Salient Cleavages, Tactical Cooperation and Violence between Armed Groups in Yemen

Abstract

p>Existing scholarship ignores relational interdependencies when attempting to understand the behaviour of non-state armed groups during civil war. This paper investigates the in- terconnected web of alliances and rivalries in the Yemen Civil War to answer the following question: why do armed groups fight each other? We employ a network approach to investigate determinants of intergroup violence. This emphasises the role of identity, arguing that operating in salient cleavages necessitates that groups align or distinguish themselves from each other. We further argue that informal cooperation incentivises violence the longer the war continues. These arguments are tested using pooled Exponential Random Graph Models to account for endogenous structures over time. Results indicate that shared identity is a significant driver of hostilities, moderating cooperation and amplifying the effects of group attributes. Robustness checks and simulations demonstrate that network models more accurately capture the underlying mechanisms to predict fighting in this case.</p

Full text

Salient Cleavages, Tactical Cooperation and Violence
between Armed Groups in Yemen
Sukayna Younger-Khan
Research Fellow
WZB Social Science Center Berlin
June 13, 2022
Abstract
Existing scholarship ignores relational interdependencies when attempting to understand
the behaviour of non-state armed groups during civil war. This paper investigates the in-
terconnected web of alliances and rivalries in the Yemen Civil War to answer the following
question: why do armed groups fight each other? We employ a network approach to in-
vestigate determinants of intergroup violence. This emphasises the role of identity, arguing
that operating in salient cleavages necessitates that groups align or distinguish themselves
from each other. We further argue that informal cooperation incentivises violence the longer
the war continues. These arguments are tested using pooled Exponential Random Graph
Models to account for endogenous structures over time. Results indicate that shared iden-
tity is a significant driver of hostilities, moderating cooperation and amplifying the effects
of group attributes. Robustness checks and simulations demonstrate that network models
more accurately capture the underlying mechanisms to predict fighting in this case.
Keywords civil war, armed groups, rivalry, identity, alliances, social network analysis
1

Introduction
Global conflict trends reveal a recent surge in multiparty civil wars, with intense fighting
between armed groups significantly contributing to intractability and claiming ever higher
fatalities (Cederman and Pengl, 2019, p. 2). Unfortunately, there is a lack of scholarship
on multidimensional explanations investigating intertwined competitive and cooperative dy-
namics between different kinds of armed groups with different motivations.
This study contributes to the literature on interactions between armed groups in civil
conflict. It addresses a significant gap by examining the effects of positive relationships (co-
operation) on negative interactions (conflict). We test a theory of violent competition that
emphasises group identity and tactical cooperation while incorporating relational dependen-
cies to address the following question: why do armed groups fight each other? The Yemen
civil war is used as a test case to examine two main mechanisms driving violence between
groups: tactical cooperation during active hostilities and shared identity based on group
motivations and salient cleavages.
Our enquiry is limited to the period from January 2015 (official start of Saudi-led in-
tervention) to March 2020 (latest pre-Covid 19 data). There are deliberate reasons for our
choice of case: first, the Yemeni conflict has been largely overlooked in existing research
due to difficulties untangling complex and volatile relationships in an evolving multiparty
environment, and second, the presence of large numbers of armed groups and their frequent
interactions enlarges the sample area, improving the efficiency and consistency of estimates
(Derpanopoulos, 2018, p. 98).
This paper employs a network approach to investigate determinants of intergroup conflict.
We focus on the role of identity, arguing that the salience of societal cleavages creates a
continuum of identity for groups to align with or distinguish themselves from each other.
Thus, moderately shared identity directly increases the likelihood of conflict between groups,
while minor similarities do not exert this influence. At the same time, we expect that the
indirect effect of shared identity on conflict increases significantly for moderately similar
dyads that fight common enemies. Thus, we also contend that informal cooperation between
groups incentivises violence the longer the war continues.
We test these arguments using Exponential Random Graph Models (ERGMs) to account
for endogenous structures arising from repeated interactions over time and explore the causal
pathway between identity, alliances, and violence. Preliminary findings indicate that shared
identity is a significant driver of hostilities, mediated by cooperation, and amplifies the
effects of group-level attributes. Using a variety of robustness measures and simulation, it is
shown that network models more accurately capture the underlying data-generating process
to predict fighting between groups in this case.
The following section discusses prominent research on interactions between armed groups,
including studies on alliance formation and rebel fratricide. Then, we present our theory of
inter-group conflict which is tested in a set of dyad-dependent ERGMs. We conclude with a
discussion of our findings and future avenues for research.
2

