event Interaction Survey (WEIS). Over the past decade,
interest in event data analysis has increased as the combination of machine-readable news reports
and automated coding have dramatically reduced the costs of generating, customizing, and
analyzing event data.
This volume is directed at researchers who are interested in employing event data to study
international behavior, though most of the information contained here is equally relevant to the
study of domestic behavior. We will cover two general topics. First, we provide a general
introduction and survey of the potentials and problems of event data analysis as it exists at the
beginning of the 21st century, with a particular emphasis on methods for automated coding of
events from electronic sources (data sets developed prior to the 1990s, in contrast, were coded
by humans from paper or microfilm sources).
Second, we will discuss a variety of techniques specifically designed for the analysis of event
data. These include an assortment of clustering algorithms, the Levenschtein metric, parallel
event sequences, and hidden Markov models. These methods, while generally well-developed in
fields distant from political science (for example speech recognition and the analysis of protein
coding sequences), are relatively inaccessible to the typical political analyst whose primary
training has been in econometric techniques. We believe that these methods can be a useful
supplement to conventional time-series analysis techniques for characterizing and forecasting
event sequences.
The bulk of this book deals with our research over the past ten years with the Kansas Events
Data System project (KEDS), or with closely related work such as that of the Protocol for the
Analysis of Nonviolent Direct Action (PANDA) project at Harvard. While much of this work
has been reported in journal-length articles in various political science publications (Schrodt 1991,
1993, 1999; Gerner et al 1994; Schrodt and Gerner 1994, 1996, 1997; Schrodt, Davis and Weddle
1994), this volume provides the opportunity to put all of the material in one place, and to update
some of our earlier observations in light of subsequent experience. For example, while we
Schrodt and Gerner
Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

Page 1-3

initially saw machine coding as an inexpensive though qualitatively inferior alternative to human
coding, but ten years of experience has convinced us that the consistency and stability of machine
coding make it superior to human-coded data.
The methods we have developed are, necessarily, illustrative rather than definitive. For
instance, virtually all of our work has focused on international behavior in the Middle East, a
region that is not necessarily typical of international politics in general, but which we've studied
and visited for many years. Nonetheless, the methods that we have used in this research, as well
as many of the problems that we have encountered, are likely to apply to many other studies.

1.1. What are event data?

In their simplest form, event data convert natural language reports of political activity to a
data set where each entry has the form
date

source

target

event

The "date" is a calendar date; typically this is accurate to within a day or two of when the event
occurred. The "source" and "target" are political actors such as nation-states, national leaders,
political parties, non-governmental organizations or guerrilla movements. The "event" is a
specific code for the type of activity that was reported. Event coding systems may have
anywhere from fewer than 20 to more than 100 distinct event categories.
Event data is created through a process of content analysis (see Krippendorf 1980; Weber
1990) and involves four steps; these are illustrated schematically in Figure 1.1. In the first step,
the researcher identifies a source of news about political interactions. This could be a newswire
service such as Reuters or Agence France Presse (AFP), an internationally-oriented newspaper
such as The New York Times or Times of London, a set of regional newspapers and news
magazines, or a news summary such as Facts on File or Deadline Data on World Affairs. As we
will discuss below, the choice of the event source can have a substantial effect on the number and
types of events reported.

Schrodt and Gerner

Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

Page 1-4

Human coding
Train
coders
Determine
news
source

Develop
or adopt
coding
scheme

Code
stories

Actors
Events
Other

Develop
dictionaries
Machine coding

Figure 1.1. Creating an event data set

Second, a coding system is developed, or a researcher may decide to adopt an existing coding
system such as WEIS or COPDAB. Table 1.1 shows a sample of the lead sentences of reports
on the Reuters newswire that preceded Iraq's invasion of Kuwait in August 1990. (The full set
of reports is considerably more extensive, particularly during the week prior to the invasion.) In
most cases, each lead corresponds to a single event, although some sentences generate multiple
events. For example, the report "July 23, 1990: Iraqi newspapers denounced Kuwait's foreign
minister as a U.S. agent Monday" corresponds to an event in the WEIS event coding scheme: the
WEIS category 122 is defined as "Denounce; denigrate; abuse." In this event, Iraq is the source
of the action and Kuwait is the target. Together, this information generates the event record
"900723 IRQ KUW 122" where "900723" is the date of the event, IRQ is a standard code for
Iraq, KUW is the code for Kuwait, and 122 is the WEIS category. Table 1.2 shows the Reuters
stories converted to WEIS events.

Schrodt and Gerner

Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

Page 1-5

Table 1.1. Reuters Chronology of 1990 Iraq-Kuwait Crisis

July 17, 1990: RESURGENT IRAQ SENDS SHOCK WAVES THROUGH GULF ARAB STATES
Iraq President Saddam Hussein launched an attack on Kuwait and the United Arab Emirates (UAE)
Tuesday, charging they had conspired with the United States to depress world oil prices through
overproduction.
July 23, 1990: IRAQ STEPS UP GULF CRISIS WITH ATTACK ON KUWAITI MINISTER
Iraqi newspapers denounced Kuwait's foreign minister as a U.S. agent Monday, pouring oil on the flames
of a Persian Gulf crisis Arab leaders are struggling to stifle with a flurry of diplomacy.
July 24, 1990: IRAQ WANTS GULF ARAB AID DONORS TO WRITE OFF WAR CREDITS
Debt-burdened Iraq's conflict with Kuwait is partly aimed at persuading Gulf Arab creditors to write off
billions of dollars lent during the war with Iran, Gulf-based bankers and diplomats said.
July 24, 1990: IRAQ, TROOPS MASSED IN GULF, DEMANDS $25 OPEC OIL PRICE
Iraq's oil minister hit the OPEC cartel Tuesday with a demand that it must choke supplies until petroleum
prices soar to $25 a barrel.
July 25, 1990: IRAQ TELLS EGYPT IT WILL NOT ATTACK KUWAIT
Iraq has given Egypt assurances that it would not attack Kuwait in their current dispute over oil and
territory, Arab diplomats said Wednesday.
July 27, 1990: IRAQ WARNS IT WON'T BACK DOWN IN TALKS WITH KUWAIT
Iraq made clear Friday it would take an uncompromising stand at conciliation talks with Kuwait, saying its
Persian Gulf neighbor must respond to Baghdad's "legitimate rights" and repair the economic damage it
caused.
July 31, 1990: IRAQ INCREASES TROOP LEVELS ON KUWAIT BORDER
Iraq has concentrated nearly 100,000 troops close to the Kuwaiti border, more than triple the number
reported a week ago, the Washington Post said in its Tuesday editions.
August 1, 1990: CRISIS TALKS IN JEDDAH BETWEEN IRAQ AND KUWAIT COLLAPSE
Talks on defusing an explosive crisis in the Gulf collapsed Wednesday when Kuwait refused to give in to
Iraqi demands for money and territory, a Kuwaiti official said.
August 2, 1990: IRAQ INVADES KUWAIT, OIL PRICES SOAR AS WAR HITS PERSIAN GULF
Iraq invaded Kuwait, ousted its leaders and set up a pro-Baghdad government Thursday in a lightning predawn strike that sent oil prices soaring and world leaders scrambling to douse the flames of war in the
strategic Persian Gulf.
Source: Reuters

Schrodt and Gerner

Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

Page 1-6

Table 1.2. WEIS Coding of 1990 Iraq-Kuwait Crisis

Date

Source

Target

WEIS Code

Type of Action

900717

IRQ

KUW

121

CHARGE

900717

IRQ

UAE

121

CHARGE

900723

IRQ

KUW

122

DENOUNCE

900724

IRQ

ARB

150

DEMAND

900724

IRQ

OPC

150

DEMAND

900725

IRQ

EGY

054

ASSURE

900727

IRQ

KUW

160

WARN

900731

IRQ

KUW

182

MOBILIZATION

900801

KUW

IRQ112

REFUSE

900802

IRQ

KUW

223

MILITARY FORCE

The coding system must specify the set of political interactions that constitute an "event,"
identify the political actors that will be coded (for example, whether nonstate actors such as
international organizations and guerrilla movements will be included in the data set) and establish
the categories of events and their codes. Some systems also specify additional information to be
coded about the event: For instance, the COPDAB data set codes whether an event is primarily
military, economic, diplomatic, or one of five other types of relationship. WEIS codes for
specific "issue arenas" such as the Vietnam War, Arab-Israeli conflict, and SALT negotiations.
The complete list of event categories for four systemsWEIS, COPDAB, BCOW and
IDEAare found in Appendix 1.
The coding systems in COPDAB and WEIS are comprehensive: they attempt to code all
political interactions by all states and some non-state actors during a period of time. In contrast,
specialized event data sets such as Hermann and Hermann's (1973) CREON (Comparative
Research on the Events of Nations) and Leng's (1987) BCOW (Behavioral Correlates of War)
focus on specific subsets of behavior, foreign policy and crises respectively. A variety of
domestic and international event data collections, usually focusing on a limited set of actions such
as uses of force, domestic violence, or changes of government, are embedded in other data sets
Schrodt and Gerner
Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

