Territorial control in civil wars: Theory and measurement using machine learning
| Date | 01 November 2020 |
| DOI | 10.1177/0022343320959687 |
| Author | Therese Anders |
| Published date | 01 November 2020 |
| Subject Matter | Regular Article |
Territorial control in civil wars: Theory
and measurement using machine learning
Therese Anders
Hertie School Data Science Lab & SCRIPTS Cluster of Excellence
Abstract
Territorial control is a central variable for civil war research – yet, we lack sufficiently detailed data to capture
subnational dynamics and offer cross-country coverage. This article advances a new measurement strategy for
territorial control in asymmetric civil wars. Territorial control is conceptualized as an unobserved latent variable
that can be estimated via observed variation in rebel tactics. The strategy builds on a theoretical model of rebel tactics,
by which rebels use terrorism less when they control a given area – preferring conventional tactics, which require
higher levels of territorial control. The latent variable, territorial control, is estimated via a Hidden Markov Model
(HMM). As an observable indicator for rebel tactics, I leverage geo-coded event data and a function of the relative
frequency of terrorist attacks and conventional war acts, weighted by time and distance. The model yields estimates
of territorial control for asymmetric civil wars at a resolution of 0.25 decimal degree minimum diameter hexagonal
grid cells. Validation of estimates for the Colombian and Nigerian civil wars suggests HMMs as a fruitful avenue to
estimate spatiotemporal variation in territorial control.
Keywords
civil war, event data, latent variables, territorial control
Territorial control in civil war influences dynamics of
violence, civilian victimization, and rebel governance.
While territorial control is universally recognized as theo-
retically important, empiricalstudies are scarce because we
lack data on who controls an area for most conflicts. For
example, the majority of research on the Colombian civil
war ignores territorial control as a variable,
1
or uses occur-
rence of rebel violence as a proxy.
2
However, armed
actors’ presence cannot be equated with the magnitude
of their rule. Rebels might attack a town market, only to
immediately retreat to remote hideouts without com-
manding control or preventing access by government
forces. Important questions, such as how rebel territorial
control affects internal displacement, cannot be ade-
quately answered without systematically produced data
on territorial control that vary temporally and
subnationally. The difficulty of measuring territorial con-
trol limits the data available for analysis, especially for
asymmetric civil wars that do not feature clearly defined
frontlines and instead exhibit ‘messy patchwork’ patterns
of control (Kalyvas, 2006: 88).
Given the difficulty of directly observing territorial
control in asymmetric conflicts, how can we estimate
changes in territorial control across time and space?
To fill the gap in availability of data, I propose a
novel measurement strategy for territorial control in
asymmetric intrastate conflict. I show that we can
estimate territorial control by translating a theory of
actor behavior into a machine learning model – lever-
aging information on variation in rebel tactics based
on event data.
Building on existing work regarding the relationship
between territorial control and insurgent tactics,
I develop a model linking the relative frequency of
Corresponding author:
t.na.anders@gmail.com
1
An exception is Arjona (2016) who collects data on historic patterns
of control in Colombia. However, the data are not publicly available
and limited to a few villages.
2
E.g. Prem, Saavedra & Vargas (2019).
Journal of Peace Research
2020, Vol. 57(6) 701–714
ªThe Author(s) 2020
Article reuse guidelines:
sagepub.com/journals-permissions
DOI: 10.1177/0022343320959687
journals.sagepub.com/home/jpr
To continue reading
Request your trial