Background
We provide a brief, high-level overview of the state of the art. Scholarly attention has been
increasingly devoted to the determinants of group interactions, primarily alliance formation
(Akcinaroglu, 2012; Christia, 2012; Bapat and Bond, 2012; Quinn, Joshi, and Melander,
2019) and fratricidal violence (Christia, 2008; Fjelde and Nilsson, 2012). Two theoretical
camps focusing on either power or identity dominate this niche. Power relations theory
conceives armed group interactions (positive/negative) as driven by rationalist motivations
to win and maximize returns (Bapat and Bond, 2012; Fjelde and Nilsson, 2012; Nyg˚ard
and Weintraub, 2015), while identity theories emphasise the role of ethnicity or ideology in
influencing group behaviour (Conrad, Greene, et al., 2021; Gade, Hafez, and Gabbay, 2019;
Kalyvas and Balcells, 2010).
Identity theorists formulate arguments based on ideological, ethnic, linguistic, or political
cleavages which shape interactions with other groups, directly and indirectly, by providing
organisational advantages that influence their behaviours (Kalyvas and Balcells, 2010) or
affecting processes of fragmentation, recruitment, and cohesion within ideological movements
(Burton, 2015). Groups with distinctive identities may be able to exploit ethnic, tribal
or religious loyalties for recruitment and mobilization of fighters (Weinstein, 2006; Eck,
2009) and benefit from greater cohesion within movements leading to decreased likelihood
of fragmentation or defection (Kalyvas and Balcells, 2010; Burton, 2015).
The role of ethnicity in enabling, encouraging, or influencing hostile interactions between
armed groups in civil wars has been heavily scrutinised. Co-ethnic rebels are significantly
prone to fratricide since they view each other as existential threats competing for control
of the same territory, constituency, and resources. They are motivated by their expectation
to quickly absorb the resources of defeated rivals instead of other groups (Pischedda, 2020).
Exploiting new, disaggregated data also reveals that groups which share an ethnic identity
are more likely to engage in rivalry than others. In this context, power asymmetry is only
significantly associated with rivalry when co-ethnic motivations are present (Conrad, Greene,
et al., 2021).
Ideological and political motivations have been examined with somewhat consistent re-
sults, albeit without concurrent consideration of shared ethnicity or religion. Ideological
divergence can arise from differences in conflict narratives (stated justifications for fight-
ing), territorial ambitions (territorial nationalism or secessionist), political order (post-war
institutional structure) (Gade, Hafez, and Gabbay, 2019; San´ın and Wood, 2014).
Armed rivals’ opposing perceptions of irreconcilable ideological divides believed to threaten
their survival within ideological organisations may incite fratricidal violence - the so-called
‘proximity-distance paradox’ (Hafez, 2017, p. 605). Gade et al. find additional evidence to
support this proposition while studying infighting, but also present compelling results show-
ing that ideological homophily is a primary determinant of cooperation in their twin case
studies of Syrian rebel groups (Gade, Hafez, and Gabbay, 2019; Gade, Gabbay, et al., 2019).
Confusingly, evidence of the opposite effect in the latter case is equally convincing; Chris-
tia (2008), Lilja and Hultman (2011) and Fjelde and Nilsson (2012) present findings that
indicate that shared identities increase the risk of violence between similar groups. These
seemingly contradictory propositions can be reconciled if we consider the context of multi-
3