Page 1-7

such as Rummel's (1972) DON (Dimensionality of Nations), the World Handbook (Taylor and
Hudson 1972), and various data sets collected by the research projects of Ted Gurr (for example
Gurr 1974). (Peterson (1975) and McGowan et al. (1988) provide a general discussion of these
data sets.)
News stories can be coded either by human coders, or with specialized software; the relative
merits of these two approaches will be discussed in chapter 2. In a project using human coders,
these coding rules are collected into a formal coding manual. In an extended project, these
manuals are often fifty or more pages in length and deal with a variety of contingencies that
coders may encounter. Coderstypically graduate students or advanced undergraduates in
political scienceneed to be trained so that a news story will be assigned the same codes by any
individual coding it. The training stage is frequently quite time consuming but, with sufficient
training, most projects train coders to the point where two coders will assign the same code to a
news report in 80% to 90% of the cases (see Burgess and Lawton 1972: 58). Human coding is a
relatively slow process and most human-coding projects have been maintained over a number of
years, with intermittent periods of re-training as coders enter and leave the project.
In contrast, in a machine-coding project, coding rules are implemented through a computer
program that employs customized dictionaries of phrases that identify actors and events and
associate these with specific codes (see Lehnert and Sundheim 1991; Schrodt, Davis, and Weddle
1994). These dictionaries are typically constructed by having a dictionary developer monitor the
codes assigned by the automated system as it goes through a large number of test sentences.
Appropriate new vocabulary is added or modified by the developer whenever the program makes
a coding error. When the level of coding accuracy reaches a level sufficient for the intended
application, the dictionaries are fixed and the entire data set (including the test cases) is re-coded,
thus insuring that the same accuracy level is found in the entire data set. The fully-automated
coding could take from a few minutes to several hours of computation, depending on the number
of stories coded.

Schrodt and Gerner

Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

Page 1-8

A completed event data set will cover a period of years and typically contain tens of
thousands of events. These are then used in statistical or computational analyses. Most
statistical studies do not work with the individual events, but instead work with weekly,
monthly or yearly aggregations of the events. An example of this type of analysis is found in
Chapter 4. This aggregation is done by assigning a numerical value to each type of event using a
standardized scale, and then summing those values. Computational analyses, in contrast, work
directly on the event sequences without the intermediate step of numerical aggregation; some of
these methods will be discussed in Chapters 5 and 6.
Figure 1.2 shows the actions that the United States directed towards the Soviet Union for the
period 1948-78 based on the COPDAB scores reported in Goldstein and Freeman (1990:162).
(In the COPDAB coding system negative numbers indicate conflictual behavior while positive
numbers indicate cooperation.) The events are coded from The New York Times and a variety of
regional newspaper sources. The COPDAB time series shows three general periods. The early
Cold War (1948-62) is characterized by uniformly negative relations, although relations are more
stable in the late 1950s than in the early 1950s. A partial thaw occurs in 1962-70 following the
Cuban Missile Crisis, with the relationship centered at zero, hence neutral. Finally, the 1970-78
period shows the rise and fall of the Nixon-Kissinger dtente policy.
The event data record of U.S.-Soviet interactions corresponds closely to the patterns one
would expect from an historical study. Moreover, the event data can also be used to fine-tune
that chronology. For example, while Nixon clearly intended to implement a dtente policy from
the beginning of his administration in 1969, there was continued disagreement between the U.S.A.
and U.S.S.R. over the U.S. involvement in Vietnam, the 1968 Soviet invasion of Czechoslovakia,
and other issues. Thus the aggregated interactions between the U.S.A. and U.S.S.R. do not
become positive until 1971. Positive interactions peak about the time of Nixon's resignation in
1974; the event data scores then decline during the two years of the Ford administration and
return to post-Cuban Missile Crisis levels by 1976.

Schrodt and Gerner

Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

Page 1-9

100
50
0
-50
-100
-150
-200

1978

1976

1974

1972

1970

1968

1966

1964

1962

1960

1958

1956

1954

1952

1950

1948

-250

Figure 1.2. Azar-Sloan-scaled series for USA USSR, 1948-78

IRN > IRQ

50
0
-50
-100
-150
-200
-250
-300
-350

Figure 1.3. Goldstein-scaled series for Iran Iraq, 1979-97

Schrodt and Gerner
Analyzing International Event Data

DRAFT: February 6, 2000

04-97

04-96

04-95

04-94

04-93

04-92

04-91

04-90

04-89

04-88

04-87

04-86

04-85

04-84

04-83

04-82

04-81

04-80

04-79

-400

International Event Data

Page 1-10

Figure 1.3 shows the events initiated by Iran and directed to Iraq, coded with the WEIS
system and aggregated using Goldsteins (1992) scale.1. The major offensives of the Iran-Iraq
war are clearly visible, as are the negotiations and overtures that Iraq made to Iran following Iraq's
invasion of Kuwait. The 1990s, as expected, are characterized by sporadic disagreements but
only limited activity compared to the 1980s.
In contrast to the largely conflictual behavior in Figure 1.3, the graph of the Saudi Arabia to
U.S.A. relationship in Figure 1.4 shows a sequence of ebbs and flows in the relations between the
two states (also note that the vertical scale is about one-tenth that of Figure 1.3). An examination
of the individual events shows that the positive values primarily correspond to meetings,
agreements and statements of policy support. Negative values correspond to diplomatic
disagreements between the two governments, for example on the controversy over the sale of
AWACS aircraft in the early 1980s, and over U.S. policy towards Iran and Israel later in that
decade. The very large negative spike in 1996 is due to the Khobar Towers bombing in 1996.
This data set did not differentiate sub-state actors, and therefore the bombing was coded as a
Saudi attack on the U.S.; additional negative events were generated by the diplomatic
disagreements between the two governments over the investigation of that incident. The only
large anomalous negative value occurs in January and February 1991, where there are a number of
incorrectly-coded uses of force. These come from reports about the Second Gulf War where the
machine coding system failed to accurately determine the appropriate actors.2

1
2

The data in Figures 1.3 and 1.4 were coded by the authors from Reuters lead sentences using KEDS. The Iraq to
Iran sequence is similar: the two correlate with r = 0.84
As we will discuss later, the KEDS coding dictionaries, which were developed to code general international
behavior, were prone to errors when coding stories dealing with military activities in the Second Gulf War.
Descriptions of these events involved vocabulary and sentence constructions not encountered in other contexts,
and sentences were sometimes misinterpreted.

Schrodt and Gerner

Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

Page 1-11

SAU > USA

20
10
0
-10
-20
-30
-40

04-97

04-96

04-95

04-94

04-93

04-92

04-91

04-90

04-89

04-88

04-87

04-86

04-85

04-84

04-83

04-82

04-81

04-80

04-79

-50

Figure 1.4. Goldstein series for Saudi Arabia U.S.A., 1979-97

As these figures illustrate, event data can be used to summarize the overall relationship
between two countries over time. The patterns shown by event data generally correspond to the
narrative summaries of the interactions found in historical sources. But unlike narrative accounts,
event data can be subjected to statistical analysis and other computerized analytical techniques.
Event data analysis relies on a large number of events to produce meaningful patterns of
interaction. The information provided by any single event is very limited, and single events are
sometimes affected by erroneous reports and coding errors. However, important events trigger
other interactions throughout the system. For example while Iraq's invasion of Kuwait by itself
generates only a single event with WEIS code 223 (military force) the invasion triggers an
avalanche of additional activities throughout the international system as states and international
organizations denounce, approve or comment on the invasion, so the crisis is very prominent in
the event record.

