Chapter Three The Statistical Characteristics of Event Data: DRAFT October 30, 2000

Chapter Three
The Statistical Characteristics of

Event Data
Most political science data derive either from random samples, or from the measurement of
entire populations. Survey research projects, which certainly provide the largest amount of data
in political science generally, rely on carefully-designed random sampling techniques that have
been refined over the past century. The sampling properties of these surveys can be analyzed
using probability theory and have well-understood characteristics. Sampling is used not just to
provide information on public opinion, but also data that are treated as if they were population
characteristics such as unemployment and inflation rates.
At the opposite end of the spectrum one finds data sets that aspire to provide data on the
entire population, rather than a sample. When that population is very large, as with the United
States census, there may still be elements of sampling, but in many instances this is almost
entirely absent. The Correlates of War project, for example, in all likelihood has identified all
historical instances of war during the 1815-1990 period that involve nation-states recognized
by the European powers. Definitional differences might remain, and some of the variables such
as GNP or battle casualties may be poorly measured, but the population itself has been
identified. Data involving international trade, treaties, borders, government type, and crises
between major states also involve little if any sampling.
Event data fall between these two categories: they are neither comprehensive, nor a random
sample. The news media report only a tiny fraction of the events that occur in any given day,
but they do not report these events at random. Reporting is affected by factor such as the type
of event (a war is more likely to be reported than a street crime), by the novelty of the event
(events at the outbreak of a war are more likely to be reported than the continuation), by the
DRAFT October 30, 2000
Statistical Characteristics of
Event Data
Page 3-2
actors involved (major states, and countries with high GDP, get far more coverage than small or
poor states), and by the accident of a reporter being at a place when something interesting
happens, and the editors putting that report into a newspaper or onto a wire service.
This chapter will explore several issues involving the statistical characteristics of event data.
We will first look at the definition of what constitute an event, an issue on which the existing
literature shows less than complete consensus. We will then systematically go through the
sources of error and uncertainty that come between events on the ground and a coded event
data set. Coding mistakes are one source of these errors, but by no means the only, nor
necessarily the greatest, source of error. Based on this model, we examine the consequences of
alternative means of coding and aggregating events, showing that in at least some circumstances,
analyses are relatively insensitive to specific coding schemes and weighting methods. Finally, we
look at how the dynamic event data approach to early warning, which will be explored in the
remaining chapters, compares to the alternative approach of structural early warning.
3.1. Defining an "event"

Despite the widespread use of event data, there is no single universally accepted definition of
what constitutes an "event." This section will first survey some existing definitions, then
propose an alternative.
The early event data projects provided relatively succinct definitions. According to Burgess
and Lawton (1976:6), for example, "events data is the term that has been coined to refer to words
and deeds i.e. verbal and physical actions and reactions that international actors (such as
states, national elites, intergovernmental organizations and NGOs) direct toward their domestic
or external environments." Azar and Ben-Dak define an event as:
some activity undertaken by an international actor (a nation-state, a major subunit of a
nation-state, an international organization) . . . at a specific time and which is directed toward
another actor for the purposes of conveying interest (even non-interest) in some issue. Thus
an event involves (1) an actor, (2) a target, (3) a time period, (4) an activity, and (5) an issue
about which the activity revolves. (1975:1; quoted in Laurance, 1990:112)
Schrodt and Gerner
Analyzing International Event Data
DRAFT: October 30, 2000
Event Data
Page 3-3
The COPDAB project designates events as:

occurrences between nations which are distinct enough from the constant flow of
"transactions" (trade, mail flow, travel and so on) to stand out against this background as
"reportable" or "newsworthy." Thus, to qualify as an "event," an occurrence has to be
actually reported in some reputable and available public source. (Azar, 1980:146; see also
Davies, 1991:3)
The WEIS codebook does not contain a clear description of an event, but an early paper by
McClelland provides the following definition:
Event-interaction is meant to refer to something very discrete and simple to the veritable
building blocks of international politics, according to my conception. The content of
diplomatic history is made up, in large measure, of event-interactions. They are the specific
elements of streams of exchange between nations. Here are a few examples for hypothetical
Nations A and B: Nation A proposes a trade negotiation, Nation B rejects the proposal,
Nation A accuses B of hostile intentions, Nation B denies the accusation, Nation B deploys
troops along a disputed boundary, Nation A requests that the troops be withdrawn, . . . Each
act undertaken by each actor as in the illustration is regarded as an event-interaction. (1967:8)
In a similar fashion, the manual for the BCOW data set (Leng, 1987) does not provide an explicit
definition of an event but its dense set of verb-oriented event categories implicitly describes the
concept.
There are at least two significant problems with these older definitions from the standpoint of
rigorously delineating event data. First, they refer to activities of international actors that an
analyst or coder almost never observes. An analyst observes the report of an activity. This
difference may seem subtle but it is important; it is the same as the difference between an attitude
or opinion unobservable mental states and the answer to a questionnaire, which is
observable.
Second, none of the definitions explicitly indicates what constitutes an "activity,"
"transaction," or "action." In some cases (e.g., BCOW), the codebook covers this quite
thoroughly. But in other instances most importantly, the ICPSR's COPDAB and WEIS
codebooks the coder is given only a set of English verbs, verb phrases, and noun phrases from
Schrodt and Gerner
Event Data
Page 3-4
which to infer the underlying coding rules. Through the use of the phrase "direct towards," some
definitions (Burgess and Lawton, 1972; Azar and Ben-Dak, 1975) also implicitly require an
assessment of motive on the part of the initiator of the event. This leaves considerable room for
ambiguity in interpretation. Lest this seem like a trivial point, one major source of ambiguity in
event coding arises from policy statements with no explicit target audience, e.g., "Iraq announced
it was raising its oil output beyond OPEC quota levels." Some coding systems do not consider
this an event because of the absence of a target; others consider it an event with "the world" as
the target.
At the risk of definition proliferation, we suggest the following formal definition of an event:
An event is an interaction, associated with a specific point in time, that can be described in a
natural language sentence that has as its subject and object an element of a set of actors and as
its verb an element of a set of actions, the contents of which are transitive verbs. 1
When applied to a specific data source, replace the words "can be" with "is." This definition
encompasses most of what is currently considered to be event data but, unlike the existing
definitions, it can be unambiguously implemented. The key elements are: time, natural language,
actors, and actions.
Time. All event data record a time or period of time when an interaction occurred. The most
common unit used is the day, although in some instances (for example, studying crisis
negotiation) a finer unit such as the hour might be appropriate. Most analyses of event data
aggregate to either months or years.
Natural language. Event data coders do not observe events; they observe reports of events
presented in a natural language such as Arabic, Chinese, English, German, or Hindi. Empirically,
events can only be defined with respect to a human language or set of languages. The event
As a matter of record, while this definition appears to be designed to justify machine coding a priori, it was
actually a consequence of some years of experiments with machine coding and the gradual realization that
interactions that do not meet this criteria are likely to be ambiguous to humans as well as machines.
Schrodt and Gerner
Event Data
Page 3-5
coding exercise converts natural language into nominal data that can be analyzed using formal
methods.
Actors. Any model of political activity will be specific to certain persons, organizations, and
places, all of which are specified by noun phrases in the language or languages used in the source.
Many of these phrases may be synonyms referring to the same actor. For example, in U.S.
political discourse, the former Union of Soviet Socialist Republics might variously be called the
USSR, the Soviets, the Soviet Union, or by the name of its current leader (e.g., Khrushchev,
Brezhnev). For most international relations research, the actors will be political entities;
however, the definition is flexible and allows for the possibility of human-nature interactions if
this is important for the research questions being studied.
Actions. An event coding scheme deals only with certain interactions between actors. All of
these interactions can be described by transitive verbs; for example, apologize, met with,
endorsed, promise, accuse, threaten, or attack. (Transitive verbs are those that can take a direct
object and indirect object. For some events, the second actor is the direct object of the sentence
("Syria accused Israel . . ."); in other cases, it is the indirect object ("Saudi Arabia promised
economic aid to Syria"). As with the nouns, multiple verbs might signify the same category of
behavior, either because the words are synonyms within the language (e.g., grant, bestow,
contribute, donate, fund, present, provide) or because the behaviors, although linguistically
distinct, are politically equivalent, a characteristic that Most and Starr (1989: chapter 5) refer to
as "foreign policy substitutability." These equivalence sets will vary with the specific problem
or the theoretical approach and in large part determine the validity of a particular coding scheme.
One objection that could be raised to the approach of focusing on a standard list of verbs and
actors is that in some unusual situations, the exact wording of an official speech will be an
important point in negotiations. For example, President Richard Nixon's use of the term
"People's Republic of China" rather than "Red China" signaled the beginning of the USA-PRC
dtente in 1970. Similarly, the United States insisted on some very precise wording in the
Schrodt and Gerner
Event Data
Page 3-6
statements by Yasir Arafat in 1988 prior to the re-establishment of U.S. diplomatic contact with
the PLO. However even a negotiation involving specific language generates a large number of
standard event verbs such as accept, reject, inform, confer, and rebuke. Precise language may be
the object of a negotiation, but rarely is it part of the process of negotiation, and even less does it
affect the reporting of the political behavior.
This does not rule out the possibility of focusing on specific phrases that are used with
specialized intent in diplomatic communication. For example, the phrase "frank discussion,"
when used by the U.S. State Department, almost always means that the discussions involved
substantial disagreement. To the extent that these words are used consistently, they can be
coded as signaling disagreement: The phrase "had frank discussions" might code to a WEIS 111
(Turn down proposal) rather than a WEIS 025 (Explain policy), despite the latter interpretation
being closer to the dictionary meaning of the phrase.
The language vary dramatically between sources. Pro-government and anti-government
sources might report the same occurrence using quite different words:
Terrorists slaughtered innocent civilians in the town of Ochos Rios before being driven off by
government troops.
Liberation forces battled occupying forces in the town of Ochos Rios, causing several
casualties before retreating.
Rhetorical flourishes terrorists versus liberation forces, slaughtered versus causing, driven off
versus retreated often say more about the source than they do about the political behavior. In
many cases, such rhetoric is used in a highly styled fashion that can actually improve coding
accuracy, once the patterns are identified and vocabulary lists customized. From either phrasing
of this event one can infer that things were not very quiet in Ochos Rios and that people were
killed. With multiple reports of this type, one can reasonably infer some sort of insurgency.
Similarly, the extensive popular literature on differences in U.S. and Japanese business
negotiating styles (e.g., Hall and Hall, 1987) points out that Japanese negotiators rarely explicitly
reject a proposal (e.g., "We will consider your proposal." means "No.") whereas U.S. negotiators
Schrodt and Gerner
Event Data
Page 3-7
tend to use extremely strong language (e.g., "We can't possibly work with that." means "We're
getting close to an agreement."). To the extent that such phrases are used consistentlyand
unless they are used consistently they have no information value under any circumstancesthe
phrases can be accommodated in any systematic coding framework, although the lists of verbs
may need to vary with the source of the statement. Because those lists are explicit and
reproducible, the source of errors due to misinterpretation can be isolated and identified. In some
cases, the terms used may be so ambiguous as to preclude coding altogether. However, if the
source itself is ambiguous, human coding will also be unreliable and the appropriate action is to
find an alternative source of information.2
3.2. Stochastic Elements in the Measurement of Event

