A Quasi-Experimental Study of Contagion and
Coordination in Urban Conflict:
Evidence from The Syrian Civil War in Damascus
Navid Hassanpour∗
[email protected]
Abstract
Motivated by dichotomous scholarly results on the conflict-communication nexus,
I compare the predictions of two competing theories of urban conflict based on coordination and contagion. I exploit a quasi-experimental intervention in the Syrian
nation-wide communications in November 2012 along with a geolocated dataset of
daily conflict locations (n = 627) in Damascus to build a panel (n = 275 days × 252
spatial units = 69, 300), to show that disruption of communications coincided with an
unprecedented increase in the geographic dispersion of conflict. Furthermore, while
compared to pro-regime atrocities rebel activity was more spread in space and time,
lack of coordination meant more spatiotemporal clustering: a combination of global
dispersion and local clustering defined the insurgency. On average, one additional prior
incident in the spatiotemporal vicinity translated to 25% higher rates of recurrence. In
addition to detecting significant spatiotemporal spillover of violence, I show that the
contagion was effectively activated during the blackout.
I would like to thank Institute for the Study of War (ISW) and Joseph Holliday for providing data
on conflict locations in Damascus, Hazem Younes and Rafal Rohozinski for assisting with primary sources
of evidence on communication manipulations in Damascus, Yale’s Geographic Information Systems Center,
particularly Stacey Maples for providing excellent expert assistance with the GIS analysis, Hussam Jefee
and Daniel Masterson for assisting with the survey on the ethnic and income distribution of Damascus, and
finally Faisal Ahmed, Ali Bangi, Dominik Duell, Robert Keohane and Kevin Mazur for helpful feedback on
previous drafts. I am also grateful for the helpful comments I received from the participants in the APSA
2014 Annual Conference and the 2014 Digital Methods Conference at Princeton University.
∗
1
Keywords: Collective Action, Contagion, Geographic Information Systems, Information &
Communication Technology, Syria, Urban Conflict
2
Is communication technology conducive to collective violence? Scholarly answers to the
dilemma differ in their conclusions. The introduction of mobile communication to the African
continent is shown to have increased the levels of civil violence (Pierskalla and Hollenbach
2013), while the Iraqi rebels are found to destroy cell towers in an attempt to diffuse the
flow of information to the authorities, and the existence of cell communication on the regional level is linked to lower levels of conflict (Shapiro and Weidmann 2014). A potential
resolution to these seemingly contradicting findings is impeded by empirical difficulties as
highly detailed data on communication and conflict in the volatile context of a live conflict
are rare. Randomized experimentation in such occasions are obviously impermissible. In
such situations, the best one can do is to extrapolate situations that share a strong resemblance to natural experiments (Dunning 2012), identify significant effects in the independent
variable of interest (here the spatiotemporal dynamics of the conflict), and then explore the
processes that underly such effects. An opportunity for such a study represented itself during
The Syrian Civil War in late November 2012.
In a recent interview, Edward Snowden, a fugitive former system administrator for the
Central Intelligence Agency (CIA) and a previous National Security Agency (NSA) contractor (Alexander 2013), with enough journalistic flair ascribed the only multi-day country-wide
disruption of the Internet in Syria in 2012–and the longest during the conflict to date–to
the U.S. National Security Agency. This claim, if true, provides a unique opportunity for
studying the effects of Information and Communication Technology (ICT) on the course of
urban conflict in the form of a natural experiment run in Syria by the NSA. Snowden claims
(Bamford 2014)
. . . One day an intelligence officer told him that TAO–a division of NSA hackers–
had attempted in 2012 to remotely install an exploit in one of the core routers at
a major Internet service provider in Syria, which was in the midst of a prolonged
civil war. This would have given the NSA access to email and other Internet
traffic from much of the country. But something went wrong, and the router was
3
bricked instead–rendered totally inoperable. The failure of this router caused
Syria to suddenly lose all connection to the Internet. . .
Comparing the above with a description of the disruption from SecDev, a data science and
open intelligence company which monitored the Syrian Internet at the time is intriguing1
. . . In Syria all backhaul for cell networks is done on IP and . . . we were able to
notice all IP infrastructure went down in Syria. That means the cell networks
did not go down because the physical layer went dark, but because the IP layers
topped fun tinning (core routers were off, and DNS was down).2
While Snowden’s claims can not be independently verified, in the following I will argue that
there is strong evidence for the singularity of the Syrian media disruption in November 2012.
***
Since the beginning of the Syrian Civil War the Assad regime has frequently used local disruptions of communication as a means of war.3 However, on November 29, 2012 the country
experienced an unprecedented and complete shutdown of the Internet and fast cell communications (Syrian Observatory of Human Rights 2012) that lasted for at least two days.4 In
the aftermath of the blackout, the daily fatalities of the war in the whole country increased.5
Later I will show that the disruption was also accompanied by an increase in the geographical
dispersion of conflictual incidents in the city of Damascus proper and its suburbs.6 A theory
of civil conflict based on coordination (Olson 1971), selective incentives (Lichbach 1995) and
punishing ambivalence (Kalyvas and Kocher 2007) alone can not explain the increase in the
1
Rohozinski (2013)
Personal correspondence with SecDev, IP stands for Internet Protocol Layer, DNS represents Domain
Name System.
3
Personal interviews with Damascus residents in southern and eastern Damascus, also see Chozick (2012)
4
Connectivity was restored on December 1, the incident is the longest in duration to date,
http://www.google.com/transparencyreport/
5
Daily counts based on reports in Syrian Observatory of Human Rights (2012) show an uptick in the
number of fatalities, Gohdes (2014) demonstrates a similar surge.
6
Similar decentralization processes were reported during the complete media shutdown of the 2011 Egyptian Revolution (Hassanpour 2014).
2
4
intensity and dispersion of conflict during the blackout. In fact, resource mobilization theories predict a de-escalation in the absence of communication (Marwell, Oliver, and Prahl
1988). Based on a similar logic authorities often disrupt communications to diffuse rebellion
(Howard, Agarwal, and Hussain 2011). Furthermore, if the Syrian government wanted to
strike the rebels at certain areas, it could have done so on a local level, as it had been doing
for the total length of the civil war to that date. The blackout even as an ominous signal
may not motivate participation on the side of the rebels, in fact it informs the population of
an imminent strike. A similar blackout of only a few hours on July 19, 2012 prompted flights
from volatile neighborhoods in Damascus (Syrian Observatory of Human Rights 2012):
Damascus: The Basateen al-Mezzah area is witnessing a large-scale migration
out of the area, they are fearing military operations from the syrian [sic] forces
who have surrounded the area.
In asymmetric conditions of the Syrian civil war such a signal discouraged direct coordinated engagement. An answer to the puzzle of decentralization and escalation during total
blackouts merely based on a logic of coordination is incomplete.
***
In the following, I outline a research design which employs the Syrian media disruption of
November 2012 as a tool for studying the influence of connectivity on collective action in the
urban environment. After identifying a significant increase in the geographical dispersion of
conflict in the Syrian Capital during the disruption, I characterize the processes of contagion
in Damascus in 2012, simply defined as proliferation of conflict from one spatial unit to an
adjacent one, and argue that proxies for both dispersion and contagion can be found. Based
on a detailed account of conflict in Damascus in the course of the last nine months of 2012,
I construct a panel dataset that documents the daily existence of violent conflict at a mileby-mile resolution in an area of 252 square miles covering Damascus and its suburbs from
April to December 2012, in conjunction with the geographical elements of the Damascus’
5
urban landscape. The pertinent GIS (Geographical Information System) dataset is used to
detect processes of spatial contagion during the conflict.
A nearest incident analysis of the location of violent conflicts shows that on average premeditated and coordinated violent incidents were less clustered in space and time compared
to the rest of the conflictual events. Anti-regime violent incidents were also more dispersed.
A combination of both effects constitutes a dispersed insurgency that was locally clustered.
I employ a global games model to show why such a situation a recipe for violent proliferation
during a media blackout. Furthermore, I find the existence of conflict in the neighboring
spatiotemporal units to be significantly linked to the occurrence of violence in a given unit
of analysis, i.e. a day-spatial window. This provides an estimate for the rate of conflict at
time t based on the number of violent incidents in the spatial neighborhood at time t − 1. It
is shown that existence of one additional incident in the immediate spatial neighborhood in
the previous day translated to a 25% increase in the rates of conflict on a given day. Furthermore, while both contagion and disruption parameters in the panel dataset are significant
predictors of violent incidents, adding the interaction term between the two reveals that the
significance of communication disruption as a facilitator of conflict is through its activation
of contagion processes.
Geometry of Contagion versus Calculus of Coordination
The role of digital communication in conflict can be cast in two distinct molds: communication links can be used to facilitate collective action via coordination (Marwell, Oliver, and Prahl
1988), digital communication can also create long bridges that are known to interact with the
proliferation of social processes in small world networks (Watts and Strogatz 1998). Such
long bridges can traverse through spatial and structural confines and spread a certain behavior through simple or complex contagion. Long bridges have been shown to hinder diffusion
via stifling action when the contagion is of complex kind. In complex contagion, induced by
threshold models of collective action (Granovetter 1978), multiple social reinforcements are
6
required for triggering action7 and the existence of further communication links can only prevent clustered reinforcement in decentralized cores of contentious (Centola and Macy 2007).
The provision of public goods for preventing free-riding is difficult in large groups (Olson
1971); in contrast, contagion is facilitated in larger groups and suppressed in smaller ones
(Axelrod 1997). Now if complex contagion is found to be a cogent explanation for the
escalation of conflict in Damascus, then the removal of all communication, i.e. long bridges in
the complex contagion terminology, should promote proliferation and escalation in dispersed
clusters of rebellion. In this paper, I also develop an independent theoretical explanation for
the escalation of conflict based on the removal of the public signal (a proxy for the complete
blackout) in a “beauty contest” model (Keynes 1936) underlying global games of collective
action (Morris and Shin 2002). Based on the model, in certain network configurations8
representative of local social interactions, the absence of public signal promotes participation,
it does not impede it.
Retrospectively, the debate between those who take communication technology to be a
catalyst of rebellion (Pierskalla and Hollenbach 2013), and others who find communication
media to be an opium of the masses of sorts (Shapiro and Weidmann 2014, Kern and Hainmueller
2009) is best understood under a framework that distinguishes coordination-based explanations from those which explain contention as a result of contagion processes. Cell communication eased coordination among African rebels, hence increased the level of rebel conflict
under the watch of weak African states. However, while communication facilitates rebel
coordination, it simultaneously opens conduits for reconnaissance and intimidation. The absence of communication also facilitates local contagion away from detectable channels, and
promotes pockets of contention that are instrumental to the diffusion of contentious behavior
based on complex contagion (Centola and Macy 2007, Centola 2010, Siegel 2009) In targeting
cell towers, Iraqi rebels disrupted information channels to the authorities, at the same time
7
The results of the dynamics are clearly the opposite if the contagion is of simple kind, i.e. when a single
contact between two individuals is enough for transmitting a behavior (Watts and Strogatz 1998).