party conflicts: as the number of active rebel groups increases, competition for resources,
recruitment and popular support intensifies, leading groups to fight to eliminate each other
and ensure their survival.
Recent work has emphasized opportunity, theorizing about the exogenous conditions of
the environment which enable actors to fight each other or seek out the means to engage in
violence (Pischedda, 2020; Mendelsohn, 2021). We focus on motivation, examining endoge-
nous factors, such as group-level attributes and dyadic connections in the wider relational
environment, making certain groups in certain interactions more violent than others. Per-
ceptions of irreconcilable incompatibilities exacerbate conflict between groups. We argue
that this effect is particularly pronounced when groups are divided by a myriad of iden-
tity cleavages, as more similar and dissimilar groups view each other with distrust in the
absence of credible information about strength and motives, aggravated by their greed and
underlying grievances.
There have been some attempts to structure the motivations of armed groups. Phillips
refers to intrafield and interfield pairs of groups that either share their primary motivation
or differ substantially (Phillips, 2015; Bacon, 2018). Weidmann (2016) implements an agent-
based computational model of cleavages and violence in civil wars wherein conflict actors
differ along a set of identity dimensions that vary in salience over time. He observes that local
divisions reinforce local violence, especially in rural communities and micro-macro alliances
increase violence indirectly between local actors.1
Conflict outcomes and government strategies to combat the threat of multiple groups
cannot be dependent on investigating singular dimensions of identity. Assuming the inde-
pendence of dyads is not reflective of multiparty civil war wherein complex interdependencies
arise from cooperative and conflict relationships between actors and fluctuate over time with
the entry and exit of divergent groups in the environment (Dorff, Gallop, and Minhas, 2019).
Intergroup conflict affects relationships with and between actors by weakening an opponent’s
military capability and emboldening challengers; failure to account for consequential effects
decreases the predictive ability and explanatory power of conventional models (Cremaschi et
al., 2020). Ignoring connections between actions and effects leads to misleading conclusions
about group behaviour, nature, and patterns of violence.
Following the lead of Gade, Gabbay, et al. (2019) and Dorff, Gallop, and Minhas (2019),
we support a paradigm shift towards the dynamic, relational approach of network science.
The intertwined nature of actors and events during war creates relational systems, cate-
gorised by multi-level observations capturing actors (conceptualised as nodes or vertices in
network terms) and relationships between actors (i.e., ties or edges), thus including monadic,
dyadic, and multilateral relationships embedded within a conflict network. Persistent pat-
terns of interactions generate endogenous dependencies through structural network features
that define, enable or restrict the behaviour of actors (Hafner-Burton, Kahler, and Mont-
gomery, 2009, p. 562).
1These model predictions are not subjected to empirical testing as “the universe of local cleavages across different cases [cannot
be mapped out before the occurrence of violence] to find out which of them actually turn violent” (Weidmann, 2016, p. 555).
4