Data
As noted in the introduction to this chapter, event data are neither a random sample nor a
description of a complete population. In the process of going from events on the ground to a
data set that can be analyzed using statistical techniques, a number of different sources of
stochastic variation are introduced. These are summarized in Figure 3.1, and each source will be
discussed in detail in this section. The section will conclude with some suggestions on how some
of these errors could be corrected, at least for the purpose of analyzing aggregate data.
3.2.1. Noise in Event Data
Any event stream contains noise in the form of random events that appear to be endogenous
to the process being studied but which in fact have been generated by other processes or are
irrelevant to the model. Another source of noise is when a source (e.g. Reuters)faced with a
slow news dayreports events that it normally would not report. The distinction between
For example, one can imagine a situation where an area had a substantial amount of economically motivated
banditry as well as politically motivated guerrilla activity but where the government press referred to all violent
activity as the product of "bandits" as a way to delegitimize the guerrillas.

Schrodt and Gerner
Event Data
Page 3-8
endogenous and exogenous events depends on the theory being studied. What is noise to one
model may be the signal of another, just as the variation in growth of a soybean plant due to
insect damage is noise in a study of fertilizers but signal to a study of insecticides.
True
Events
Reported
Events
Noise
Events
Events in
system
Censoring
Source
text
Assign
codes to
text
Misclassification
errors
Coded
Events
Final data
set
Figure 3.1. The Process of Generating Event Data

In correlational studies, noise is usually assumed to be normally distributed. In an event data
study, the default model for noise would be statistical independence and a Poisson distribution
since the Poisson is the only temporal probability distribution that is memory-free. Ideally, one
could ascertain the Poisson intensity parameter for the distribution of noise affecting each event
code and use that information in statistical studies, just as the mean and variance of normally
distributed errors are used in correlational studies. As in correlational studies, the actual noise
will often not be distributed according to the assumptions of the model, and might exhibit nonPoisson behavior such as cyclicity or statistical interdependence.
3.2.2. Censoring
Censoring means that an event occurs in the system and does not appear in the data. The
term is meant in the statistical, rather than political, sense: while overt censoring of information is
certainly a factor in event data, far more problematic are the editing and coverage biases
Schrodt and Gerner

Event Data
Page 3-9
introduced in journalistic and historical sources; these have been discussed exhaustively in the
event data literature.
Censoring occurs nonuniformly. Any event data set has a vector dC (i.e. a vector indexed on
the set of classification codes C)3 that is the ratio of the frequency of codes in the observed data
set to their frequency in an ideal data set where all events occurring in the system were reported.
These ratios can be quite lowit is unlikely that existing data sets capture more than a few
percent of all political activity except for extreme events such as the outbreak of war. To the
extent that some events are more important than others in determining international behavior,
censoring is probably inversely proportional to importance: the more important an event, the
more likely it will be reported.
The converse of censoring is "disinformation": strings in the source text concerning events
that did not actually occur. These might be introduced deliberatelyfor example the deception
campaigns that preceded the US ground offensive against Iraq in 1991but they are more
commonly generated by rumors and second-hand information. While disinformation from
deception and rumor is probably not a major component of event data, it is more problematic
than pure noise because it is very non-random and is specifically designed to make the system
appear as though it is operating under a different process than that actually occurring.
In particular, rumors must convey plausible patterns of human behavior or they will not be
voluntarily transmitted. For example, in the week following the June 1989 Tiananmen Square
massacre in Beijing, many rumors circulated about pro-democracy military forces preparing to
move against the city. These rumors eventually proved to be completely groundless but were
credible: when a comparable situation occurred in Romania in December 1989, military units did
turn against the government. The nature of rumors and story-telling means that a credible
sequence of events is generated; if information is missing from the story, it will be provided in a
In Figure 1, d is placed prior to the source text because that is where most censoring occurs in the real world
events occur but are not reported. For reasons that will be clear momentarily, it is more convenient to index d
on event codes rather than original events.

Schrodt and Gerner
Event Data
Page 3-10
fashion that makes sense to the teller and listener, rather than generated randomly. Rumors, as a
consequence, may be more likely to fit models of political behavior than will actual events. Note
that if regularity did not exist in the international system, disinformation and strategic deception
could not exist since the entity being deceived must be able to fit the information to a pre-existing
model of international behavior.
3.2.3. Misclassification
Misclassification occurs when the code assigned to a text string does not correspond to the
code that was intended when the coding scheme was designed. Any event data set has an error
matrix M where eij gives the probability of misclassifying vj as vi. This error matrix incorporates
errors due to both validity and reliability problems.
The coding errors generated by human and machine coding are quite different. Most human
coding errors either involve relatively subtle differences in behavior (e.g. deciding whether a
negative statement should be coded into WEIS's "comment", "reject" or "protest" categories) or
missing the event altogether. As a consequence, in human coding M has a block structureone
can arrange the row and columns of M is a fashion where non-zero entries are most likely to
occur in adjacent cells with the remainder of the matrix being zero: A trade agreement may be
confused with a cultural exchange but it is unlikely to be confused with a war.
(The single most common human coding error, however, is a form of censoring rather than
misclassification: human coders tend to miss some of the dyadic events in a sentence. For
example the simple phrase "Representatives from France and Russia will visit Baghdad and
Teheran next week" generates eight WEIS eventsfour dyadic visits (WEIS 032) and four dyadic
hostings (WEIS 033)and human coders tend to miss some of the combinations. Humans are
also more likely than machines to miss events in compound sentences.)
Machine coding, in contrast, will occasionally generate event codes that are completely
implausible (uses of force between allies) due to sentences that have unusual grammatical
constructions or that use metaphors mistakenly interpreted as events. While these are relatively
rare and can usually be avoided with an appropriate filter, the statistical techniques appropriate
Schrodt and Gerner
Event Data
Page 3-11
for machine-coded data may be different from those appropriate for human-coded data. In
machine coding, the structure of M generally varies with the event category.
For example, the widespread use of military metaphors to describe conflictual but nonviolent
effects (e.g. "At the United Nations, Iraq's ambassador blasted the U.S. policy on continued
sanctions" versus "Israeli warplanes blasted suspected Hizballah positions in southern Lebanon")
means that the WEIS force code is often erroneously coded. The WEIS accuse and deny
categories, in contrast, are almost never incorrectly miscoded, and typically the word used in the
news report will be "accuse" or "deny." The major WEIS category promise is usually correctly
identified, but the subcategory may by incorrectly specified if the dictionaries do not have
contain the appropriate direct object (for example "promised to send emergency aid" is in our
dictionaries; "promised to send trained dogs to find earthquake victims" is not).
When machine coding is used, the rate of classification errors can be substantially reduced by
the use of simple filters that tell the program to skip sentences that appear too complex or do not
have sufficient information to code correctly; this process of "complexity filtering" is discussed
in Chapter 2. For example a sentence such as
Prince Sadruddin Aga Khan, a veteran troubleshooter, was named on Tuesday to oversee all
U.N humanitarian operations in Iraq and Kuwait, and on Iraq's borders with Turkey and Iran,
where hundreds of thousands of refugees have fled in recent weeks.
contains references to five different actors and at least three verb phrases is much more likely to
be incorrectly coded (by a machine or a human) than a simple "Syria accused the United States
of..." sentence.
3.2.4. Schematic Error
A final source of classification error is "schematic error." This occurs when the coding
system combines two sets of behavior that have distinct natural-language representationsand
which should remain distinct for theoretical or analytical reasonsinto a single category. The
system may also do the oppositeseparate two sets of behavior that could be combinedas in
WEIS's notoriously overlapping warn and threaten categories.
Schrodt and Gerner
Event Data
Page 3-12
Event structures allow multiple events to serve the same function. Coding errors within these
substitution sets will have no effect on the fit of the structure, since all events within the set are
equivalent. The existing coding schemes of COPDAB and WEIS implicitly use a high degree of
substitution, since they map many distinct text strings into the same event code. However, the
substitution mapping is done at the coding stage rather than the modeling stage.
If one were dealing with a set of models where the substitution sets were always the sameif
a pair of events v1 and v2 were found in one substitution set they would always be found in a
substitution set whenever one or the other were presentthen the problem of determining the
details of a coding scheme would be solved. Again, existing event sets implicitly do this already.
However, the substitution sets probably vary across models. For example, the event
[move an aircraft carrier from the Mediterranean to the Gulf]
and the event
[move an aircraft carrier from the Pacific Ocean to the Gulf]
are roughly equivalent as far as US relations with Iraq are concerned, but have very different
implications for relations between the US and Korea. The specific event structure being
analyzedthe context of the eventis important.
Schematic error also occurs when certain event types are ignored altogether. WEIS and the
international scale of COPDAB were designed for the coding of inter-state events, and implicitly
focused on the militarized conflicts of the Cold War. Trade was considered routine and not
coded, but even trade disputes received little attention. Several categories of politically salient
activity that have become very important in the late- and post-Cold War periodfor example
refugee flows and human rights violationsdo not have separate codes, nor, as the IDEA project
has pointed out, are several types of general political activity such as the adjudication of a
dispute.
Schrodt and Gerner