Schrodt and Gerner

Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

Page 1-12

1.2. The Development of Event Data Analysis in

International Relations Research
The historical development of event data concepts and collections dates back to nearly the
beginning of "scientific" studies of international politics. The early theoretical development of
WEIS is thoroughly discussed in a series of papers by McClelland (1967a, 1967b, 1968a, 1968b,
1969,1970)3; the foundations of the COPDAB project are fairly well documented in a series of
papers coming out of Azar's Michigan State University event data conferences during 1969-71
(Azar, Brody and McClelland 1972; Azar and Ben-Dak 1975; Azar et al. 1972; Azar and Sloan
1975).
Most of these early efforts were motivated by attempts to develop statistical early warning
indicators of international and domestic instability. The Department of State experimented with
coding event data for a small set of states in 1971 in its Foreign Relations Indicator Project (see
Lanphier 1975). The Pentagon's Defense Advanced Research Project Agency (DARPA)
sponsored a large-scale project in the 1970s to develop event data models for crisis forecasting
and management. In the early years of the Reagan administration, the National Security Council
staff in the White House undertook a major event data collection and analysis effort. MITs
case-oriented CASCON project was also receiving considerable attention in the policy
community during this period, including the U.S. Department of Defense, State Department and
Arms Control and Disarmament Agency, and the United Nations (Bloomfield and Moulton 1998:
chapter 8).
These efforts apparently had little long-term impact on the formulation of foreign policy,
although the data produced continued to be used in academic research. Laurance (1990) analyzes
the reasons for the limited impact of event data on policy, including the failure to coordinate the
event data projects with the analysts and policy-makers who were supposed to use the data, the
absence of user-friendly analytical tools, and the absence of guidelines on how event data could

We are indebted to Harold Guetzkow for an extensive collection of the early WEIS memoranda.

Schrodt and Gerner

Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

Page 1-13

be used with traditional, non-statistical sources of information. Because of these problems, new
global event data collection efforts ceased in the 1980s, although the COPDAB and WEIS data
continued to be refined, and data sets such as CREON were used in academic research. A small
number of new data sets focusing on international crisesnotably Leng's BCOW (Leng 1987)
and Sherman's SHERFACS (Sherman and Neack 1993)were developed during this time.
Large-scale event data efforts were revived in the early 1990s in the second phase of the
National Science Foundation's "Data Development in International Relations" project (DDIR),
directed by Dina Zinnes and Richard Merritt (Merritt, Muncaster, and Zinnes 1993). Rather
than simply extending the work of the 1970s, DDIR emphasized the development of new
approaches, with particular emphasis on exploiting the computing power available in personal
computers and using machine-readable news sources. The Global Event Data System (GEDS) at
the University of Maryland grew out of this project, as did the early work on KEDS and several
other experimental projects. The DDIR project marked a transition between the DARPA-style
event data research and contemporary approaches, and the various articles in Merritt, Zinnes and
Muncaster (1993) show a mix of old and new techniques. These methods slowly diffused into
the policy community, and by the end of the decade, event data were employed in the
development of experimental early warning systems at the U.S. Department of Defense, in the
dynamic modeling phase of the State Failures Project, and in Switzerland by the FAST project
(Krummenacher and Schmeidl 2000).4
By the late 1990s, three general changes were apparent in event data analysis. First, there
was a move away from newspaper sources to wire-service sources, particularly Reuters, and

We are currently aware of event data projects using the VRA automated codera commercial spin-off from the
PANDA projectat the Joint Warfare Analysis Center (JWAC) and CINCPAC at the U.S. Department of
Defense, UNICEF in the United Nations, and FAST in Switzerland. KEDS-based data are being used at JWAC
and a U.S. government project on forecasting genocide/politicide; GEDS-based data on accelerators was used
in the U.S. State Failures II project (Esty et al 1998). All of these projects have focused on the forecasting of
state breakdowns and humanitarian crises. It is quite possible that additional projects exist that we would not
know about, particularly if they used software from the MUC experiments (Lehnert and Sundheim 1991) rather
than software derived from DDIR projects.

Schrodt and Gerner

Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

Page 1-14

projects moved from relying on paper to using electronic databases, notably NEXIS.5 The
electronic systems were more convenient than the paper sources; for example the NEXIS system
provided an elaborate Boolean search facility that allowed a researcher to narrowly focus a
request for information. Wire service reports were more comprehensive than reports from
newspapers, particularly as newspapers such as the New York Times began to place greater
emphasis on soft news such as lifestyle and popular culture, at the expense of their international
coverage.
Second, machine-coding replaced human coding for almost all applications. There were some
initial experiments in the early 1990s to determine whether machine-assisted coding
systemswhich employed user-friendly software to simplify the clerical aspects of event
codingcould produce dramatic improvements in efficiency compared to the paper-and-pen
systems of the 1970s. This did not prove to be the case, and the computer-assisted human
coding projects experienced the same inefficiencies caused by coder turnover, boredom, and
training costs that had plagued earlier efforts. In the meantime, dramatic reductions in the cost
and increases in the capacity of personal computers gave an overwhelming economic advantage to
the machine coding projects. As the decade progressed, other advantages to machine coding
became apparent for example its transparency, stability and flexibilityand consequently
machine coding became the preferred approach except in a few specialized applications.
Third, the focus of most event data collections shifted from the global approach to regional
approaches. New data sets tended to focus on a specific geographical area, and decisions were
made to code specific sub-state actors based on the political circumstances. The end of the Cold
War also shifted the focus of most event data efforts from superpower conflict (Ashley 1980;
Goldstein and Freeman 1990; Ward and Rajmaira 1992) to regional and even substate conflicts
(Huxtable and Pevehouse 1996; Kinsella 1995, 1998; Reuveny and Kang 1996a, 1996b; Bond et
al. 1997; Goldstein and Pevehouse 1997; Schrodt and Gerner 1994, 1997; Schrodt 1999, 2000).

Prior to 10 June 1997, the NEXIS service, which was available to academic institutions at a relatively inexpensive
rate, contained daily updates from various Reuters services.

Schrodt and Gerner

Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

Page 1-15

While the COPDAB coding system was maintained and expanded by the GEDS project
(http://www.bsos.umd.edu/cidcm/geds/), WEIS continues to be the most widely employed
system, albeit it has often been extended to provide greater detail in the coding of domestic
conflict events. The most notable of these WEIS extensions was done by the Protocol for the
Analysis of Nonviolent Direct Action (PANDA) project at Harvard in the mid-1990s
(http://data.fas.harvard.edu/cfia/pnscs/panda.htm), which produced a global, Reutersbased event data set covering 1984 through early 1995. To accommodate domestic events,
PANDA more than doubled the number of WEIS categories, while providing a systematic table
for translating PANDA codes to WEIS codes. More recently, this effort has been extended to
the IDEA coding system (Taylor, Jenkins, and Bond 1999), which is designed to be used in the
next edition of the World Handbook.

1.3. Event Data Sets

Event data sets fall into two general categories. Actor-oriented data sets record all interactions
among a set of actors for a specific period of time, for example the Middle East 1979-99.
Episode-oriented sets look only at the events involved in a specific historical incident, usually an
international crisis or use of force.
The objective of most academic event data research is to find theoretically-informed statistical
regularities, so event coding systems are closely linked to a theory or set of theories about
international behavior. The detailed reports of the event data collection efforts (for example Azar,
Brody and McClelland 1972; Azar and Ben-Dak 1975; Burgess and Lawton 1972; Hermann et al
1973; Merritt, Muncaster and Zinnes, 1993)although not necessarily the codebooks for those
datashow a deep awareness of the linkage between theory, coding and data collection.
The WEIS and COPDAB schemes, for example, were constructed in the milieu of the
international relations theory of "realism" that placed primary emphasis on diplomatic and
military behavior. In contrast, the CREON data set focuses on the elements of foreign policy
behavior identified by the theories developed in James Rosenau's Inter-University Comparative
Schrodt and Gerner
Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

Page 1-16

Foreign Policy Project (see Hermann et al 1973:8-15). While both of these data sets provide
good indicators of conflict behavior, neither are particularly useful in studying contemporary
international economic or environmental issues. Similarly, when the PANDA project sought to
study non-violent direct action in domestic conflict, they found that they needed to make
substantial extensions to the WEIS coding scheme.
1.3.1. Actor-Oriented Data Sets
WEIS
The WEIS coding scheme classifies events into 63 specific categories; these are organized into
22 general categories such as "Consult", "Reward", "Protest" and "Force" (see Table 1.3). The
general categories form a very rough cooperation-conflict continuum. WEIS coding was the de
facto standard used by the U.S. government-sponsored projects during the 1970s, and
consequently a number of the data sets available in the ICPSR archive use the WEIS scheme.
The WEIS data set available at the ICPSR covers only eleven years (1966-77) and contains
only about 90,000 events; the source text is The New York Times. Data after 1977 have
continued to be coded by McClelland and several of his studentsmost recently Rodney
Tomlinson at the US Naval Academy (Tomlinson 1993)but the full series is not available in
the public domain at the present time.
Because most common statistical routines, such as regression analysis, use numerical rather
than categorical data, WEIS events are often aggregated into numerical scores before being
analyzed. Vincent (1979) and Goldstein (1992) provide two such scales that assign numbers on a
cooperation-conflict continuum to each WEIS category; Table 1.3 shows examples of Goldstein
scores for several WEIS categories. As we will demonstrate in Chapter 3, widely-varying scales
can produce similar analytical results in many problems, because much of the variance in event
data is due simply to the presence or absence of an event. WEIS codes can also be translated into
the COPDAB scale, although one cannot translate from COPDAB to WEIS because COPDAB
makes fewer distinctions in the type of event.

