Data-based Computational Approaches

to Forecasting Political Violence

Philip A. Schrodt, James Yonamine, and Benjamin E. Bagozzi

1 Introduction and Overview

The challenge of terrorism dates back centuries if not millennia. Until recently, the
basic approaches to analyzing terrorismhistorical analogy and monitoring the
contemporary words and deeds of potential perpetratorshave changed little:
the Roman authorities warily observing the Zealots in first-century Jerusalem could
have easily traded places with the Roman authorities combatting the Red Brigades
in twentieth century Italy.
This has changed with the exponential expansion of information processing
capability made possible first by the development of the digital computer, followed
by the phenomenal growth in the quantity and availability of machine-readable
information made possible by the World Wide Web. Information that once circulated
furtively on hand-copied sheets of paper (or papyrus) is now instantly available
for good or illon web pages which can be accessed at essentially no cost from
anywhere in the world. This expansion of the scope and availability of information
in all likelihood will change the dynamics of the contest between organizations
seeking to engage in terrorism and those seeking to prevent it. It is almost certainly
too early to tell which group will benefit moremany of the new media are less
than a decade oldbut the techniques of processing and evaluating information
will most certainly change.
This chapter provides an extensive overview of inductive statistical and computational methodologies used in the analysis and forecasting of political violence, and
some of the challenges specific to the issue of analyzing terrorism. It is intended for
the non-specialist, but assumes a general familiarity with data and computational
methods. Our purpose is not to exhaustively explore any of these methodseach

P.A. Schrodt () J. Yonamine B.E. Bagozzi

Political Science, Pennsylvania State University, University Park, PA 16801, USA
e-mail: schrodt@psu.edu; jxy190@psu.edu; beb196@psu.edu
V.S. Subrahmanian (ed.), Handbook of Computational Approaches to Counterterrorism,
DOI 10.1007/978-1-4614-5311-6 7,
Springer ScienceCBusiness Media New York 2013

129

130

P.A. Schrodt et al.

technique would typically require tens or hundreds of pagesbut instead to provide

a sufficient introduction to the basic concepts and vocabulary, as well as a very
extensive sets of references, so that the reader can explore further on his or her own.
The psychologist and philosopher William James, in his Lowell Institute lectures
in 1906, subsequently published under the title Pragmatism: A New Name for Some
Old Ways of Thinking notes that the fundamental split in philosophy, dating to the
very origins of the field, is between rationalists who seek to find an intellectual
structure that will reveal a proper order in the world, and empiricists, who take
the disorder of the observed world as a given and simply try to make as much sense
of it as they can. More than a century later, we find exactly the same split in formal
approaches in the social sciences: The rationalist position is expressed in deductive
approaches such as game theory, expected utility models, systems dynamics and
agent-based models, which seek to explain behavior from a set of a priori firstprinciples and their consequent emergent properties. The empiricist approach is
found in inductive statistical and computational data-mining approaches which
extract structured information from large sets of observed data. Both approaches
are well-represented in this volume [38, 48, 103, 105, 117, 123, 151, 152] but this
review will focus only on the empirical approaches.
Throughout this chapter, we will generally be looking at models which focus
on forecasting and understanding political violence in general, not just approaches
to terrorism per se. This is done for two reasons. First, due to a combination
of data limitations and a paucity of interest in the research community prior to
the 2000s, the number of quantitative studies of terrorism was quite limited and
focused on a relatively small number of approaches [149]. However, the large-scale
research efforts did consider various forms of political violence more generally.
Second, most of those more general methods are at least potentially applicable to the
study of terrorismin fact, many are generally applicable to almost any behavior
for which large amounts of data are availablealbeit we will frequently caveat
these possibilities with concerns about some of the atypical aspects of terrorism, in
particular the distinctions between methods appropriate to rare-events analysis and
those dealing with high-frequency behaviors. Finally, in many instances, there is a
close relationship between situations of general political violence such as civil war
and state failure, and conditions which encourage the formation and maintenance of
terrorist groups, so political instability is of considerable interest on its own.
We will first provide a general overview the types of data commonly used
in technical forecasting models, then consider the two broad categories of computational models: statistical and algorithmic. Within statistical modeling, we
assess the strengths and weaknesses of the widely-used ordinal least squares and
logit methods, conventional time series approaches, vector-autoregression models,
event-history models, and a variety of techniques involved in rare-events analysis.
Within computational modeling, we consider supervised classification methods and
unsupervised clustering methods, and then look more specifically at the issue of
using event sequences for forecasting. Finally, we will briefly consider some of the
recent developments in computational network analysis.

Forecasting Political Violence

131

1.1 The Development of Technical Political Forecasting

There is both a push and a pull to the current interest in technical political
forecasting. As with many aspects of the Information Revolution of the past several
decades, some of this is driven purely by technology and the availability of data,
computational power and readily accessible software. We can do things that we
couldnt do before, and clearly the impact of intellectual curiosity and technical
itches to scratch is a factor.
But if the technological factors were the only driver, we could satisfy that by
clicking on virtual cows and broadcasting our prowess on Facebook (http://www.
wired.com/magazine/2011/12/ff cowclicker/all/1). Instead, at the level of serious
policy analysismess up counter-terrorism, and real people die, with no option
to resetthe motivation is different: humans are appallingly bad at predicting
human behavior. The pathbreaking studies of Philip Tetlock [161], who assessed
the performance of 284 expert political forecasters in the assessments of more than
80,000 predictions over a 20 year period, found that their accuracy was barely better
than chance, and the record for some of the most publicly-visible forecasters, who
play to a crowd anxious for dramatic predictions, is actually worse than what could
be achieved by a dart-throwing chimpanzee. Tetlocks results confirmed for the
political realmalbeit in much greater detailsimilar findings in economics and
psychological assessment [94, 111] but were nonetheless a stunning indictment of
the subject matter expert as a prognosticator.
Tetlocks work was particularly timely given the intellectual climate for the
development of statistical and computational models of political violence. The
earliest systematic studies of political violence using modern statistical methods
and standards for data collection date from the middle of the twentieth century, in
the work of British meteorologist Lewis Richardson [131]. This approach gained
considerable momentum in the 1960s with the behavioral revolution in political
science, which used the emerging technologies of social science statistics, digital
computing, and machine-readable data to begin the systematic assessment of various
theoretical propositions about political violence.
These found, for the most part, that the patterns in the qualitative wisdom
literature were far less generalizable than their advocates believedperhaps
unsurprising in a world that had experienced two civilization-shattering wars in
only three decadeswith the results on the democratic peace [12, 15, 81] being
the notable exception. This period saw the initial development of many large-scale
databases relevant to the study of political violence, including the Correlates of
War (http://www.correlatesofwar.org/) on international and civil war, Polity (http://
www.systemicpeace.org/polity/polity4.htm) on state political characteristics, and
early event data sets [8, 36, 99, 110] on political interactions in general. The data
sets collected specifically for the study of political behavior were supplemented
by readily-available global data on demographic, development, economic and trade
characteristics of states provided by organizations such as the United Nations (UN),
International Monetary Fund (IMF) and World Bank.

132

P.A. Schrodt et al.

The U.S. Department of Defense Advanced Research Projects Agency (DARPA)

funded a series of projects for the development of statistical forecasting in the 1970s
[4, 44, 85], and parallel efforts, using both statistical and computational (artificial
intelligence) methods continued in the political science community under National
Science Foundation funding [39, 76, 87, 160]. However, these early efforts were not
particularly successful in the production of robust, timely, and accurate forecasts.
In retrospect, neither the data nor methods available at the time were sufficient for
the task. By the 1980s U.S. government efforts had largely ended, though the basic
research in the academic community continued, as did some efforts by international
non-governmental organizations [45, 75, 79, 132].
The U.S. government re-engaged with the development of technical political
forecasting in the mid-1990s, motivated in part by the considerable progress in
the development of methods and data since the earlier efforts, and by dramatic
intelligence failures by human analysts on such events as the end of the Cold
War [64, 65], state failure in Somalia and the former Yugoslavia, and the Rwandan
genocide. The most conspicuous effort was the U.S. multi-agency State Failures
Project [60, 61], later renamed the Political Instability Task Force (PITF) [74], a
long-term collaboration between government analysts, contractors, and academic
researchers. After some initial missteps involving efforts which were overly complex, PITF developed models that predicted various forms of political instability
with 7080% accuracy1 with a 2-year time horizon and global coverage. In the mid2000s, DARPA returned to the issue with the Integrated Conflict Early Warning
System (ICEWS) [117, 118], which achieved similar levels of accuracy.
As a consequence of these various efforts, technical forecasting of political
violence is now quite well developed and the subjects of numerous efforts in both
North America and Europe [18, 27, 33, 8284, 118, 165, 167]. While most of these
studies look at violence and irregular political activity such as coups, a quantitative
literature has also emerged on the topic of terrorism; extended reviews of this
literature can be found in [66, 126, 127, 150, 169].

2 Data Sources
As noted in the previous section, one of the major factors fueling the development of
computation methods for the analysis and forecasting of political violence has been
the availability of machine-readable data sets. Collection of data sets has now been
underway for close to five decadesthe Correlates of War Project, for example,
began in 1963and are now the product of thousands of hours of careful research,
refinement and coding, and have been used in hundreds of studies. In other cases
automated coding of atomic event data and analysis of the new social mediathe

Defined here in terms of the percentage of conflict-onset cases correctly classified.

Forecasting Political Violence

133

collections are relatively new and unexplored, but potentially provide very large
amounts of data in near-real-time. These data are generally available on the web,
either through academic archives such as the Inter-University Consortium for
Social and Political Research (http://www.icpsr.umich.edu/icpsrweb/ICPSR/) and
Harvards Dataverse Network (http://thedata.org/), through government and IGO
sources such as USA.gov (http://www.usa.gov/Topics/Reference-Shelf/Data.shtml)
and the United Nations Statistics Division (http://unstats.un.org/unsd/databases.
htm) orincreasinglythrough individual web sites established by the projects
collecting the data.
In this section, we will briefly describe some general types of data that have been
used in models of political violence, with comments on a few of the strengths and
weaknesses of each type. As with any typology, not all of the data sets fit clearly
into a single category, but most will.

2.1 Structural Data

Structural data refer to characteristics of a single political unit: typically this is a
nation-state but it could also be a province or city, or a militarized group. At the
nation-state level, classical structural measures include socio-economic variables
such as population, literacy and infant mortality rates, gross domestic product
(GDP), GDP per capita, percent urban population, ethnic composition, and the Gini
index measure of income inequality. Specialized political science data sets such
as Polity [104] and Freedom House [62] provide an assessment of a large number
of dimensions of governance, often at a very high level of detail: for example the
Institutions and Elections Project [130] has coded 127 characteristics of government
and electoral institutions for all countries with populations over 500,000 from
19722005.
At the (militarized) group level, structural data is available for a wide variety of
terrorism-related actors [6, 112, 153]. For instance, the Minorities at Risk (MAR)
and Minorities at Risk Organizational Behavior (MAROB) datasets each provide
extensive structural information on the social characteristics (e.g., ideology, religion,
population) of politically active communal groups throughout the world [6, 112].
Finally, note that structural data can be used either directly or indirectly. GDP
and population are good examples of direct measures: they are primarily used to
adjust for the economic capacity and number of people in a country. Probably the
best known indirect measure is infant mortality rate (IMR), which has consistently
emerged as a one of the most important predictors of political instability. This is
not due to any tendency of individuals whose children have died to immediately go
out and revolt, but rather because IMR has proven, in multiple independent studies,
to be a good indirect measure of a combination of poverty and the failures of a
government to deliver basic services.