8
One prominent example is small world networks.
7
they enabled local network relations among potential recruits, a communication network
that was far from the eyes of the Iraqi state. Forecasts of the logic of contagion are expected
to be different from those of coordination. For example coordination in the open requires
safe havens from the authorities as well as means for far reaching communication, provided
by higher geographical elevation and tools for digital communication. In contrast, contagion
benefits from interactions in closed and crowded urban spaces in the plains, and contagion
processes from radical cores suffer from long communication ties provided by excessive usage
of digital communication. I use geolocated conflict data to test and confirm a number of
these conjectures on the role of elevation and influence of communication disruption in urban
conflict.
Peripheral Rebellion versus Urban Warfare
In the absence of communication technology, patterns of visibility in urban landscape dictate
the size of the audience for rebel communication. The spatial configuration of neighborhoods
in densely populated areas shape the dynamics of conflict. When there is a possibility of
observing others’ action in close quarters and close-knit social networks, contagion processes
become prominent. On the other hand, coordination mechanisms are likely to play an important role in sparsely populated social geographies. The dichotomy between the two processes
leads to a research plan that is informed by the geography of conflict as well as the nature
of the means of digital communication. The logic of civil conflict in urban environments
combines insights from civil war studies with those of contentious politics, at times violent,
in urban spaces. The topology of geographical terrain is known to influence the odds of
insurgency (Fearon and Laitin 2003), urban planning has been used as a means of political
control (Gould 1995, Scott 1998). Social networks transform during civil wars (Wood 2008,
Lyall 2010), and the underlying social network structure influences the nature of collective
action (Metternich, Dorff, Gallop et al. 2013, Parkinson 2013).
The Syrian Civil War stands as an example of conflict in the context of Damascus’
8
convoluted urban landscapes, as well as those of Hama, Homs, and Aleppo. Both sides
of the fight implemented an array of communication strategies to gain the upper hand in
a conflict that relies on mediated narration as a means of war (Lynch, Freelon, and Aday
2014). The urban staging of The Syrian Civil War, particularly in Damascus, along with
variations in the communication regime during the conflict provide a plausible testing ground
for theories of coordination and contagion in urban warfare.
The Syrian Civil War and Modes of Connectivity
The Arab Spring’s diffusion across the Middle East reached Syria in March 2011 (Lynch
2013). Civilian protests in mostly Sunni areas faced uncompromising military suppression
and soon escalated to an all out war. In suppressing the protests without mercy, Bashar Al
Assad was replicating a similar episode from 1982, when his father Hafiz Al Assad responded
to a local rebellion in the city of Hama with full force leaving fatalities in the order of
thousands (Seale 1988, Van Dam 2011). Following the indiscriminate killings, the authorities
remodeled Hama:
Heavily damaged old quarters were bulldozed away, roads were cut through where
once no car could pass, squares and gardens were laid out. The whole of Hama
was reshaped on a grand scale . . . (Seale 1988) p. 334.
The possibility of personal publication and transmission of news in 2011 meant the recent
atrocities could be reported to the outside world, a feat impossible thirty years ago during
the obscure battle of Hama. Following the escalation to a civil war, the fighting afflicted
one Syrian city after another. Fighting occurred both in rural and urban areas and the
newly salient ethnic and religious allegiances fueled local contention (Petersen 2013, Christia
2013, Holliday and Lynch 2012). Dense urban areas in Damascus, Aleppo, Hama, Homs,
and Dara became grounds for clashes. Rebels used digital means of communication for
organization; simultaneously these tools became venues for surveillance and tracking on the
9
side of the regime (Chozick 2012). Less than a year after the start of the uprising, Damascus
became a scene of clashes between the rebels and the regime’s forces. A fragile ceasefire
between the two sides was in place between April and June of 2012, the end of which
culminated in a rebel offensive in Damascus starting on July 14. On July 18, a bomb attack
against the National Security headquarters killed the Syrian defense Minister, his deputy, a
former defense minster, and Syria’s national security chief. The Syrian government staged
the Battle of Damascus against the rebels in the following two weeks. South and Eastern
areas of Damascus were scenes of the regime’s counterattack in the ensuing two months.
The second major rebel offensive in Damascus in 2012 occurred in November when the
rebels attempted at capturing Damascus International Airport in southeast of the city on
November 28. On the 29th major clashes happened in east of Damascus and both sides of
the conflict claimed they had the airport under their control (Reynolds 2012). The clashes
continued and escalated in December.9
The communication disruption that Edward Snowden refers to as a technical glitch enacted by NSA occurred during the clashes over the airport in the southeastern corner of
Damascus. Internet communications were shut down all together across Syria around noon
local time (GMT+2) on November 29 (SecDev 2012) and did not return to normal till the afternoon of December 1. Accompanied with the Internet access, fast mobile communications
were also disrupted during the blackout,10 influencing rebel communications; e.g. for mobilization and coordination purposes, rebels extensively used Skype (Chozick 2012)11 which
relies on network protocols distinct from those used for voice calls. The software on mobile
devices could not operate in the absence of fast cellular communications.
In the following, I exploit the very same nation-wide intervention in communications in
late November along with the geolocated daily account of atrocities during the Syrian Civil
9
The timeline of the atrocities are quoted from the New York Times and Syrian Observatory for Human
Rights’ Facebook page. For an account of the regime’s counterinsurgency see (Holliday 2013).
10
All backhaul for the Syrian cell networks flows through the Internet Protocol layer, and a disruption of
the Internet means a simultaneous shut down of the 3G cell network (Rohozinski 2013).
11
For Syria’s Rebel Movement, Skype Is a Useful and Increasingly Dangerous Tool, available at the following link: http://goo.gl/J8yZuR. Skype is a voice over IP software.
10
War in Damascus in 2012 as an identification instrument for exploring the communication–
conflict nexus.
Data Description and Research Design
I parsed Damascus’ urban sprawl and its main suburbs into a one-mile by one-mile grid
approximately 18 miles long and 14 miles wide. On this grid, I superimposed a dataset of
violent conflict locations in Damascus and its suburbs on a daily basis during 2012 (Holliday
2012).12 The data covers the date, the high resolution longitude and latitude for each event,
as well as information on the orientation (anti or pro-regime) and type of the conflictual
incident (see below), and the precision of the available information on the event. For each
incident to be coded once per day per location it had to involve at least one of the following
“types”:
1) Direct fire engagement between regime and rebel forces (fire fight with two parties involved), 2) Bombing attack against regime troops or facilities, 3) Airstrikes,
4) Indirect fire (i.e. shelling) only if reported casualties are greater than 15, 5)
Assassinations or kidnappings of regime personnel, 6) Regime raids or arrests
that involve a major troop movement (Holliday 2012).
Using a strategy similar to (Pierskalla and Hollenbach 2013) and (Shapiro and Weidmann
2014)13 I intersected the conflict data with the available information on the levels of connectivity in the spatial window under consideration. Figure (1) depicts the spatial distribution of
conflictual incidents in Damascus and its suburbs (Reef Damascus). Total monthly counts of
conflictual events for the months of April to December 2012 were (29, 29, 43, 81, 50, 87, 53, 82, 173),
respectively.
12
Two main sources for coding conflict locations were Syrian Observatory for Human Rights, as well as
the state-sponsored Syrian Arab News Agency (SANA) Holliday (2012)
13
albeit with a much higher resolution
11
' OpenStreetMap (and) contributors, CC-BY-SA
Figure 1: Top: Number of distinct conflictual incidents in each 1-square mile cell, June–
December 2012, Damascus and suburbs. N = 600, April and May not Included. Bottom:
Conflict locations January-December 2012 across Damascus, superimposed on the 1-mile
grid and an OpenStreetMap baseline. Some locations contain more than one incident.
12
High resolution panel data on connectivity in Damascus neighborhoods on par with
the detailed conflict location information is difficult to obtain, because unlike events which
are by nature spatially marked, levels of connectivity, often controlled from afar, are not
particularly tied to the spatial domain in question.14 I conducted interviews on the nature of
local disruptions that could not be captured in an aggregate measure such as the countrywide
traffic. The interviews provided qualitative data on the nature of local disruptions the Assad
regime applies in conjunction with military tactics. It became clear that neighborhoods in
the south and east of Damascus, such as Dayr el Asafir, Yarmouk, East Ghouta and Daraya
endured extended and selective disruptions of mobile and Internet communications, at times
for months, particularly those close to military facilities and those under an ongoing military
attack.15
In the absence of daily and mile by mile data on connectivity I exploited nation-wide
disruptions of cell and internet media as the intervention of interest. There were two incidents
of countrywide Internet and mobile disruptions in Syria in 2012, one occurred on July 19
and lasted for approximately an hour. The The extended disruption mentioned in Snowden’s
interview occurred on November 29 in midday local time and lasted till December 1 into the
late afternoon. Prior to the multi-day long disruption of November 2012, the only daylong
blackout occurred on Friday June 3, 2011 at the beginning of the Syrian uprising, shortly
after the Egyptian experimentation with a nation-wide blackout in January 2011. The
disruption in November 2012 is the longest in the 2011-2014 period, and the only multiday blackout in 2012.16 A comparison between the two occasions of country-wide blackout
depicted in figure (2) shows the different technical nature of the two. While the one in 2011
14
I tried to extrapolate daily average connectivity for each of square mile spatial windows in figure (1)
based on geolocated tweets produced in Damascus, but an inquiry showed that the number of geocoded
tweets in the spatiotemporal window of interest was too few. See appendix SI. 1.
15
Personal interviews with ex-rebels who spent the year 2012 in Damascus were conducted via Skype. Interview transcripts provide limited neighborhood-level topical data on connectivity disruptions in Damascus
in 2012. Because the data is not complete on a daily basis and in all quarters of Damascus, it is impossible
to draw definitive inferences merely from qualitative accounts.
16
For a complete list of traffic disruptions of Google products and services and by proxy the Internet see
http://www.google.com/transparencyreport/traffic/disruptions/
13
went into effect gradually, the disruption on November 29 was abrupt. A gradual switching
off of communication links is unlikely to have caused the global failure of communications
in November 2012.
Figure 2: A depiction of full Internet disruption on November 29, 2012 (Top) and June 3, 2011
(Bottom), Google services, fraction of worldwide traffic, normalized, source: Google Transparency Report http://goo.gl/5sNjN and http://goo.gl/7G089, respectively. Mobile
communications were shut down in parallel to the Internet. Comparing the two disruptions
shows they were of different technical nature.