Schrodt and Gerner

Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

Page 1-17

Table 1.3. Examples Of WEIS Event Codes

WEIS Category

Goldstein value

11. REJECT
111
Turn down proposal; reject protest, demand or threat
112
Refuse; oppose; refuse to allow

-4.0
-4.0
-4.0

12. ACCUSE
121
Charge, criticize, blame, disapprove
122
Denounce, denigrate, abuse

-2.0
-2.2
-3.4

13. PROTEST
131
Make complaint (not formal)
132
Make formal complaint or protest

-2.0
-1.9
-2.4

17. THREATEN
171
Threat without specific negative sanctions
172
Threat with specific nonmilitary negative sanctions
173
Threat with force specified
174
Ultimatum: threat with negative sanctions and time limit specified

-6.0
-4.4
-5.8
-7.0
-6.9

18. DEMONSTRATE
181
Non-military demonstration; walk out on
182
Armed force mobilization, exercise and/or display

-6.0
-5.2
-7.9

COPDAB
The COPDAB data set is substantially larger in size and scope than WEIS, with about
350,000 international events for the period 1948-78. COPDAB uses a number of different news
sources rather than depending solely on The New York Times; in particular it uses a variety of
regional sources to cover events outside of North America and Europe.6 In contrast to the
categories in WEIS, COPDAB uses an ordinal coding scheme that goes from 1 to 16 (see Table 4)
supplemented by a numerical cooperation-conflict intensity scale developed by Azar and Sloan
(1975). COPDAB coding also classifies an event into one of eight typesfor example symbolic,

Because WEIS and COPDAB are based on different sources, they do not have a high degree of overlap:
International Studies Quarterly (1983) contains two analyses of this problem along with a commentary by
McClelland; Reuveny and Kang (1996b) deal with the issue of splicing the two data sets.

Schrodt and Gerner

Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

Page 1-18

political, military, economic or cultural. The GEDS project has been augmenting the original
COPDAB data with more recent data on selected dyads during the 1990s, using machine-assisted
coding methods.
Table 1.4. Examples Of COPDAB Event Codes
09

Nation A expressed mild disaffection toward B's policies, objectives, goals, behaviors with A's
government objection to these protestations; A's communiqu or note dissatisfied with B's policies
in third party. [scaled vale: 6]

10

Nation A engages in verbal threats, warning, demands and accusations against B; verbal, hostile
behavior. [scaled vale: 16]

11

Nation A increases its military capabilities and politico-economic resources to counter Nation B's
actions or the latter's contemplated actions; A places sanctions on B or hinders B's movement in
waterways or on land and attempts to cause economic problems for B. [scaled vale: 29]

________________________________________________________________________
CREON
The Comparative Research on the Events of Nations data set (Hermann et al 1977; East,
Salmore and Hermann 1978) is designed for the study of foreign policy interactions. Its basic
event coding scheme is similar to that of WEIS, but CREON also codes over 150 variables dealing
with the context of the event, related actions, and internal decision-making processes. Unlike
WEIS and COPDAB, CREON does not code all interactions during a period of time: instead it
covers a random sample of time periods during 1959-68 and a stratified sample of 36 nationstates that contains a disproportionate number of developed and English-speaking countries. The
purpose of CREON is to study the foreign policy process, rather than foreign policy output. In
practice this means that CREON is better suited than WEIS or COPDAB to studying the
linkages between the foreign policy decision-making environment and foreign-policy outputs for
specific decisions, but it cannot be used to study policy outputs over a continuous period of time
or for countries not in the sample. CREON is no longer maintained, but is available through the
ICPSR.

Schrodt and Gerner

Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

Page 1-19

PANDA
The most current global event data set comes from the PANDA project at the Program for
Nonviolent Sanctions and Cultural Survival at the Center for International Affairs at Harvard
(Bond, Bennett and Vogele, 1994). PANDA data set was automatically coded from Reuters lead
sentences using the KEDS computer program and employs a superset of the WEIS coding
scheme that provides greater detail in internal political events. It contains about 500,000 events
covering the entire world for the period 1984 to early 1995; the data are available on the Web at
http://data.fas.harvard.edu/cfia/pnscs/DOCS/datafiles.htm.

Table 1.5. PANDA versus WEIS codes for the WEIS "Reject" category
WEIS

PANDA

111

Turn down proposal; reject protest,

demand or threat

411

turn down proposal, reject nonjudicial appeal

112

refuse; oppose; refuse to allow

412

refuse to allow; ban or prohibit action

413

defy law, customs or norms

414

flee, protest emigration

415

ignore, isolate, ostracize,

disclose information, outing

421

voice opposition, disagreement

Other Actor-Oriented Event Data Sets

While WEIS, COPDAB, CREON and PANDA are the largest actor-oriented data sets, a
variety of smaller sets exist. The ICPSR has several regionally-specific, WEIS-coded data sets
dating from the 1970s, and additional regional data sets are being collected at the present time.
The South Africa Event Data set (SAFED; van Wyk and Radloff 1993) is a WEIS-coded
collection focusing on southern Africa for the period 1977-88; it has unusually dense coverage of
non-state actors such as guerrilla movements. Ashley (1980) assembled a data set focusing only
on the interactions of the superpowersthe USA, USSR and PRCfor 1950-72; this contains
about 15,000 events and is coded with a COPDAB-like scale.
Schrodt and Gerner
Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

Page 1-20

1.3.2. Episode-Oriented Data Sets

BCOW
The Behavioral Correlates of War data set (Leng 1987) codes a sample of 45 major
international crises over the period 1816-1979; roughly half of these crises culminated in war and
the other half were resolved without war. Most of the crises are in the 20th century; about a
third are post-WWII; and many of the crises preceding WWI and WWII are included in the
sample. BCOW's event codes are an expanded version of the WEIS scheme containing about 100
categories and differentiating more clearly between verbal, economic and military behavior. Leng
(1993b) contains an extensive analysis of this data set.
BCOW uses multiple sources of information, including newspaper accounts, diplomatic
histories, and chronologies (Leng 1987:1). The number of events in each crisis range from 120
events in the 1889-90 British-Portugal crisis in southern Africa to 2352 events in the 1956 Suez
crisis. The ICPSR data set is accompanied by a very extensive coding manual that would allow a
researcher to code additional crises in a manner consistent with the original data; it also includes
some specialized software that can be used to analyze the data.
Table 1.6. Examples Of BCOW Event Codes
Military Actions (sample from a total of 36 categories)
11212
11333
21143
31133

International Peacekeeping Force

Alert
Change in Combat Force Level
Fortify Occupied Territory

Diplomatic Actions (sample from a total of 35 categories)

12121
12362
12213
32151

Negotiate
Declare Neutrality
Punish of Restrict Foreign Nationals
Grant Independence to Colony

Economic Actions (sample from a total of 20 categories)

13121
23121
23231

Economic Negotiation
Sell or Trade
Pay for Goods or Services

Unofficial Actions (sample from a total of 11 categories)