134

P.A. Schrodt et al.

2.2 Dyadic Data

Dyadic data deal with relations between two entities. As above, these are usually
states but could also be individuals or organizations. Trade datanotably the IMFs
Direction of Trade [89]are some of the most commonly used. Again, depending
on the application, these can be used as either direct or indirect measures: trade
is important in its own right, but measures like bilateral trade can also reflect the
extent to which two states have common interests. Data on shared alliances, joint
international organization (IGO) membership and military involvement (as allies or
antagonists) are also readily available and are among the commonly used concepts
used to test traditional international relations theories.
In recent years, the most important development in dyadic measures has been
geospatial, which we discuss in Sect. 5.2. While the concept of political geography
is quite old, only recently have political-geographic measures reached a level where
they can be used systematically. However, political violence has a very strong
geospatial componentshared borders are consistently the single most important
predictor of whether states will engage in conflict, and overrides issues such as
commonalities or differences in religion, language and culture [15], and high levels
of terrorism tend to strongly cluster in both time and spaceand these methods are
gaining increasing attention [34, 35, 37, 129, 162].

2.3 Atomic Event Data

Atomic events datausually called simply event data in the literature are basic
units of political interaction coded from news sources that provide the date, source,
target, event-type of an interaction. These have a long history in the quantitative
analysis of conflict, with the early large data sets, the World Event Interaction
Survey (WEIS) [109, 110] and the Conflict and Peace Data Bank (COPDAB) [8, 9]
dating to the 1960s. While the original data were laboriously coded by (bored)
students working with paper and microfilm, starting in the 1990s collection shifted
to automated coding from machine-readable news wires [20, 67, 123, 138, 143],
which resulted in data sets coded in near-real-time containing millions of events
[72, 118, 143]. Along these lines, a number of more specialized event data sets
have recently been developed to specifically code terrorism events, and employ a
mixture of human and automated coding techniques to do so. Most notably, the
National Counterterrorism Centers Worldwide Incidents Tracking System (WITS),
the Global Terrorism Database (GTD), and the Terrorism in Western Europe Events
Data set (TWEED) all provide wide coverage and detailed codings of terrorist
attacks and related events [114, 120, 163].
Most event data sets use systematic typologies to reduce the nearly infinite
variety of event descriptions possible in natural language reports to a manageable
number of categories that a computational model can be process. All of these
typologies have about 20 macro categories and around 100 more detailed

Forecasting Political Violence

135

subcategories. While the original event data collections focused almost exclusively
on nation-states, contemporary systems provide higher levels of substrate aggregation, sometimes down to the level of individuals [68, 147]. While most studies in
the past aggregated the events into a single cooperation-conflict dimension using a
scale [73], more recent approaches have created composite events out of patterns of
the atomic events [88, 140] or looked at sequences of events [50].

2.4 Composite Event Data

Composite data are those which code the characteristics of an extended set of
events such as a war, crisis, or terrorist incident: the Correlates of War project
(COW; http://www.correlatesofwar.org/) is the archetype; International Crisis Behavior (http://www.cidcm.umd.edu/icb/) is a more recent example as is the Global
Terrorism Database [reference to The Global Terrorism Database, 19702010
Gary LaFree, Laura Dugan, Chap. 1 in this volume]. And multiple characteristics
of the incident are coded. Composite event data are typically coded from a number
of reports of the eventas distinct from atomic event data, which generally looks at
single sentencesand typically code a large number of characteristics of the event.
At present, composite event data are usually coded by humans, though machineassisted coding approaches are becoming increasingly prominent due primarily to
automated data field extraction methods, able to rapidly locate information from
text, such as the number of individuals killed or the amount of aid promised.
In some instances, it may also be possible to define composite events such as
civil war by using patterns of the atomic events [88]. This would also make the
differences between definitions used by various project unambiguous (or at least
comparable) and allow the composite events to be easily constructed out of existing
data sets rather than starting every new project from the beginning, and dramatically
reduce the cost of this type of data.

2.5 Social Media and Other Unstructured Data Sources

The new social mediaweb sites, blogs, chat rooms, Twitter, Facebook, and other
information easily available on the webrepresent an emerging frontier in data collection. The advantages of these sources are clear: they can provide direct access to
information about groups through the materials that they voluntarily make available,
they are easily accessible at nearly zero cost, and with automated natural language
processing, they can be analyzed in near-real-time. Social communication media
such as Twitter and Facebook also provide extremely finely-grained, minute-byminute data. In addition, a number of analysts have suggested that the social media
themselves are changing the character of political mobilization, as demonstrated by
the events of the Arab Spring, particularly in Tunisia and Egypt.

136

P.A. Schrodt et al.

Skeptics, however, have pointed out that social media also have some disadvantages. While access to the Web is increasing rapidly, it is still far from universal, and
social media in particular tend to be disproportionately used by individuals who are
young, economically secure, and well-educated. In areas with strong authoritarian
regimes, notably China, there are substantial (though not uniformly successful)
efforts to control the use of these media, and a government agency with even modest
resources can easily create a flood of false posts, sites and messages. While there
is some political content in these media, the vast bulk of the postings are devoid of
political contentOMG, Bieber fever!!!and what relevant content does
exist may be deliberately or inadvertently encoded in a rapidly-mutating morass of
abbreviations and slang almost indecipherably to conventional NLP software. (This
contrasts to the news stories used to encode atomic event data, which generally are
in syntactically correct English.) Finally, a number of analysts have argued that the
critical communications development for political mobilization is the cell phone,
both for voice and texting, rather than the Web-based media.

2.6 The Challenges of Data Aggregation

In most instances, conflict data comes pre-aggregated: no assembly required.
However, researchers interested in analyzing specific events extracted form the text
(which is common in event data studies as well as social network analysis), must
make critical aggregation decisions in three areas to convert the raw text or event
triplets into a format suitable to their models of choice: (1) actors (2) actions,
(3) temporal. (In contrast, techniques such as sentiment analysis and unsupervised
clustering algorithms (see Sect. 4.2.2), are equipped to analyze text in its raw format.

2.6.1 Actors
Most major event data datasetsincluding WEIS, CAMEO, ICEWS, and VRA
code the source and target actors for each event in the data set. However, many
of these actors may be irrelevant to the specific outcome-of-interest. For example,
a study focusing on Israeli-Palestinian conflicts would not want to include events
between Aceh rebels and the Indonesian army, as these are not relevant to the
conflict of interest. Although excluding Indonesian rebel activity is obvious in this
case, more difficult decisions exist, such as whether or not to include events between
members of the Lebanese and Syrian armies, or the governments of the United
States and Iran in a study of conflict between Israel-Palestine. Yonamine [170]
provides a more detailed description of event data aggregation.

Forecasting Political Violence

137

2.6.2 Actions
Event datasets use a numerical code to reflect the specific type of event that is occurring between the two actors. Since the numerical codes do not carry intrinsic value,
researchers manipulate the code to reflect meaningful information. The majority
of extant event data literature either scales all events, assigning them a score on a
conflict-cooperation continuum or generates event counts reflecting the number of
events that occur within conceptually unique categories. The Goldstein Scale [73],
which is the most commonly used scaling technique within the event data literature
[73, 77, 125, 136, 138], assigns a value from a 10 to C10 conflict/cooperation
scale, with 10 reflecting the most conflictual events and 10 indicating the most
cooperative.
Despite the preponderance of the Goldstein scale, a number of other studies
[140, 141, 144] utilize count measures. Duvall and Thompson [51] put forth the
first event data count model by placing all events into one of four conceptually
unique, mutually exclusive categories: verbal cooperation, verbal conflict, material
cooperation, material conflict. Although this count approach is more simplistic than
scaling methods, [135, 144] find strong empirical results using this count method of
action aggregation.

2.6.3 Temporal
Finally, scholars must temporally aggregate data in order to perform empirical
analyses at levels appropriate for their theory or empirical models of choice. All
of the previously mentioned event data sets code the exact day on which events
occur. As the specific time-of-day that events occurred is not reported, events must
at the very minimum be aggregated to the daily level [125,135,144], though weekly
[26, 148], monthly [136, 165], quarterly [93], and annual level aggregations are
common within the literature. By aggregated, we mean that the events occurring
within the selected temporal length must be jointly interpreted. Common approaches
are to calculate the sum or the mean of events that occur within the chosen temporal
domain.

3 Statistical Approaches
Most of the work on forecasting political conflict has used statistical modeling, since
this has a much longer history in political science than algorithmic and machine
learning approaches. While the bulk of these studies have focused on simply interpreting coefficient estimates within the frequentist mode of significance testing, a
method which has proven to have little utility for predictive modeling [165], more
recent work has taken prediction seriously, both using classical time series models
and more recently a substantial amount of work using vector autoregression (VAR)

138

P.A. Schrodt et al.

models. In addition, recent work has focused on one of the most challenging aspects
of forecasting, particularly when applied to counterterrorism: the fact that these
events occur very rarely. While this presents a serious challenge to the method, a
number of sophisticated methods have been developed to deal with it.