A Network Theory of Cooperation, Identity and Con-
flict
Violence between armed groups is not simply perpetrated along the macro-cleavage. Instead,
the wartime environment encourages the settling of old feuds and the creation of new rivalries
through violent means amongst armed groups other than insurgents or militants (Kalyvas,
2003). While outbidding theory may explain how high numbers of competitors can affect
rates of violence, this explanation is often limited to terrorism or civilian targeting during
civil wars with contradictory findings (Findley and Young, 2012; Cunningham, Bakke, and
Seymour, 2012; Conrad and Spaniel, 2021). We develop a multi-layered theoretical argument
about the perceived salience of shared identity, which both directly, and indirectly through
the mechanism of cooperation, motivates violence in interactions.
Salience of Shared Identity Cleavages
Drawing on Kalyvas’ proposition that micro-cleavages, such as tribal or clan rivalries, sub-
stantially influence civil war dynamics, we argue that the salience of cleavages is equally
deterministic of patterns of violence between groups (Kalyvas, 2003). Actors involved in
a conflict generally share some identity dimensions and differ along others, e.g., actors are
ethnolinguistic groups but adhere to different religions. However, similarities or differences
on their own are not indicative of which cleavages matter to actors. Only cleavages perceived
to be important by the actors will motivate groups to violence.
Salience is rendered through social, political, or economic relevance of identity dimensions
to actors in the environment, either through exclusion or perceived threats to group interests
(Cederman, Wimmer, and Min, 2010). Armed groups emphasise some dimensions more
than others, e.g., Salafist militants may consider religion the most crucial cleavage while
pro-government groups might emphasise political ideology. The salience of cleavages varies
across actors and over time, and it may be triggered by political elites activating in-group vs
out-group dynamics to reinforce and weaponise group identification (Tokdemir et al., 2021).
The organisational structure of armed groups may dictate their selection of tactics, strate-
gies and behaviour (Weinstein, 2006; Bultmann, 2018). As cleavages become more salient,
group identification among members increases. In-group cohesion improves as members
band together in their collective actions, and further differentiate themselves from outsiders
(Oakes, 2002; Turner et al., 1987). Consequently, perceived threats to group identity and
status make members more willing to incur fighting costs to reinforce their social influence
and weaken their competitors (Tokdemir et al., 2021).
The chaos of war empowers armed groups to engage in hostilities with others to settle
disputes and capture resources. But fighting incurs costs in blood and treasure, and rational
actors seek to maximise their gains by capitalising on windows of opportunity and vulner-
ability to pick their battles (Pischedda, 2020). There is more to be gained from competing
with groups that share ethnicity, religion, or tribal constituencies, operate within the same
cleavages and place equal or greater importance on maintaining their positions within those
cleavages.
Shared identity increases the perceived distance between actors; more similar groups
5

compete for the same resources, constituency, and loyalty. To differentiate themselves in
the conflict marketplace, they may adopt a more extreme stance, engage in violence against
civilians or fight each other to eradicate the threat of losing their status and relevance (Hafez,
2017; Asal et al., 2019). Similar groups thus face credible threats from each other as their
presence dilutes popular support, exacerbates divisions within aggrieved populations through
policing of in-group boundaries, and worsens competition for dominance, representation and
relevance in the political sphere (Christia, 2008; Burton, 2015). Armed groups that view
survival as a zero-sum game harbour greater hostility towards similar groups. Such groups
may deliberately target or attack others, actively participate in protracted hostilities, or
employ brutal, indiscriminate tactics in their interactions with each other.
Bultmann notes that “patterns of violence do not simply relate to pre-war motives or
ideologies but are, to a large extent, also the result of emergent dynamics of warfare itself”
Bultmann (2018, p. 609). The occurrence of violence between identity-proximate actors
indirectly influences neighbouring actors by signalling to them the importance of shared
cleavages. This reinforces the salience of specific cleavages to observers, whether close allies
or competitors and entrenches violence along those cleavages, e.g., constant fighting between
tribal militias signals to others that tribal identity is significantly more important than
national identity and empowers local groups to resort to violence to resolve tribal feuds and
grievances.
A well-studied endogenous property of social networks is homophily, often expressed as
“birds of a feather flock together” or the tendency of actors with common attributes to form
more links with each other (Taraktas, 2019; Goodreau, Kitts, and Morris, 2009). Groups
with shared identities may engage in greater conflict interactions that present in the conflict
network, clustering along group identity and higher transitivity.2We focus on five salient
cleavages along which armed groups distinguish themselves politically or socially in the on-
going war: ethnicity, ideology, religion, tribal/communal and pro-government. Actors are
similar across some dimensions while differing across others, e.g., belonging to the same
ethnic group but adhering to different religions. It is expected that groups are more likely
to compete and fight others with shared identities; in local communities, clashes frequently
occur between tribal/ethnic groups than with foreign-sponsored militias, while political mili-
tias frequently engage in violence against each other over political relevance or control of
territory and resources.
Most cleavages in the Yemeni social fabric are cross-cutting: members of ethnic groups
are also members of religious sub-groups and further divided into tribes and communities in
different geographic regions. Reinforcing cleavages played a significant role in the onset of
the current war and continuing violence between the marginalised Zaydi Shia Houthi tribe
and the affluent Sunni dominated government (Abduljaber, 2018). These politically salient
cleavages emerged along historical ethnic and tribal dimensions, the natural consequence of
festering grievances stemming from the 1990 unification, 1994 civil war and reunification of
republican North Yemen and socialist South Yemen states (Ghanem, 2019).
Rising tensions reignited historical rivalries between tribal groups and triggered infighting
within regional coalitions. In central and northern Yemen, burgeoning defiance of Houthi
2If actors aand band band care linked, then the concept of transitivity holds that aand cwould also be linked (Maoz, 2012,
p. 349).
6