Schrodt and Gerner
Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

14251
14213
14152

Page 1-21

Proforeign Demonstration
Antiforeign Demonstration
Hostage Taking

CASCON
The Computer-Aided System for the Analysis of Local Conflicts system (CASCON;
Bloomfield and Moulton 1989, 1997) codes the characteristics of 85 internal and international
conflicts during the post-World War II period. The analytical framework is based on a study by
Bloomfield and Leiss (1969) and is organized around six predefined conflict phases ranging from
the issues leading to the initiation of the dispute to the resolution of the dispute. CASCON
codes 571 "factors" for each crisis; some of these describe specific types of events, others
describe contextual characteristics of the crisis such as whether the parties to the conflict are
dependent on outside aid.
The current version of CASCON is an integrated "decision support system" designed to help
decision-makers and students compare current crises with the historical data on the crises; it
accompanies by an undergraduate textbook on international conflict. The system runs on
personal computers and an earlier version won a prestigious EDUCOM/NCRIPTAL award for
excellence in educational software. The system contains the conflict data set, a variety of
analytical tools that can be used to compare conflicts, and a subsystem for entering new cases
into the database.
SHERFACS
The SHERFACS data set (Sherman and Neack 1993) codes over 700 international disputes
and almost 1,000 domestic disputes in the 1945-1984 period. It combines several different
coding schemes, including COPDAB event codes, the CASCON crisis phase structure, and a
variety of conflict management variables originally used in the Butterworth (1976) data set on
crisis mediation. SHERFACS is particularly strong on coding non-state actors such as ethnic

Schrodt and Gerner

Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

Page 1-22

groups, transnational actors such as intergovernmental organizations, and non-national actors

such as multinational corporations.
An early version of SHERFACS is available from the ICPSR (Alker and Sherman 1982,
1986). A much more extensive version of the data set was planned and partially implemented,
but this project was cut short by Frank Shermans untimely death in 1996. Nonetheless, much
of this information is available at a site maintained by Hayward Alker
http://www.usc.edu/dept/ancntr/Paris-in-LA/Database/sherfacs.htmland

is in the

public domain.
Other Episode-Oriented Event Data Sets
Several other data collections available from the ICPSR such as The World Handbook (Taylor
and Hudson 1972) contain some limited amounts of event data. Another example is the PRINCE
Project data set (Coplin, O'Leary, and Shapiro, n.d.). This data set was originally collected in
conjunction with a computer simulation project and contains a small set of event data dealing
with political issue positions for the period 1 January 1972 to 30 June 1972. Other data sets
have been collected for the study of a specific crisis. For example Lebovic (1993) coded events
during the period prior to the 1991 Gulf War (2 August 1990 to 16 January 1991 in order to
analyze the impact of foreign policy "momentum" in that crisis. The International Political
Interactions project (Moore and Davis 1998;
http://garnet.acns.fsu.edu/~whmoore/ipi/ipi.html)

provides conflict event information

on selected dyads for 1979-92. The International Crisis Behavior data set (Brecher and
Wilkenfeld 1997; http://www.colorado.edu/IBS/GAD/spacetime/data/ICB.html), while
not strictly event data, is another example of a data set focusing on the characteristics of conflict
episodes from 1918 to 1994;

1.4. Sources of news reports

One of the perennial problems in event data analysis has been the choice of which news
reports to code. Because most event data sets have been coded in the United States, they have
Schrodt and Gerner
Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

Page 1-23

tended to use reports written in English. Because human coding is very labor-intensive, those
projects have also tended to favor news sources that are readily available and indexed. For
example, WEIS was coded from the New York Times; CREON from Deadline Data on World
Affairs, which abstracts 46 international sources (Hermann et al 1973:18).
There has always been concern that news sources could have substantial regional biases. An
early study by Doran, Pendley, and Antunes (1973) showed dramatic differences between New
York Times coverage of violence in Central America and the levels of violence reported by Central
American sources. Because of this, some data sets have used regional news sources. For example,
the BCOW codebook (Leng, 1987) lists dozens of periodical and historical sources and Azar
(1980: 146) states that COPDAB is based on "events reported in over 70 sources."
During the 1990s, most event data projects (whether human or machine coded) shifted to the
Reuters newswire source. At present, Reuters is archived from 15 April 1979 to 10 June 1997 in
the NEXIS data service, and material subsequent to that date can be obtained from the Reuters
Business Briefing service. Reuters issues about a thousand stories per day, and therefore
provides far greater coverage than the Times or any other single regional source. While Reuters
has some regional biasesit devotes greater coverage to areas that are of interest to institutions
that can afford its services but which are not covered by other mediathe fact that it is global
rather than based in a single city or country reduces these substantially.7
Reuters and other news services such as Agence France Presse (AFP) and the BBC World
Service have been criticized as providing a "white European male" perspective. Based on our
unsystematic encounters with newswire personnel in the Levant, we find this criticism itself to
be a white male perspective: Most of Reuters correspondents ("stringers") are local to the areas
they are reporting from, and quite a few of them (as well as much of the Reuters editorial staff)
7

The status of the availability of Reuters has been in flux over the past three or four years. For a period of time after
dropping NEXIS, Reuters was only available directly from Reuters Business Briefing, and only for a ten year
period. In mid-1999, it joined with Dow Jones Interactive to create a service that initially promised to provide
the full Reuters archives, possibly back as far as the early 1970s. By early 2000, however, that combined service
had yet to materialize, though the Web-based Dow Jones service has itself emerged as an attractive alternative to
both NEXIS and Reuters. Given the current wave of media mergers, acquisitions, and spin-offs, anything might
happen in the future.

Schrodt and Gerner

Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

Page 1-24

are not male. Most stringers have had a disproportionate exposure to Western education and
normsthey are elites, not peasantsbut they typically grew up with the local culture and
language.
The distinction here is important: the global news agencies are based in the capitals of the
hegemonic powers, but the reporting itself (and some of the editing) comes from a very
decentralized network of observers who are usually native to the areas from which they are
reporting. In contrast, the hegemonic "papers of record," such as The New York Time rely much
more heavily on their own nationals to provide coverage. At best these individuals experience a
long learning curve, as the autobiographical accounts of correspondents such as Thomas Friedman
and Harrison Salisbury attest; at worst they are Graham Greene wannabees without a long-term
commitment to a region whose understanding of the local politics has invariably been filtered
through the alcoholic haze of too many expense-account dinners at five-star hotels. The
differences in the quality of reporting can be profound.)
1.4.1. Regional versus Global Sources
Data services such as NEXIS and Dow Jones Interactive carry the text of literally hundreds
of regional news sources. As these sources became easily available, a number of researchers
(including ourselves) thought that the regional sources could be used to fill in the gaps of the
reporting of global news sources. provide details as necessary. We assumed that services such as
Reuters and AFP would contain a selective subset of the events reported in the regional sources,
and anything really important in a regional source also would be reported by the global source.
Alas, it isnt so, at least for Reuters. Reuters and the regional sources are supplementary,
rather than complementary. A number of different studies in different regions of the world have
shown that Reuters reports events that are not reported in the regional sources (and vice versa).
These studies include
Europe

Gerner et al 1994, Huxtable and Pevehouse 1997

Middle East

Gerner et al 1994

Africa

Huxtable and Pevehouse 1997, Moore 1997

Schrodt and Gerner

Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

Southeast Asia

Page 1-25

Howell and Barnes 1993

The same is true when global sources are compared: For example in his dissertation research
on West Africa, Huxtable (1997) assumed that the English-language sources Reuters and BBC
would focus on Anglophone states such as Nigeria, Ghana and Sierra Leone, while the Frenchlanguage AFP would focus on Francophone states such as Senegal, Niger and Cte d'Ivoire. This
did not prove to be the case: Reuters would sometimes pick up major events in Francophone
states that were missed by AFP, and AFP would sometimes provide better coverage of
Anglophone states. In some cases, it was almost possible to reconstruct the travel itineraries of
individual Reuters, BBC and AFP reporters as they worked their way through West African
capitals, producing a flurry of temporary detail on areas that would receive no coverage for
another year.
The fact that event data is coded from a finite number of sources has been criticized by Alker
(1988), among others, as privileging some interpretations of history over others. While this may
be true, it is no more or less the case than the situation facing traditional studies of political
behavior. Short of descending into a post-modernist quagmire where nothing can be assumed,
concluded or explained, any political analysis must assume that certain events, conditions,
motivations and coalitions occurred, and others did not. The traditional method of composing
accounts of political activities using a variety of documentary and autobiographical sources is one
way of doing this; the processes by which text is selected for event coding is another. Each
method is subject to selection bias and varying interpretation.
In retrospect, the event data community under-estimated the sheer volume of "events" that
occur in the world. At 1000 stories per day, Reuters initially looked like a major improvement
over the ever-diminishing international coverage of the New York Times. But a bit of reflection
will show that even 1000 stories is only a tiny fraction of all of the political events that occur in
the world on any given day. Any news source is going to nonrandomly sample only a small
number of these events, so the question is whether that sample is useful for a specific analytical
task.
Schrodt and Gerner
Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