3.1 Cross-Sectional Regression and Logit

By far the most common statistical model in contemporary researchprobably
accounting for at least 80% of the studiesare variants on ordinary least squares
regression and the closely-related logit model. Ordinary least squares regression
uses equations of the form
Yi D 1 xi1 C    C p xip C "i D xi0 C "i ;

i D 1; : : : ; n;

(1)

Logistic regression, in contrast, is used to predict a values between 0 and 1

typically interpreted as a probabilityand does this by using an equation of the
form
1
ez
Yi D z
D
(2)
e C 1 1 C e z
where the variable z is usually defined as
z D 0 C 1 x1 C 2 x2 C 3 x3 C    C k xk ;

(3)

While these approaches have been used extensively in the academic literature,
their practical utility is, unfortunately, quite limited due to a reliance on the
frequentist significance testing approach. Extended discussions of this issue can
be found in [3, 70, 96, 137] but briefly, significance testing was originally developed
in the early twentieth century to study problems where the null hypothesis of
a variable having no effect was meaningful (the archetypical example is whether
a new medicine has a different effect than a placebo). In some political science
applications, this is still validfor example, determining whether a forthcoming
election is a statistical tie based on opinion pollingbut in most models a variable
is generally not included unless there are theoretical (or common sense) reasons to
assume that it will have at least some effect. Because this is all that the significance
test is assessinga non-zero effect, not an effect that has some meaningful impact
the information it provides is limited and all but negligible for predictive problems
[165].
Although the frequentist approach can be useful in weeding out variables that
might seem to be important but in fact are not, in contemporary models that
tend to be complex, even this should be interpreted with caution. Linear models
are particularly susceptible to problems of colinearity when the independent
variables xi are correlatedas is often the case, particularly in models where

Forecasting Political Violence

139

a researcher is experimenting with several measures of a diffuse concept such

as economic development or authoritarianismand in the extreme (but not
uncommon) case, colinearity can cause the estimated direction of a variable to be the
opposite of what it would be in the absence of the colinearity. Specification error
leaving out a variable that in fact has a causal link to the dependent variable Yi is
another issue, and has the unfortunate impact of inflating the estimated coefficients
of the variables which are in the model.
Finally, the existing literature tends to use complex models, often with a dozen
or more variables, despite the fact that a large set of literature, going back as early
as the 1950s [111], with the lesson repeated by [3] about a decade ago, indicates
that simple models are preferable, for at least two reasons. First, complex models
tend to fit the error, providing overly-optimistic assessments of the accuracy of
the model based on the existing data, with those estimates decreasing the accuracy
of the model once new data are available. Second, the nearly inevitable presence
of colinearity in non-experimental social science variables tends to increase the
variance of the estimated coefficients as the number of independent variables
increase.
On the positive side, the strengths and weaknesses of linear models have
been studied for decades and are well-understood. The problems noted above are
widespread but not inevitable, and models which have confirmed the same result
in a wide variety of formulationsfor example on the impact of geographical
contiguity, joint-democracy and trade on interstate conflict, and the impact of
economic development on political instabilityare probably robust. There has
also been a gradual shift away from the classical frequentist approaches to the
Bayesian approach [71, 92], which simply uses the data to adjust estimate of
the distribution of the coefficient, rather than employing either/or significance tests,
as well as methods such as matching cases within a sample [121] which can improve
the accuracy of the coefficient estimates.

3.2 Classical Time Series

Classical time series models predict the future value of a continuous variable based
on some combination of the past values, usually with a focus on reducing the
systematic error in the model to white noise, that is, errors with a mean of zero
and the correlation between any "t and "s is equal to zero. The classical reference
is [22], an open-source introductory is available at http://statistik.mathematik.uniwuerzburg.de/timeseries/, and an extended treatment can be found in [78]. Examples
in the literature on forecasting political conflict include [79, 125, 145, 146]; [124]
provides a general reference with respect to applications in political science.
The notation AR(p) refers to the autoregressive model of order p. The
AR(p) model is written

140

P.A. Schrodt et al.

Xt D c C

p
X

'i Xt i C "t

(4)

i D1

where '1 ; : : : ; 'p are the parameters of the model, c is a constant and "t is white
noise. In most applications, the error terms "t are assumed to be independent,
identically-distributed, and sampled from a normal distribution.
The notation MA.q/ refers to the moving average model of order q:
Xt D  C "t C

q
X

i "t i

(5)

i D1

where the 1 ; : : : ; q are the parameters of the model,  is the expectation of Xt

(often assumed to equal 0), and the "t ; "t 1 ; : : : are white noise error terms.
The notation ARMA.p; q/ refers to the model with p autoregressive terms and q
moving-average terms, where a moving average is defined by . This model contains
the AR.p/ and MA.q/ models,
Xt D c C "t C

p
X
i D1

'i Xt i C

q
X

i "t i

(6)

i D1

In practical terms, classical time series models are useful for phenomena where
the current value of the variable is usually highly dependent on past values. This
incremental behavior is characteristic of a great deal political behaviors, notably
public opinion, budgeting, and most conflict behavior: once a civil war or insurgency
gets going, it is likely to continue for a while, and once peace is firmly established,
that is also likely to the maintained. In such situations, the best predictor of the
variable at time t is its value at time t  1 Consequently, these models are very
widely used.
Incremental models, however, have a critical weakness: by definition, they
cannot explain sudden change, whether isolated incidents or the onset and cessation
of protracted behaviors. An autoregressive model can have a very good overall
predictive record but still miss the beginning and end of a conflict, and these may be
of greatest interest to decision-makers.
The classical time series literature places a great emphasis on the characteristics
of the error terms. While these are treated in the literature as random variables, in
practical terms, much of the error is simply the effects of variables that were not
included in the model. Because most social and demographic indicators and many
economic indicators are also very strongly autoregressivefor example indicators
such as infant mortality rate and literacy rarely change more than a percentage
point year to yearthese errors will strongly correlate with their lagged values,
hence the interest in the MA and ARMA models. Unfortunately, disentangling
the effects of autoregressive variables and autoregressive errors is extraordinarily
difficult in many circumstances, which in turn has led to the development of vectorautoregressive models, discussed in Sect. 3.3.

Forecasting Political Violence

141

Finally, one will occasionally encounter questions about co-integration and

non-stationarity in time series models [113]. This issue generally arises in
situations where exponential growth is occurring in the variable, as can happen
in situations of hyper-inflation, individuals who are really lucky in their choice
of stocks or really unlucky in their choice of real estate markets. While conflict
data will switch between modespeace, crisis and warand this is a form of nonstationarity (which refers to a change in the statistical characteristics of the errors),
this is not an issue of co-integration. Unfortunately, the tests for co-integrated time
series have a very high probability of incorrectly identifying it in situations where
the autoregressive coefficient is close to 1.0that is, the variables change very
little from time period to time periodand hence conflict time series can easily
be incorrectly identified as co-integrated.
A second common problem encountered in time series with conflict data are
time series cross sections (TSCS), where data are collected across a number of
cases across a period of time (for example, incidents of terrorism by country-year).
Because these cases are not independentin particular, contemporary trans-national
terrorist groups operate in multiple countriesthis will affect the estimates of the
coefficients and methods that do not adjust for this can give misleading results
[10, 11].

3.3 Vector Autoregression Models

A relatively recent modification of the classical time series approach has been
vector autoregression (VAR) models. Originally developed by Sims [154] and
used primarily for financial forecasting, VAR models were also used in some of
the earliest applications of sophisticated econometric modeling to assessment
of terrorist strategies [5254, 58] and have been used in a number of policyoriented studies in the post-9/11 period [28, 29, 5557, 59], as well as more general
applications to conflict early warning [26, 30, 63].
A p-th order VAR, denoted VAR.p/, is
yt D c C A1 yt 1 C A2 yt 2 C    C Ap yt p C et

(7)

where c is a k  1 vector of constants, Ai is a k  k matrix (for every i D 1; : : : ; p)

and et is a k  1 vector of error terms.
The distinctive characteristic of VAR approaches is their emphasis on response
functions rather than on the interpretation of individual coefficients. VAR takes
as a given the fact that if the specified model is correct, the relations between
the variablesor technically speaking, their colinearitywill make the uncertainty
in the estimated coefficients very large even when the predictive accuracy of the
system as a whole is quite high. Consequently, rather than trying to interpret
those coefficients, an analyst looks at the effectsincluding error bandsof
an exogenous change propagating over time through a series of inter-dependent

142

P.A. Schrodt et al.

variables. The downside of VAR is that it is very data-intensive and requires data
measured consistently across a large number of time points.
A more recent variation on VAR models has been the addition of switching
components, which allow for the impact of distinct regimes on the behavior of the
system [155]. While originally developed for the study of foreign exchange markets,
this is clearly applicable to political conflict situations where the system is likely
to behave differently in times of peace than during times of war. Bayesian versions
have also been developed [26] and applied to cases such as the conflict in the Middle
East [25, 27].

3.4 Event-History and Survival Models

A different approach to conventional time series and VAR, but somewhat similar
to the probabilistic interpretation of logit models, is the event-history or survival
model approach [24, 86]. Rather than trying to predict the value of a behavior at a
specific time, these models predict probabilities of events, and in particular look at
the change in probability over time of the next event occurring. The models were
thoroughly developed in the bio-medical literature to model disease prognosis
hence the term survivalprior to their application in political science. See
[23, 41, 42] for applications to political conflict.
The survival function is defined as S.t/ D Pr.T > t/ where t is time, T is a
random variable indicating when the eventfor example the next terrorist attack, or
the end of a terrorist organizationoccurs, and Pr is probability. In many models,
the focus of the analysis is the hazard function
.t/ dt D Pr.t  T < t C dt j T  t/ D

f .t/ dt
S 0 .t/ dt
D
S.t/
S.t/

(8)

which the event rate at time t conditional on survival until time T , T  t. From
these functions, information such as the average number of events in a given period
or average lifetime can be derived.
Of particular interest in survival analysis is the overall shape of the hazard
function. For example, a survival function which levels off indicates that once
an interval has occurred without an event occurring, it is less likely to occur
again, whereas a survival function that declines linearly means that the occurrence
or nonoccurrence provides no information. Other forms are also possiblefor
example many biological organisms exhibit a U-shaped hazard function, meaning
that mortality is high in the early stages of life, then becomes low for a period of
time, then rises again as the organism reaches its natural age limit.
A key downside of survival models is the difficulty that humans have in working
with probabilities [94, 161]. In addition, the models are relatively new to the study
of political conflict, due to earlier constraints of computational feasibility and the
availability of appropriate estimation software. However, with these constraints
removed, there are likely to be more such applications in the future.

Forecasting Political Violence

143

3.5 Rare-Events Models

In most studies of political violence, the occurrence of violence is rare relative to
the instances wherein violence did not occur. This poses a number of challenges to
the empirical analysis of terrorism. When the absolute number of observations (or
actual terrorism-events) in a sample is low, the rarity of terrorism can bias coefficient
estimates and underestimate the probability of terrorist events. Even When one
has a large sample of terrorist incidents, the relative rarity of these events and the
preponderance of zeroes that this rarity entails can pollute samples with non-event
observations that effectively have zero likelihood of ever experiencing terrorist
events. This zero inflation contaminates terrorism-event data with multiple (latent)
data generating processes, which when unaccounted for, can similarly bias estimates
and predicted probabilitiesoften again by understating the true occurrence of
terrorism.
Within logit models of binary terrorism data (see Sect. 3.1 above), scholars have
shown that information content deficiencies of these sorts are primarily a function of
two dynamics: the total number of observations within ones sample (N), and the
mixture of 1s and 0s therein [97]. In such instances, estimates will be biased downward due to an underestimation of the variance around P r.ZjY D 1/ and common
methods for computing predicted probabilities will exacerbate this bias through
their disregard for estimation uncertainty [97]. While such concerns may not at
first appear to be detrimental to big data analyses, they are nevertheless highly
relevant to large-N conflict studies given the growing importance of in-sample and
out-of-sample forecasting techniques in the discipline [12, 166]. Indeed, even for
large-N event data sets, the practices of subdividing data sets into smaller training,
test and validation-data sets will frequently necessitate that scholars address these
concerns in order to achieve stable and comparable estimates of conflict forecasts
across sub-samples.
SOMA behavior models and CAPE modelswhich we discuss in more detail in
Sect. 4.4 belowprovide one potential solution to this problem [107, 157]. These
models account for behavioral changes, and have been found to be especially useful
in modeling the behaviors of terrorist groups under the specific rare-event contexts
outlined above. Within the political science literature, an alternative approach is the
rare-events logit model, which addresses these challenges by treating rare-events
as a sampling design problem. Provided that the dataset has a representative (or
complete) sample of actual terrorism events, this approach calls for researchers to
then augment their samples of observed terrorism events with a comparable, random
sub-sample of all corresponding non-events. When this strategy is consistent with
the assumptions of case-control or case-cohort sampling approaches, unbiased
coefficient estimates and predictions can be derived by adjusting logit estimates
and predicted probabilities for the true ratio of events to non-events within the
population (or an estimate thereof) during model estimation [97]. This makes
such an approach particularly advantageous for empirical studies of terrorism that
(i) employ in-sample and out-of-sample forecasting techniques (ii) focus on a