Event Data
Page 3-13
3.2.5. Statistical Corrections for Coding and Censoring Errors

The overall error structure of an event data set can be specified by combining the elements of
noise, censoring and misclassification. Let r be the true frequency of the codesthe frequency
generated from an ideal uncensored data source coded using E without misclassification errors.
Let n be the frequency of the noise (both random noise and disinformation); let
D = diag(dC)
that is, the matrix with the elements of dC on the diagonal. Then x, the observed frequency of
events in the data set, is
x = MD (r + n)
Using this relationship, it is possible to make several statistical corrections to an event set if
one is interested only in the aggregate frequency of events; frequency is the key concern is most
correlational and descriptive studies.
Assume that the misclassification matrix M can be estimated and let t = (r + n) be the true
event frequency vector. If there is no censoring, then the observed frequency vector x is simply
x =Mt
One can correct for misclassification and get an improved estimate of t by adjusting x using
x* = M-1x
Under certain circumstances, M is very straightforward to compute. Suppose one has two
coding systems, one very slow but accurate (e.g. coding by a principal investigator or well-trained
and well-motivated graduate student coders), and the other fast but less accurate (e.g. machine
coding or poorly trained, supervised or motivated work study students). Assuming that both
coding processes are consistent, then M can be estimated by comparing two coding results on a
suitably large and representative set of texts.
If the censoring vector d were also known, the correction can be extended further:
x* = (MD)-1x
d is less likely to be known with any degree of confidence than M, although efforts could be made
to approximate it by comparing multiple sources.
Schrodt and Gerner
Event Data
Page 3-14
The standard technique for estimating the size of an unknown population is to sample the
population twice and compare the number of cases captured in both samples4. Let N be the true
population size, n1 and n2 the sizes of two independent random samples from N, and m the
number of cases that occur in both random samples. The probability of a case being in sample 1
is p 1 = n1 N; probability of a case being in sample 2 is p2 = n2 N, so
n1 n2
m = p 1 p 2 N N =
m
Once N is known, the number of cases being censored can be estimated by comparing the number
of events found in a source to the estimated population size N.
The weakness in this approach is the requirement that the two samples be random. All text
sources of event data are biased to report certain events while ignoring others, rather than
randomly sampling from all possible events. However, comparing two sources that are
attempting to provide equivalent coveragefor example the New York Times and the Los Angeles
Timeswould provide a rough estimate of the censoring probabilities. This technique could also
be used to ascertain which event categories are more frequently censored.
3.3. Interval vs. Discrete Approaches to Event Data

Analysis
As noted in Chapter 1, Charles McClelland's early work assumed that event data would be
used for the systematic study of sequences of behaviors. McClelland (1970:6) notes, in the
quotation that opens Chapter 1, that event data could form a bridge between the then-prevalent
general systems theories of international behavior, and the understanding of political behavior
through textual history: a starting point [for event data research] is provided as readily by the
ordering principle of classical diplomatic history as by the basic concepts of general system
analysis.
This method is typically used to estimate fish or insect populations; it has also been used to estimate the
undercount in the U.S. Census.

Schrodt and Gerner
Event Data
Page 3-15
In McClelland's assessment, however, this transition away from the general systems
approachwhich was rooted in continuous-time dynamics and interval-level variablesfailed.
After some years of work with event data focusing on several crises, he concluded:
It proved relatively easy to discern event patterns and sequences intuitively. We found we
could follow the successions of action and response in flow diagram form. Stages of crisis
and the linkage of event types to temporary status quo situations also were amenable to
investigation. We were defeated, however, in the attempt to categorize and measure event
sequences. This was an early expectation that was disappointed by the data which showed
too few significant sequences to support quantitative or systematic treatment. (McClelland,
1970:33)
As a consequence of this failure, McClelland's "World News Index" project, published in the
mid-1970s, used interval-level variables in its measures. With the hindsight of two decades, the
failure of a discrete event approach appears due to a paucity of data and processing capability.
McClelland writes of analyzing hundreds or at most thousands of events; a contemporary event
data researcher has available tens of thousands of events and computer power sufficient to work
with millions.
After this early definition of international politics as event sequences, the field of quantitative
IR moved rapidly to analyzing events with interval-level techniques. This change was probably
due to the general shift in quantitative IR in the late 1960s away from historical approaches
towards theories based on the model of the physical sciences and economics. By the 1970s a
Kuhnian split was underway in international relations with the traditionalist and behavioralist
camps proudly speaking totally different languages, whereas when McClelland's 1961 World
Politics article was written, this split was not apparent. For example, Rummel (1972) proposes
a science of international politics similar to meteorology, using interval-level metaphors such as
field theory and interval-level techniques such as factor analysis. Azar, while using WEIS as the
model for COPDAB, abandoned McClelland's nominal categories in favor of an interval-level
measure and approached coding as a scaling problem. Azar and Sloan (1975) consists entirely of
interval-level data and Azar emphasizes
Schrodt and Gerner
Event Data
Page 3-16
quantitative aggregations, called here 'analytic data', [which] are summaries of the weighted
frequencies of interactions. They describe the amount of conflict or cooperation exchanged
between or within nation-states over some unit of time. (Azar 1980:150)
This conversion of discrete entities to interval-level data is somewhat puzzling from a
statistical standpoint. The reason probably is due as much to paradigmatic developments in
quantitative international relations as in the nature of the data. During the 1970s, data-based
studies of international behavior saw the ascendancy of correlational analyses, particularly
regression. The mathematics behind these techniques had been fully developed by
econometricians and could be easily applied to international relations data using SPSS and other
statistical packages; comparable tools were not available for sequence analysis. The successful
formal theories were continuous-variable models such as the Richardson arms race model and
DYNAMO-like global models. Rational choice models preserved some discrete variables,
particularly in game theory, but even these models used continuous variables in expected utility
calculations5.
The emphasis on crisis, initiated by McClelland and expanded in a large event-based crisis
management literature (e.g. Hoople, Andriole and Freedy 1984, Azar et al. 1977) also contributed
to the emphasis on interval-level variables. Implicit in most of the crisis models is either a simple
distinction between crisis and non-crisis or a unidimensional ordinal set of "steps" to crisis (e.g.
Hoople, 1984). Crisis forecasting is reduced to a problem of monitoring some activityusually
some aggregated measureto ascertain when the system is going to change states. The crisis
literature also frequently emphases the concept of the "intensity" of events, another interval
measure.
The advantage of this approach is that a wide variety of methods are readily available. The
clear disadvantage is that the process of reducing behavior to a single dimension through scaling
Another factor favoring continuous formulations may be the ease with which one can draw a line representing a
continuous variable and call it a "pattern" (for example Brody, 1972) whereas discrete sequences are more
difficult to visualize and explain (McClelland, 1961).

Schrodt and Gerner
Event Data
Page 3-17
loses a great deal of information and introduces a large number of free parameters (the issue of the
choice of weights will be discussed in detail below). For example in principle (although almost
never in practice), a month characterized by a large amount of conflict in the first two weeks
(negative numbers on most scales), followed by a large amount of reconciliation in the last two
weeks (positive numbers) could aggregate to value close to zero, which is the same value that
would occur in a month where nothing happened.
A second, more subtle, problem occurs with aggregation: it removes the analysis a step
further from the cognitive and organizational processes that are generating the events. While
decision-makers may do some aggregationone of the most commonly used metaphors in
political analysis is indicating whether a situation is "heating up" or "cooling down"detailed
political responses are usually triggered by specific sets or sequences of events, not by the
crossing of some numerical threshold.
In political activity, unlike economic activity, both the stimuli and responses are likely to be
discrete, not continuous. Prices of stocks or the levels of interest rates, for example, move in
predictable adjustments and when they fail to move continuously across that range (as in an
investigation of NASDAQ trading a couple years ago), suspicions are triggered. Furthermore,
small changes in the price will almost always result in proportionally small changes of supply
and demand.
Political events, in contrast, move in jumps that are predicated on the prior state of the
system. The fall of a single rocket following a period of peace will trigger a major response,
whereas the fall of a single rocket during a period of war usually will go unnoticed. A model that
can maintain the event data in its disaggregated form is, ceteris paribus, more likely to be
successful in predicting actual behavior.
Despite these disadvantages, one of the most common techniques used in event data analysis
is to aggregate events over timetypically by week, month or yearusing a numerical scale.
This changes the data set from an irregular, nominal-level time series to a regular, interval-level
time series. An event data scale assigns a numerical value to each event category found in the
Schrodt and Gerner
Event Data
Page 3-18
coding scheme. Table 3.1 shows subsets of the Azar-Sloan (1976) and Goldstein (1992) scales,
which apply to the COPDAB and WEIS coding schemes respectively; the full scales can be
found in the event coding appendices of this volume.
Table 3.1. Examples of Event Scales