Temporal Analysis of the Dynamics of Dispersion based on Pairwise
and Pairwise Normalized Distance, Event Count
The full disruption of communication in November 2012 coincided with an unprecedented
increase in the dispersion of the conflict. To demonstrate the effect, I define three different
measures of conflict dispersion (see equation (1)) and examine the temporal profile of the
14
dispersion measure before, during, and after the blackout in November.
Using the longitude and latitude of the incidents on each day I calculated the daily sum
of pairwise distances between every possible pair of incidents that happen on the same day.
For all incidents, i and j on day t, the pairwise distance is defined as dij (t), this gives the
first definition of dispersion in equation (1). One can also normalize the sum of pairwise
distances using the daily count of incidents, Nt , finally the count on every day itself can be
used as a measure of conflict dispersion.
D1 (t) =
Nt
X
dij (t), D2 (t) =
Nt
X
dij (t)/Nt , D3 (t) = Nt
(1)
ij
ij
Figure (3) shows the temporal profile of the geographical dispersion of violent incidents
in Damascus and Reef-Damascus (i.e. suburbs of Damascus) based on the three definitions
in equation (1). all three definitions portray an unprecedented increase in conflict dispersion
during the blackout. A simple examination of temporal correlations using an Ordinary Least
Squares (OLS) regression for calculating correlations in table (1) shows that dispersion at
time t was correlated with t−1, and days of full disruptions (coded as dummies on November
29, 30, and December 1) are robustly linked to a decentralization of the conflict. The Fridays
do not induce an increase in dispersion, and their role is not statistically significant.
In the absence of definitive evidence on the verity of Snowden’s claims, I examine a few
potential explanations in the following sections in order to show that they do not provide
convincing alternative reasons on why the conflict was decentralized in the aftermath of the
disruption on the scale seen in figure (3).
Alternative Explanations
There are two distinct alternative explanations for the increase in the dispersion of the
conflict in Damascus. First, one could argue that the increase in dispersion would have
happened regardless of the disruption. In the next section, I will examine such a possibility
15
2000000
1500000
1000000
500000
0
1e+05
8e+04
6e+04
4e+04
2e+04
0e+00
20
15
10
5
0
Figure 3: Dispersion of conflictual incidents in meters, Damascus and suburbs, April 1, 2012
to December 31, 2012 for three dispersion definitions in equation (1). The three days in
16
which Syria was experiencing a complete blackout, November 29, 30, and December 1, are
in color.
Table 1: Dependent variable: dispersion at time t: three definitions, OLS regressions
(1)
(2)
(3)
Dispersion at t
Dispersion at t
Dispersion at t
Sum Distance Sum Distance Normalized
Count
∗∗∗
∗∗∗
Dispersion at time t − 1
0.166
0.261
0.315∗∗∗
(0.0447)
(0.0510)
(0.0518)
Disruption
1025193.1∗∗∗
(71164.5)
49011.8∗∗∗
(5402.1)
9.683∗∗∗
(1.272)
Fridays
-21675.6
(21098.7)
-1773.2
(1550.4)
-0.582
(0.371)
cons
25656.7∗∗
(8056.0)
275
0.474
4356.2∗∗∗
(656.1)
275
0.339
1.537∗∗∗
(0.179)
275
0.298
N
adj. R2
Standard errors in parentheses
∗
p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001
in detail by constructing a predictor of conflictual incidents based on spatiotemporal as well
socioeconomic patterns of conflict and show that even after controlling for socioeconomic,
temporal and spatial trends, disruption is a statistically significant independent variable in
engendering conflict.17
One could also ascribe the increase in the dispersion to a coordinated campaign by
the regime to stifle rebel activity on the 30th and 1st. Based on this explanation, the
regime decided to disrupt communications and preempt rebels in different areas of Damascus
afterwards. However, there are two issues that cast doubt on such a solution to the puzzle:
first, the significant increase in dispersion did not happen on the first day of disruption, i.e.
November 29. Instead it took a day for the clashes to spread to different parts of the city.
17
Furthermore, a simple examination of the daily count trend in the five months of July to November
2012 does not support the assertion: the increase in dispersion at the end of November was not indicative
of the regular dynamics of the conflict. A linear regression of daily location counts over the location counts
from the past seven days, based on the 153 days in July to November, predicts location counts of 3.25 and
4.14 for November 30 and December 1 respectively. Real counts were 9 and 13, well beyond the 95% CI of
predicted values. A linear predictor of the number of conflict locations at time t can be built based on counts
at times t − 1, . . . , t − k similar to the Auto Regressive (AR) procedure for event data prediction outlined in
(King, Pan, and Roberts 2013).
17
If the regime had planned to ambush the rebels in different areas of the city immediately
after the disruption, it should have done so on November 29 simultaneously with the full
disruption going into effect, not during the next two days. Second issue with this explanation
is based on the temporal pattern of the events in locations that experienced unrest during
the November blackout. First incidents are defined as those occurring for the first time
during the year 2012 at a certain location. Among all first incidents occurring in 31 days of
the November 1st to December 1st, 64%, i.e. 9 out of 14, happened between November 29
to December 1, in 3 days alone, when the cell and high-speed Internet communications were
absent. From the total 13 neighborhoods experiencing conflict on December 1, 5 of them
had never seen conflict during that year. Similarly, out of 9 neighborhoods experiencing
conflict on November 30, 3 locations were the scene of violence for the first time in the past
330 days. If the regime was planning to attack rebel strongholds beforehand, conflict would
have been more likely to happen in areas with a history of violence during the past eleven
months, not many new locations. The same pattern as the rest of November should have
prevailed. Furthermore and finally, although the main battles in late November 2012 took
place over the Damascus International Airport (Reynolds 2012), majority of clashes in the
ensuing blackout did not happen in the vicinity of the airport. Clashes had spread to other
parts of the city on November 30 and December 1. Table (2) includes the neighborhoods
experiencing clashes before, during, and after the blackout.18
Detecting Clustering, Contagion, Processes of Escalation:
Spatiotemporal Patterns of Urban Conflict in Damascus
In this section I further parse the geographical and temporal elements of the conflict in
Damascus in 2012 in three steps in order to detect localization and contagion patterns in
18
Source: (Syrian Observatory of Human Rights 2012, Holliday 2012). The frequency of first incidents
during 11/01–12/01/2012 is included in Appendix SI. 2.
18
Date
November
November
November
November
November
26
27
28
29
30
December 1
December 2
December 3
December 4
Conflict Locations
2: Set Zaynab/ Darayya
2: Jaramana (approximate)/ Darayya
5: Douma/ Harasta/ Beit Sahem/ Set Zaynab/ Darayya
5: Irbin/ Babbila/ Beit Sahm/ Kafarsouseh/ Jisreen
9: Al Mouadamyeh/ Darayya/ Kafar-Sousah/ Yalda/ Beit Sahem/
Az Zyabeyeh/ Jisreen/ Saqba/ Douma
13: Mouadamiya/ Douma/ Jisreen/ Saqba/ al-Thiyabiya/ Daraya,
Bebila/ Yalda/ Beit Sahm, Aqraba(2)/ Kafarsouseh/ A’sh al-Warwar
4: Jisreen/ Maliha/ Nashabiya/ Deir al-Sleiman
4: Beit Sahem/ Daraya/ al-Nabek/ Barzeh
2: Tadamun/ al-Fahama
Table 2: Major neighborhoods in Damascus and suburbs experiencing violent incidents, late
November–early December
time and space. First, I use nearest past incident data to detect clustering patterns during
the conflict. In particular, I use the distinction between coordinated acts of violence and
the rest of the incidents to show that coordinated incidents were less clustered. Along the
same lines, pro-regime incidents show higher levels of spatiotemporal continuity. Anti-regime
activity was more dispersed. Second, I employ a panel dataset of 252 × 275 geographical
cell-days to show that after controlling for structural parameters such as proxies for population, elevation, density of throughways, and ethnic and economic composition, as well as
spatial indices, still the disruption and the events of neighboring locales in space and time
had a significant impact on the occurrence of violence in a certain location. Such ties across
space and time were not endogenous results of some overarching spatiotemporal process.
Even after controlling for both time and geographical locations, as well as regular structural
components of conflict, disruption shows a significant impact on the escalation of conflict.
Finally, using interaction terms between a spatiotemporal lagged index of conflict and the
disruption variable I demonstrate the role of disruption in staging contagion from one geographical unit to the other resulting in escalation of the conflict. In particular, the contagion
is shown to had been activated during the disruption.
19
The Construction of the GIS Dataset
I used the conflict data in Damascus in 2012 in conjunction with GIS processing and a survey
study of socioeconomic factors in Damascus to build the panel dataset.19 The distance from
the nearest incident in space and time, as well as the number of incidents in the immediate
spatial and temporal lag neighborhood of each unit of study (cell-day) provide efficient ways
for measuring the spatiotemporal contagion of conflict from time t − T to t and from a cell to
its neighboring vicinity. I calculated both of these parameters for each conflictual incident,
and cell-day respectively.
For building a spatiotemporal predictor of conflict, a rectangular grid was superimposed
on the area under study. I tested several grid sizes (0.5, 1, 2 miles).20 The 1 mile grid
is the only parsing mechanism that simultaneously captures the significance of elevation,
population, and the sum length of streets in engendering urban conflict in each geographical
unit.21
The average elevation for each cell, a measure of population on a mile-by-mile basis, and
the total length of roads situated inside each cell comprise the main structural elements
of interest: the elevation level influences the nature of insurgency in each area, the size of
population in urban areas influences the organization of collective action, finally the density
of roads in neighborhoods represents proxy levels for visibility in the geographical terrain,
roads are also spatial conduits for street battles and their density is expected to have a
significant role in altering the dynamics of contagion.22
19
Geographic information embedded in event data facilitates detection of conflictual patterns
(Cederman, Weidmann, and Gleditsch 2011), and prediction of conflict in space and time after further deconstruction to spatial and temporal elements (Gleditsch and Weidmann 2012, Ward and Gleditsch 2002).
For the purpose of event modeling and regression analysis, spatial regression methods take the inherent
correlation between units of analysis into account (Ward and Gleditsch 2008).
20
The details of the test and the selection criteria for the one-mile grid are included in table (SI.1) in
Appendix SI. 3.
21
Average elevation and total length of streets were calculated using the ArcGIS software
and digitized maps of Damascus available on OpenStreetMap http://www.openstreetmap.org/
total population in each cell was calculated using a proxy for population via Landhttp://web.ornl.gov/sci/landscan/
(Dobson, Coleman, Durfee et al.
2000,
Scan
project
LandScan Global Web Applications 2013).