Page 1-26

1.5. Example: KEDS Data for the Arab-Israeli Conflict

This section discusses our major validity test of the KEDS system, published as Schrodt and
Gerner (1994). It is not the only validity test we have done but it is the most systematic. The
test focuses on the Levant for the period 1982-92, and compares the KEDS-generated data with
the general textual record of "events on the ground." (In Chapter 2, we will also compare this
machine-coded data with the closest comparable human-coded data set, Tomlinson's (1993)
extension of the original WEIS set.)
To create this data set, we downloaded the first sentences (leads) from Reuters News
Service stories available from the NEXIS data service. NEXIS is searched using keywords that
can be arranged into Boolean statements. To create this data set, the search command was
HEADLINE (ISRAEL! OR JORDAN! OR EGYPT! OR LEBAN! OR SYRIA! OR PLO OR
PALEST!).
The "!" is a wild card character that matches any word beginning with the preceding letters;
"PALEST!" picks up "Palestinian," "Palestinians," and "Palestine." 8 We examined only those
dyads in which both the source and the target of the event were among the seven actors of
interest. A total of 23,127 events are included in this eleven year data set; daily events reports
were aggregated to the monthly level prior to analysis. Israeli actions toward Palestinians, Israeli
actions toward Lebanon, and Palestinian actions toward Israel account for the greatest number of
events. The fewest events are recorded for Lebanese actions toward Jordan, Lebanese actions
toward Egypt, and Jordanian actions toward Lebanon. 9
In order to obtain a general sense of these regional interactions, we began by examining the
conflict-cooperation patterns of each directed dyad during the eleven year period. (A directed

In later phases of the KEDS project, we further refined this search statement to eliminate sports stories, historical
chronologies, monthly news summaries and other information that we did not want to code.
9 The small number of reported events involving Lebanon can probably be accounted for by two factors. First, the
situation in Lebanon was tremendously unstable during the 1982-92 period. Therefore, it is likely that few
routine interactions occurred; instead the data consist largely of military events and diplomatic discussions about
Lebanon. Second, the instability probably decreased reporting of "ordinary" events: the few media
representatives in the country focused their attention on the domestic and international conflicts.
Schrodt and Gerner
Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

Page 1-27

dyad refers to the actions of X toward Y and is represented as XY.) We were particularly
interested in reciprocal directed dyads (X's actions toward Y and Y's actions toward X) as well as
any directed dyad with an usually high or usually low net cooperation score. Net cooperation
was calculated by weighting each WEIS according to the Goldstein (1992) scale and totaling the
events for each month. This produces a single numerical score on a conflict-cooperation
dimension similar to that used in COPDAB; negative scores indicate conflict and positive scores
indicate cooperation..
The average conflict or cooperation scores shown by the KEDS data across the entire period
are consistent with narrative accounts describing the specific relationships. Between 1982 and
1992, most analysts describe interactions among the United States, Israel, and Egypt as relatively
harmonious, albeit with some tensions. In contrast, interactions among Israel, Lebanon, Syria,
and the Palestinians were quite strained. These differences show up clearly in the net
cooperation measures. For example, IsraelPalestinians has the lowest average net cooperation
score of any directed dyad over the eleven year period. IsraelPalestinians are roughly twice as
conflictual as those of any other directed dyad examined. Other highly conflictual directed dyads
(in order from most conflictual) include IsraelLebanon, PalestiniansIsrael, LebanonIsrael,
PalestiniansLebanon, LebanonPalestinians, IsraelSyria and SyriaLebanon.
Fifty-five percent of the directed dyads have a mean positive net cooperation score. The
variation in the extent of cooperative actions among these 23 directed dyads is not great: most
have net cooperation scores that are only slightly positive. The most cooperative directed dyad
is USAIsrael. Other relatively cooperative directed dyads include IsraelUSA, USAEgypt,
IsraelEgypt, USAJordan, and EgyptUSA.
We also looked at monthly net cooperation scores for key reciprocal directed dyads to
determine whether the data were accurately reflecting major events in the region such as the 1982
Israeli invasion of Lebanon, the Syrian military presence in Lebanon, the onset and evolution of
the intifada, and the regional peace talks that began in Madrid in October 1991. We briefly
summarize three dyads here; additional cases are discussed in Gerner (1993).
Schrodt and Gerner
Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

Page 1-28

1.5.1. Face Validity

As Figure 1.5 illustrates, Israeli-Lebanese relations have been characterized almost entirely by
hostility throughout the eleven-year period. The most severe conflict occurred during the 1982
war in Lebanon that began in June and continued through mid-August; this raised the
IsraelLebanon hostility score by a factor of two. Conflict peaks again in mid-1983, before
Israel had withdrawn its troops from the Bekaa Valley or the Chouf Hills in southern Lebanon; in
early 1985, after Israel had decided to pull out of most of southern Lebanon but before the
evacuation had been completed; at the beginning of 1987, as Palestine Liberation Organization
(PLO) fighters moved back into Lebanon; and late in 1991, when Israel launched a series of
artillery attacks against Iranian-backed Hizballah forces and Lebanese villages. There is no overall
improvement in the relationship in the fifteen months following the Madrid talks; this is
consistent with the increased Israeli-Lebanese military hostilities that accompanied the early
negotiations.

0
-50

Net Cooperation

-100
-150
-200
-250
-300
-350
ISR>LEB
LEB>ISR

-400
-450

Figure 1.5: Goldstein series for IsraelLebanon, 1982-1992

Schrodt and Gerner
Analyzing International Event Data

DRAFT: February 6, 2000

Jul.92

Jan.92

Jul.91

Jan.91

Jul.90

Jan.90

Jul.89

Jan.89

Jul.88

Jan.88

Jul.87

Jan.87

Jul.86

Jan.86

Jul.85

Jan.85

Jul.84

Jan.84

Jul.83

Jan.83

Jul.82

Jan.82

-500

International Event Data

Page 1-29

United States actions toward Israel show the greatest amount of cooperative behavior of any
directed dyad in the data set (see Figure 1.6). This can be attributed in part to the consistently
high level of consultation between the two countries. Thirty-five percent of reported
USAIsrael events and 36 percent of IsraelUSA events fall into the WEIS "consult" category,
with peaks in 1982-83 and again in 1991-92. The negative net cooperation of USAIsrael
during 1982 is the result of an unusually large number of U.S. actions in response to Israel's
invasion of Lebanon that fall into the WEIS "accuse" and "reject" categories. In this same period,
Israel initiated a number of cooperative consultations, which accounts for the generally positive
IsraelUSA pattern. During the intifada we see the same phenomenon: United States
consultations with Israel help moderate the impact of U.S. accusations and rejections of Israeli
proposals. Although Israeli-U.S. interactions occasionally move into the net conflict area, this
pattern is far less strong than it would have been without the positive consultative activities.