Azar-Sloan Scale
COPDAB Category (Azar 1982)
COPDAB Code
Scale value
Military, economic and strategic support
31
Mild verbal support; exchanges of minor officials
Diplomatic-economic hostile actions
11
29
Full scale war
15
102
WEIS Code
Scale value
Goldstein Scale
WEIS Category
Praise
41
3.4
Promise Policy Support
51
4.5
Extend Military Aid
72
8.3
Criticize
121
-2.2
Ultimatum
174
-6.9
Military Engagement
223
-10.0
The Azar-Sloan and Goldstein scales are the systems most commonly used in the literature,
but other systems exist; their development is summarized in Goldstein (1992:373-374). The
Goldstein scale has been used in a number of recent studies, including Huxtable and Pevehouse
1996, Goldstein and Pevehouse 1997; Schrodt and Gerner 1997, 1998, Bond et al. 1997; Kinsella
1995, 1998; Reuveny and Kang 1996a, 1996b. Prior to this, Vincent's (1990) scale for
aggregating WEIS events was used in a number of studies, most notably Goldstein and Freeman's
(1990) book-length study of superpower interactions.
All of these scales have been constructed by querying panels of experts about the relative
intensity of various event categories. In most cases, the weights have been assigned on a single
Schrodt and Gerner
Event Data
Page 3-19
cooperation-conflict dimension. Beyond this, there is no consistency in the scalingfor example

the Goldstein weights are roughly proportional to the logarithm of the Azar-Sloan weights.
The weighting schemes appear to work fairly well. While in principle the uni-dimensional
cooperation-to-conflict scaling should be problematicfor example the USA-Canada or USAJapan relationships are characterized by high levels of both cooperation and political conflictin
practice this hasn't prevented the scaled data from being used successfully in a variety of studies.
This may be due in part to the fact that event data have been primarily employed to study highly
conflictual situations such as the Cold War (Ashley 1980; Goldstein and Freeman 1990; Dixon
1986) and the Middle East (Azar 1972; Azar et al. 1979; Schrodt and Gerner 1997, 1998) where
"cooperation" is largely expressed as a reduction of conflict.
Nonetheless, there are a couple of clear problems with scaling. First, aggregating events is
controversial: the "folk criticism"6 of the Azar-Sloan scale is "3 riots equals a nuclear war." This
debate goes back to the earliest event data discussions (e.g. Azar and Ben-Dak 1975; Azar, Brody
and McClelland 1972) and has continued over time: see exchanges between Howell (1983) and
McClelland (1983) or Vincent (1990) and Dixon (1990). Second, the assignment of weights by
panels of experts is arbitrary, atheoretical and detached from any specific empirical context. For
example, why should the same set of weights should apply to a dispute such as Israel-Lebanon,
where military exchanges are very common, and USA-USSR, where military exchanges were
virtually nonexistent? Finally, it is unclear how much effect the choice of a particular weighting
scheme has on the results of the studyminor differences in weights could lead to major
differences in results.
The next two subsections report the results of three sets of experiments we have done on the
effects of changing the scaling and aggregation in event data. These experiments were done in the
context of substantive research, rather than as basic experiments on event data in general, but
As in "folk theorem": we've heard this phrase many times over the years but have no idea who originated it. The
Azar-Sloan value for "inciting of riots" (COPDAB category 12) is equal to 44; "full-scale war" is 102.
Schrodt and Gerner
Event Data
Page 3-20
consistently show that event data seem to be relatively insensitive to these changes. This would
help explain why coding and scaling systems such as the Goldstein scale and the WEIS categories
have been effective for political analysis despite their rather ad hoc nature, and also provides
some indication of the relative importance of censoring, coding, and schematic errors discussed in
Section 3.1.
3.3.1. The Effects of Simplifying Scales
In this subsection, we will examine the effect of simplified weighting schemes on the
delineation of phases of political behavior in the Middle East (1979-1995) is done in Chapter 4,
using the Goldstein (1992) scale as the reference point.7 The weights will be progressively
simplified to give decreasing levels of differentiation between the event categories. These
simplified scales are shown in Table 3.2; in the analysis the data were reduced to monthly
aggregations using six different systems that derive progressively less information from the daily
events.. By comparing the results of these various weighting schemes we can evaluate the extent
that the results of an analysis are dependent on the choice of a particular weighting scheme.
Table 3.2. Alternative Scales
Goldstein:
Goldstein weights
difference:
cooperative events = 1; conflictual events = -1.
total:
all events = 1.
conflict:
cooperative event = 0; conflictual events = 1.
cooperation:
cooperative event = 1; conflictual events = 0.
report:
1 if any event was reported in the month, 0 otherwise
The details of the analytical method will not be repeated here; these are discussed in the various sections of
Chapters 4 and 6.
Schrodt and Gerner
Event Data
Page 3-21
Table 3.3 reports the effects of these alternative weights on a discriminant analysis for the
assignment of crisis phase (see Chapter 4, section 4.3). The results are rather striking: There is
almost no difference between the weighting systems: all behave almost identically in various
measures of the discriminant analysis. In fact, the Goldstein scale is actually the least effective
weighting system for differentiating between the a priori clusters of behavior. The most effective
measures seem to be the number of cooperative events, and the total number of events, though
the differences in classification accuracy are not large.
Table 3.3. Discriminant Analysis Results

Weighting
scheme
%correct(1)
variance
explained(2)
canonical
corr.(3)
Wilks' signif
# factors
Goldstein
85.6%
76.3%
0.85
0.008
<.001
difference
89.7%
74.7%
0.85
0.007
<.001
total
94.4%
83.0%
0.93
0.001
<.001
conflict
88.2%
76.9%
0.86
0.007
<.001
cooperation
92.3%
82.2%
0.91
0.002
<.001
report
89.2%
73.6%
0.87
0.008
<.001
random date(4)
61.0%
69.5%
0.66
0.131
.37
random dyad(5)
57.4%
68.8%
0.67
0.119
.18
Notes:
1. Percentage of the cases classified correctly.
2. Variance explained by the first three factors of the discriminant analysis.
3. Canonical correlation for the first discriminant function. The canonical correlation is equivalent to the Pearson
product moment of the value of the discriminant function regressed on a set of dummy variables representing
each of the categories in the data set
4. This set used the Goldstein weights but randomized the order of the dates in the data set. (4)(5)4. is, the data
had no correspondence to the actual political phases, but the5. isd the information in eac to the patterns observed
in a given month.
5. This set randomly ordered the information in each dyads so that the marginal characteristics of the data set remain
the same, but the data are randomly order and have no relation to either the actual phase or to each other.
Schrodt and Gerner
Event Data
Page 3-22
The discriminant analysis on the two random data sets produce quite different results from
the analysis of the actual events. The random data have almost a 70% accuracy rate in phase
classification, probably because of the contrast between the very high dimension (54) of the
variable space and the relatively small number of clusters being classified (7). However, the
Wilks' measurement correctly adjusts for thisdespite the high classification accuracy,
shows that the none of the discriminant functions are statistically significant. 8
A second set of tests was done with using the LML clustering measure discussed in Chapter
4, Section 4.5. Figure 3.2 shows the cluster boundaries generated by each of the weighting
vectors, using the LML difference level =0.15 (except the Goldstein weight, where =0.309).
Table 3.4 shows the number of cases where the boundaries determined by the simplified vectors
are within 3 months of the a priori boundaries discussed in Chapter 4, Table 4.1, and the
Goldstein boundaries.
These results are generally similar to those found in the discriminant analysis: in general the
simplified weights do as well or better than the Goldstein scale in matching the a priori
boundaries, and there is little difference in the results obtained with the various simplified
weighting schemes. The "total" vectorwhere all event types are weighted equallyalso has a
high level of correspondence with the divisions found by the other methods; the "difference"
variation on this does as well in matching the a priori transitions but (unsurprisingly)
corresponds more closely to the Goldstein divisions than does the constant vector. In contrast to
the discriminant analysis, the vector with the least correspondence to the other transitions
This is an instructive example of why inferential statistical methods, rather than gross classification accuracy, is
needed to evaluate these models.
The higher value of in the Goldstein weights is a result of those weights being between -10.0 and 10.0, whereas
the remaining weights are between -1 and 1. Because the high number of zero values in the vectors more or less
forces the regression line through zero, the data are effectively not invariant to a change of scale despite the use of
a correlation measure, and the higher variance of the Goldstein scale leads to higher correlation values.
Schrodt and Gerner
Event Data
Page 3-23
considers only the cooperative events; in particular this has almost no correspondence with the a
priori transitions.
Goldstein
difference
total
cooperation
conflict
report
Figure 3.2. Cluster boundaries under various weighting systems
Table 3.4. Correspondence between the boundaries generated by default vectors and
other techniques
Genetic
(N=12)
Goldstein
(N=11)
a priori
(N=6)
difference
total
cooperation
conflict
report
Goldstein
--
Schrodt and Gerner

Event Data
Page 3-24
3.3.2. Optimizing the Scale

The Goldstein scale was not designed for the purpose of clustering political behavior, and it is
possible that some other set of weights might be superior for this purpose. To assess this
possibility, optimal weights for the LML clustering procedure were estimated using a genetic
algorithm (GA)described in Appendix 3.Athat maximized the following clustering measure:
Fc =
average distance between adjacent clusters

average distance within clusters
where "distance" is defined by the correlation metric and the "average distance" is calculated as
the average distance between points. Fc is similar to the F-ratio maximized in discriminant
analysis except that only the distance between adjacent clusters is considered and measure uses
the total distance between points, rather than group variances.
The GA works reliably and most of the experiments produced similar sets of cluster
boundaries. The boundaries are, however, dependent on the value of in the LML criterion:
Higher values of consistently produce higher F c values, but fewer clusters, because of the
stricter threshold for establishing a new cluster.
The results of a number of different experiments with the genetic algorithm are shown in
Figure 3.3. The thick lines above the X-axis are a histogram of the cluster boundaries identified
by 100 GA runs for =0.15; Fc in these runs ranged from 1.70 to 1.51 compared to Fc = 1.08 for
the Goldstein weights.10 The solid squares on or near some of these lines show the number of
cluster boundaries found for 25 GA runs with =0.25. The solid triangles just below the X-axis
show the location of the cluster boundaries generated by the Goldstein weights for =0.30. As
indicated in Table 2, the relationship between the simplified weighting schemes and the divisions
10
For =0.25 the range of Fc for 25 GA experiments is 2.35 to 1.90; for the Goldstein scale Fc = 1.30. By
comparison, the Fc values for the other weighting systems are difference = 1.18; total = 1.46; cooperation =
1.26; conflict = 1.29; report = 1.42.