22
Average elevation and total length of streets were calculated using the ArcGIS software
20
The socioeconomic condition of each neighborhood is also a cogent component of the civil
conflict. The lower income areas are expected to experience more conflict (Collier and Hoeffler
2004), and the ethnic composition of the areas under study influences the rates conflict in
each geographical unit (Horowitz 2000). Lacking official data on socioeconomic parameters
on the microlevel,23 I surveyed Damascus current and previous natives on these parameters.24
Fourteen expert coders with in depth knowledge of Damascus coded ethnicity and income
levels for each neighborhood. Ethnicity–Religion was chosen between categories Sunni, Alawite, Shia, Christian, Druze, and Kurdish. The income level was coded on a tripartite level
Rich, Middle Class, Poor. I was also cognizant of the fact that Syria is witnessing large scale
intra and international migration (Khaddour and Mazur 2013), hence I emphatically asked
the survey respondents to answer these questions for the year 2012 alone.
The socioeconomic survey questions were asked on the neighborhood level to provide the
best level of cognition for our survey subjects. The neighborhoods themselves were outlined
based on existing maps of Damascus and the output from three Damascus experts.The map
in figure (4) shows the imposition of these neighborhoods on the 1-mile grid. Later I used the
same pattern to cluster the standard errors in Poisson regressions based on neighborhoods,
or control for the neighborhood effect.25
In the following, I conduct two distinct set of examinations. First, I find that the distance from the nearest incident was significantly correlated with the nature of violence being
and digitized maps of Damascus available on OpenStreetMap http://www.openstreetmap.org/
total population in each cell was calculated using a proxy for population via LandScan
project
http://web.ornl.gov/sci/landscan/
(Dobson, Coleman, Durfee et al.
2000,
LandScan Global Web Applications 2013). In Appendix SI. 3. see figure (SI. 2) for patterns of population density and figure (SI. 2) for the elevation patterns of the spatial window.
23
There are datasets such as G-Econ (Nordhaus, Azam, Corderi et al. 2006) and (Nordhaus 2006) that
provide economic data on microlevel, however the highest resolution of the dataset was in the order of 50
miles, which did not apply to the 1-mile resolution.
24
There is no detailed and recent information on the ethnic composition of Damascus available similar to
what exists for Baghdad in (Izady 2013).
25
Among the neighborhoods included in the coding were Al-Midan, Al-Qabun, Al-Salihiyah, Al-Shaghour,
Barzeh, Dummar, Jobar, Kafr Sousa, Mezzeh, Muhajreen, Old City, Qadam, Qanawat, Sarouja and Rukn
Eldin in Damascus proper and Al-Assad, Aqraba, Darayya, Douma, Hajar-al-Aswad, Harasta, Irbin, Jdaydet,
Jaramana, Mazzeh 86, Mouadamiyah, Qudssaya, Saqba, Sayyida Zeinab, Yalda, Yarmouk, Zabdean and
Zamalka in Reef-Dimashgh.
21
Figure 4: One-mile grid with the approximate neighborhood delineation superimposed.
coordinated or not prearranged, and being pro or anti-regime. Second, I employ the panel
dataset in conjunction with the spatiotemporal and socioeconomic variables to build a predictor for the occurrence of violence in a cell on a certain day; the results reveal a significant
interaction between media disruption and contagion.
Clustering, Nearest Past Incident & Detecting Spatiotemporal Localization
To further illustrate the spatiotemporal dynamics of the conflict I calculated the distance
between each violent incident with the nearest event, both in time and space, in that order,
and used the distance as a measure of clustering (Clark and Evans 1954). For each of the
627 incidents, this measure provides a proxy for the spatiotemporal correlation in the course
of the process in the last nine months of 2012.
A histogram of the minimum distances in figure (SI.x) shows that the distribution is far
from the Poisson distribution which is characteristic of random spatial distribution (Diggle
22
2013). In fact the most frequent distances in the histogram are concentrated below 2 kilometers.
Furthermore, figure (5) shows the temporal profile of the minimum distances across the
last nine months of 2012 in Damascus. There is a conspicuous anomaly on days November
30 and December 1. Although these two days presented the highest levels of geographical
dispersion in figure (3), they show a bipolar distribution concentrated in unusually large
and small distances. The daily average values for distance from nearest incident as a proxy
for spatial clustering is depicted in figure (5). The dominance of small nearest distances,
in conjunction with high levels of dispersion suggest that dense but dispersed conflictual
clusters were prominent during the disruption. Later I set forward a series of tests to verify
the simultaneous existence of the two dynamic forces, i.e. escalation and clustering on the
local level, and increasing dispersion on the global level.
The results of simple OLS regressions with distance from the nearest past incident as
dependent variable (N = 627) and two indices on the orientation of events, and the level of
pertinent premeditation are included in table (3). For each of the violent incidents included
in the dataset orientation index determines if the incident was anti (positive dummy)
or pro-regime (negative dummy). Furthermore, coordination index provides a measure of
coordination in the context of each of the conflictual incidents. The incidents are identified as
“coordinated” if they are in one or more of following categories: Assassination, Execution,
Improvised Explosive Device (IED), Ambush/Raid, Rotary wing strike , or Car bomb. In
contrast, instances of Direct fire are coded as a category separate from the aforementioned
classes. The main aim of such a classification is to use this crude proxy of coordination in
conjunction with spatiotemporal dynamics. The distinction may lend itself to a meaningful
distinction between the dynamics attributed to the coordinated events versus the rest. In fact
in the following I will show that it is the case. There are statistically significant differences
between the dynamics of coordination and contagion processes.
The results of regressions in tables (3) and (4)–disruption index included, an indicator
23
20000
15000
10000
5000
0
0
5000
10000
15000
20000
04/03/12 06/27/12 08/05/12 09/17/12 11/01/12 12/01/12 12/17/12
04/03/12 05/06/12 06/07/12 07/02/12 07/21/12 08/12/12 09/04/12 09/23/12 10/11/12 11/07/12 11/30/12 12/18/12
Figure 5: Average distance from nearest neighboring incident in meters, April to December,
2012. Daily average values represent a measure of spatial clustering of conflictual incidents.
Note the small average values, i.e. high clustering, during the disruption, dates in x-axis
weighted by the number of daily events. Note the overall decreasing pattern, and the plateau
at the end of November. Top: daily averages. Bottom: daily values all included.
24
showing if the event occurred during the disruption–show two intriguing qualities of the
events in relation to the spatiotemporal dynamics: first, in line with the dispersion analysis in the previous section, anti-regime incidents were more spread in space and time. In
contrast pro-regime attacks showed higher levels of continuity in space and time. Second,
non-coordinated incidents were more clustered in space and time. Both of these effects are
statistically significant on the 5% level or less. Figure (6) shows the regression multipliers,
all 95% confidence intervals are above zero.
To summarize, clustering is linked with non-coordination, and anti-regime activity is
more dispersed. Both coordinated and government-baed attacks are likely to show more
continuity in space an time. Compared to the rebels, the regime was more likely to strike at
the same place in short durations of time.26
Table 3: Components of distance from the nearest incident, OLS regression
orientation index
Coef.
Std. Err.
631.5734 243.5537
(1)
min dist t i
t
P > |t| [95% Conf. Interval]
2.5932 .0097 [153.2892,1109.8580]
coordination index
537.8554
2.4186
N
222.3852
.0159
[ 101.1413,974.5695]
627
There is also a negative and significant relation between coordination and anti-regime
nature of an event, i.e. coordinated events were likely to be pro-regime, see table (5).27 this
means anti-regime activity was not as coordinated as pro-regime operations. Therefore, in
table (6) the two quadrants (I) and (III) represent the pertinent combinations of coordination and orientation indices and their impact on distance from the nearest incident. In both
of these situations, there are two processes that simultaneously work in two opposite directions, for example, anti-regime incidents tend to induce events with higher nearest distances,
26
This property facilitates the identification of coordination and contagion processes in the next section,
because this means controlling for regularities in space and time does away with explanations of escalation
based on pro-regime activity.
27
The indices are not multicollinear, 1 − R2 = 0.82.
25
Table 4: Components of distance from the nearest incident, Disruption Index included, OLS
regression
coordination index
Coef.
462.1994
Std. Err.
219.3350
(1)
min dist t i
t
P > |t|
2.1073
.0355
orientation index
476.8578
241.8253
1.9719
.0491
[1.9664,951.7492]
disruption index
-3564.9670
760.5896
-4.6871
.0000
[-5058.5970,-2071.3370]
N
[95% Conf. Interval]
[31.4739,892.9250]
627
Table 5: The relation between orientation and coordination indices, as expected pro-regime
operations were more coordinated
coordination index
cons
N
R2
Coef. Std. Err.
-.3895
.0330
(1)
orientation index
t
P > |t| [95% Conf. Interval]
-11.7917 .0000
[-.4544,-.3246]
.4294
13.6828
.0314
.0000
[.3678,.4911]
627
0.182
Anti-Regime
Pro-Regime
Coordination No Coordination
+ + (II)
+ – (I)
– + (III)
– – (IV)
Table 6: Four possibilities, coordination and orientation, quadrants of interest, (I) and (III),
contain majority of the events. The signs show the effect of the two parameter on clustering
(distance from nearest incident) of conflictual events. + represents an increase in the nearest
distance, – a decrease.
26
Coefficient Plot
Coefficient
event_data$disruption_index
event_data$coordination_index
event_data$orientation_index
-4000
-2000
0
Value
Figure 6: OLS coefficients, distance from nearest incident for each event (N = 627) regressed
over indices for coordination, orientation (pro-anti regime), communication disruption
27
while events that are not coordinated show more spatiotemporal proximity to the nearest
incident. The tension between the two effects is also in display in the conflict data during
the communication disruption, which shows more clustering on the local level accompanied
with an increase in the dispersion of incidents. The end result is escalation on the local level,
and increasing dispersion of the clusters on the global level.
As the next step, I test for the existence of spatial contagion on the local level. For doing
so I extend the analysis from events to cell-days to detect potential mechanisms of contagion
in the panel.
Contagion: Detecting Spillover in Moore Neighborhoods,28 Panel
Dataset
I combined the account of conflictual incidents with spatial controls to produce a panel GIS
dataset of size 252 × 275 = 69300 entries. Dependent variable is an indicator that is 1 if at
least one incident happened inside the relevant cell on a specific day, and 0 otherwise.
In the previous section I enumerated a number of confounding parameters which complicate the conclusions on the definitive impact of the communication disruption in November
2012 on the dispersion and escalation of conflict across Damascus. A panel dataset covering
the 252 square-mile spatial window in Damascus for the 275 days in April to December
2012 can test the validity of such alternative hypotheses. For example controlling for the
index of the cells ameliorates the concerns with the characteristics of the locale that may
have caused certain patterns of conflict. In addition to the concerns regarding the structural
components, the examination of the panel dataset also provides a test for the alternative
hypothesis of preplanned attacks. In the previous section, I demonstrated that pro-regime
attacks showed high degrees of continuity in space and time, therefore controlling for spatiotemporal trends of the conflict can capture such endogenous components. Note that the
government’s decision to attack a certain neighborhood relies on the history of the unrest, a
28
In a rectangular grid, Moore neighborhood for a given cell is the collection of 8 cells surrounding it.