144

P.A. Schrodt et al.

specific region or country over a short period of time, or (iii) require additional
variable coding or data collection that, due to time or money constraints, cannot be
completed for an entire sample of interest.
A second challenge to the analysis of rare-event data arises from the corresponding preponderance of zero (non-event) observations within such data sets. As recent
studies note [126], many zero-observations in terrorism-event data sets correspond
to country or region cases that have no probability of experiencing a terrorist-event
in any period of interest. In these instances, empirical analyses of conflict-event data
risk conflating two distinct types of zero-observations; one for which the probability
of terrorism is non-zero but terrorism-events nevertheless didnt occur, and one
wherein the probability of terrorist-events is consistently zero. Given that the latter
set of zero-cases often arise as the result of covariates that correlate or overlap with
ones primary independent variables of interest, ignoring zero-inflation processes of
these sorts not only leads to an underestimation of terrorist events generally, but can
also bias specific coefficient estimates in indeterminate directions. To correct for
these biases, one must conditionally model both the zero-inflation process and the
main outcome of interest.
Zero-inflated mixture-models specifically address these very problems, most
notably within the contexts of event-count models such as Poisson or negative
binomial estimators. In essence, these models employ a system of two equations
to estimate the combined probability of an observation (i) being inflated and
(ii) experiencing an event-outcome of interestusually by including separate but
overlapping covariates as predictors for each respective equation. For example, one
could expand the logit model presented in Eq. 13 above to the zero-inflated logit
approach by incorporating a second, inflation-stage logit equation as so,


1Yi

1
1
1
f .z; w/ D
1
C 1
1 C e w
1 C e w
1 C e z


Yi
1
1
 1
(9)
1 C e w
1 C e z
where the variable w represents the additional set of inflation-stage covariates;
w D 0 C 1 v1 C 2 v2 C 3 v3 C    C k vk ;

(10)

which may or may not overlap with z [13]. Zero-inflated models thereby add
an additional layer of nuance to the empirical modeling of conflict-events by
estimating both the propensity of ever experiencing an event of interest and the
likelihood of experiencing an event of interest conditional on being able to do so.
This allows one to use ex-ante observable and theoretically informed covariates
to account for the probability that a given zero observation is inflated, and to
then probabilistically discount these zeroes leverage within ones primary analysis
without dropping these observations entirely. While such zero-inflated modeling
approaches have been most extensively applied to political violence count-data
[14,40,126], zero-inflated models have also recently been developed and applied by
conflict-researchers to a variety of other limited dependent variables [83, 159, 168].

Forecasting Political Violence

145

4 Algorithmic Approaches
Although more traditional statistical models still dominate quantitative studies
of political conflict, algorithmic approaches have proven effective, thus gaining
momentum not just within political science [21, 139] but also in other disciplines.
For example, computer scientists have developed the CONVEX [157], CAPE [106],
and SOMA [107] tools to forecast terrorist group behavior. While we will follow
common practice in using the statistical vs. algorithmic distinction to differentiate between methodologies, there is overlap between the two definitions. For
example, linear regression is consider a canonical statistical approach, but as we
describe in Sect. 4.1.1, it is also a straightforward example of a supervised linear
algorithm.
In general, by algorithmic approaches, we refer to specific models (such as neural
networks or random forests) or techniques (like bagging and boosting) that attempt
to leverage computational power to specify and train models through iterative
resampling techniques and to build and assess out-of-sample predictions, rather than
obtaining a single estimate of the model coefficients. Algorithmic approaches can
provide a number of benefits over statistical models and are particularly relevant to
forecasting terrorism for at least the following four reasons.
First, machine learning algorithms are often better suited than many traditional
statistical models at handling big data data sets with large numbers of independent variables that potentially exceed the number of observations. Second, these
algorithms are also less dependent on rigid assumptions about the data generating
process and underlying distributions. Third, as opposed to some statistical models,
many machine learning algorithms were specifically designed to generate accurate
predictions, and do this exceedingly well. Finally, a number of the algorithmic
approaches approximate the widely used qualitative method case-based reasoning
[95,108,116] which match patterns of events from past cases to the events observed
in a current situation, and then use the best historical fit to predict the likely outcome
of the current situation; [134, Chap. 6] gives a much more extended discussion
of this approach. This similarity to the methods of human analysts accounted for
these methods originally being labeled artificial intelligence in some of the early
studies.
Indeed, major trends in the empirical study of political violence, such as the
big data revolution and an increasing interest in predictive models, mean that
algorithmic approaches will likely become increasingly popular in the coming
years. In the following sections, we address some of the most relevant machine
learning algorithms for forecasting political violence. Following standard practices,
we divide algorithmic approaches into two general, though not mutually exclusive
categories, supervised and unsupervised algorithms, with an additional discussion
of sequence analysis techniques.

146

P.A. Schrodt et al.

4.1 Supervised Cross-Sectional Classification Methods

Using the language of political science, supervised algorithms require an observed
dependent variable valuebe it binomial, continuous, or categoricaland at least
one independent variable for each observation. The vast majority of quantitative
studies of political violence employ data that meet the requirements of supervised
algorithms. In the following section, we discuss the two supervised algorithmic
approaches that have received the most attention in the political violence literature,
linear models and neural networks, and also discuss tree-based approaches that have
yet to take hold in the study of political violence but have proven successful at
forecasting in other disciplines.
4.1.1 Linear Approaches
Though generally thought of as a statistical approach, liner regression is also one of
the most common and straightforward examples of a supervised learning algorithm.
The dominance of linear regression among the broader group of existing supervised
linear approaches is not fully unwarranted, since linear regression contains nice
attributes, such as being BLUEthe Best Linear Unbiased Estimator. Additionally, and perhaps more importantly, linear regression approaches are widely taught,
straightforward to interpret, and easy to implement with one line of code (or a few
clicks of a drop down menu) in virtually every statistical package.
Consistent with the widespread use of the linear method, other supervised
approaches similarly utilize a line (in <3 dimensions) or a hyper plane (in >2
dimensions), such as linear discriminant analysis (LDA) and support vector machines (SVM). LDA is a supervised linear classification algorithm that produces
nearly identical results to linear regression when Y takes on two classes, but can
outperform multinomial logit and probit models as the number of classes exceeds 2
[80, pp. 106114]. SVMs are an additional linear classification supervised algorithm
that is most commonly used in problems such as text classification when the number
of covariates (i.e. words in the text example) exceeds the number of observations.
[49,142] demonstrate the strength of SVMs in classifying articles for the Militarized
Interstate Disputes 4.0 (MID4) project, while [90] uses SVM to extract events of
terrorism from text.
4.1.2 Neural Networks
Aside from traditional linear regression approaches, neural networks have arguably
received the most attention in quantitative political violence literature among
supervised learning algorithms due almost entirely to Beck, King and Zengs
important article on forecasting interstate conflict [12]. (Neural networks can also
be applied in an unsupervised setting: see [98].) Neural networks were originally
created to model complex data characterized by a large number of nodes that receive

Forecasting Political Violence

147

and process input functions and then transfer information to other nodes, such as
neurons in the case of the human brain.
Although most work with neural networks seems far removed from terrorism,
[12] articulately explain how supervised neural networks are not only an appropriate algorithmic approach to predicting violence but can also be applied as a
straightforward extension to logistic regression. Like a logistic regression, neural
networks can be applied to a traditional TSCS dataset with a binary dependent
variable to generate predicted probabilities that are interoperated identically to the
 parameter of logistic regression. However, the primary advantage of a neural
network approach is that they are able to account for the potential of massive
nonlinear interaction effects that may causally link the independent variables to
the outcome of interest without having to directly specify interactive or non-linear
terms to the model as required in a logistic. [12] demonstrate that the neural network
approach consistently outperforms logistic regression in out-of-sample accuracy
measures. Though we are unaware of neural networks being applied in studies of
terrorism, it is likely that doing so could yield similar improvements in predictive
accuracy.
4.1.3 Tree-Based Algorithms
Though yet to gain traction in the political violence literature, tree-based approaches
[80, pp. 305317, 587604] are commonly used to generate accurate predictions
in a host of other disciplines, including finance, medicine, computer science, and
sociology. In brief, tree-based algorithms operate by iteratively partitioning the data
into smaller sub-sets based on a break-point in a particular independent variable.
This process is repeated with the goal of creating bins of observations with similar
Yi values (for continuous data) or class (for categorical data).
We highlight three important factors that contribute to the strong predictive
accuracy of tree-based approaches. First, trees can be used to forecast both continuous (i.e. regression trees) and binomial (i.e. classification trees) dependent
variables. Second, tree-based approaches, such as random forests, are not sensitive
to degrees of freedom and can handle more independent variables than observations. Third, leading tree-based approaches incorporate iterative re-sampling,
weighting, and model averaging strategies like bagging and boosting techniques
[156], which tend to enhance accuracy and stability vis-`a-vis other supervised
learning classification and regression algorithms. In social science applications of
tree-based algorithms, [16] and [17] demonstrates that random forests can help
generate accurate forecasts of violent crime rates in the United States. Despite the
scarcity of tree-based approaches in political science, scholars in other disciplines
ranging from ecology (see [43]) to transportation studies (see [158]) have generated
accurate predictions using this approach. As the quantitative political violence
literature continues to progress, it will behoove scholars to continue experimenting
with supervised forecasting algorithms that have demonstrated their value in other

148

P.A. Schrodt et al.

disciplines. We believe that tree-based approaches, especially boosted regression

trees (BRTs) and boosted random forests, are particularly useful for scholars
attempting to forecasting terrorism.

4.2 Unsupervised Methods

Unlike supervised learning algorithms that train a model based on relationships
between a matrix of covariates and a corresponding vector of observed dependent
variables for each observation, unsupervised approaches are applied to datasets for
which dependent variables are latent and therefore not directly provided. Though
the concept of unsupervised algorithms may seem abstract, useful applications exist
in the study of political violence. We describe two types of particularly relevant
unsupervised learning algorithms in the following sections dimensionality reduction
and clustering approaches. (Hidden Markov Models (HMMs) can also be classified
as an unsupervised learning algorithm, but we present it in Sect. 4.3 as a sequencebased approach.)