Schrodt and Gerner
Event Data
Page 3-25
found by the GA are generally similar to the relationship between those divisions and the
Goldstein divisions.
Figure 3.3. Cluster boundaries under genetic algorithm optimization

While the cluster boundaries determined by the GA were quite consistent across the various
experiments this was not true of the scale weights generating those boundaries: in fact there was
no consistent patterns whatsoever in those values. The distribution of the 4,950 correlations
between the weight vectors generated by the experiments is generally Normal with a mean of
0.077 and standard deviation of 0.307; the number of significant correlations do not exceed the
number expected by chance. The average correlation of the GA weights with the Goldstein
weights is r=-0.008 with a standard deviation of 0.188, and there is virtually no difference in the
average correlation within the best-fitting 50% of the vectors (r=-0.006) and within the remaining
50% (r=-0.009).
Examination of the intermediate results produced by the GA showed that despite this
diversity, the GA is working correctly to generate and select vectors that increase the value of Fc;
there are simply a lot of different ways to do this. The lack of convergence of the weight vectors
is probably due to a problem comparable to the effect that collinearity in a linear model has on
Schrodt and Gerner
Event Data
Page 3-26
increasing the variance of the estimates of the parameters of a regression equations: Because the
correlation distance is invariant with respect to a linear transformation of the weights, very
different sets of weights can be used to produce essentially the same distances. This is
consistent with the fact that there is relatively little variance in the delineation of the clusters but
huge variance in the values of the weights producing those clusters. Consequently, producing an
"ideal" set of weights via an estimation procedureexpert panel, genetic algorithm, or
otherwiseis probably a hopeless task.
A genetic algorithm was also used to maximize the separation of the clusters, given the a
priori cluster boundaries in Table 4.1 in Chapter 4. The purpose of this experiment was to see
whether it would be possible to find a better set of weights than those provided by Goldstein for
that set of cluster transitions.11 Twenty experiments were done, with the GA allowed to run for
128 generations.
The results of this experiment differed substantially from the experiments where both the
weights and cluster boundaries are allowed to vary. In particular, there is a significant correlation
at the 0.01 level between the weight vectors in about 30% of the cases. None of those weight
vectors, however, correlate significantly with the vector of Goldstein weights. Furthermore, the
11
We also tried to estimate optimal weights in a linear fashion using discriminant analysis, where the independent
variables were the WEIS event counts (by 2-digit category) totaled across all of the dyads by month. While this
technique removes the information on which dyads were interactingto differentiate the dyads and event
categories would involve 1188 variables and we have only 208 data pointsthis method could provide a rough
estimate for the event weightings appropriate for the full system.
The results were generally disappointing. The classification accuracy with all variables was only 73%, and
the first three discriminant functions explained only 75% of the variance. There was no discernible pattern to the
weights or functions. When stepwise discriminant was used the accuracy dropped to 60% but the variables
chosen tended to be those with a high density of events: 01 (Yield), 02 (Comment), 03 (Meet), 11 (Reject), 12
(Accuse), 13 (Protest), 21 (Seize) and 22 (Force). The discriminant weights mirrored those of the Goldstein
scale to some extent: 01 = 0.65, 02 = 0.70, 03 = 0.23, 11 = -0.48, 12 = -0.09, 13 = 0.22, 21 = -0.32 and 22 =
-0.83.
Schrodt and Gerner
Event Data
Page 3-27
value of Fc for the optimized vectors is 1.58 while the Fc value for the constant vector is 1.54, so
the optimization provides very little improvement in the separation of the clusters.
The variance of the weight in many of the optimized vectors is quite small: Half have a
standard deviation between 0.5 and 1.0; the remainder have a standard deviation between 5.0 and
7.0 (the Goldstein weights have a standard deviation of 4.34). This bi-modal distribution is
entirely a function of whether or not the vector contains both positive and negative weights: the
low-variance vectors have only positive weights.
The high correlation between these optimal vectors and the constant vector suggested one
additional experiment: computing the distance between points by correlating the frequencies of
the 2-digit WEIS events without applying any weighting (in principle this method could also be
applied to 3-digit categories). The LML measure was computed as before, with the only
difference being that the correlation was computed on vectors containing counts of the twentytwo 2-digit WEIS events for each dyad-month, so each correlation used 22 x 56 = 1232 points
rather than the 56 points of the vectors containing the Goldstein scores.
Figure 3.4 shows the LML t curves from this analysis, and Figure 3.5 shows the cluster
boundaries. In general, the event count measure produces results quite similar to those of the
Goldstein measurethe correlation (r) between the Goldstein LMLt and event count LML t is
0.63 particularly in terms of matching the a priori cluster boundaries. (The =0.15 threshold is
used for the event count boundaries, which in part explains the higher number of clusters for that
measure.) Once again this suggests that the clustering method is not strongly dependent on the
Goldstein weights, and the frequency of coded events alone is the primary factor that
differentiates the major political features of the data. In another experiment that parallels the
"report" scale used in subsection 3.3.1, LML curves were computed for a data set that replaced
all event counts that were greater than zero with a value of 1in other words, a set that
measured only the presence of events rather than their quantity. This produces credible cluster
breaksfor example it correlates at the 0.78 level with the LML curve from the true data set
Schrodt and Gerner

Event Data
Page 3-28
when the constant vector is usedthough unsurprisingly the variation of the curve is attenuated
compared to that produced by the actual data.
0.4
0.3
0.2
0.1
0
-0.1
-0.2
-0.3
0.4
Event Count
Goldstein
Figure 3.4. Comparison of LML for the event count and Goldstein weights
Figure 3.5. Cluster breaks for the event count and Goldstein weights
3.3.3. Simplifying the Nominal Coding Scheme

The final experiment involves simplifying the coding system at the level of categories. This
will use the hidden Markov model technique discussed in Chapter 6, which uses disaggregated
Schrodt and Gerner
Event Data
Page 3-29
nominal variables rather than scaled data. The substantive problem involved forecasting conflict
in the former Yugoslavia for the period 1990-1999.
The original motivation for this experiment was, in fact, pragmatic. Hidden Markov models
(HMM) involve a large number of parameters and are estimated with a numerical optimization
technique that produces a large number of local maxima. The models had originally been
estimated using the 22 WEIS cue categories (plus a non-event code), but the resulting
parameter estimates had come nowhere near converging to a single set of estimates, and were
proving very difficult to interpret.
Based on the earlier experiments with the discriminant and LML analyses on the Middle
East, it seemed possible that this lack of convergence might be due to the model trying to derive
too much detail from the data. For example, if any incidences of violence in the region resulted in
events distributed across the WEIS violence categories expel, seize and force, and
likewise cooperation involved a combination of meetings, promises, agreements and requests,
then attempting to draw subtle distinctions between these categories would be futile. While
estimation of an HMM uses a non-linear method, one likely result of this co-occurrence of event
types would be an effect similar to that of co-linearity in linear regression: the standard errors of
the parameter estimates would increase substantially, and models with very similar degrees of fit
might have quite different parameters, as was being found. The models were therefore reestimated using the following five-category system:
0. Non-event
1. Verbal cooperation (WEIS categories 02, 03, 04, 05, 08, 09,10)
2. Material cooperation (WEIS categories 01, 06, 07)
3. Verbal conflict (WEIS categories 11, 12, 13, 14, 15, 16, 17)
4. Material conflict (WEIS categories 18, 19, 20, 21, 22)
This reduced the total number of parameters in the model by about a factor of 5. It is also likely
to reduce the effect of coding variance and coding error somewhat: Several of the verbal
Schrodt and Gerner
Event Data
Page 3-30
conflict codes in WEIS are ambiguous even for human coders, and the automated coding
probably generates some misclassification in those categories.
The results of this experiment are given in Table 3.5; the rows refer to various combinations
of optimization weight (P vs. N) estimated at 1, 3 and 6-month forecast lags (see Chapter 6). 32
Monte-Carlo genetic algorithm estimates were done for each set of experimental parameters. For
purposes of comparison, Table 3.5a presents the results for the 23 event-category model; Table
3.5b shows the results for the 5 event-category model in the same format.
Table 3.5a. Accuracy for 23-Category Coding System

Experiment
%accuracy
%high
correct
%low
correct
%high
forecast
%low
forecast
P1
77.6
29.3
89.5
40.8
83.7
P3
76.0
29.0
87.9
37.9
82.9
P6
76.9
25.9
90.6
42.6
82.0
N1
54.2
92.7
45.3
28.1
96.4
N3
49.0
88.1
39.6
25.9
93.3
N6
47.7
88.5
37.4
26.3
92.8
Table 3.5b. Accuracy for 5-Category Coding System

Experiment
%accuracy
%high
correct
%low
correct
%high
forecast
%low
forecast
P1
74.4
46.2
81.5
38.9
85.6
P3
71.7
44.1
78.9
35.4
84.4
P6
71.4
44.2
78.8
36.4
83.8
N1
61.9
90.7
54.6
33.7
95.8
N3
57.8
87.0
50.2
31.4
93.6
N6
56.8
85.9
48.8
31.5
92.7
Schrodt and Gerner