rule has led to violent clashes between tribes with conflicting loyalties to local elites and
the Houthis (Carboni and Nevola, 2019). Secessionist groups continue to fight other rebels
perceived as competition in fractured Southern governorates (Roy and Carboni, 2019).
This leads us to the following hypothesis:
H1 Shared identity increases the likelihood of conflict between armed groups. The likelihood
of conflict is higher for dyads with more salient cleavages in common than other dyads.
Informal Cooperation As Fighting Strategy
Drawing from the logic that contemporary civil wars are characterised by ‘shifting coali-
tions of groups with malleable allegiances and at times divergent interests, only some of
whom actually engage in violence at any given point in time’, we reasonably assume that
the existence of cooperation influences the conflict environment as a whole (Pearlman and
Cunningham, 2011, p. 4). We focus on tactical cooperation between armed groups, defined
as friends in combat or otherwise ‘[coordinating] joint military attacks against a common
foe’ (Akcinaroglu, 2012, p. 885).
Rational groups are motivated to win the war and achieve maximum gains (material or
political). In highly fragmented wars, these are competing motivations as the presence of
more groups increase the candidate pool for cooperation. In this environment, groups favour
”marriages of convenience” that increase joint fighting capacity in the immediate future, lead
to greater cumulative gains and lower losses and improve chances of winning. Concurrently,
commitment problems caused by lack of trust or fear of exploitation inhibit prolonged coop-
eration through coalition-building and incentivise switching to relatively balanced alliances
for weaker groups.
Informal cooperation is primarily dictated by fight and flight – groups are momentarily
aligned against common enemies, either part ways after battles are over and spoils distributed
or switch sides to improve their chances. This leads to constant side-switching, defection, and
fractionalisation in the environment, increasing the chances of former allies facing each other
on the battlefield. Naturally, the presence of such transient prior cooperation (competition)
does not preclude future competition (cooperation).
Macro-micro alliances play an essential role in identity-based conflicts. Supralocal groups
fighting along the master cleavage often support subnational or local groups by providing re-
sources and training. In exchange, the latter benefit from the endorsement of tribal leaders,
local networks, and bigger pools for combatant recruitment (Kalyvas, 2003). The provision
of support is conditional on local groups fighting onside in the interests of the supranational
group; however, in practice, interests are unlikely to overlap fully or over more extended
periods. Emboldened local groups use resources to pursue their agendas by quashing local
rivals, improving their status and significance. Ironically, consolidating strength and influ-
ence within local networks may increase information asymmetries and commitment problems
in such alliances, and thriving local groups may threaten their ally’s interests in the region.
Weidmann refers to these as adverse selection and moral hazard problems (Weidmann, 2016,
p. 885)).
We propose that cooperation dynamics affect conflict ties in two main ways. First,
allied groups benefit from positive economies of scale (Steinwand and Metternich, 2020)
7