60
50

Net Cooperation

40
ISR>USA

30

USA>ISR

20
10
0
-10
-20
-30

Figure 1.6. Goldstein series for Israel-United States, 1982-1992

Schrodt and Gerner

Analyzing International Event Data

DRAFT: February 6, 2000

Jul.92

Jan.92

Jul.91

Jan.91

Jul.90

Jan.90

Jul.89

Jan.89

Jul.88

Jan.88

Jul.87

Jan.87

Jul.86

Jan.86

Jul.85

Jan.85

Jul.84

Jan.84

Jul.83

Jan.83

Jul.82

Jan.82

-40

International Event Data

Page 1-30

Net Cooperation

-100
-200
-300
-400
-500

ISR>PAL
PAL>ISR

Figure 1.7. Goldstein series for Israel-Palestinians, 1982-1992

Finally, Israeli and Palestinian net cooperation over time is particularly interesting (see Figure
1.7). The net cooperative measure picks up a number of critical shifts in the overall attitude of
each actor toward the other. For instance, the 1982 Israeli invasion of Lebanon is distinctly
marked by a dramatic increase in conflictual Israeli actions toward Palestinians (both in Lebanon
and in the West Bank and Gaza) preceding the invasion and the sharp drop-off in such events
once Israeli troops had withdrawn from much of Lebanon later the same year. Net conflictual
actions by Palestinians toward Israel are also higher than average during 1982, although the
intensity is much less dramatic than that seen in Israeli actions toward the Palestinians. Israeli
conflictual events begin to increase again in 1986, after Yitzhak Shamir replaced Shimon Peres as
prime minister in the 1984 Labor-Likud Unity government. The data correctly show the
dramatic increase in conflictual actions by both actors, particularly Israel, at the beginning of the
intifada. The intifada continues to affect the pattern of Israeli and Palestinian actions throughout
the next several years. The decrease in conflictual actions by each actor matches the decline in

Schrodt and Gerner

Analyzing International Event Data

DRAFT: February 6, 2000

Jul.92

Jan.92

Jul.91

Jan.91

Jul.90

Jan.90

Jul.89

Jan.89

Jul.88

Jan.88

Jul.87

Jan.87

Jul.86

Jan.86

Jul.85

Jan.85

Jul.84

Jan.84

Jul.83

Jan.83

Jul.82

Jan.82

-600

International Event Data

Page 1-31

the intensity of the intifada; the renewal of strongly conflictual Israeli actions toward Palestinians
during the second half of 1992 is also recorded.
The asymmetry in the reported events from Israel to Palestinians versus Palestinians to Israel
is probably partly due to the political asymmetry of the situation and partly due to difference in
reporting. As the occupying power in the West Bank, Gaza and southern Lebanon, Israel was
more likely to initiate action than the Palestinians. For example, during the 1982-92 period,
Israeli police frequently broke up Palestinian demonstrations, but there were no instances of
Palestinian police breaking up Israeli demonstrations (or, for that matter, Palestinian police). But
in other instances, particularly those involving Israeli activity in Lebanon and during the intifada,
some of the asymmetry may be the result of reporting styles. Many Reuters reports were based
on Israeli sources, and consequently the lead sentence would tend to emphasize the Israeli
actionIsraeli troops raided Palestinian guerilla baseswithout necessarily reporting the
reciprocal behavior that would occur in a military clash.
An additional check on the face validity of these event data is the relationship between the
number of Palestinians shot and killed by Israeli soldiers and settlers during the intifada, as
recorded by a source independent of Reuters, and net cooperative Palestinian and Israeli actions.
While the record of fatalities is a more specific type of behavior than the aggregate net
cooperation measure, one would expect the two series to covary. Figure 1.8 illustrates the
relatively close relationship between these phenomena. Palestinian deaths by shooting are
strongly and negatively correlated with Israeli net cooperative actions toward Palestinians (r = 0.51; significant at the .01 level). In other words, months in which Israeli actions toward
Palestinians are less cooperative according to the Goldstein-scaled measure tend to be the months
in which a higher number of Palestinians are shot to death by Israeli forces. Shooting deaths are
also negatively correlated (r = -0.47; significant at the .01 level) with Palestinian net cooperative
actions toward Israel.10
10

The data on Palestinian shooting deaths by Israeli occupation forces comes from the Jerusalem-based Palestine
Human Rights Information Center. Deaths that occur in the context of a military operation or when there is no
clear human rights violation are excluded from these figures, as are deaths due to beating, tear gas inhalation, or

Schrodt and Gerner

Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

Page 1-32

Net Cooperation/Deaths

0
-100
-200
-300
-400
-500

Shootings

ISR>PAL

Dec.92

Jun.92

Dec.91

Jun.91

Dec.90

Jun.90

Dec.89

Jun.89

Dec.88

Jun.88

Dec.87

-600

PAL>ISR

Note: Number of shooting deaths was multipled by -10.

Figure 1.8: Palestine-Israel Net Cooperation and Palestinians Shot to Death by Israeli Forces
during the Intifada
In short, this quick examination of some of the directed dyads indicates that the KEDS ArabIsraeli conflict data set accurately reports key events in the region during the eleven year period.
There were no unpleasant surprises. Both cooperative and conflictual patterns show up where
expected and there are no major unexplained clusters of events..

1.6. Comparison of Lead and Full Story Coding

Most of our KEDS data sets are coded using only the first sentencethe "lead"of a
newswire report. Following standard journalistic practice, the lead sentence usually summarizes
the story that follows, and commonly has a relatively simple declarative structure, as illustrated

other non-bullet causes. The level of association between Palestinian deaths and the actions of the two directed
dyads is actually stronger if beating and tear gas inhalation deaths are included. Gerner (1990, 1991) discusses
these data in greater detail.
Schrodt and Gerner
Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

Page 1-33

earlier in Table 1.1. This approach contrasts with earlier human-coded projects, which coded
entire stories.
This emphasis on leadswhich typically contain less than 10% of the text of the
storywould appear to ignore a great deal of useful information, and this has been one of the
common criticisms of the KEDS approach. In practice, however, very little seems to be lost, at
least when Reuters is being coded.
In a typical Reuters report, the body of the story adds very few events beyond those in the
lead, and coding the full story substantially complicates automated coding methods. The body of
the story will often repeat the event reported in the lead several different times, provide
background information on events that occurred earlier in time, and provide extensive direct
quotations that are very difficult to code correctly. When a single Reuters story contains
multiple codeable eventsfor example an outbreak of violence accompanied by condemnations
and offers of mediationthese almost always generate multiple stories, each with an appropriate
lead. In the Reuters text-stream, in fact, one is far more likely to encounter the problem of the
same event being reported in multiple leads than the problem of a significant event not being
reported in any lead.
Theses characteristics of Reuters are, however, more pronounced in an areas such as the
Levant, which is reported very intensely. When our colleague Phillip Huxtable was working on a
data set for West Africa (Huxtable 1997)a region that does not receive intense coverage in the
international presshe noted that Reuters would sometimes append a series of event reports to
the end of a story that were unrelated, except in their regional focus, to the primary focus of the
story. The underlying editorial model appeared to be "Hey, if you're sufficiently interested in
this out-of-the-way region to read this far, you'll probably enjoy this other stuff as well."
Huxtable concluded that it might be possible to increase the density of events by full-story
coding.
Huxtables observation led us to run some specific tests comparing lead-sentence and fullstory coding on a data set dealing with the Persian Gulf region, which receives more sporadic
Schrodt and Gerner
Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

Page 1-34

coverage than the Levant. After downloading the full stories from NEXIS, we first filtered the
text to remove all sentences appeared to be direct quotes.11 Lengthy direct quotations tend to be
very difficult to code because spoken language is less systematic than journalistic language, and
since the main point that the speaker is making is usually summarized by the Reuters text outside
of the quotations, little is lost by removing quotes. The remaining sentences were coded using
the same dictionaries and complexity filter used to code the lead sentences.
The full-story coding (and downloading) was substantially more time-consuming than leadsentence coding. Coding the lead sentences required about 2-hours on an 80 Mhz Macintosh
7100; coding the full stories required a full 24-hours.12 The full-story coding generated 264,421
events, as opposed to the 48,721 events generated from the leads.
Table 1.7 shows the correlation (r) between the monthly Goldstein series generated from the
lead-sentence and full-story coding for 30 directed dyads. With the exception of dyads involving
the United Arab Emirates (UAE), the correlations of the two series are quite high, usually above
0.75. This suggests that in most statistical studies involving linear models, similar results will be
obtained with either approach. More generally, lead-sentence coding is probably quite adequate
for exploratory work, given the much greater investment of time required to download complete
stories.
Table 1.7. Correlation of Goldstein-score time series generated with leads and full
stories
Target
Source

11

IRN

IRQ

SAU

USA

KUW

UAE

The filter skipped sentences containing a double-quote character (ASCII 34) that were either preceded by a period
(.) or followed by a comma (,). This eliminated all correctly formatted sentences in Reuters that quote a speaker,
while retaining sentences that contain short phrases placed inside quotes such as:
Palestinian diplomacy has ended Lebanon's bloody "camps war" but analysts say it is likely to prompt a
confrontation between Israel and Amal.

12

These times are substantially longer than those required on a contemporary personal computer, which would
have a much higher clock speed.