Event Data
Page 3-31
In general, the results of the 5-category analysis are comparable to those of the 23-cateogry
analysis. In both schemes, the drop-off in accuracy with the increasing forecasting lag is
smallabout 4% from the 1-month to 6-month forecast lagthough consistently there is a small
decrease. The overall accuracy decreases about 4% for the P models and increases about 8% for
the N models. The largest difference in the results occurs with respect to the accuracy of the
high-conflict predictions in the P modelsthese average about 18% better in the percentage of
the observed high week that were correctly forecast, albeit at the cost of an 8% decrease in the
corresponding percentage of the observed low weeks that were correctly forecast. The N model
shows an 11% increases in the percentage of the observed low weeks that were correctly forecast
and a 5% increase in the percentage of forecasts of high conflict that actually had high conflict.
All of the remaining statistics differ from the original model by less than 3%.
This analysis clearly supports the hypothesis that the use of simplified event coding systems
at worst involves only a small penalty in terms of predictive accuracy, and at best can actually
improve the accuracy, probably through the reduction of noise. This is particularly important
when automated coding is being used, since automated coding is generally less capable of making
subtle distinctions between event categories, but generally is quite good at making large
distinctions such as the difference between cooperative and conflictual behavior.
(Meanwhile, the intuitive insight that motivated this experiment did not prove to be correct:
The simplified coding system reduced the variance of the HMM parameter estimates somewhat,
but not to the point where these converged to a single, readily-interpreted set of estimates. This
could either be due to the collinearity remaining in the datawhich looked at the activity of four
actors in this subsystemor to intrinsic indetermancy of HMM parameter estimates.)
3.3.4. Analysis
Based on these experiments, it appears that most of the information that can be used to
differentiate political behavior is found in the event counts themselves, rather than in the detailed
classification or weighting of events. This could be due to at least two factors.
Schrodt and Gerner
Event Data
Page 3-32
First, almost all event data contain a large number of time points where no activity has been
reported; these non-events are unaffected by any change in the weighting or classification
schemes. Around 50% of the dyad-months in the Levant data set have zero values, as are about
50% of the dyad-days in the most intensely-active dyad in the Balkans data set, Serbia=>Bosnia.
Second, the existence of any activity in a dyad may be a signal that Reuters reporters or
editors think that the political behavior in the dyad is important: this is particularly true with
respect to verbal activities where Reuters has often has an option of reporting the activity (in
contrast to physical actions such as demonstrations and military clashes, which are reported
more frequently).
This in turn means that results of event data analyses are likely to be very robust, rather than
being dependent on any particular idiosyncratic choice of scale or coding scheme. The
experiments here are by no means definitive, since they have examined only two geographical
regions and three analytical techniques. The Levant and Balkans are definitely atypical of dyads
in generalthey have greater levels of conflict, and are much more thoroughly covered by the
news mediabut they may be representative of the conflict-prone regions that are most
frequently analyzed using event data. Similarily, the correlational methods used in discriminant
analysis and the LML clustering are typical of most statistical analysis.12 It therefore seems
unlikely that the development of a complex and highly differentiated scale to be a magic bullet
12
These tests have all used machine-coded data, so the insensitivity might be due to the errors found in automated
coding. However, this seems somewhat unlikely given the magnitude of the effect, the fact that the overall error
rate in machine coding is comparable to that of human coding, and the fact that many of the categories that are
ambigous in machine coding are also ambiguous to human coders.
Along a similar line, a journal referee for an earlier version of this research suggested that the results should
not be accepted until we replicated them on all dyads in the international system. While we concur that this
would be interesting, there are roughly 40,000 such dyads, the optimized weight and HMM experiments each
required about 24-hours of computer time, and we were disinclined to spend the next two centuries doing that
exercise. Readers might, however, find it useful to do comparable experiments with whatever region and
analytical method they are working on.

Schrodt and Gerner
Event Data
Page 3-33
that will suddenly reveal features of the data that were completely invisible to simpler
techniques. While one can not necessarily conclude that Less is more in event data, there is
also little evidence that More is more.
The downside of this dependence on event reports is that we know event coverage to be
inconsistent across geographical regions and sources. In one of the earliest studies of regional
source effects, Doran, Pendley and Antunes (1973) found substantial differences in reports of
domestic violence in Central America depending on whether regional or North American sources
were used. More recently, Davenport and Galaich (1998) found very substantial differences
across geographical regions in reports of human rights violations found in the New York Times.
When Schrodt, Huxtable and Gerner (1996) compared Reuters-based data on the Levant and
West Africa, missing data was more of a problem in the latter region than in the former, and a
discriminant analysis using the Goldstein scale classified a priori behavioral phrases with only
about 75% accuracy in West Africa versus the 90% accuracy in the Levant.
This inconsistency does not necessarily imply that event data cannot be used for political
analysisto the contrary, as increasingly sophisticated techniques are employed in the analysis
of event data, the results are becoming ever stronger and more consistent. However, if those
results are strongly dependent on the frequency of the reports rather than on their content,
comparisons across geographical regions and across different sources may be more difficult than
had been anticipated when the event data exercise began. Any source (or combination of sources)
of eventswhether a global source such as Reuters, Associated Press or Agence France Presse, a
hegemonic source such as the New York Times or Times of London, or the various capital-city
regional sources found in COPDABpresents only a tiny subset of the "events" that occur in a
day. The question for the event data analyst is not whether that subset is comprehensiveit
never will bebut whether it is useful for the research task at hand.
Finally, this lack of sensitivity to event weights and categorization has an important
implication for the use of machine-coded data. While machine-coding is more consistent over
time than human-coded data, machine-coding is less sensitive to nuances of reported political
Schrodt and Gerner
Event Data
Page 3-34
behavior, and it is possible that those nuances could be very important. This question has, in
fact, driven much of the debate about event data coding from the earliest developments of the
technique.
Our analysis does not support the conclusion that nuance is important: Because similar
results can be obtained from huge differences in the weighting of event categories and high level of
aggregation in the nominal coding of events, there is little evidence that subtle differences in the
coding of events would have a major difference on statistical measures based on that data.
Consequently, inexpensive machine-coded data is likely to contain most of the relevant
information that a vastly more expensive human-coded data set would contain. In particular, the
most common machine-coding error is confusing the object of an action with an indirect object or
a location. Coding programs will not, however, create a actor that is not mentioned in the text.
For example, if a series of events involves Israel, Syria, Lebanon and the Palestinians, some
actions of Israel towards Syria might be incorrectly coded as applying to Lebanon or the
Palestinians., but it would never create an extraneous Egypt=>Jordan event from these texts. If
most of the information found in event data comes from "who" rather than "what," relatively
simple coding systems that can be efficiently coded with automated techniques will prove quite
functional.
3.4. Statistical Approaches to Early Warning

The discussion so far has focused on general characteristics of how event data is collected and
coded. In this section, we will consider some issues dealing with how event data is used in
political analysis. We will consider specifically on the problem of statistical early warning,
which has been the focus of most of our work, as well as much of the work that motivated the
initial development of event data.
Schrodt and Gerner