or synergistic benefits (Akcinaroglu, 2012, p. 885) which allow them to fight more or fight
back common enemies. In line with the axiom, “the enemy of my enemy is my friend”,
this would establish structural balance within the conflict environment. Second, although
these groups may engage in less conflict with their partners in the short term, the temporary
nature of tactical cooperation between groups with diverging goals and interests leads to the
dissolution of marriages of convenience, thus generating conflict between these groups in the
long term.
As conflict ties may likely be consequences of cooperation between groups, these create
second-order dependencies within the observed network, akin to reciprocity or retaliation in
directed networks (Dorff, Gallop, and Minhas, 2019). Since most tactical cooperation can
only occur between groups within the same geographical region, violent breakdown of these
arrangements can have third-order (network) effects by raising tensions in other dyads and
uncommon neighbours, diffusing conflict in that region.
Such volatile tactical alliances are omnipresent in Yemen; third-order effects arose from
the tumultuous relationship between the Supreme Political Council controlled by the Houthis
and the Republican Guard and Special Forces who rallied behind former President Ali Ab-
dullah Saleh in 2015. The two groups had coordinated military operations and conducted
attacks against the incumbent regime and Saudi-led coalition forces until 2017, when the
alliance finally disintegrated under the weight of deep mistrust and constant infighting. The
split led to six days of violence between multiple armed groups in the Sanaa Governorate
and hundreds of deaths, culminating in the assassination of Saleh and his top officials and
the Houthis seizing the capital (Carboni, 2018).
Thus, we formulate the following hypothesis:
H2 Cooperation increases the likelihood of conflict between armed groups. Armed groups
are more likely to form a conflict tie if they also share a cooperative tie with another
grou
Complementary logics
We account for group and dyad-level characteristics that influence intergroup conflict. Group
strength may have both first-order and second-order effects, e.g., stronger groups form more
conflict ties, and they are especially attractive opponents for equally matched or weaker
groups. Existing studies find support for the proposition that geographical location, territo-
rial control and foreign sponsorship affect absolute and relative power of groups during war
(Buhaug, Gates, and Lujala, 2009; de la Calle and S´anchez-Cuenca, 2015; Popovic, 2018).
We also control for years of active presence and coalition membership to accurately test our
hypotheses as these attributes may directly or indirectly spur the formation of ties within
the observed network.
8

Research Design
Data
In the absence of an agreed definition, non-state armed groups are best described in negative
terms: armed groups operating within the conflict territory, excluding state military and
armed forces defined in international law (Crawford, 2002). This definition allows for the
broad inclusion of different actors that challenge the state’s monopoly on violence and/or
are not aligned with the internationally-recognised government (Gleditsch et al., 2002).
Dependent Conflict Network
We use ACLED event data to construct the conflict network in Yemen from January 2015
to March 2020 (Raleigh et al., 2010). Compared to UCDP, which records actors involved in
hostilities exceeding 25 battle deaths per year, ACLED does not adhere to fatality counts,
thus recording more (and smaller) armed groups. No other dataset records the war for the
full period of interest. Battle observations are extracted to construct a network of conflict
interactions between groups. Due to difficulties associated with untangling attackers and
defenders, Curiel, Walther, and O’Clery (2020) consider ACLED event data as observations
of symmetric events resulting in an undirected network.
Conflict ties are conceptualised as active engagement in battles between groups, with
group-government dyads excluded during data pre-processing. These are weighted by battle
frequency of each dyad. A distinction is made between primary participation (direct fighting
party) and secondary participation (assisting or associating with primary parties), with only
the former constituting ties. The data is arranged as a weighted sociomatrix; the observed
network is made of 189 nodes and 2664 ties, consisting of 150 mutual dyads that engaged
in at least one battle over the 2015-2020 period (excluding infighting loops and unidentified
actors).
Independent Cooperation Network
We operationalise ties of cooperation as tactical assistance or coordination between armed
groups during attacks on their fighting partner(s), following Akcinaroglu (2012, p. 885). This
network is made of 1121 ties of tactical cooperation between the same 189 actors forming the
conflict network. It consists of 100 cooperative dyads which fought on the same side in at
least one battle. We verify the presence of this collaboration through secondary sources, i.e.
local, regional and international news reports in Arabic and English, UN Security Council
reports and policy briefs. We employ the completed network in Figure 1 as an independent
network to evaluate H1.
9