4.2.1 Dimension Reduction

Factor analysis is a general algorithmic approach to reduce the dimensionality of
matrix of potential covariates by identifying latent attributes. Principal components
analysis (PCA) is a more restricted form of factor analysis and [91] is among the
oldest (circa 1901) and most commonly employed dimension reduction tools in the
social science [69]. In brief, the PCA algorithm works by analyzing relationships
among the covariates and using these to create a new set of orthogonal variables
called principal components that reflect the latent attributes and can be used in
place of the original set of covariates in a quantitative model. Certain models, like
linear regression, are sensitive to degrees of freedom and cannot operate when a
dataset contains more covariates than observations. In these instances, a researcher
can implement PCA or a similar approach to sufficiently reduce the dimensionality
of covariates to enable the linear regression to converge [2].

4.2.2 Clustering
Clustering approaches such as k-means and Latent Dirichlet Allocation are similar
to dimension reduction in that they attempt to identify latent classes amongst a
set of observations, but differ in that they identify discrete, rather than continuous
solutions like PCA and FA [69]. Within the machine learning literature, k-means
approaches are among the most commonly used clustering algorithms, though their
application to the study of political violence has been scarce [80, pp. 509520].

Forecasting Political Violence

149

However, [139] demonstrate that k-means can successfully identify latent clusters
within event data that identify phases of violence in the Middle East.
Latent Dirichlet Allocation [19] is another clustering algorithm that is primarily
applied to raw text. In the typical Latent Dirichlet Allocation application to
document classification, each document is assumed to be a mixture of multiple,
overlapping latent topics, each with a characteristic set of words. Classification is
done by associating words in a document with a pre-defined number of topics most
likely to have generated the observed distribution of words in the documents.
The purpose of LDA is to determine those latent topics from patterns in the data,
which are useful for two purposes. First, to the extent that the words associated with
a topic suggest a plausible category, they are intrinsically interesting in determining
the issues found in the set of documents. Second, the topics can be used with other
classification algorithms such as logistic regression, support vector machines or
discriminant analysis to classify new documents.
Despite the surface differences between the domains, the application of Latent
Dirichlet Analysis to the problem of political forecasting is straightforward: it is
reasonable to assume that the stream of events observed between a set of actors
is a mixture of a variety political strategies and standard operating procedures
(for example escalation of repressive measures against a minority group while
simultaneously making efforts to co-opt the elites of that group). This is essentially
identical to the process by which a collection of words in a document is a composite
of the various themes and topics, the problem Latent Dirichlet Analysis is designed
to solve. As before, the objective of Latent Dirichlet Analysis will be to find those
latent strategies that are mixed to produce the observed event stream. These latent
factors can then be used to convert full event stream to a much simpler set of
measures.
The Latent Dirichlet Analysis approach is similar in many ways to the hidden
Markov approach (Sect. 4.3). In both models, the observed event stream is produced
by a set of events randomly drawn from a mixture of distributions. In an HMM,
however, these distributions are determined by the state of a Markov chain, whose
transition probabilities must be estimated but which consequently also explicitly
provides a formal sequence. An Latent Dirichlet Analysis, in contrast, allows any
combination of mixtures, without explicit sequencing except to the extent that
sequencing information is provided by the events in the model.

4.3 Sequence Development: Hidden Markov Models

Hidden Markov models (HMM) are a type of stochastic signaling model that has
become increasingly popular in computing the probability that a noisy sequence
was generated by a known model, most notably in speech recognition and protein
sequence comparison. A detailed discussion and application of the method to the
problem of forecasting terrorism can be found in [123] in this volume; other
applications to political forecasting include [21, 133, 135, 144].

150

P.A. Schrodt et al.

An HMM is a variation on the well-known Markov chain model, one of the most
widely studied stochastic models of discrete events. Like a conventional Markov
chain, a HMM consists of a set of discrete states and a matrix A D aij of
transition probabilities for going between those states. In addition, however, every
state has a vector of observed symbol probabilities, B D bj .k/ that corresponds to
the probability that the system will produce a symbol of type k when it is in state j.
The states of the HMM cannot be directly observed and can only be inferred from
the observed symbols, hence the adjective hidden.
In empirical applications, the transition matrix and symbol probabilities of an
HMM are estimated using an iterative maximum likelihood technique called the
Baum-Welch algorithm which finds values for the matrices A and B that locally
maximize the probability of observing a set of training sequences. Once a set of
models has been estimated, they can be used to classify an unknown sequence by
computing the probability that each of the models generated the observed sequence.
The model with the highest such probability is chosen as the one which best
represents the sequence.
The application of the HMM to the problem of classifying international event
sequences is straightforward. The symbol set consists of the event codes taken
from an event data set such as IDEA [123] or CAMEO [143]. The states of the
model are unobserved, but have a close theoretical analog in the concept of crisis
phase [101]. Different political phases are distinguished by different distributions
of observed events from the event ontology. Using CAMEO coding as an example, a
stable peace would have a preponderance of cooperative events in the 0110 range
which codes cooperative events; a crisis escalation phase would be characterized by
events in the 1115 range (accusations, protests, denials, and threats), and a phase of
active hostilities would show events in the 1822 range, which codes violent events.
An important advantage of the HMM is that it can be trained by example
rather than by the deductive specification of rule. Furthermore, HMMs require no
temporal aggregation. This is particularly important for early warning problems,
where critical periods in the development of a crisis may occur over a week or
even a day. Finally, indeterminate time means that the HMM is relatively insensitive
to the delineation of the start of a sequence: It is simple to prefix an HMM with
a background state that simply gives the distribution of events generated by a
particular source (e.g. Reuters/IDEA) when no crisis is occurring and simply cycle
in this state until something important happens.

4.4 Sequence Analysis: Sequence Matching

Sequence analysis is a method that assesses the degree of similarity between two
or more 1xN vectors, commonly populated by numerical or categorical inputs.
Scientists first developed the approach as a tool to locate matches between different
genomic D.N.A. sequences [31, 46] and it has since been applied to various
disciplines including sociology [1] and demographics [102]. Although existing

Forecasting Political Violence

151

Table 1 Archetype sequence the precedes an attack against the base

Variable
Level of civilian casualties
Number of ambush attempts
against patrol units
Attack on base?

Weekt4
1
3

Weekt3
2
3

Weekt3
3
3

Weekt1
4
0

Weekt
N/A
N/A

No

no

No

No

Yes

studies using sequence analysis primarily focus on matching rather than forecasting,
Martinez et al. [106], Silva et al. [157], and Dorazio et al. [50], do utilize sequence
analysis-based algorithms for the explicit goal of prediction.

4.4.1 Archetypal Sequence Matching

Dorazio et al. suggest an archetype-driven approach to sequence analysis with
the explicit goal of predicting political violence that uses sequence analysis to
build features on which a statistical or algorithmic model produces out-of-sample
forecasts. For this archetype approach be succeed, it requires the existence of
a distinct pattern of events (i.e. an archetype) that tends to precede a given
outcome-of-interest. If such an archetype exists, then a model may be able to
forecast the outcome of interest based on the extent to which a sequence whose
outcome is unknown (for example, events being observed in real time) is similar
to the archetype sequence. The process of using this archetype-driven sequence
analysis approach for prediction follows four main steps:
1.
2.
3.
4.

Define the archetype(s)

Calculate distances
Train a model
Build out-of-sample forecasts

To more clearly explain this approach, we use a hypothetical example to

demonstrate each of the four steps above. Imagine our goal is to predict a binary
variable indicating whether or not an American military base will be attacked in
the upcoming week. Looking back, officers notice patterns in the level of civilian
casualties (scale of 14, 4 being most severe) and the number of attempted ambushes
on patrol missions that have occurred in peaceful weeks that precede an attack on
the base. Table 1 completes <Step 1> as it reflects the pattern of events that
tends to precede an attack against the base in terms of archetypical sequences. If
officers observe 4 week progressions in real time that are sufficiently similar to the
archetypical sequence of the level of civilian casualties and the number of ambush
attempts on patrols, then they can confidently predict that an attack against the base
will occur in the next week.
Just how similar must the observed sequences be to the archetypical sequences
in order to predict that the following week will experience an attack against the
base? To address this, a researcher must first complete <Step 2>, calculating
the mathematical degree of similarity (i.e. distance) between the archetypes and

152

P.A. Schrodt et al.

Fig. 1 Building co-variates with Euclidean distance

set of training sequence of equal length (i.e. 4). Euclidean distance is one of the
most common and robust measures, though many other distance measures exist.
Figure 1 demonstrates how Euclidean is applied to calculate distanceswhich serve
as the covariates in <Step 3>between training sequences and the archetypical
sequences.
The two distances, Di stanceci vi li ancasualty and Di stanceambushat t empt s are
calculated for every observation (i.e. 4 week sequence) in the training set. To
complete <Step 3>, we choose a logistic regression, which is suitable to a
parsimoniously specified model with a binary dependent variable. To train the
logistic model, estimate the values that maximize the likelihood function below.
L.jy/ D

N
Y

ni
y
i i .1  i /ni yi
y
.n

y
/
i
i
i
i D1

(11)

To complete <Step 4> and build actual forecasts based on the archetypedriven sequence analysis approach, we apply the two estimates that result from
(Eq. 6) to the logistic regression formula in order to calculate f .z/, which reflects
the likelihood that the week following the 4-week period used to generate the
Di stanceci vi li ancasualty and Di stanceambushat t empt s distances from the archetypes
will experience an attack on the base.
z D 0 C 1 Di stanceci vi li ancasualty C 2 Di stanceambushat t empt s
f .z/ D

1
1 C e z

(12)
(13)

To the extent that archetypical sequences that tend to precede an outcome-ofinterest exist, sequence analysis may be an effective tool of prediction that can
provide leverage over other approaches.

4.4.2 Convex Algorithms

In addition to the sequential archetype approach, Martinez et al. [106] put forth two
alternative, sequence-based algorithms (ConvexkN N and ConvexMerge) that have
proven successful at predicting terrorist activities. The key difference between these
algorithms and the previously discussed archetypical approach is that the former do
not assume the presence of an archetype. Thus, rather than calculating the distances
between out-of-sample sequences with unknown outcomes and an archetypical

Forecasting Political Violence

153

sequence, the Convex algorithm calculates the various distance measure between
each out-of-sample sequence and all in-sample sequences. In order to form out-ofsample predictions, the Convex algorithm determines the K-number of the most
similar sequences based on various distance measures. If all of the k-nearest
neighbor sequences have the same dependent variable, then the predicted value
for that out-of-sample sequence is assigned as the dependent variable value of
the single nearest neighbor. When this occurs, the ConvexkN N and ConvexMerge
approaches generate identical predictions. However, when the dependent variable
values of the nearest neighbor sequences are not identical, ConvexkN N rounds to the
nearest integer while ConvexMerge assigns more value to more similar sequences.
In addition to Martinez et al., Sliva et al. [157] apply both variants of the Convex
algorithm to terrorist group behavior.

5 Network Models
We label our final category network models: these are approaches that specifically
consider the relationships between entities, either according to their interactions
(SNA models) or location (geospatial models). As with many of the algorithmic
applications, network models have only become feasible in the last decade of
so, as they require very substantial amounts of data and computational capacity.
Consequently the number of applications at present is relatively small, though these
are very active research areas.