Schrodt and Gerner

Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

IRN
IRQ
SAU
USA
KUW
UAE

--.94
.77
.89
.68
.38

Page 1-35

.90
--.74
.96
.94
.38

.73
.98
--.75
.43
.35

.85
.98
.64
--.55
.65

.76
.93
.52*
.88
--.46

.47
.70
.31
.52
.35
---

N=119
* Excluding Feb-91, which contains a number of incorrect codes associated with the Second Gulf War.
With Feb-91 included, the correlation is 0.17

The lower correlations associated with the UAE are consistent with Huxtable's observations
about West Africa: states perceived as peripheral by the international news media are more likely
to be discussed only in the body of a story and not in the lead. For example, the SAU UAE
dyad contains 200 reported events in the full-story series, but only 6 (!) in the lead sentences.
The contrast between full-story and lead-sentence coding in IRN UAE series is less dramatic,
but still has 126 events in the full series versus 21 in the leads.
Table 1.8 shows the dyadic reciprocity correlation (X Y x Y X; see Dixon 1996 and
Goldstein and Freeman 1990) for the lead and full-story sequences. These correlations show a
very clear pattern, with the full story reciprocity being higher in all but two cases. Once again,
the correlations for the minor actors are substantially lower than those for major actors; in some
cases they are not even statistically significant. Some of the reciprocity in the full stories may be
artificial because the full story is more likely to present "the other side" and thus generate a
"reciprocal" event that would not have been present in the absence of the Reuters reporter.
However, the full story is also more likely to present secondary events that occurred but which
did not by themselves justify a separate story and thus a lead. Furthermore, the fact that these
series are aggregated by month should reduce the likelihood that the observed reciprocity is
simply an artifact of Reuters editorial guidelines.
Table 1.8. Dyadic reciprocity in Goldstein-score monthly aggregations generated with
leads and full stories

Schrodt and Gerner

Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

Page 1-36

Dyad
IRN-IRQ
IRN-SAU
IRN-USA
IRN-KUW
IRN-UAE

Full
.95
.85
.80
.71
.25

Leads
.84
.61
.71
.32
.06

IRQ-SAU
IRQ-USA
IRQ-KUW
IRQ-UAE

.96
.97
.84
.00

.78
.92
.77
.02

SAU-USA
SAU-KUW
SAU-UAE

.73
.63
.76

.53
.35
.93

USA-KUW
USA-UAE

.27
.22

-.02
.27

KUW-UAE

.86

.64

Finally, the similarities between the full-story and lead-sentence series can also be found in
more complex measures of the series. Figures 1.9 and 1.10 give examples of the autocorrelation
functions and cross-correlation functions of some of these series. The shapes of the curves are
similar, although the strongest correlations are found in the full-story data.

Schrodt and Gerner

Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

Page 1-37

Autocorrelation for SAU ->IRQ

0.6
0.5
0.4
0.3
0.2
0.1

SAU>IRQ

12

-0.1

11

10

ALL SAU>IRQ

Figure 1.9. Autocorrelation function of Goldstein series for Saudi Arabia Iraq for fullstory and lead-sentence events

Cross-correlation for IRN->SAU x USA->IRN

0.7
0.6
0.5
0.4
0.3
0.2
0.1
-12 -11 -10 -9

-8

-7

-6

-5

-4

-3

0
-2 -0.1
-1 0

-0.2
LEADS

ALL

Figure 1.10. Cross-correlation function of Goldstein series for Iran Saudi Arabia with
USA Iran for full-story and lead-sentence events

Schrodt and Gerner

Analyzing International Event Data

DRAFT: February 6, 2000

10 11 12

International Event Data

Page 1-38

In summary, this analysis of the Gulf finds that for major actors there is generally a high
correlation between the Goldstein series generated only with the lead sentences and those
generated with full stories. Given that full-story coding involves substantially more downloading
time, lead-sentence coding is probably sufficient for most exploratory work. However, full-story
coding provides dramatically more information on minor actors, which is consistent with
Huxtable's conclusions about West Africa.13 Information is "out there" in the wire service
reports, even if one has to dig for it, and minor actors are sometimes dramatically underrepresented unless the entire story is coded.
Curiously, the marginal frequencies in human-coded data sets such as WEIS and COPDAB
generally look more like the frequencies of lead-sentence coding than the frequencies of full-story
coding, even if the coders were supposed to be working with the entire story. For example,
COPDAB contains about 350,000 international events for the period 1948-78. Our full-story
data set on the Gulf therefore records, on an annual basis, about the same number of events as
COPDAB records for the entire world, despite the fact that COPDAB was generated by fullstory coding from multiple regional sources.
In some early unpublished experiments that we did comparing KEDS automated coding to
that of graduate student coders, we found that the human coders were more likely to miss
secondary events reported in a story, as well as some of the combinations created by multipleactor meetings. If the human coder, rather than meticulously coding every single sentence, relies
(explicitly or cognitively) on a summary of the story, the secondary events involving minor
actors will be missed.

13

We also found this pattern in a data set we coded for Central Asia, 1989-99. The correlation between lead-story
and full-story monthly scores aggregated using the Goldstein scale was very high for Afghanistan, which was
covered continuously by the news media during the period. The correlations were substantially lower for minor
actors such as Kyrgistan and Uzbekistan.

Schrodt and Gerner

Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

Page 1-39

1.7. Conclusion
There are two fundamental reasons for using event data in political analysis. First, politics
does not have the convenient numerical measures such as location, momentum, and temperature
found in physics, or variables such as price, interest rates, and GNP found in economics.
Political activity instead consists largely of discrete actions and communications directed from
one actor to another over time. McClelland's (1970) original observations on the potential utility
of event data as a method of addressing this problem still hold.
Second, human analysts have a limited ability to absorb vast quantities of largely redundant
material. The text of NEXIS news wire leads covering only Israeli-Palestinian interactions for
1989 runs to some 300 pages. The full articles would fill perhaps 2000 pages; we suspect that
few researchers would read all of these. The task becomes even more formidable if one is dealing
with a long time series such as the Cold War: just what were the U.S.A. and U.S.S.R. doing on 16
August 1955? While most human analysts can memorize the day-to-day details of a short time
period such as the Cuban Missile Crisis, or the major events of a long period such as the Cold
War, we are skeptical about the human ability to memorize, much less analyze, day-to-day
details for a long time period.
Event data fill that gap. The text of the journalistic sources provide memory and a variety
of statistical and other computational methods can provide analysis. Between the text and
analysis, one needs something similar in content to event data.
Science magazine once surveyed how new techniques in the physical and biological sciences
sometimes revolutionized not just the methodologies, but also the theories, within their fields:
Not everybody appreciates the importance of technique. Many scientists, in fact, are
"theory snobs" who dismiss technique as a kind of blue-collar suburb of science. . . . [But
there is,] clearly, enormous transforming power in techniques. In the absence of an essential
technique, a researcher or a field flounders, developing elegant theories that cannot be
decisively accepted or rejectedno matter how many intriguing circumstantial observations
are available. But with a key technique in hand, the individual and field move ahead at almost
terrifying speed, finding the right conditions to test one hypothesis after another.
Schrodt and Gerner
Analyzing International Event Data

DRAFT: February 6, 2000

International Event Data

Page 1-40

Conversely, new techniques often uncover new phenomena that demand new theories to
explain them. (Hall 1992: 345)
The research in international relations, and much of comparative politics, is arguably theory
rich and data poor. Too many theories are chasing too few facts, and for large sectors of those
communities, research tools still consist of CNN, the New York Times, a copy of Thucydides
and a snifter of brandy. At the same time, the interactions in international system are becoming
more complex with the end of the Cold War and the need is greater than ever to be able to
systematically study alternative theoretical explanations for that behavior.
One of the favorite parables employed by evangelical preachers is that of a sailing ship
becalmed for weeks in the Atlantic, its crew slowly dying of thirst. Sighting a passing vessel,14
the beleaguered crew appeals frantically for water. The other ship replies, "Throw down your
buckets; you are surrounded by fresh water!": they are resting in the outflow of the mighty
Amazon River.
The quantitative international relations community has often felt becalmed with respect to
data. We have no American National Election Study, no U.S. Census or National Institutes of
Justice data, and only so many ways one can analyze the World Handbook, Correlates of War,
WEIS and COPDAB. But in fact, we are sitting amid a river of political databoth eventoriented and contextualflowing past us every day from journalistic sources. Those sources are
increasingly machine-readable, and if we can find a means of tapping them using the natural
language capabilities of contemporary computers, we will find ourselves awash in data.

14

As with many parables of evangelical preachers, the movement of this second vessel under windless conditions is
not explained...

Schrodt and Gerner

Analyzing International Event Data

DRAFT: February 6, 2000