Abidah Tribal Militia
Al Humayqani Tribal Militia
Al Islah Party
Al Qaeda in the Arabian Peninsula
Giants Brigade
Militia Abu Hammam al Yafei
National Resistance Forces
Popular Resistance
Saiqa Brigades
Security Belt Forces
Shabwani Elite Forces
Southern Resistance
Southern Transitional Council
Supreme Political Council
General People's Congress
cooperation
100
200
300
400
500
cleavage
1
2
3
4
eigen_centrality
0.00
0.25
0.50
0.75
1.00
Figure 1: Independent cooperation network for armed groups in Yemen from 2015-2020 (excluding isolates). Node colour indicates
identity cleavages; node size indicates Eigenvector centrality scores which measures the amount of influence a node has within the
network based on popularity of its connections.3Thickness of ties indicates intensity of cooperation within dyads. Stress majorization
algorithm used for graph layout.
10

Covariate of Interest
In addition, we construct a dyadic covariate to test our hypotheses on the formation of
conflict ties between armed groups.
Shared Identity While identity scholarship is ubiquitous, this fascination has not yielded
any conceptual or methodologically clarity to measure identity as a variable (Abdelal et al.,
2009). To circumvent the lack of consensus, we employ a constructivist understanding of
identity based on the salience of societal cleavages, which is fluid and dynamic to accommo-
date the diversity of identity dimensions across countries and cases. This dyadic covariate is
a similarity matrix of shared identity cleavages per dyad. Using dichotomous indicators for
five politically and socially salient cleavages in the Yemeni civil war (ethnicity, religion, ideol-
ogy, tribal/communal, pro-government nationalism), we compare groups in pairs to identify
common cleavages. The edgewise attribute ranges from one to four based on the number
of shared dimensions and we do not observe any groups with all five cleavages in common.
This similarity measure allows for examination of dynamic group identities which emerge
relative to other groups in the immediate neighbourhood or wider conflict environment.
Controls
Power Projection In the absence of accurate troop size data on most recorded actors
in our network, we construct the power projection covariate as a proxy for group strength.
The measure is derived from active engagement in hostilities and geographical range of
armed groups. It assumes that conflict participation across multiple locations, whether
simultaneously or over time, has a deleterious effect on fighting capacity by diluting military
strength and over-extending limited troops. Through penalising armed presence in multiple
administrative regions, this measure better distinguishes between weak, weaker, strong and
stronger groups for a more accurate representation of strength.
Foreign Sponsorship This dichotomous covariate identifies whether the group received,
or presently receives, foreign state sponsorshi The involvement of third countries, especially
Saudi Arabia, UAE and Iran, have exacerbated the proliferation of armed groups with fluid,
capricious allegiances and affiliations.
Territorial Control This nodal attribute measures whether the group seized or overtook
territory during conflict events over the course of the observed period. The dichotomous
indicator is drawn from event data and supplementary sources, and excludes observations
involving state or state-affiliated groups recapturing territory; we are interested purely in
the effects of territorial control on the propensity of actors to form conflict ties with other
groups.
Coalition Group This is a dichotomous covariate for armed groups composed of multiple
sub-groups or factions, whether loosely organised or regimented and hierarchical. Where
these coalitions comprise distinct sub-groups who also engage independently in hostilities, the
sub-groups are included in the node set. For example, National Resistance Forces is primarily
11