5.1 Social Network Analysis Models

Social network analysis (SNA) in a type of graph theory that models systems in
terms of nodes, which are the unit of analysis and can be anything from states in the
international system to members of a terrorist cell, and edges that connect the nodes.
The edges can reflect any type of interactions that occur between the nodes, such as
levels of international trade if states are the nodes or amount of communication
if the nodes reflect members of a terror cell. Though still an emerging method,
SNA is becoming particularly attractive approach for studying terrorism [122]. For
example, [7] use SNA to model the organizational structure of the terrorist groups
responsible for the 2008 Mumbai attacks and [47] develop an edge-based network
to forecast the source and target locations of transnational terrorist attacks. We
speculate three potential explanations for its rapid growth.
First, terrorist scholars are often interested in information regarding withingroup dynamics of terrorist organizationssuch as chains of command, lines of
communication, and policy preferences of leadershipwhich SNA may reveal more
effectively than statistical or algorithmic approaches. Second, the rise of social
networking platforms like Facebook, Twitter, less publicized online forums, and

154

P.A. Schrodt et al.

any other modes of communication open source or otherwise, has greatly increased
the amount of information that can be used to populate a social network. Third, SNA
has been almost simultaneously embraced by the intelligence, policy, and academic
communitiesa rare feat for a new methodology. These factors, as well as its ability
to deliver unique insights, is likely to continue driving its growth among quantitative
studies of terrorism.

5.2 Geo-spatial Models

Similar to our discussion of temporal dynamics above (Sects. 3.2 and 3.3), the
incorporation of spatial dynamics into statistical models of terrorism can often
enhance modeling and forecasting capabilities. Political violence frequently clusters
or diffuses in geographic space [35, 164, 166], and failure to account for these
dependencies can bias inferences [5, 164]. Spatial statistics allow one to model
the existence of dependencies or influences across spatial units with a dependency
matrix that measures geographic distances between units, which can then be
used to specify a spatial autoregressive model with spatially lagged dependent
variable or a spatially correlated error structure model [5, 164]. A number of recent
terrorism studies employ such approaches [115, 119]. Scholars have also tailored
these spatial statistics to the study of binary dependent variablesakin to the
binary terrorism events mentioned abovethrough the use of (e.g.) autologistic
models that incorporate distance-based connectivity matrices [164, 166]. Finally,
spatial forecast methods employing feature reduction techniques have recently been
applied to terrorism events in efforts to incorporate the explanatory power of high
dimensional geographic spaces into probabilistic forecasting models, and show
great promise in this regard [32, 128].

6 Conclusion
In 1954, the psychologist Paul Meehl [111] published a path-breaking analysis of
the relative accuracy of human clinical assessments versus simple statistical models
for a variety of prediction problems, such as future school performance and criminal
recidivism. Despite using substantially less information, the statistical models either
outperformed, or performed as well as, the human assessments in most situations.
Meehls work has been replicated and extended to numerous other domains in the
subsequent six decades and the results are always the same: the statistical models
win. Meehl, quoted in Kahneman [94, Chap. 21], reflecting on 30 years of studies,
said, There is no controversy in the social sciences which shows such a large body
of qualitatively diverse studies coming out in the same direction as this one.
This has not, of course, removed the human analysts from tryingand most
certainly, claimingto provide superior performance. These are justified in the face

Forecasting Political Violence

155

of overwhelming empirical evidence that they are inaccurate using a wide variety of
excusesdare we call them pathologies?that Kahneman [94, Part 3] discusses
in great detail in the section of his book titled, appropriately, Overconfidence.
It is not only that simple statistical models are superior to human punditry, but as
Tetlock [161] established, the most confident and a well-known human analysts
actually tend to make the worst predictions. Furthermore, in most fields, the
past performance of analystsnotoriously, the well-compensated stock-pickers of
managed mutual fundsprovides essentially no guide to their future performance.
At present, there is very little likelihood that human punditry, particularly the
opinionated self-assurance so valued in the popular media, will be completely
replaced by the unblinking assessments of computer programs, whether on 24-hour
news channels or in brainstorming sessions in windowless conference rooms.
Humans are social animals with exquisite skills at persuasion and manipulation;
computer programs simply are far more likely to provide the correct answer in an
inexpensive, consistent and transparent manner.
Yet with the vast increase in the availability of data, computational power, and
the resulting refinement of methodological techniques, there is some change in the
works. While the covers of investment magazines are adorned with the latest wellcoffed guru who by blind luck has managed to have an unusually good year, in
fact algorithmic trading now accounts for all but a small fraction of the activity on
financial exchanges. Weather forecasts are presented on television by jocular and
attractive individuals with the apparent intelligence of tadpoles, but the forecasts
themselves are the result of numerical models processing massive quantities of data.
In the political realm, sophisticated public opinion polling replaced the intuitive
hunches of experts several decades ago, and polls are so rarely incorrect that it is a
major news event when they fail. Even that last bastion of intuitive manliness, the
assessment of athletes, can be trumped by statistical models, as documented in the
surprisingly popular book and movie Moneyball [100].
Similar progress in the adoption of models forecasting political violence, particularly terrorism, is likely to be much slower. As we have stressed repeatedly, one
of the major challenges of this research is that political violence is a rare event,
and acts of terrorism in otherwise peaceful situationsthe bolt out of the blue of
Oklahoma City, 9/11/2001, Madrid 2004 and Utya Island, Norway in 2011are
among the rarest. Consequently even if the statistical models are more accurate
and there is every reason to believe that they will beestablishing this will take
far longer than is required in a field where new predictions can be assessed by the
day, or even by the minute. In addition, a long series of psychological studies have
shown that human risk assessment is particularly inaccurateand hence its validity
overestimatedin low probability, high risk situations, precisely the domain of
counter-terrorism. Getting the technical assessments in the door, to say nothing of
getting them used properly, will not be an easy task, and the initial applications
will almost certainly be in domains where the events of interest are more frequent,
as we are already seeing with the success of the Political Instability Task Force.
But as this chapter has illustrated, the challenges are well understood, a plethora of

156

P.A. Schrodt et al.

sophisticated techniques await experimentation, and the appropriate data is readily

available, all that is needed is the will and skill to apply these.
Acknowledgements This research was supported in part by a grant from the U.S. National
Science Foundation, SES-1004414, and by a Fulbright-Hays Research Fellowship for work by
Schrodt at the Peace Research Institute, Oslo (http://www.prio.no).

References
1. Abbott A (1995) Sequence analysis: new methods for old ideas. Ann Rev Sociol 21:93113
2. Abdi H, Williams LJ (2010) Principal components analysis. Wiley Interdiscip Rev 2(4):
433459
3. Achen C (2002) Toward a new political methodology: microfoundations and ART. Ann Rev
Pol Sci 5:423450
4. Andriole SJ, Hopple GW (1984) The rise and fall of events data: from basic research to
applied use in the U.S. Department of Defense. Int Interact 10(3/4):293309
5. Anselin L (1988) Spatial econometrics: methods and models. Kluwer, Dordrecht
6. Asal V, Pate A, Wilkenfeld J (2008) Minorities at risk organizational behavior data and
codebook version 92008. http://www.cidcm.umd.edu/mar/data.asp
7. Azad S, Gupta A (2011) A quantitative assessment on 26/11 mumbai attack using social
network analysis. J Terror Res 2(2):414
8. Azar EE (1980) The conflict and peace data bank (COPDAB) project. J Confl Resolut 24:
143152
9. Azar EE, Sloan T (1975) Dimensions of interaction. University Center for International
Studies, University of Pittsburgh, Pittsburgh
10. Beck N, Katz. JN (1995) What to do (and not to do) with time- series cross-section data. Am
Pol Sci Rev 89:634634
11. Beck N, Katz JN, Tucker R (1998) Taking time seriously: time-series-cross-section analysis
with a binary dependent variable. Am J Pol Sci 42(4):12601288
12. Beck N, King G, Zeng L (2000) Improving quantitative studies of international conflict: a
conjecture. Am Pol Sci Rev 94(1):2135
13. Beger A, DeMeritt JH, Hwang W, Moore WH (2011) The split population logit (spoplogit):
modeling measurement bias in binary data. http://ssrn.com/abstract=1773594. Working paper
14. Benini AA, Moulton LH (2004) Civilian victims in an asymmetrical conflict: operation
enduring freedom, afghanistan. J Peace Res 41(4):403422
15. Bennett DS, Stam AC (2004) The behavioral origins of war. University of Michigan Press,
Ann Arbor
16. Berk R (2009) The role of race in forecasts of violent crime. Race Soc Probl 1(4):231242
17. Berk R, Sherman L, barnes G, Kurtz E, Ahlman L (2007) Forecasting murder within a
population of probationers and parolees: A high stakes application of statistical learning.
J R Stat Soc 172(1):191211
18. Besley T, Persson T (2009) Repression or civil war? Am Econ Rev 99(2):292297
19. Blei DM, Ng AY, Jordan MI (2003) Latent dirichlet allocation. J Mach Learn Res 2:9931022
20. Bond D, Jenkins JC, Taylor CLT, Schock K (1997) Mapping mass political conflict and civil
society: Issues and prospects for the automated development of event data. J Confl Resolut
41(4):553579
21. Bond J, Petroff V, OBrien S, Bond D (2004) Forecasting turmoil in Indonesia: an application
of hidden Markov models. Presented at the International Studies Association Meetings,
Montreal