28
decision made based on the pattern of previous and ongoing rebellious activity in the area.
For the very same reason, pro-regime attacks, compared to opposition violence show higher
levels of spatiotemporal clustering (see table (6)). Moreover, controlling for a lagged version of the dependent variable as well as indices for both spatial window and temporal unit
(square mile and day) ensures that regularities in space and time are deduced away from the
treatment of interest, i.e. communication disruption.
Furthermore, table (8) includes the results of the same poisson regression, only with the
parameters changing by time in the panel. Again disruption and lagged dependent variable
are significant, and these effects are robust.
The results of the Poisson count regressions are included in table (7). The existence
of conflict (as a binary index) is regressed over a number of control variables, including
indicators for communication disruption and the existence of conflict in the previous day
in the immediate Moore neighborhood. Across all four models the results on the controls,
i.e. a daily lagged version of the dependent variable, an index for the spatial cell index, the
date index, average elevation of the square-mile cell, population and sum street length, a
neighborhood index (ranging over 36 neighborhoods), and the ethnic and income composition
of these neighborhoods are robust.
An examination of control variables in table (7) shows that unlike peripheral rebellions
(Fearon and Laitin 2003) urban conflict in Damascus in 2012 was less likely to occur in high
elevations. Instead lower altitudes were more likely to be a scene of conflict. The effect is
robust and significant across all models. The sum length of streets in each cell was also
strongly related to the rates of violent occurrences in Damascus. Streets represent conduits
for warfare and throughways for visibility. Urban conflict is staged in public venues, including
streets and alleyways; hence, the sum street length’s expected robust and significant positive
impact on conflict. The index for neighborhood location is also positive and significant,
showing that the borders of the 36 neighborhoods capture the building blocks of a conflictual
puzzle. The significance of neighborhood indices in the presence of other control variables
29
Table 7: Panel Dataset Regressions (Poisson), Rates of Violent Conflict in a Cell-Day
(1)
vid
(2)
vid
(3)
vid
(4)
vid
vid lagged
1.656∗∗∗
(0.123)
1.638∗∗∗
(0.123)
1.605∗∗∗
(0.126)
1.592∗∗∗
(0.125)
cell index
0.00417∗∗∗
(0.000879)
0.00413∗∗∗
(0.000879)
0.00420∗∗∗
(0.000881)
0.00416∗∗∗
(0.000881)
date
0.00547∗∗∗
(0.000579)
0.00519∗∗∗
(0.000588)
0.00529∗∗∗
(0.000585)
0.00503∗∗∗
(0.000593)
average elevation
-0.00389∗∗∗
(0.000734)
-0.00386∗∗∗
(0.000735)
-0.00379∗∗∗
(0.000734)
-0.00377∗∗∗
(0.000735)
population
0.0000143∗∗∗ 0.0000144∗∗∗ 0.0000143∗∗∗ 0.0000143∗∗∗
(0.00000259) (0.00000259) (0.00000259) (0.00000259)
sum street length
0.0000165∗∗∗ 0.0000166∗∗∗ 0.0000161∗∗∗ 0.0000163∗∗∗
(0.00000418) (0.00000419) (0.00000418) (0.00000418)
neighborhood index
0.0209∗∗∗
(0.00536)
0.0211∗∗∗
(0.00535)
0.0205∗∗∗
(0.00540)
0.0207∗∗∗
(0.00539)
ethnicity
0.0900
(0.146)
0.0881
(0.145)
0.0888
(0.146)
0.0867
(0.146)
income
1.647∗∗∗
(0.143)
1.644∗∗∗
(0.143)
1.631∗∗∗
(0.143)
1.630∗∗∗
(0.143)
disruption
0.764∗∗∗
(0.218)
spatial temp lag d
cons
N
Log-Likelihood
χ2
-4.829∗∗∗
(0.521)
69300
-2567.2
1237.2
-4.817∗∗∗
(0.521)
69300
-2562.2
1247.2
Standard errors in parentheses
∗
p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001
30
0.753∗∗∗
(0.218)
0.262∗
(0.114)
0.254∗
(0.114)
-4.881∗∗∗
(0.521)
69300
-2564.7
1242.2
-4.869∗∗∗
(0.521)
69300
-2559.8
1251.9
Table 8: Panel Dataset Regressions (Poisson), Rates of Violent Conflict in a Cell-Day, Fixed
Effects
(1)
vid
vid lagged
0.800∗∗∗
(0.125)
disruption
1.139∗∗∗
(0.215)
spatial temp lag d
0.601∗∗∗
(0.118)
16225
-1994.6
123.5
N
Log-Likelihood
χ2
Standard errors in parentheses
∗
p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001
suggests that the civil war in damascus was framed on the level of local urban life. The
results confirm the importance of spatial neighborhood composition on the intensity and
nature of civil conflict (Gould 1995). Ethnicity, defined as the difference between the Sunni
and Shia population proportions does not show a significant effect. Poor neighborhoods are
more likely to experience conflict.
The results in table (7) show that the spatiotemporal lag parameter29 was significantly
correlated with the likelihood of conflictual incidents. Finally, the impact of the disruption
on the likelihood of violent conflict is also positive and statistically significant, even after
controlling for spatiotemporal lags, cell index and temporal trends. The government’s decisions on attacking certain neighborhoods are based on their history of unrest, their overall
geographical characteristics and history of violence. These parameters are included as control variables. Despite the significance of all these parameters, communication disruption
shows a positive and significant influence on the rates of conflict. Figure (7) shows the effect
29
Defined as a binary variable indicative of the existence of conflict in the previous day in the Moore
neighborhood, i.e. 8 squares surrounding each square mile.
31
of multipliers in table (7). All 95% confidence intervals, other than those of the ethnicity
parameter, are distinctly above or below zero and are significant.
To ensure that the proliferation from a square-mile cell to the neighboring ones was not
a usual endogenous trend, I examined the 9-month average statistics in Appendix SI. 5 to
show that unlike the panel data, the merely spatial information, averaged in time, does not
show significant contagion patterns in Moore neighborhoods.
Similar results hold when instead of a binary indicator for the existence of conflict,30 the
number of conflictual events in the Moore neighborhood in the previous day is used. To
further illustrate the importance of spatiotemporal lag as a predictor for political violence,
figure (8) shows the increase in the likelihood of conflict in a square-mile cell as a function of
the number of incidents in its Moore neighborhood in the previous day. One more conflictual
incident in the vicinity in the previous day translates to an approximately 25% increase in
the likelihood of a violent incident in a given cell.31
Escalation: Contagion was Activated by the Disruption
Finally for examining the potential influence of communication disruption on patterns of contagion across space and time, I included an interaction term (Kam and Robert J. Franzese
2007) (spatiotemporal lag × disruption) along with explanatory parameters in table (7). The
interaction term is a proxy for the influence of disruption as a treatment on the likelihood of
violent incidents, this time via contagion processes. The results are included in table (9). All
control variables in table (7) are also included in the procedure, but the results are similar
and omitted for the sake of brevity.
The results in table (9) culminate in a noteworthy finding. While in the absence of the
interaction term both the contagion proxy and the disruption index are significantly influencing the odds of conflict, when the interaction term is included, it is only the interaction
term that is significant. This means the influence of disruption on the odds of conflict in
30
31
spatial temp lag d
The number of previous incidents in the Moore neighborhood varied between 0 and 4.
32
Coefficient Plot
spatial_temp_lag
Coefficient
disruption
income
ethnicity
vid_lagged
0.0
0.5
1.0
1.5
2.0
Value
Coefficient Plot
neighborhood index
Coefficient
average elevation
date
cell index
0.00
0.01
0.02
0.03
Value
Coefficient Plot
Coefficient
sum street length
population
0.0e+00
5.0e-06
1.0e-05
1.5e-05
2.0e-05
2.5e-05
Value
Figure 7: Coefficients from Poisson regressions in table (7)
33
#
0.006
0.005
+
0.004
of violent incident on day t
+
0.003
+
+
+
+
+
0
1
2
3
4
# of incidents in 1 mile Moore neighborhood on day t-1
Figure 8: An increase of one in the number of neighboring incidents in the previous day
translates to approximately 25 % higher rates of a violent incident in one’s 1 mile spatial
Moore neighborhood.
34
Table 9: Panel Regressions (Poisson), Rates of Violent Conflict in a Cell-Day, Interaction
(1)
vid
(2)
vid
disruption
0.753∗∗∗
(0.218)
0.557
(0.286)
spatial temp lag d
0.254∗
(0.114)
0.217
(0.119)
disruption# spatial temp lag d
N
Log-Likelihood
χ2
1.304∗∗∗
(0.324)
69300
-2559.8
1251.9
69300
-2559.1
1253.4
Standard errors in parentheses
∗
p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001
this regression is captured through its interaction with contagion. In other words, the media
disruption exacerbated revolutionary unrest via activating spatiotemporal contagion.32
Conclusion: Social Value of Information Depends on
Social Network Structure
In the course of this study I examined the puzzling outcome of a blackout during the Syrian
Civil War in order to contrast and compare two theories of collective action in urban warfare,
coordination and contagion. The results of a detailed GIS study provided evidence for the
significance of spatial spillovers, particularly during the blackout. Detecting contagion processes means blackout should have influenced the contention the way it did: complex contagion does suffer from the introduction of long communication ties (Centola and Macy 2007).
32
Fixed effect regression results corroborate the evidence on the significance of the variables of interest,
and are omitted for the sake of brevity.
35
If contagion was prevalent, then public communication could have been a hindrance not a
facilitator, as it is for coordinated collective action based on selective incentives (Lichbach
1995). It is likely that in the absence of coordination social facilitators such as lack of leadership, urge to find news, and the compulsion of an ominous blackout forced more face-to-face
interactions in pubic spaces and motivated unintended contacts with instigators of conflict
already operating in the spatial neighborhood. The result was local clustering and global
dispersion. Contagious spillovers were shown to be effective in general, and were aggravated
during the media shutdown, in particular.
Yet there is another concrete way to show why removing means of communication can
increase rates of participation. To offer a theoretical explanation for the phenomenon, I
employ a “beauty contest” model, first introduced in (Keynes 1936), and later used as
a model of collective social behavior (Morris and Shin 2002, Angeletos and Pavan 2007,
Bergemann and Morris 2013), and more recently protests (Little 2014).
Consider a situation in which each individual i is receiving two signals, one public y, and
the other private xi . Each of these signals is a noisy version of the state of the affairs θ ∈ R.