Event Data
Page 3-35
Statistical approaches to early warning can be classified into two broad categories: structural
and dynamic. 13 The structural category consists of studies that use events (or more typically, a
specific category of event such as a civil or international war) as a dependent variable and explain
these using a large number of exogenous independent variables. In the domain of domestic
instability, this approach is exemplified by the work of Gurr and his associates, most recently in
the "State Failure Project" [SFP] (Esty et al. 1995, 1998); Gurr and Lichbach (1986) and Gurr
and Harff (1996) provide surveys of these methods more generally. In the field of international
instability, the structural approach is illustrated by the work of Bueno de Mesquita and his
associates, and more generally by the Correlates of War project; Wayman and Diehl (1994),
Gochman and Sabrosky (1990) and Midlarsky (1993) provide general surveys. These
approaches have tended to use standard multivariate linear regression models, although recently
the research has branched out to other techniques; for example, the SFP uses logistic regression,
neural networks and some simple time series methods.
In contrast to the structural approach, in dynamic early warning models event data measures
are used as both the independent and dependent variables. Most of the event data projects of the
late 1970s classified dyads with respect to the likelihood of a crisis based on a set of event-based
empirical indicators. For instance, the Early Warning and Monitoring System (EWAMS),
developed with funding from the U.S. Defense Advanced Research Projects Agency (DARPA;
see Hopple 1984; Laurance 1990), evaluated three WEIS-based indicators (conflict, tension, and
uncertainty) to determine an alert status for any dyad. Azar et al. (1977) use a similar approach
based on whether behaviors measured with the COPDAB event scale fall outside a range of
"normal" interactions for the dyad.
13
This discussion will not consider the large literature on non-statistical (qualitative) approaches to forecasting
Contemporary surveys of qualitative approaches can be found in Rupesinghe & Kuroda (1992), Gurr & Harff
(1994), and Adelman & Schmeidl (1995). We also will not deal with the topic of long-range forecasting using
formal methods, which is primarily done using simulation. Ward (1985) and Hughes (1993) provide surveys of
that literature.
Schrodt and Gerner
Event Data
Page 3-36
Researchers justify the dynamic approachwhich is at odds with most statistical modeling
in political science in using only lagged endogenous variablesin three ways. The first rationale
is that many of the structural variables that are theoretically important for determining the
likelihood of conflict do not change at a rate sufficient for use in an early warning indicator; in fact
many are essentially fixed (e.g. ethnic and linguistic heterogeneity; historical frequency of
conflict; natural resource base). Data on variables that are changingfor example unemployment
rates, economic and population growth ratesare often reported only on an annual basis, and the
quality of these reports tends to be low in areas under political stress.
The second justification for the dynamic approach is that it reduces the information required
by the model. The data collection effort of the first phase of SFP, for example, measured 75
independent variables (Esty et al 1995:9); this requires a large amount of information from a vast
number of sources.14 In contrast, the event data collections used in dynamic models focus on
reported political interactions that can be collected systematically in real-time, which increases
the predictive utility of the model.
The final justification for dynamic modeling involves the nature of political events
themselves: the approach assumes that the effects of exogenous variables used in the structural
models will be reflected in the pattern of events prior to a major change in the political system.
As illustrated in Figure 3.6, the dynamic approach effectively uses the lagged values of the events
as a substitute for the structural variables.
14
However, the final models developed in the project found that most of the forecasting power could be accounted
for with only three variablesinfant mortality, trade openness, and democracy (Esty et al 1998:viii). Phase II of
the SFP involves some limited analysis of dynamic variables, and suggests expanding this approach in future
studies.
Schrodt and Gerner
Event Data
Exogenous
variables
Page 3-37
Crisis
Event
Exogenous
variables
Lagged
Events
Crisis
Event
Estimated Model
Structural Modeling Approach
Dynamic Modeling Approach
Figure 3.6. Comparison of the structural and dynamic approaches to early warning
To take a concrete illustration, Gurr (1995: 7) notes "We think, for example, that ethnic
heterogeneity probably is most significant for state failure when it coincides with lack of
democracy and low regime durability." Consequently, the SFP includes measures for those three
variables: ethnolinguistic diversity, regime democracy, and regime durability.
A dynamic approach, in contrast, would not measure these aspects of a political system
directly, but would instead assume that each would be reflected in the types of events picked up
by the international media. The presence of democracy, for instance, would be reflected not only
in periodic elections but in a large number of reports of disagreements between the government
and the elected opposition. A low level of regime durability would be reflected in coups and
attempted coups. To the extent that ethnicity was an important political factor, it would be
reflected in ethnically-oriented political rallies, outbreaks of violent ethnic conflict and similar
events. A suitably-designed event coding scheme should detect the presence or absence of these
events and make the appropriate forecast, without directly measuring the underlying variables.
At a theoretical level, the dynamic-events approach accepts the importance of exogenous
structural variables: Ceteris paribus, countries with a high level of ethnic heterogeneity will have
a different propensity for conflict than those with a low level; democracies are likely to be
different than autocracies, and so forth. The difference between the early warning approaches is
a matter of measurement: the structural modeling approach seeks to measure these variables
directly, whereas the dynamic approach assumes that to the extent that the variables are relevant
Schrodt and Gerner
Event Data
Page 3-38
for early warning problems, they can be measured indirectly through the patterns of events the
variables generate.
This is an optimistic, but not wholly implausible, assumption. For example, in the Reutersbased data with which we have been working, there is a clear contrast between Israel and Syria with
respect to the presence of a democratic opposition and between Lebanon and Egypt with respect
to the importance of ethnicity: The ethnic conflict in Lebanon is one of the most conspicuous
features of the data set. Our impression is that the increase in democracy in Jordan, and the
fluctuations in the Egyptian government's acceptance of a democratic opposition, would also be
reflected in the activities reported in Reuters, although we have not attempted to analyze this.
An econometric analogy to this is found in the distinction between "technical" and
"fundamental" analysis of stock prices. A fundamental analysis attempts to predict price
changes on the basis of underlying factors such as marketing, management, raw material prices,
and macroeconomic trends. Technical analysis, in contrast, assumes that these factors will be
reflected in the patterns of the movements of the price of a stock (or set of stocks) and therefore
analysis of those prices alone will provide sufficient information for forecasting. Fundamental
analysis corresponds to the structural approach to modeling political events; technical analysis to
the dynamic.
Until relatively recently, technical analysis generally had a bad reputation, consisting as it
did largely of statistically-dubious patterns based on small samples, wishful thinking, and gurus
whose fortunes were based more on the sale of books than on trading stock. With the increase in
computing power in the 1980s, the situation changed, and "programmed trading systems" can
now process sufficiently large amounts of information to generate profits (and periodically throw
the market into chaos) working solely with information endogenous to the market itself. The
increased information processing capacity in the 1990s in contrast to that available in the 1970s
may have a similar effect on event data analysis.
Because of the labor-intensive character of human event coding, the primitive statistical
methods available at the time, and institutional factors (Daly & Andriole 1980; Andriole &
Schrodt and Gerner
Event Data
Page 3-39
Hopple 1984; Laurance 1990), the event-based early warning research was largely discontinued
during the 1980s. Nonetheless, a small set of dynamic modeling efforts continued. These
employed increasingly-advanced econometric time-series methods that modeled an interval-level
measure of events as an autoregressive time series with disturbances. Goldstein & Freeman
(1990) provide a book-length example of this approach; Ward (1982), Dixon (1986), Ward &
Rajmaira (1992), Lebovic (1994) and Goldstein and Pevehouse (1996) illustrate the continued
development of dynamic models of events, although these studies generally used event data to
explore political interactions rather than for forecasting.
Unfortunately, standard econometric time series methods have only limited utility in the
problem of early warning. In general, time series analysis seeks to determine a function
y t+k = f(yt, y t-1..., Xt, Xt-1...) for some k>0
In English, the fundamental problem of time series is to determine the future values of a variable
y given some present and past values of that variable and (possibly) the present and past values
of a vector of exogenous variables X. Due to the importance (and potential financial rewards) of
accurate economic forecasts, there is a massive literature on time series estimation in
econometrics (see Hamilton 1994).
In contrast, the problem of statistical early warning consists of finding a time T such that
y t - ys > t > T > s
for some indicator variable y. In English, this means that the variable y has consistently higher or
lower values after time T than it had prior to time T, which would occur in aggregated event data
following a qualitative shift in the type of political behavior in which a dyad was engaged.
An additional distinction is that econometric time series generally are highly autoregressive
(e.g., GNP, unemployment, prices of consumer goods, and inflation rates) or at least have an
autoregressive component combined with generally random noise (e.g., stock prices; exchange
rates). The GNP or unemployment rate of a major industrialized economy has tremendous
Schrodt and Gerner
Event Data
Page 3-40
inertia. For instance, while the stock market crash of October 1929 was sudden, the high
unemployment rates of the Great Depression required two or three years to fully develop.
Furthermore, most econometric time series are measured continuously rather than episodically,
so missing data is less of an issue.
Early warning, however, focuses on shifts in the time-series that are not autoregressive, even
although the series taken as a whole might be autoregressive. An autoregressive model of warand-peace will be very accurate, as illustrated by the presumably apocryphal story about the
European political analyst who said "Every day from 1910 to 1970 I predicted that Europe
would remain at peace when at peace, and remain at war when in war, and I was only wrong four
times." This type of model is not, however, very useful. (More technically, such a measure
succeeds according to a frequency-based measure but fails according to an entropy-based measure,
which places higher weight on the prediction of low-probability events.) The econometric
problem most comparable to political early warning is forecasting sudden economic shifts such as
those observed in exchange rate fluctuations (e.g., the collapse of the Mexican peso or the
European Exchange Rate Mechanism). These problems are similar to political early warning in
the sense that they are primarily psychological and do not reflect a major change in the
underlying physical reality: the economic fundamentals of the Mexican or European economies
did not change dramatically during the days of the exchange-rate crises, but the perceptions of the
future values of the relevant currencies did change.
Despite these complications, it should be noted that in two very important respects
prediction is an easier problem than the typical econometric estimation problem. First,
forecasting models have right-and-wrong answers, or at least their accuracy can be evaluated
probablistically. Coefficient estimation problems, in contrast, do not have answers: one can
always specify an error structure, prior probability or alternative model structure that places the
estimated emphasis on different variables, and there is no empirical method of deciding between
these specifications. Secondand closely related to the first issueforecasting problems are not
affected by collinearity, which is the bane of coefficient estimation in the social sciences because
Schrodt and Gerner
Event Data
Page 3-41
every behavior tends to be linked to every other behavior. Coefficient estimates with low
standard errors are clearly useful for obtaining a theoretical understanding of a situation, but they
are not essential for the pragmatic purposes of forecasting (Wonnacott & Wonnacott 1979:81).
For this reason, it is not surprising that models with very diffuse coefficient structuresfor
example neural networks and VARincreasingly are found in early warning research.
Most of the examples of early warning used in this book employ some form of sequence
analysis. As discussed in Chapter 4, the sequence analysis approach has a long history in
political scienceat the most fundamental level, it is simply a systematic rendition of the "case
study" or "lessons of history" technique that has been used by decision-makers since time
immemorial (see May 1973, Mefford 1985, Neustadt & May 1986, Vertzberger 1990, Khong
1992) . History is considered relevant to decision-makers because they assume that when a
particular set of events and circumstances observed in the past is observed again, the resulting
events from that prior case can also be expected to apply in the new case, all other things being
equal.
This simple observation is both reinforced and attenuated by the fact that it is reflexivethe
methods that decision-makers use to interpret the past have an impact on how they create the
future. If decision-makers act consistently on the "lessons of history", then history will in fact
have lessons.
By itself, however, belief in the importance of historical examples is insufficient to create
empirical regularities because of "Van Crevald's Law"15: A conspicuously successful strategic
innovation is unlikely to succeed twice precisely because it was successful the first time. More
generally, work of the Santa Fe Institute on the so-called the "El Farol Problem" (see Casti 1997)
has demonstrated that systems of adaptive utility maximizers generally do not exhibit regularized
15
"...war consists in large part of an interplay of double-crosses [and] is, therefore, not linear but paradoxical. The
same action will not always lead to the same result. The opposite, indeed, is closer to the truth. Given an
opponent who is capable of learning, a very real danger exists that an action will not succeed twice because it has
succeeded once." (Van Creveld 1991:316; italics in original).

Schrodt and Gerner
Event Data
Page 3-42
behavior because they look at history. In computer simulations, such agents tend to show
quasi-chaotic behavior that is not predictable. If the political world consists solely of rational
adaptive agents, there is little point in trying to make predictions based on past behaviors.16
There are undoubtedly some forms of international behavior (for example international exchangerate behavior) for which this is true.
But it is not be true in all cases. Situations of international conflict usually involve
organizational behavior rather than individual behavior, and for a variety of reasons both
theoretical and practical, organizations are substantially less likely to engage in rapidly adaptive
behavior than are individuals. Mature organizations instead are likely to rely on rule-based
standard operating procedures (SOPs) that are designed to insure that a specific set of stimuli will
invoke a specific response (Cyert and March 1963, Allison 1971). A classical Weberian
bureaucracy, unlike the adaptive maximizer of complexity theory, is virtually designed to assure
the success of a sequence analysis approach.
The SOPs are themselves adaptivethey are designed to effectively solve problems and many
are acquired through historical experience. But in a situation of the protracted interaction, two
organizations with SOPs are coadaptive: each responds in part to the environment created by the
other.17 In most circumstances, this eventually brings their SOPs into a Nash equilibrium within
the space of possible SOPs where neither can change strategies unilaterally without a loss of
utility. This is more likely to occur when the same organizations have been interacting over a
period of time, and when the payoff environment has been relatively stable. This is found,
16
Predictions could still be made on the basis of other characteristics of the systemfor example the effects that
economic or technological changes have on the utility functions of the actors, and even predictions about the
range of strategic outcomes. But in the absence of a completely specified model and complete information,
there is little point in trying to make point predictions in a chaotic system.
17
A detailed discussion of the concept of coadaptation is beyond the scope of this chapter, but general discussions
from a natural science perspective can be found in Maynard-Smith (1982) and Kauffman (1993); Anderson,
Arrows and Pines (1988) discuss a number of social science applications, and Schrodt (1993) applies the concept
to the issue of international regimes.