Forecasting Political Violence

157

22. Box GE, Jenkins GM, Reinsel GC (1994) Time series analysis: forecasting and control, 3rd
edn. Prentice Hall, Englewood Cliffs
23. Box-Steffensmeier J, Reiter D, Zorn C (2003) Nonproportional hazards and event history
analysis in international relations. J Confl Resolut 47(1):3353
24. Box-Steffensmeier JM, Jones BS (2004) Event history modeling: a guide for social scientists.
Analytical methods for social research. Cambridge University Press, Cambridge
25. Brandt PT, Colaresi MP, Freeman JR (2008) The dynamics of reciprocity, accountability and
credibility. J Confl Resolut 52(3):343374
26. Brandt PT, Freeman JR (2006) Advances in Bayesian time series modeling and the study of
politics: Theory testing, forecasting, and policy analysis. Pol Anal 14(1):136
27. Brandt PT, Freeman JR, Schrodt PA (2011) Real time, time series forecasting of inter- and
intra-state political conflict. Confl Manage Peace Sci 28(1):4164
28. Brandt PT, Sandler T (2009) Hostage taking: understanding terrorism event dynamics.
J Policy Model 31(5):758778
29. Brandt PT, Sandler T (2010) What do transnational terrorists target? Has it changed? Are we
safer? J Confl Resolut 54(2):214236
30. Brandt PT, Williams JT (2007) Multiple time series models. Sage, Thousand Oaks
31. Brochet J, Marie-Paule-Lefrance, Guidicelli V (2008) Imgt/v-quest: the highly customized
and integrated system for ig and tr standardized v-j and v-d-j sequence analysis. Nucl Acids
Res 36:387395
32. Brown D, Dalton J, Hoyle H (2004) Spatial forecast methods for terrorist events in urban
environments. Intell Secur Inform 3073:426435
33. Buhaug H (2006) Relative capability and rebel objective in civil war. J Peace Res 43(6):
691708
34. Buhaug H, Lujala P (2005) Accounting for scale: measuring geography in quantitative studies
of civil war. Pol Geogr 24(4):399418
35. Buhaug H, Rod JK (2006) Local determinants of African civil wars, 19702001. Pol Geogr
25(6):315335
36. Burgess PM, Lawton RW (1972) Indicators of international behavior: an assessment of events
data research. Sage, Beverly Hills
37. Cederman LE, Gleditsch KS (2009) Introduction to special issue of disaggregating civil
war. J Confl Res 24(4):590617
38. Choi K, Asal V, Wilkenfeld J, Pattipati KR (2012) Forecasting the use of violence by
ethno-political organizations: middle eastern minorities and the choice of violence. In:
Subrahmanian V (ed) Handbook on computational approaches to counterterrorism. Springer,
New York
39. Choucri N, Robinson TW (eds) (1979) Forecasting in international relations: theory, methods,
problems, prospects. Freeman, San Francisco
40. Clark DH, Regan PM (2003) Opportunities to fight: a statistical technique for modeling
unobservable phenomena. J Confl Resolut 47(1):94115
41. Cunningham DE (2011) Barriers to Peace in Civil War. Cambridge University Press,
Cambridge
42. Cunningham DE, Gleditsch KS, Salehyan I (2009) It takes two: a dyadic analysis of civil war
duration and outcome. J Confl Resolut 53(4):570597
43. Cutler DR, Edwards TC Jr, Beard KH, Cutler A, Hess KT, Gibson J, Lawler J (2007) Random
forests for classification in ecology. Ecology 88(11):27832792
44. Daly JA, Andriole SJ (1980) The use of events/interaction research by the intelligence
community. Policy Sci 12:215236
45. Davies JL, Gurr TR (eds) (1998) Preventive measures: building risk assessment and crisis
early warning. Rowman and Littlefield, Lanham
46. Dereux J, Haeberli P, Smithies O (1984) A comprehensive set of sequence analysis programs
for the vax. Nucl Acids Res 12(1):387395

158

P.A. Schrodt et al.

47. Desmarais BA, Cranmar SJ (2011) Forecasting the locational dynamics of transnational terrorism: a network analysis approach. In: Paper presented at the 2011 European
Intelligence and Security Conference, Athens. http://ieeexplore.ieee.org/xpls/abs all.jsp?
arnumber=6061174&tag=1
48. Dickerson JP, Simari GI, Subrahmanian VS (2012) Using temporal probabilistic rules to
learn group behavior. In: Subrahmanian V (ed) Handbook on computational approaches to
counterterrorism. Springer, New York
49. DOrazio V, Landis ST, Palmer G, Schrodt PA (2011) Separating the wheat from the chaff:
application of support vector machines to mid4 text classification. Paper presented at the
annual meeting of the midwest political science association, Chicago
50. DOrazio V, Yonamine J, Schrodt PA (2011) Predicting intra-state conflict onset: an events
data approach using euclidean and levenshtein distance measures. Presented at the midwest
political science association, Chicago. Available at http://eventdata.psu.edu
51. Duval RD, Thompson WR (1980) Reconsidering the aggregate relationship between size,
economic development, and some types of foreign policy behavior. Am J Pol Sci 24(3):
511525
52. Enders W, Parise GF, Sandler T (1992) A time-series analyssis of transnational terrorism. Def
Peace Econ 3(4):305320
53. Enders W, Sandler T (1993) The effectiveness of anti-terrorism policies: a vectorautoregression-intervention analysis. Am Pol Sci Rev 87(4):829844
54. Enders W, Sandler T (2000) Is transnational terrorism becoming more threatening? A timeseries investigation. J Confl Resolut 44(3):307332
55. Enders W, Sandler T (2005) After 9/11: is it all different now? J Confl Resolut 49(2):259277
56. Enders W, Sandler T (2006) Distribution of transnational terrorism among countries by
income classes and geography after 9/11. Int Stud Q 50(2):367393
57. Enders W, Sandler T (2006) The political economy of terrorism. Cambridge University Press,
Cambridge
58. Enders W, Sandler T, Cauley J (1990) Assessing the impact of terrorist-thwarting policies: an
intervention time series approach. Def Peace Econ 2(1):118
59. Enders W, Sandler T, Gaibulloev K (2011) Domestic versus transnational terrorism: Data,
decomposition, and dynamics. J Peace Res 48(3):319337
60. Esty DC, Goldstone JA, Gurr TR, Harff B, Levy M, Dabelko GD, Surko P, Unger AN
(1998) State failure task force report: phase II findings. Science Applications International
Corporation, McLean
61. Esty DC, Goldstone JA, Gurr TR, Surko P, Unger AN (1995) State failure task force report.
Science Applications International Corporation, McLean
62. Freedom House (2009) Freedom in the world. http://www.freedomhouse.org/reports
63. Freeman JR (1989) Systematic sampling, temporal aggregation, and the study of political
relationships. Pol Anal 1:6198
64. Gaddis JL (1992) International relations theory and the end of the cold war. Int Secur 17:558
65. Gaddis JL (1992) The United States and the end of the cold war. Oxford University Press,
New York
66. Gassebner M, Luechinger S (2011) Lock, stock, and barrel: a comprehensive assessment of
the determinants of terror. Public Choice 149:235261
67. Gerner DJ, Schrodt PA, Francisco RA, Weddle JL (1994) The machine coding of events from
regional and international sources. Int Stud Q 38:91119
(2009) Conflict and mediation event observations
68. Gerner DJ, Schrodt PA, Yilmaz O
(CAMEO) Codebook. http://eventdata.psu.edu/data.dir/cameo.html
69. Ghahramani Z (2004) Unsupervised learning. In: Advanced lectures on machine learning.
Springer, Berlin/New York, pp 72112
70. Gill J (1999) The insignificance of null hypothesis significance testing. Pol Res Q 52(3):
647674
71. Gill J (2003) Bayesian methods: a social and behavioral sciences approach. Chapman and
Hall, Boca Raton

Forecasting Political Violence

159

72. Gleditsch NP (2012) Special issue: event data. Int Interact 38(4)
73. Goldstein JS (1992) A conflict-cooperation scale for WEIS events data. J Confl Resolut
36:369385
74. Goldstone JA, Bates R, Epstein DL, Gurr TR, Lustik M, Marshall, MG, Ulfelder J, Woodward
M (2010) A global model for forecasting political instability. Am J Pol Sci 54(1):190208
75. Gurr TR, Harff B (1994) Conceptual, research and policy issues in early warning research: an
overview. J Ethno-Dev 4(1):315
76. Gurr TR, Lichbach MI (1986) Forecasting internal conflict: A competitive evaluation of
empirical theories. Comp Pol Stud 19:338
77. Hamerli A, Gattiker R, Weyermann R (2006) Conflict and cooperation in an actors network
of Chechnya based on event data. J Confl Resolut 50(159):159175
78. Hamilton J (1994) Time series analysis. Princeton University Press, Princeton
79. Harff B, Gurr TR (2001) Systematic early warning of humanitarian emergencies. J Peace Res
35(5):359371
80. Hastie T, Tibshirani R, Friedman J (2009) The elements of statistical learning. Springer,
New York
81. Hegre H, Ellingson T, Gates S, Gleditsch NP (2001) Toward a democratic civil peace?
democracy, political change, and civil war, 18161992. J Policy Model 95(1):3448
82. Hegre H, Ostby G, Raleigh C (2009) Property and civil war: a disaggregated study of liberia.
J Confl Resolut 53(4):598623
83. Hill DW, Bagozzi BE, Moore WH, Mukherjee B (2011) Strategic incentives and modeling
bias in ordinal data:the zero-inflated ordered probit (ziop) model in political science. In: Paper
presented at the new faces in political methodology meeting, Penn State, 30 April 2011. http://
qssi.psu.edu/files/NF4Hill.pdf
84. Holmes JS, Pineres SAGD, Curtina KM (2007) A subnational study of insurgency: farc
violence in the 1990s. Stud Confl Terror 30(3):249265
85. Hopple GW, Andriole SJ, Freedy A (eds) (1984) National security crisis forecasting and
management. Westview, Boulder
86. Hosmer DW, Lemeshow S, May S (2008) Applied survival analysis: regression modeling of
time to event data. Series in probability and statistics. Wiley, New York
87. Hudson V (ed) (1991) Artificial intelligence and international politics. Westview, Boulder
88. Hudson VM, Schrodt PA, Whitmer RD (2008) Discrete sequence rule models as a social
science methodology: an exploratory analysis of foreign policy rule enactment within
Palestinian-Israeli event data. Foreign Policy Anal 4(2):105126
89. International Monetary Fund. Statistics Dept. (2009) Direction of trade statistics yearbook,
2008. http://www.imf.org/external/pubs/cat/longres.cfm?sk=22053.0
90. Inyaem U, Meesad P, Haruechaiyasak C (2009) Named-entity techniques for terrorism event
extraction and classification. In: Paper presented at the eighth international symposium on
natural language processing. http://ieeexplore.ieee.org/xpls/abs all.jsp?arnumber=5340924&
tag=1
91. Jackman RW, Miller RA (1996) A renaissance of political culture? Am J Pol Sci 40(3):
632659
92. Jackman S (2009) Bayesian analysis for the social sciences. Wiley, Chichester
93. Jenkins CJ, Bond D (2001) Conflict carrying capacity, political crisis, and reconstruction.
J Confl Resolut 45(1):331
94. Kahneman D (2011) Thinking fast and slow. Farrar, Straus and Giroux, New York
95. Khong YF (1992) Analogies at war. Princeton University Press, Princeton
96. King G (1986) How not to lie with statistics: avoiding common mistakes in quantitative
political science. Am J Pol Sci 30(3):666687
97. King G, Zeng L (2001) Logistic regression in rare events data. Pol Anal 9(2):1254
98. Lee K, Booth D, Alam P (2005) A comparison of supervised and unsupervised neural
networks in predicting bankruptcy of korean firms. Expert Syst Appl 29:116
99. Leng RJ (1987) Behavioral correlates of war, 18161975. (ICPSR 8606). Inter-University
Consortium for Political and Social Research, Ann Arbor

160

P.A. Schrodt et al.