Similar to (Morris and Shin 2002) and (Bergemann and Morris 2013) assume that the noise
is additive,
y = θ+η
(2)
xi = θ + ǫi ,
η ∼ N (0, ση ), and ǫi ∼ N (0, σx ). Each individual i chooses action ai ∈ R (Note the
advantage of this model over the binary action–non-action model, taking action at the equilibrium, either on the side of the regime or the rebels becomes possible). The state of the
world, e.g. an indicator of state power, θ, is distributed uniformly on an interval ∈ R. If
the action profile for the whole population is taken to be a, then the utility of action ai in
state θ for individual i is a function of two terms, the first is a penalty based on the distance
36
between ai and the state of the world θ. The other is a “conformity” utility term,
ui(a, θ) = −(1 − r)(ai − θ)2 − r(Li − L̄),
where 0 < r < 1, and
Li =
Z
1
2
(ai − aj ) dj,
L̄ =
0
Z
(3)
1
Lj dj.
(4)
0
The first term is the conventional penalty for action deviation from θ. The second term is
a “conformity penalty”. Individual i incurs a smaller loss from acting if the action distance
profile from others, captured in the scalar parameter Li is close to the average distance profile
L̄ of everybody. All are trying to guess other’s guesses, while staying close to the real state
of the affairs. Each individual i decides between the utility from participation in collective
action setup, i.e. gaining the utility in equation (3) at equilibrium (on either side of the
contention), and abstaining from it altogether (ui = 0).
Proposition: In small world networks (Watts and Strogatz 1998), and in the framework
outlined in equations (2) and (3) removing the public signal increases the rates of participation in collective action.
Proof: See Appendix SI. 6.
A comparison of the two configurations in figure (9) reveals the relevance of the above
result. In the previous section, the spatial distribution of conflict in Damascus during the
blackout was found to represent dispersed clusters of contention, close-knit on the local and
widespread on the global level (top schematic in figure (9)); this is a network configuration
similar to a small network model (bottom schematic in figure (9)) in which local connectivity
is interspersed with a small number of long network bridges. Based on the utility and
signaling models outlined above, such networks induce an increase in the rates of collective
action in the absence of the public signal, itself caused by a shutdown of communication
37
media from similar to the Syrian blackout of November 2012.33
The results emphasize the distinctive nature of seemingly leaderless and spontaneous
mobilization (Kuran 1989). When collaboration in large groups is infeasible, contagion on
the mass scale operates effectively. Interpersonal transfer of attitudes, including rebellion,
is much more likely in urban environments compared to rural theaters of insurgency. However, clustering of rebellion is mainly studied at the center (Oliver, Marwell, and Teixeira
1985) not elsewhere, because leaders are often assumed to be socially central. In contrast,
the findings of this study invite more attention to the importance of peripheral clustering
of contention and the distinctive dynamics of collective action cascades originating from
marginal hotbeds of urban conflict.
33
The disruption itself is not a public signal, as it can either alarm the population about an imminent
strike or spread the impression of a weak state. The net psychological effect of such rare and consequential
events is not clear. It certainly does not help to generate an accurate public signal among the disconnected
public, instead it causes it to wither.
38
Figure 9: Top: schematic depiction of decentralization and localization processes after the
disruption, each dot represents a conflictual incident. Bottom: a small world network with
random bridges (dashed) and connectivity radius r > 1, here r = 2.
39
References
Alexander, Keith. 2013. National Security Agency Data Collection Programs. C-SPAN
http://www.c-span.org/video/?313429-1/nsa-chief-testifies-damage-surveillance-leaks.
Angeletos, George-Marios, and Alessandro Pavan. 2007. Efficient use of information and
social value of information. Econometrica 75 (4):1103–1142.
Axelrod, Robert. 1997. The Dissemination of Culture: A Model with Local Convergence
and Global Polarization. Journal of Conflict Resolution 41 (2):203–226.
Bamford,
James.
2014.
The
Most
Wanted
Man
in
the
World.
Wired
http://www.wired.com/2014/08/edward-snowden.
Bergemann, Dirk, and Stephen Morris. 2013. Robust Predictions in Games with Incomplete
Information. Econometrica 81:1251–1308.
Cederman, Lars-Erik, Nils B. Weidmann, and Kristian Skrede Gleditsch. 2011. Horizontal
Inequalities and Ethnonationalist Civil War: A Global Comparison. American Political
Science Review 105 (3):478–495.
Centola, Damon. 2010. The Spread of Behavior in an Online Social Network Experiment.
Science 329 (5996):1194–1197.
Centola, Damon, and Michael W. Macy. 2007. Complex Contagions and the Weakness of
Long Ties. American Journal of Sociology 113 (3):702–734.
Chozick, Amy. 2012. For Syria’s Rebel Movement, Skype Is a Useful and Increasingly Dangerous Tool. The New York Times A12. http://goo.gl/J8yZuR.
Christia, Fotini. 2013. What Can Civil War Scholars Tell Us About the Syrian Conflict?
In The Political Science of Syria’s War, POMEPS Briefings 22, edited by Marc Lynch.
Washington, DC.
40
Clark, Philip J., and Francis C. Evans. 1954. Distance to Nearest Neighbor as a Measure of
Spatial Relationships in Population. Ecology 35 (4):445–453.
Collier, Paul, and Anke Hoeffler. 2004. Greed and grievance in civil war. Oxford Economic
Papers 56 (4):563–595.
Diggle, Peter J. 2013. Statistical Analysis of Spatial and Spatio-Temporal Point Patterns.
New York, NY: Chapman and Hall.
Dobson, J.E., P.R. Coleman, R.C. Durfee, and B.A. Worley. 2000. LandScan: a global
population database for estimating populations at risk. Photogrammetric Engineering and
Remote Sensing 66 (7):849–857.
Dunning, Thad. 2012. Natural Experiments in the Social Sciences: A Design-Based Approach. New York, NY: Cambridge University Press.
Fearon, James D., and David D. Laitin. 2003. Ethnicity, Insurgency, and Civil War. American
Political Science Review 97 (1):75–90.
Gleditsch, Kristian Skrede, and Nils B. Weidmann. 2012. Richardson in the Information
Age: Geographic Information Systems and Spatial Data in International Studies. Annual
Review of Political Science 15:461–481.
Gohdes, Anita R. 2014. Pulling the Plug: Network Disruptions and Violence in the Syrian
Conflict Working paper.
Gould, Roger V. 1995. Insurgent Identities: Class, Community, and Protest in Paris from
1848 to the Commune. Chicago, IL: Chicago University Press.
Granovetter, Mark S. 1978. Threshold Models of Collective Behavior. The American Journal
of Sociology 83 (6):1420–1443.
Hassanpour, Navid. 2014. Media Disruption and Revolutionary Unrest: Evidence from
Mubarak’s Quasi-Experiment. Political Communication 31 (1):1–24.
41
Holliday, Joseph. 2012. Syrian Civil War’s Location Dataset. Washington, DC: Institute for
the Study of War.
———. 2013. The Assad Regime: From Counterinsurgency To Civil War. The Institute for
the Study of War .
Holliday, Joseph, and Michael Lynch. 2012. The Battle for Damascus: The Current State of
Play in Syria. ISW-Institute for the Study of War .
Horowitz, Donald L. 2000. Ethnic Groups in Conflict. Berkeley, CA: University of California
Press, 2nd edn.
Howard, Philip N., Sheetal D. Agarwal, and Muzammil M. Hussain. 2011. When Do States
Disconnect Their Digital Networks? Regime Responses to the Political Uses of Social
Media. The Communication Review 14 (3):216–232.
Izady, Michael. 2013. The Gulf/2000 Project http://gulf2000.columbia.edu/maps.shtml.
Kalyvas, Stathis N., and Matthew Adam Kocher. 2007. How “Free” Is Free Riding in
Civil Wars? Violence, Insurgency, and the Collective Action Problem. World Politics
59 (2):177–216.
Kam, Cindy D., and Jr. Robert J. Franzese. 2007. Modeling and Interpreting Interactive
Hypotheses in Regression Analysis. Ann Arbor, MI: University of Michigan Press.
Kern, Holger L., and Jens Hainmueller. 2009. Opium for the Masses: How Foreign Media
Can Stabilize Authoritarian Regimes. Political Analysis 17 (4):377–399.
Keynes, John Maynard. 1936. The General Theory of Employment and Money. London,
UK: Macmillan.
Khaddour, Kheder, and Kevin Mazur. 2013. The Struggle for Syria’s Regions. POMEPS
Briefings 20 43 (269):2–11.
42
King, Gary, Jennifer Pan, and Margaret Roberts. 2013. How Censorship in China Allows
Government Criticism but Silences Collective Expression. American Political Science Review 107 (2):326–343.
Kuran, Timur. 1989. Sparks and Prairie Fires: A Theory of Unanticipated Political Revolution. Public Choice 61 (1):41–74.
LandScan Global Web Applications. 2013.
LandScan Global Web Applications
http://goo.gl/bfwQND.
Lichbach, Mark Irving. 1995. The Rebel’s Dilemma. Ann Arbor, MI: University of Michigan
Press.
Little, Andrew. 2014. Communication Technology and Protest. Proceedings of Annual Meeting of the Midwest Political Science Association .
Lyall, Jason. 2010. Are Coethnics More Effective Counterinsurgents? Evidence from the
Second Chechen War. American Political Science Review 104 (1):1–20.
Lynch, Marc. 2013. The Political Science of Syrias War. POMEPS Briefings 22 .
Lynch, Marc, Deen Freelon, and Sean Aday. 2014. Blogs and Bullets 3: Syria’s Socially
Mediated Civil War. United States Institute of Peace .
Marwell, Gerald, Pamela E. Oliver, and Ralph Prahl. 1988. Social Networks and Collective
Action–A Theory of the Critical Mass. III. American Journal of Sociology 94 (3):502–534.
Metternich, Nils W., Cassy Dorff, Max Gallop, Simon Weschle, and Michael D. Ward. 2013.
Antigovernment Networks in Civil Conflicts: How Network Structures Affect Conflictual
Behavior. American Journal of Political Science 57 (4):892–911.
Morris, S., and Hyun Song Shin. 2002. Social Value of Public Information. American
Economic Review 92:1521–1534.
43
Nordhaus, William. 2006. Geography and Macroeconomics: New Data and New Findings.
Proceedings of the National Academy of Sciences 103 (10):3510–3517.
Nordhaus, William, Qazi Azam, David Corderi, Kyle Hood, Nadejda Makarova Victor, Mukhtar Mohammed, Alexandra Miltner, and Jyldyz Weiss. 2006.
tailed description of derivation of G-Econ data.
De-
Yale University, Typescript
http://gecon.yale.edu/sites/default/files/gecon_data_20051206_1.pdf.
Oliver, Pamela, Gerald Marwell, and Ruy Teixeira. 1985. A Theory of the Critical Mass.
I. Interdependence, Group Heterogeneity, and the Production of Collective Action. The
American Journal of Sociology 91 (3):522–556.
Olson, Mancur. 1971. The Logic of Collective Action. New York, NY: Schocken Press.
Parkinson, Sarah Elizabeth. 2013. Organizing Rebellion: Rethinking High-Risk Mobilization
and Social Networks in War. American Political Science Review 107 (3):418–432.
