The shape of things to come? Expanding the inequality and grievance model for civil war forecasts with event data
| DOI | 10.1177/0022343316684192 |
| Author | Daina Chiba,Kristian Skrede Gleditsch |
| Date | 01 March 2017 |
| Published date | 01 March 2017 |
| Subject Matter | Research Articles |
The shape of things to come? Expanding
the inequality and grievance model for civil
war forecasts with event data
Daina Chiba
Department of Government, University of Essex
Kristian Skrede Gleditsch
Department of Government, University of Essex & Peace Research Institute Oslo (PRIO)
Abstract
We examine if dynamic information from event data can help improve on a model attempting to forecast civil war
using measures reflecting plausible motivation and grievances. Buhaug, Cederman, and Gleditsch predict the risk of
civil war using a horizontal inequality model with measures reflecting motivation and relevant group characteristics at
the country level. The predictions from their model outperform in an out-of-sample forecast conventional country-
level models of civil war, emphasizing vertical inequality and country characteristics. However, most grievance
measures change little over time. We surmise that a model reflecting potential motivation for conflict can be
improved with more dynamic information on mobilization and the behavior of actors. Our conjecture receives
some support in the empirical analysis, where we consider both conflict onset and termination over territorial and
governmental incompatibilities in the Uppsala/PRIO Armed Conflict Data, and find some evidence that event data
can help improve forecasts. Moreover, models with the original grievance measures do better than purely event based
models, supporting our claim that both structure and event based components can add value to conflict prediction
models. However, the contribution of events to improving predictive power is modest and not entirely consistent,
and some types of conflict events seem easier to forecast than others.
Keywords
civil war, prediction
Introduction
We examine if more dynamic information on interac-
tions from event data can help improve the predictive
ability of models focusing on grievances and inequality.
Buhaug, Cederman & Gleditsch (2014) argue that
efforts to predict civil wars can be improved by more
attention to actors and their grievances. They propose
better theoretically informed measures at the country
level, reflecting motivation and group characteristics rel-
evant to civil war, and show that the predictions from the
proposed model outperform conventional country-level
models of civil war, emphasizing country characteristics,
both for in-sample classification and in an out-of-sample
forecast. However, since most of the grievance measures
change little over time, the model identifies primarily
‘structural risk’ or potential motivation and opportunity
for conflict. We surmise that such a model can be
improved with more dynamic information on mobiliza-
tion and the behavior of actors. Stated more poetically,
attention to actors and their grievances may help us
appreciate the shapes of different types of pegs and holes
when it comes to the motivation for conflict, but more
attention to change and the behavior of actors may help
identify the shape of things to come in their interaction.
To anticipate, we will present results that partially
Corresponding author:
ksg@essex.ac.uk
Journal of Peace Research
2017, Vol. 54(2) 275–297
ªThe Author(s) 2017
Reprints and permission:
sagepub.co.uk/journalsPermissions.nav
DOI: 10.1177/0022343316684192
journals.sagepub.com/home/jpr
support our conjecture that event data can improve out-
of-sample forecasts of civil war events over governmental
and territorial incompatibilities in the Uppsala/PRIO
Armed Conflict Data. Moreover, models with the orig-
inal grievance measures do better than purely event based
models, supporting our claim that both structure and
event based components can add value to conflict pre-
diction models. However, the contribution of events to
improving predictive power is modest and not entirely
consistent across samples, and some types of conflict
events seem easier to forecast than others.
Structure and events in forecasting conflict
The ability to forecast conflict has long been an impor-
tant aspiration (see e.g. Choucri & Robinson, 1978).
The last decades have seen many efforts to predict civil
war or domestic instability and to develop more systema-
tic global risk assessments, with the Political Instability
Task Force as a prominent example (see Esty et al., 1995;
Goldstone et al., 2010; King & Zeng, 2001). Many early
efforts to predict civil war relied largely on off-the-shelf
political and economic country characteristics such as
GDP per capita, democracy, and income inequality (see
Ward, Greenhill & Bakke, 2010). Drawing on previous
disaggregated studies of civil war (see Cederman &
Gleditsch, 2009), Buhaug, Cederman & Gleditsch
(2014) argue that understanding and predicting civil war
can be improved by a more actor oriented approach to
motivation and opportunities, and focus on group
based (or ‘horizontal’) inequalities rather than individ-
ual (or ‘vertical’) inequalities or country average or
aggregate measures. They train a model on observations
for 1960–99 and use the estimated probabilities of con-
flict for 1999 to create a forecast for civil war onset over
the next decade, 2000–09. They find that the suggested
‘horizontal’ inequalities model correctly identifies eight
out of the 26 conflicts over the period while a tradi-
tional model emphasizing only ‘vertical’ inequalities
only identifies four.
We extend the Buhaug et al. (hereafter ‘BCG’) model
by looking at how information on interactions between
the government and dissidents in prior months influ-
ences our ability to predict civil war events. Whereas
BCG looked only at conflict onset, we consider a more
general transition model with both conflict onset and
termination over territorial and governmental incompat-
ibilities. We also extend the original BCG data, both in
terms of resolution and coverage.
Traditional episodic conflict data identify violence
starting at one date tand ending at tþk. Disaggregated
event data identify individual interactions between
specific actors and targets (i.e. who did what and to
whom). Event data have often been used as the response
or to identify specific conflict events of interest. Here we
consider whether event data can inform predictions of
conflict episodes through information reflecting heigh-
tened tension between actors likely to precede a conflict
outbreak.
1
For example, event data may help identify
dissident mobilization prior to an outbreak of lethal con-
flict exceeding the conventional battle-deaths threshold.
Likewise, increasing repression by the government
against non-state actors could reflect anticipated conflict
or motivate armed dissident mobilization. Event data
may also help predict the termination of ongoing armed
conflict. For example, conciliatory acts by the govern-
ment and rebel organizations may signal improved pros-
pects for a peace agreement or augur an imminent end to
armed hostilities.
There is a long tradition of event data models of
political conflict (see e.g. Freeman & Goldstein, 1989;
Goldstein & Pevehouse, 1997). However, many studies
are country or dyad specific, an d focus on evaluating
propositions rather than prediction per se.
2
Moreover,
many traditional collections such as the Conflict and
Peace Data Bank (COPDAB) and World Event Interac-
tion Survey (WEIS) have not been updated in ways that
allow generating or evaluating real-time forecasts. Many
recent event data projects use automated coding from
news media sources (see e.g. Bond et al., 2003; Schrodt
& Gerner, 1994). This bears promise of data on a near
real-time basis, and machine coding can help avoid com-
mon problems associated with human coders, who often
classify events differently (see e.g. King & Lowe, 2003;
Ruggeri, Gizelis & Dorussen, 2011).
Following interest in the predictive power of existing
civil war models (see Ward, Greenhill & Bakke, 2010;
Weidmann & Ward, 2010), researchers have proposed
global forecasting models of civil war integrating struc-
tural characteristics and behavior gleaned from events.
Ward et al. (2013) develop a model relating the risk of
civil wars in the Uppsala/PRIO data to conflictual events
collected from news media sources, also considering a
host of country characteristics as well as random effects
and spatial clustering. They report high predictive accu-
racy, both in-sample and out-of-sample. In another
1
Gohdes & Carey (2017) use a similar approach, where they treat
killing of journalists as a precursor to political repression.
2
An important early exception is Pevehouse & Goldstein (1999),
who use evidence from Bosnia and Herzegovina to predict ex ante the
effects of NATO involvement on the Kosovo conflict.
276 journal of PEACE RESEARCH 54(2)
To continue reading
Request your trial