Schrodt and Gerner
Event Data
Page 3-43
notably, in the situation of protracted conflicts and enduring rivalries. These are situations
characterized by exactly the competitive SOP "lock-in" that we've outlined aboveantagonists
fight, on repeated occasions, over the same issues, often over the same territory, and without
resolution.
To summarize, sequence-based prediction will not work in all circumstances, but it will work
in a significant number of cases. In addition, those instances where it will not workrapid and
complex adaptationare frequently situations where other methods are not going to work either.
This relevance of event sequences may also explain in part why study of history remains popular
with politicians and diplomats despite our best efforts to divert them to the study of game
theory and statistics.
3.7. Conclusion
One of the anonymous referees who reviewed the validity studies in Schrodt and Gerner
(1994) observed "Because human coding of events is so miserable, why should you even try to
duplicate it with a machine?!". Good point: if WEIS can only be coded with 83% inter-coder
reliability, perhaps we should be looking for an alternative system that could be coded at, say,
95% reliability.
McClelland (1983) notes that WEIS was intended to be the first cut at developing an event
data coding scheme. Instead WEIS became the final word and the de facto U.S. government
standard. WEIS has obvious problems such as the ambiguous warning/threat distinction, and
similar events in distinct cue categories (e.g. 013 "admit wrongdoing; retract statement" and 061
"express regret; apologize"), and in addition harbor other application-specific cases of ambiguity
that become obvious only when one is training coders.
The original event data researchers were intent on getting the maximum amount of information
out of a newspaper story. Information was relatively scarce; and the investment of time required
first to train a coder and then have the coder locate and read a story was large in comparison to
the time require to extract some additional information. We also suspect that the extent to which
Schrodt and Gerner
Event Data
Page 3-44
the coder's cognitive biases would cause systematic errors was underestimated: The political
science research community in the 1960s was not exactly known for a diversity of perspectives,
and insights from cognitive psychology on the effects that preconceptions had on interpretation
did not become widely incorporated into the discipline for another decade.
This balance has changed with the availability of machine coding. The initial costs in
generating a new set of event data involve developing dictionaries, but once these are complete,
the incremental cost of coding additional stories is effectively zero. Consequently we might be
better off throwing away information that cannot be coded reliably and focusing instead on
sentences that can be reliability categorized. For example there is probably little point in
attempting to code the affective content of carefully worded official statements whose underlying
premise is "Diplomacy is the art of telling someone to go to hell in such a fashion that they look
forward to the trip." These are the sort of sentences that even humans can't reliably code, and we
should not expect machines to reliably code them . Coding systems that are designedand
testedfor reliability might result in event sequences that give a clearer signal of political events.
At the very least, any project that intends to invest a substantial amount of effort in
dictionary development (whether in the refinement of the event coding vocabulary or in the
refinement of the coding scheme itself) should probably first ascertain the extent to which those
refinements will make a difference given the analytical technique being used. For example, the
report weighting scheme discussed in subsection 3.3.1 can be tested on a set of data generated
using a standard .verbs dictionary, and the resulting data set should give at least a rough idea of
how much explanatory power is found in the existence of reports alone. (The detection of
reports is still sensitive to the .actors dictionary, and the actor vocabulary should almost always
be modified to accurately code any new region or time period.)
One of us recently reviewed a project that had spent a great deal of time and money
meticulously human-coding events using a detailed framework whose coding manual was over
one-hundred pages in length. At the analytical phase, however, the data, were subsequently into
simple verbal-physical/cooperation-conflict categories similar to those used in section 3.3.3.
Schrodt and Gerner
Event Data
Page 3-45
They were then used to estimate a neural network model that was insensitive to missing values
and coding error, and had a very diffuse parameter structure that virtually precluded the analysis
of the effect of the individual event category. This is the data analytical equivalent of buying a
very expensive red wine and adding it to a tomato sauce that will be simmered for three hours.
Maybe your guests will notice the difference, but in all probability, they wont, and an ordinary
table wine would have produced indistinguishable results. Similar economies are possible with
machine-coded event data.
Schrodt and Gerner

Event Data
Page 3-46
Appendix 3.A: A Genetic Algorithm for Estimating

Optimal Event Weights
Genetic algorithms (Holland 1975; Grefenstette 1987; Goldberg 1989) are a general purpose
optimization method that is particularly effective in situations where there are a large number of
local maxima. Because the LML clustering algorithm determines the cluster boundaries as well as
determining the weights, the problem is non-linear and required the use of a numerical
optimization method rather than an analytical optimization method such as discriminant analysis.
A cluster break was any point that met the following conditions:
1. LML t >
2. No cluster boundaries in the previous 8 months (i.e., minimum cluster size of 8 months)
I experimented with several values of in the range =0.15 to =0.30. The number of clusters
found is inversely proportional to the value of and the LML t 0.20 threshold is comparable to
the level found to produce cluster boundaries corresponding to the a priori clusters when the
Goldstein weights are used. A minimum cluster size is necessary because a sharp change in
behavior will produce several consecutive months where LMLt is high.
The genetic algorithm is straightforward: the optimization operates on a vector of weights for
the twenty-two WEIS 2-digit categories:
w = [w 1,...,w22]
For a given set of weights, an aggregated monthly score is computed for each dyad
22
XYt = wct =
wi cit
where
i=1
cit = number of events in WEIS 2-digit category i directed from X to Y in month t

Once these scores are calculated, the LML measure is computed, the boundaries between clusters
are determined using the LMLt > threshold and minimum size rules discussed above, and the
fitness measure
average distance between adjacent clusters
Fc =
average distance within clusters
Schrodt and Gerner
Event Data
Page 3-47
is computed with "distance" || xi - xj || is defined by the correlation metric and the "average
distance" is calculated as the average distance between points:
Between cluster distance =
Within cluster distance =
1
N1N2
|| xi - xj ||
iC 1jC 2
2
N1(N1-1)
|| xi - xj ||
iC 1 j> i
where Ni = number of points in cluster i. The measurement of the points in adjacent clusters
rather than comparing the distance of a cluster to all other clusters is done to allow the possibility
of the system returning to an equilibrium behavior, so that clusters that are separated in time
might occupy the same space.
The genetic algorithm uses 32 w vectors that are initially set randomly to numbers between
-10.0 and +10.0, the same range as the Goldstein weights. After the fitness of each vector is
computed (a "generation" in the genetic algorithm), the vectors are sorted according to the value
of Fc and the 16 vectors with the lowest fitness are replaced with new vectors created by
recombination and mutation of the top 16 vectors.18 The probability of a vector becoming a
"parent" is proportional to the relative fitness of the vector (in other words, vectors with higher
fitness are more likely to be used to produce new vectors). Mutation involves adding a random
number between -1 and +1 to the weight, and mutation is done on 50% of the weights in the new
vectors.
Most of the results we report are based on runs where the system ran for 48 generations: this
was usually sufficient to find a vector that showed little or no change. The system was
18
One new vector was generated by taking the average weight of the top 16 vectors, on the logic that weights that
were not important in the distance calculations (notably those for codes that occur infrequently in the data set)
would go to zero as the random weights canceled out. These average vectors were tagged so that their survival in
future generations could be tracked. They were rarely selected in the first set of experiments where both the
weights and cluster breaks were allowed to vary, but were frequently selected in the second set of experiments
where the weights were optimized for a given set of cluster divisions.
Schrodt and Gerner
Event Data
Page 3-48
occasionally allowed to run for 128 generations; this produced essentially the same results as the
shorter runs. This system was implemented in a C program; the source code is available from the
authors.
Schrodt and Gerner


Chapter Three The Statistical Characteristics of Event Data: DRAFT October 30, 2000

Uploaded by

Document Informationclick to expand document information

Copyright:

Available Formats

Chapter Three The Statistical Characteristics of Event Data: DRAFT October 30, 2000

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Chapter Three The Statistical Characteristics of Event Data: DRAFT October 30, 2000

Uploaded by

Copyright:

Available Formats

Chapter Three

The Statistical Characteristics of

DRAFT October 30, 2000

3.1. Defining an "event"

DRAFT: October 30, 2000

The COPDAB project designates events as:

DRAFT: October 30, 2000

DRAFT: October 30, 2000

DRAFT: October 30, 2000

3.2. Stochastic Elements in the Measurement of Event

activity as the product of "bandits" as a way to delegitimize the guerrillas.

DRAFT: October 30, 2000

Figure 3.1. The Process of Generating Event Data

Schrodt and Gerner

DRAFT: October 30, 2000

on event codes rather than original events.

DRAFT: October 30, 2000

DRAFT: October 30, 2000

DRAFT: October 30, 2000

Schrodt and Gerner

DRAFT: October 30, 2000

3.2.5. Statistical Corrections for Coding and Censoring Errors

DRAFT: October 30, 2000

3.3. Interval vs. Discrete Approaches to Event Data

undercount in the U.S. Census.

DRAFT: October 30, 2000

DRAFT: October 30, 2000

difficult to visualize and explain (McClelland, 1961).

DRAFT: October 30, 2000

DRAFT: October 30, 2000

Table 3.1. Examples of Event Scales

Military, economic and strategic support

Mild verbal support; exchanges of minor officials

Diplomatic-economic hostile actions

Full scale war

Promise Policy Support

Extend Military Aid

DRAFT: October 30, 2000

cooperation-conflict dimension. Beyond this, there is no consistency in the scalingfor example

Table 3.2. Alternative Scales

cooperative events = 1; conflictual events = -1.

cooperative event = 0; conflictual events = 1.

cooperative event = 1; conflictual events = 0.

1 if any event was reported in the month, 0 otherwise

DRAFT: October 30, 2000

Table 3.3. Discriminant Analysis Results

DRAFT: October 30, 2000

Figure 3.2. Cluster boundaries under various weighting systems

Schrodt and Gerner

DRAFT: October 30, 2000

3.3.2. Optimizing the Scale

average distance between adjacent clusters

1.26; conflict = 1.29; report = 1.42.

DRAFT: October 30, 2000

Figure 3.3. Cluster boundaries under genetic algorithm optimization

DRAFT: October 30, 2000

DRAFT: October 30, 2000