Petersen, Roger. 2013. Roles and Mechanisms of Insurgency and the Conflict in Syria. In
The Political Science of Syria’s War, edited by Marc Lynch. Washington, DC.
Pierskalla, Jan H., and Florian M. Hollenbach. 2013. Technology and Collective Action: The
Effect of Cell Phone Coverage on Political Violence in Africa. American Political Science
Review 107 (2):207–224.
Reynolds, James. 2012. Syria conflict: ’Fierce clashes’ near Damascus airport. BBC News
http://www.bbc.com/news/world-middle-east-20547799.
Rohozinski, Rafal. 2013. personal correspondence .
Scott, James C. 1998. Seeing Like A State: How Certain Schemes to Improve the Human
Condition Have Failed. New Haven, CT: Yale University Press.
Seale, Patrick. 1988. Asad of Syria, The Struggle for The Middle East. New York, NY: I.
B. Tauris.
44
SecDev. 2012. Syria Goes Offline .
Shapiro, Jacob N., and Nils B. Weidmann. 2014. Is the Phone Mightier than the Sword?
Cell Phones and Insurgent Violence in Iraq. International Organization Forthcoming.
Siegel, David A. 2009. Social Networks and Collective Action. American Journal of Political
Science 53 (1):122–138.
Syrian Observatory of Human Rights. 2012.
Syrian Observatory of Human Rights
https://www.facebook.com/syriaohr.
Van Dam, Nikolaos. 2011. The Struggle for Power in Syria: Politics and Society under Asad
and the Ba’th Party. New York, NY: I. B. Tauris.
Ward, Michael, and Kristian Skrede Gleditsch. 2008. Spatial Regression Models. Thousand
Oaks, CA: Sage.
Ward, Michael D., and Kristian Skrede Gleditsch. 2002. Location, Location, Location: An
MCMC Approach to Modeling the Spatial Context of War. Political Analysis 10 (3):244–
260.
Watts, Duncan J., and Steven Strogatz. 1998. Collective dynamics of ’small-world’ networks.
Nature 393 (6684):440–442.
Wood, Elisabeth J. 2008. The Social Processes of Civil War: The Wartime Transformation
of Social Networks. Annual Review of Political Science 11:539–561.
45
Supporting Information
SI 1: Micro-level Data on Connectivity in Spatial Units
SI 2: First Incidents During the Blackout
SI 3: The Construction of the GIS Dataset
SI 4: Further Clustering Analysis
SI 5: Confirming Lack of Average Spillover in Moore Neighborhoods
SI 6: Public Information and Protest Participation Dynamics
46
Supporting Information for A Quasi-Experimental
Study of Contagion and Coordination in Urban
Conflict: Evidence from The Syrian Civil War in
Damascus
1
SI 1: Micro-level Data on Connectivity in Spatial Units
In my search for reliable connectivity proxies I examined the statistics of geocoded tweets
generated in the spatial window containing Damascus and its suburbs with the hope of
creating data on the levels of online activity on a daily basis and to detect irregularities on
both spatial and temporal levels. This strategy, if applicable, could yield the most detailed
account of connectivity trends in the Syrian Capital. However, a preliminary inquiry1 showed
that the total number of geocoded tweets generated inside the 14 by 18 mile rectangle
containing Damascus and its suburbs during the whole year of 2012 did not exceed 4000.
But increasingly this is not the case. With the rise of geocoded online publication, more
network patterns of collective behavior are coming to light. Such detectable spatiotemporal
patterns are expected to promote the analysis and prediction of collective behavior on a
massive scale. This study utilized the available data from the Syrian conflict in order to
detect processes of contagion and escalation. Similar methods can effectively apply to the
increasingly rich information on closely watched conflicts in the urban terrain.
1
Historical tweet search by Gnip Inc. on the author’s behalf.
2
SI 2: First Incidents During the Blackout
The pattern of first incidents during the month of November and the first day of December
is depicted in figure (1). Including the conflict history of the ten months prior to November
allays concerns for temporal boundary effects. The temporal distribution of the 14 first
incidents that happened in November and December 1 in figure (1) reveals the unprecedented
concentration of these incidents during the final days of the blackout, November 30 and
December 1. These are locations that had not experienced violent conflict during the past
eleven months at all.
0.20
0.15
0.00
0.05
0.10
Proportion
0.25
0.30
0.35
First Incidents, 11/01 - 12/01/2012
Date, 11/01 - 12/01/2012
Figure 1: Distribution of first incidents at a particular location during the month of November and December 1st. The blackout persisted from November 29 to December 1. November
30 and December 1, the last two days in the plot, show an unprecedented number of such
occurrences, and are the only two data points at least 2σ away from the mean.
3
SI 3: The Construction of the GIS Dataset
Choice of Grid Size: The average elevation for each cell, a measure of population on
a mile-by-mile basis, and the total length of roads situated inside each cell comprise the
main structural elements of interest: the elevation level influences the nature of insurgency
in each area, the size of population in urban areas influences the organization of collective
action, finally the density of roads in neighborhoods represents proxy levels for visibility in
the geographical terrain, roads are also spatial conduits for street battles and their density
is expected to have a significant role in altering the dynamics of contagion.2
Elevation shows a negative correlation with urban conflict in Damascus across the board
with the three grid sizes in table (1). Unlike peripheral rebellions that are more likely to
take root in remote and inaccessible terrains, urban warfare was more likely to happen in the
plains. Total population is expected to be positively correlated with rebellion: the denser a
neighborhood’s population is, the more likely it is to experience urban conflict. Implicit in
this conclusion is the higher likelihood of contagion processes that potentially contributed
greatly to the conflict in Damascus’ urban areas. Along the same lines, total length of
streets in each grid cell is robustly linked to the number of incidents in the relevant area.
Streets are conduits for proliferation of open air conflict, as well as links of visibility in the
urban space. A spatial window with no population and no streets is unlikely to produce
any violent incident during the Syrian Civil War. Based on the same criterion, I found the
1-mile grid preferable to the 2 and .5-mile grids, because the latter two did not capture the
significance of the two main control variables of interest (population and total street length)
on par with the 1 mile grid. The 1 mile grid is the only parsing mechanism among the three
that simultaneously captures the significance of elevation, population, and the sum length
of streets in engendering urban conflict in each geographical unit.
2
Average elevation and total length of streets were calculated using the ArcGIS software
and digitized maps of Damascus available on OpenStreetMap http://www.openstreetmap.org/ total population in each cell was calculated using a proxy for population via LandScan project
http://web.ornl.gov/sci/landscan/ (Dobson et al. 2000, LandScan Global Web Applications 2013).
4
On the 1-mile grid average control parameters are the following. The average elevation
was 761.7 meters, average total population in each cell was 11546 for a total population
of approximately 2,900,000 (this includes the suburbs as well as the city). Note that the
official figures were approximately 2.6 million in 2004. Finally an average of 16382 meters of
throughways exist in each cell. In the year 2012 an average of 5.1 violent incidents happened
in each one-mile by one-mile cell covering Damascus and its suburbs.
Table (1) contains Poisson count regressions for grids of size 2, 1, and half-mile. Figure
(2) contains patterns of population density and elevation for the spatial window of interest.
Table 1: Dependent variable: number of violent incidents in each cell, Poisson count regression
Average Elevation
in each cell
(1)
2-mile Grid
-0.00353∗∗∗
(0.000493)
(2)
1-mile Grid
-0.00390∗∗∗
(0.000528)
(3)
Half-mile Grid
-0.00378∗∗∗
(0.000401)
Total Population
in each cell
0.00000139
(0.00000103)
0.0000125∗∗∗
(0.00000212)
0.0000227∗∗∗
(0.00000624)
Sum Length of Streets 0.0000138∗∗∗
in each cell
(0.00000147)
0.0000312∗∗∗
(0.00000342)
-0.0000294∗
(0.0000134)
2.828∗∗∗
(0.382)
252
-1034.3
649.1
2.184∗∗∗
(0.294)
1188
-2205.8
126.2
cons
N
Log-likelihood
χ2
3.549∗∗∗
(0.356)
84
-569.3
631.6
Standard errors in parentheses
∗
p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001
Survey on Socioeconomic Composition of Damascus: Fourteen respondents with an
in depth knowledge of Damascus coded ethnicity and income levels for each neighborhood.
Ethnicity–Religion was chosen among categories Sunni, Alawite, Shia, Christian, Druze, and
Kurdish. The income level was coded on a tripartite level Rich, Middle Class, Poor. I was
5
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
202
203
204
205
206
207
208
209
210
211
212
213
214
215
184
185
186
187
188
189
190
191
192
193
194
195
196
197
166
167
168
169
170
171
172
173
174
175
176
177
178
179
149
150
151
152
153
154
155
156
157
158
159
160
161
131
132
133
134
135
136
137
138
139
140
141
142
143
113
114
115
116
117
118
119
120
121
122
123
124
125
96
97
98
99
100
101
102
103
104
105
106
107
78
79
80
81
82
83
84
85
86
87
88
89
60
61
62
63
64
65
66
67
68
69
70
71
43
44
45
46
47
48
49
50
51
52
53
25
26
27
28
29
30
31
32
33
34
35
8
9
10
11
12
13
14
15
16
17
198
180
162
144
126
108
90
72
54
36
18
0
199
181
163
145
127
109
91
73
55
37
19
1
200
182
164
146
128
110
92
74
56
38
20
2
201
183
165
147
129
111
93
75
57
39
21
3
148
130
112
94
76
58
40
22
4
95
77
59
41
23
5
42
24
6
7
Figure 2: Top: Population proxy from (LandScan Global Web Applications 2013), baseline
roads from OpenStreetMap are also included. Bottom: Elevation patterns of the spatial
window
6
also cognizant of the fact that Syria is witnessing large scale intra and international migration
(Khaddour and Mazur 2013), hence I emphatically asked the survey respondents to answer
these questions for the year 2012 alone.
7
SI 4: Further Clustering Analysis
A histogram of minimum distances in figure (3) shows that the distribution is far from
Poisson distribution which is characteristic of a random spatial distribution (Diggle 2013).
In fact the most frequent distances in the histogram are concentrated below 2 kilometers.
0
50
Frequency
100
150
Histogram of Min_Distance
0
5000
10000
15000
20000
Min_Distance
Figure 3: Histogram of distances from the nearest incident in space and time
8
SI 5: Confirming Lack of Average Spillover in Moore
Neighborhoods
To ensure the identification of contagion in the panel data was not confounded by parameters
existing beyond the temporal dynamics of the events during the nine months covered by the
analysis, I checked the insulation of each of the 1-square mile geographical cells from its
neighbors in temporal average. Table (2) provides the results of Poisson regressions on the
number of conflictual incidents in each of the 14 × 18 cells of 1 square miles over the same
control variables used in other regression analyses, and a dummy for the existence of conflict,
as well as total number of conflictual incidents, both in the Moore neighborhood of a given
cell.