100. Lewis M (2011) Moneyball. Norton, New York

101. Lund MS (1996) Preventing violent conflicts: a strategy for preventive diplomacy. United
States Institute for Peace, Washington, DC
102. Malo MA, Munoz-Bullon F (2003) Employment status mobility from a life-cycle perspective:
a sequence analysis of work-histories in the bhps. Demogr Res 9:119162
103. Mannes A (2012) Qualitative analysis & computational techniques for the counter-terror analyst. In: Subrahmanian V (ed) Handbook on computational approaches to counterterrorism.
Springer, New York
104. Marshall MG, Jaggers K, Gurr TR (2010) Polity iv project: political regime characteristics
and transitions, 18002010. Center for Systemic Peace, Vienna, VA
105. Martinez MV, Silva A, Simari GI, Subrahmanian VS (2012) Forecasting changes in terror
group behavior. In: Subrahmanian V (ed) Handbook on computational approaches to counterterrorism. Springer, New York
106. Martinez V, Simari G, Sliva A, Subrahmanian V (2008) Convex: similarity-based algorithms
for forecasting group behavior. IEEE Intell Syst 23:5157
107. Martinez V, Simari G, Sliva A, Subrahmanian V (2008) The soma terror organization portal
(stop): Social network and analytic tools for the real-time analysis of terror groups. In:
Liu H, Salerno JJ, Young M (eds) Social computing, behavioral modeling, and prediction.
Springer/Wien, Norderstedt, pp 918
108. May ER (1973) Lessons of the past: the use and misuse of history in American foreign
policy. Oxford University Press, New York
109. McClelland CA (1967) World-event-interaction-survey: a research project on the theory and
measurement of international interaction and transaction. University of Southern California,
Los Angeles, CA
110. McClelland CA (1976) World event/interaction survey codebook (ICPSR 5211). InterUniversity consortium for political and social research, Ann Arbor
111. Meehl P (1954) Clinical and statistical prediction: a theoretical analysis and a review of the
evidence. University of Minnesota Press, Minneapolis
112. Minorities at Risk Project (2009) Minorities at risk dataset. Center for International
Development and Conflict Management, College Park. http://www.cidcm.umd.edu/mar/
113. Murray MP (1994) A drunk and her dog: an illustration of cointegration and error correction. http://www-stat.wharton.upenn.edu/steele/Courses/434/434Context/Co-int%egration/
Murray93DrunkAndDog.pdf
114. National Consortium for the Study of Terrorism and Responses to Terrorism (START) (2011)
Global terrorism database [data file]. http://www.start.umd.edu/gtd
115. Neumayer E, Plumper T (2010) Galtons problem and contagion in international terrorism
along civilizational lines. Confl Manage Peace Sci 27(4):308325
116. Neustadt RE, May ER (1986) Thinking in time: the uses of history for decision makers. Free
Press, New York
117. OBrien S (2012) A multi-method approach for near real time conflict and crisis early warning. In: Subrahmanian V (ed) Handbook on computational approaches to counterterrorism.
Springer, New York
118. OBrien SP (2010) Crisis early warning and decision support: Contemporary approaches and
thoughts on future research. Int Stud Rev 12(1):87104

119. Ocal
N, Yildirim J (2010) Regional effects of terrorism on economic growth in turkey:
A geographically weighted regression. J Peace Res 47(4):477489
120. Oskar Engene J (2007) Five decades of terrorism in europe: the tweed dataset. J Peace Res
44(1):109121
121. Pearl J (2009) Understanding propensity scores. In: Causality: models, reasoning, and
inference, 2nd edn. Cambridge University Press, Cambridge/New York
122. Perliger A, Pedahzur A (2011) Social network analysis in the study of terrorism and political
violence. Pol Sci Pol 44(1):4550

Forecasting Political Violence

161

123. Petroff V, Bond J, Bond D (2012) Using hidden markov models to predict terror before it hits
(again). In: Subrahmanian V (ed) Handbook on computational approaches to counterterrorism. Springer, New York
124. Pevehouse JC, Brozek J (2010) Time series analysis in political science. In: BoxSteffensmeier J, Brady H, Collier D (eds) Oxford handbook of political Methodology. Oxford
University Press, New York
125. Pevehouse JC, Goldstein JS (1999) Serbian compliance or defiance in Kosovo? Statistical
analysis and real-time predictions. J Confl Resolut 43(4):538546
126. Piazza J (2011) Poverty, minority economic discrimination and domestic terrorism. J Peace
Res 48(3):339353
127. Piazza J, Walsh JI (2009) Transnational terrorism and human rights. Int Stud Q 53:125148
128. Porter MD, Brown DE (2007) Detecting local regions of change in high-dimensional criminal
or terrorist point processes. Comput Stat Data Anal 51(5):27532768
129. Raleigh C, Linke A, Hegre H, Karlsen J (2010) Introducing ACLED: an armed conflict
location and event dataset. J Peace Res 47(5):651660
130. Regan P, Clark D (2010) The institutions and elections project data collection. http://www2.
binghamton.edu/political-science/institutions-and-elections%-project.html
131. Richardson LF (1960) Statistics of deadly quarrels. Quadrangle, Chicago
132. Rupesinghe K, Kuroda M (eds) (1992) Early warning and conflict resolution. St. Martins,
New York
133. Schrodt PA (2000) Pattern recognition of international crises using hidden Markov models.
In: Richards D (ed) Political complexity: nonlinear models of politics. University of Michigan
Press, Ann Arbor, pp 296328
134. Schrodt PA (2004) Detecting united states mediation styles in the middle east, 19791998.
In: Maoz Z, Mintz A, Morgan TC, Palmer G, Stoll RJ (eds) Multiple paths to knowledge in
international relations. Lexington Books, Lexington, pp 99124
135. Schrodt PA (2006) Forecasting conflict in the Balkans using hidden Markov models. In:
Trappl R (ed) Programming for peace: computer-aided methods for international conflict
resolution and prevention. Kluwer, Dordrecht, pp 161184
136. Schrodt PA (2007) Inductive event data scaling using item response theory. Presented at the
summer meeting of the society of political methodology. Available at http://eventdata.psu.edu
137. Schrodt PA (2010) Seven deadly sins of contemporary quantitative analysis. Presented at the
american political science association Meetings, Washington, DC. http://eventdata.psu.edu/
papers.dir/Schrodt.7Sins.APSA10.pdf
138. Schrodt PA, Gerner DJ (1994) Validity assessment of a machine-coded event data set for the
Middle East, 19821992. Am J Pol Sci 38:825854
139. Schrodt PA, Gerner DJ (1997) Empirical indicators of crisis phase in the Middle East,
19791995. J Confl Resolut 25(4):803817
140. Schrodt PA, Gerner DJ (2004) An event data analysis of third-party mediation. J Confl Resolut
48(3):310330
Simpson EM (2001) Analyzing the dynamics
141. Schrodt PA, Gerner DJ, Abu-Jabr R, Yilmaz, O,
of international mediation processes in the Middle East and Balkans. Presented at the
American Political Science Association meetings, San Francisco
142. Schrodt PA, Palmer G, Hatipoglu ME (2008) Automated detection of reports of militarized
interstate disputes using the svm document classification algorithm. Paper presented at
American Political Science Association, San Francisco
143. Schrodt PA, Van Brackle D (2012) Automated coding of political event data. In: Subrahmanian V (ed) Handbook on computational approaches to counterterrorism. Springer, New York
144. Shearer R (2006) Forecasting Israeli-Palestinian conflict with hidden Markov models.
Available at http://eventdata.psu.edu/papers.dir/Shearer.IP.pdf
145. Shellman S (2000) Process matters: Conflict and cooperation in sequential governmentdissident interactions. Journal of Conflict Resolution 15(4), 563599
146. Shellman S (2004) Time series intervals and statistical inference: the effects of temporal
aggregation on event data analysis. Secur Stud 12(1):97104

162

P.A. Schrodt et al.

147. Shellman S, Hatfield C, Mills M (2010) Disaggregating actors in intrastate conflict. J Peace
Res 47(1):8390
148. Shellman S, Stewart B (2007) Predicting risk factors associated with forced migration: An
early warning model of Haitian flight. Civil Wars 9(2):174199
149. Silke A (ed) (2004) Research on terrorism: trends, achievements and failures. Frank Cass,
London
150. Silke A (2009) Contemporary terrorism studies: issues in research. In: Jackson R, Smyth MB,
Gunning J (eds) Critical terrorism studies: a new research agenda. Routledge, London
151. Silva A, Simari G, Martinez V, Subrahmanian VS (2012) SOMA: Stochastic opponent
modeling agents for forecasting violent behavior. In: Subrahmanian V (ed) Handbook on
computational approaches to counterterrorism. Springer, New York
152. Simari GI, Earp D, Martinez MV, Silva A, Subrahmanian VS (2012) Forecasting grouplevel actions using similarity measures. In: Subrahmanian V (ed) Handbook on computational
approaches to counterterrorism. Springer, New York
153. Simi P (2010) Operation and structure of right-wing extremist groups in the united states,
19802007. http://www.icpsr.umich.edu/icpsrweb/NACJD/studies/25722/detail
154. Sims CA (1980) Macroeconomics and reality. Econometrica 1, 6344
155. Sims CA, Waggoner DF, Zha TA (2008) Methods for inference in large multiple-equation
Markov-switching models. J Econom 146(2):255274
156. Siroky DS (2009) Navigating random forests and related advances in algorithmic modeling.
Stat Surv 3:147163
157. Sliva A, Subrahmanian V, Martinez V, Simari G (2009) Cape: Automatically predicting
changes in group behavior. In: Memon N, Farley JD, Hicks DL, Rosenorn T (eds) Mathematical methods in counterterrorism. Springer/Wien, Norderstedt, pp 253269
158. Sun S, Zhang C (2007) The selective random subspace predictor for traffic flow forecasting.
IEEE Trans Intell Transp Syst 8(2):367373
159. Svolik MW (2008) Authoritarian reversals and democratic consolidation. Am Pol Sci Rev
102(2):153168
160. Sylvan DA, Chan S (1984) Foreign policy decision making: perception, cognition and
artificial intelligence. Praeger, New York
161. Tetlock PE (2005) Expert political judgment: how good is it? How can we know? Princeton
University Press, Princeton
162. Urdal H (2008) Population, resources and violent conflict: A sub-national study of india
19562002. J Confl Resolut 52(4):590617
163. U.S. National Counterterrorism Center (2009) Worldwide incidents tracking system (wits).
http://wits.nctc.gov/
164. Ward MD, Gleditsch KS (2002) Location, location, location: an mcmc approach to modeling
the spatial context of war and peace. Pol Anal 10(3):244260
165. Ward MD, Greenhill BD, Bakke KM (2010) The perils of policy by p-value: predicting civil
conflicts. J Peace Res 47(5):363375
166. Weidman NB, Ward MD (2010) Predicting conflict in space and time. J Confl Resolut
54(6):883901
167. Weidmann NB, Toft MD (2010) Promises and pitfalls in the spatial prediction of ethnic
violence: a comment. Confl Manage Peace Sci 27(2):159176
168. Xiang J (2010) Relevance as a latent variable in dyadic analysis of conflict. J Pol 72(2):
484498
169. Young JK, Findley MG (2011) Promise and pitfalls of terrorism research. Int Stud Rev
13:411431
170. Yonamine JE (2012) Working with even data: a guide to aggregation. Available at http://jayyonamine.com/wp-content/uploads/2012/06/Working-with-Event-Data-AGuide-to-Aggregation-Choices.pdf