The insignificance of the spatial lag index, and the negative and small value of the count
lag parameter demonstrate that spatial spillover is not a discernible phenomenon in the
average data. If that were not the case, the contagion analysis on the spillover of contention
from one spatial window to another could have been caused by a latent or implicit process
simultaneously facilitating conflict in adjacent cells. The results in table (2) show that on
average it is not the case. Significant components of conflict in the panel dataset result from
the temporal dynamics, not from latent parameters.
9
Table 2: Dependent variable: number of violent incidents in each cell, Poisson count regression
count
cell index
average elevation
(1)
count
(2)
count
(3)
count
0.00431∗∗∗
(0.000785)
0.00426∗∗∗
(0.000789)
0.00439∗∗∗
(0.000780)
-0.00327∗∗∗
(0.000613)
-0.00312∗∗∗
(0.000636)
-0.00372∗∗∗
(0.000635)
population
0.0000148∗∗∗ 0.0000146∗∗∗ 0.0000146∗∗∗
(0.00000241) (0.00000241) (0.00000242)
sum street length
0.0000169∗∗∗ 0.0000162∗∗∗ 0.0000192∗∗∗
(0.00000376) (0.00000383) (0.00000391)
neighborhood index
0.0186∗∗∗
(0.00491)
0.0184∗∗∗
(0.00492)
0.0183∗∗∗
(0.00482)
ethnicity
0.0968
(0.131)
0.105
(0.131)
0.130
(0.131)
income
1.632∗∗∗
(0.128)
1.612∗∗∗
(0.130)
1.764∗∗∗
(0.136)
spatial lag d
0.142
(0.151)
spatial lag count
cons
N
Log-Likelihood
χ2
-0.00469∗∗
(0.00163)
1.494∗∗∗
(0.433)
252
-848.1
1021.6
Standard errors in parentheses
∗
p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001
10
1.299∗∗
(0.482)
252
-847.6
1022.5
1.860∗∗∗
(0.451)
252
-843.8
1030.1
SI 6: Public Information and Protest Participation Dynamics: The Effect of Network Structure on Global
Games
Building on a model of public and private information usage introduced in (Morris and Shin
2002), in this section I demonstrate sufficient network conditions under which lack of public
information results in higher rates of collective action. I will show that the social value of
public information is contingent on the underlying social network structure, a connection that
has not been explored in the context of global games based on “beauty contests” (Keynes
1936).
Public Signaling and the Decision to Protest
Consider a situation in which each individual i is receiving two signals, one public y, and
the other private xi . Each of these signals is a noisy version of the state of the affairs θ.
Similar to (Morris and Shin 2002) and (Bergemann and Morris 2013) assume that the noise
is additive,
y = θ+η
xi = θ + ǫi ,
η ∼ N (0, ση ), and ǫi ∼ N (0, σx ). There is a continuum of individuals, each indexed on
the unit interval [0, 1]. Each individual i chooses action ai ∈ R. The state of the world, e.g.
an indicator of state power, θ, is distributed uniformly on an interval in R. If the action
profile for the whole population is taken to be a, then the utility of action ai in state θ for
individual i is a function of two terms, the first is a penalty based on the distance between
ai and the state of the world θ. The other is a “conformity” utility term,
11
ui(a, θ) = −(1 − r)(ai − θ)2 − r(Li − L̄),
where 0 < r < 1, and
Li =
Z
1
2
(ai − aj ) dj,
L̄ =
0
Z
(1)
1
Lj dj.
0
Individual i gains higher utility from acting if the action distance profile captured in the
scalar parameter Li is close to the average profile L̄. This is a setup similar to the beauty
contest (Keynes 1936) where each contestant is trying to guess other contestants’ guesses,
while staying close to the real state of the affairs.
***
Lemma: Introduction of network structure into global games can be done in multiple ways.
For example, it is possible to show that imposing a network correlation ρi between private
signals xi s and y does not alter the linear equilibrium strategy ((xi , y) are jointly Gaussian).
Based on the setup in (Bergemann and Morris 2013) the proof is straight forward and is
omitted for the sake of brevity. In this case the unique action equilibrium is
ai−eq =
αi y + βi (1 − r)xi
αi + βi (1 − r)
(2)
where
1
ρi
−
2
ση
σx ση
ρi
1
−
=
2
σx σx ση
αi =
βi
ρi is the correlation factor between the public signal y and the private signal xi . The
addition of the correlation factor ρi allows for the study of a more involved information
structure.
***
12
In the following, I outline a proof for the proposition on the relation between lack of the
public signal and rates of participation.
Proposition: In small world networks, removal of the public signal increases the rates of
collective action.
Proof: I compare the two utility terms in equation (1) in two situations 1) public and private
signals, y and xi , both are available 2) only private signal, xi , is accessible. I will show that
the utility from taking action is dependent on the quality of signals and the distribution of
signals dictated by the network structure (through its effect on pairwise distance between
private signals). In making a decision to act, each agent i is comparing the utility in (1)
to 0–the utility from staying neutral. Positive changes in utility incite taking action ai .
Negative changes in utility result in abstaining from taking action on either side. I examine
the two terms of the utility, conformity and fidelity equations, in turn.
1. Utility from Conformity (Consensus), −r(Li − L̄)
a. When there exist private and public signals, xi and y
Actions at the equilibrium are a linear combination of y and xi (Morris and Shin 2002),
see equation (2). It is easy to show that when an individual decides to take action, the
absolute value of conformity utility |Li − L̄| is a sum of terms of the following form,
2
γ(xi − xj ) =
β(1 − r)
α + β(1 − r)
2
(xi − xj )2 .
When the public signal is removed (e.g. in the case of a blackout), the above term is
multiplied by 1/γ, γ ≤ 1 :
b. Only private signals xi s–y is removed
In this case, actions ai = xi and the absolute value of the conformity utility term scales
with
(xi − xj )2 .
13
The consequences of the the removal of the public signal are noteworthy. If the term
Li − L̄ is negative, the absence of the public signal increases the overall utility by an order of
1/γ, making it more likely for the agents to participate in action. If the the term Li − L̄ is
positive then removing the public signal makes it less likely for the relevant agent i to take
part in action on either side. However the sign of Li − L̄ is a function of signal distribution
in one’s network. For node i in close-knit sub-communities where signals are closely related
P
and highly similar, the term Li that is a linear function of j (xi − xj )2 is smaller than
L̄ = Ei (Li ).
2. Fidelity Term, −(1 − r)(ai − θ)2
Similarly one can show that the change in the average difference between the equilibrium
action and the state of the world, (ai − θ)2 ,3 when the public signal is removed is equal to
(uf −p for utility only with private signal, uf −pp for the case with both public and private
signals)
∆uf = uf −p − uf −pp = (1 − r)(2r − 1 − σǫ2 /ση2 )
As expected the loss of utility resulting from the blackout is considerable when the public
signal is of high quality, i.e. ση2 is small. When that is not the case, the removal of public
signal does not drastically change the decision calculus.4
To summarize, the distribution of signals in one’s ego network effectively influences one’s
participation mechanism. The participation rates of those in close-knit communities with
below average signal-distances are likely to rise in the wake of the omission of a public signal.
Hence, in communities with strong communal and local connections media disruptions incites
action, while in loosely connected subcommunities the result is the opposite. See figure (4).
Network Configurations
3
Averaged over noise terms η and ǫi .
Assuming ση > σǫ and putting r = 1/2 + (1/2)(σǫ2/ση2 ) suffices to do away with the change in the fidelity
term.
4
14
Figure 4: Consider the above illustration of a four-member signal space (of xi s) in a twodimensional space. There is a close-knit community on the left, members of which all have
Li s that are smaller than L̄, hence they benefit from suppression of the public signal. The
change in utility after the signaling intervention is in the opposite direction for the node on
the right. It is possible to extend this analysis to general signal spaces.
Scheme 1.
Dense Locales Consider a version of signal network in Figure (4) where there are two
locales: one community in which all n signals are equidistant with distances d1 , a similar
community is located D signal units away, with all n members of that community being d2
units apart from each other, assume d2 > d1 , i.e. the second community is loosely connected.
It’s clear that the members of the first community have Li s smaller than the average distance
L̄, while all the members of the second community have Li s larger than the total average.
Hence when the public signal is shut down the members of the first community become more
likely to take part in collective action, while the opposite is true of the second community:
their loosely connected members are reliant on the public signal for deciding to take action
or to abstain.
Scheme 2.
Small world networks resemble spatially confined social networks among individuals
(Watts 2002). Each person is connected to a few neighbors, while some individuals are
connected to others across the universe of all individuals, i.e. everybody is connected to 2r
15
neighbors, while there are a small number of random diametrical bridges in the network, see
Figure (5). Take the signal distance (|xi − xj |) between neighbors in local neighborhoods
to be d. It is clear that the average signal distance in the whole network is d + ǫ, where
the small positive term ǫ exists because of rare diametrical bridges. Also for the majority
of individuals in a small world network, Li = d < L̄, which means in a small world network
an absolute majority are more likely to take part in collective action in the case of a public
signal blackout.
Figure 5: A small world network with random bridges (dashed) and connectivity radius
r > 1, here r = 2.
Conclusion
The underlying signal network dictates the results of a public signal shutdown. Those
positioned in close-knit signal communities, i.e. with a negative Li − L̄, benefit from the
disruption of the public signal and are more likely to take action. In contrast, those with
positive Li − L̄s lose utility when the public signal disappears, and are less likely to take part
after a blackout. Also note that a public signal of high fidelity (large α) makes the consensus
term irrelevant.
16
References
Bergemann, Dirk, and Stephen Morris. 2013. “Robust Predictions in Games with Incomplete
Information.” Econometrica 81: 1251–1308.
Diggle, Peter J. 2013. Statistical Analysis of Spatial and Spatio-Temporal Point Patterns.
New York, NY: Chapman and Hall.
Dobson, J.E., P.R. Coleman, R.C. Durfee, and B.A. Worley. 2000. “LandScan: a global
population database for estimating populations at risk.” Photogrammetric Engineering
and Remote Sensing 66(7): 849–857.
Keynes, John Maynard. 1936. The General Theory of Employment and Money. London,
UK: Macmillan.
Khaddour, Kheder, and Kevin Mazur. 2013. “The Struggle for Syria’s Regions.” POMEPS
Briefings 20 43(Winter): 2–11.
LandScan Global Web Applications. 2013. “LandScan Global Web Applications.”.
Morris, S., and Hyun Song Shin. 2002. “Social Value of Public Information.” American
Economic Review 92: 1521–1534.
Watts, Duncan J. 2002. “A simple model of global cascades on random networks.” Proceedings of the National Academy of Sciences 99(April): 5766–71.
17