composed of fighters from three sub-groups who also engage in hostilities independently;
Guardians of the Republic (pro-Saleh ex-military forces), Amaliqah Brigade/Giants Brigade
(anti-Houthi tribal fighters from Aden and Lahij) and Tihama Resistance (local tribal fighters
from Tihama). We coded NRF as a coalition group and included all four groups as nodes.
Geographical Presence This nodal attribute is a simple count of distinct administra-
tive regions where the group engaged in hostilities, whether anti-government or other groups,
drawn from ACLED and checked against secondary sources. As armed groups travel and
engage in conflict in multiple regions over the five years under observation, we assume that
this distinct count is an approximate indicator of the group‘s actual fighting range.
Years Active This accounts for temporal dependencies created by active involvement of
groups over time. It is coded for the number of years an armed group directly or indirectly
participates in at least one hostile incident at the intensity of an armed clash, battle or
offensive attacks. Where an armed group displays as non-active in one year antecedent
and precedent, between years of high-intensity activity, we follow UCDP best practices and
include the gap year in the count (Gleditsch et al., 2002).
Imputation of missing data
Missing data, particularly on node attributes and covariates, poses difficult problems for
verification and validity of results, with ‘commission errors‘ potentially leading to model
misspecification (Borgatti, Everett, and Johnson, 2018). As only 5.3% data are missing at
random (MAR) with no missing nodes or ties, multiple imputation is carried out using the
MICE package and including derived covariates as predictors (i.e. transform, then impute
approach (van Buuren, 2018, ch6). Following Rubin‘s rules (Rubin, 1987), predictive mean
matching is first performed on missing values for 50 iterations to generate m= 5 imputed
datasets, which are then used to estimate statistical network models. We pool all estimates
to arrive at final parameters, and evaluate MCMC and GOF diagnostics before proceeding
with the analysis (see Appendix).
Endogenous network statistics
Endogenous structures are structural features of the dependent network that have important
roles in the underlying data-generating process. We use the statnet package to specify
exponential random graph models (Handcock et al., 2019). These network models require
that researchers clearly specify endogenous dependencies which are functions of dyad states
in the network, identification and specification of these dependencies is essential to avoid
model misspecification and biased estimation (Cranmer, Leifeld, et al., 2017).
The growing popularity of inferential network analysis has not extended to conflict re-
search, particularly civil war scholarship (Hafner-Burton, Kahler, and Montgomery, 2009).
There is no consensus on common endogenous properties of conflict networks, and little
research into armed group networks. While we devote some attention to identifying and ex-
plaining network structures in our specifications, theoretically these statistics only function
as controls for certain endogenous dependencies.
12

Popularity and sociality effects: We first consider the generalised endogenous effect
of k-stars, referring to armed groups with conflict ties to kother groups in the network. The
simplest configuration of k-stars is the two-star found where an actor iengages in conflict
with two other actors jand kwho are not themselves in conflict. This statistic seeks to
capture a popularity effect: ‘the number of times in the network where two [actors] are [in
conflict] with the same [other actor]’ (Cranmer and Desmarais, 2011, p. 81). In undirected
networks, popularity and sociality refer to the same effect of forming many ties with other
nodes.
We use the geometrically weighted degree distribution statistic (gwdegree) to avoid de-
generacy and capture the anti-preferential attachment or ‘inverse popularity’ of more sociable
groups by imposing an increasing weight on higher-degree counts. The decay parameter of
α= 0.2 weights the likelihood of ties attaching or repelled from higher-degree nodes in the
network (Levy, 2016). Given that the observed network includes multiple groups fighting
mutual enemies, we expect to see a significant positive effect across specifications.
Triadic closure and transitivity effects: The triangle structure occurs when three
actors each have ties with each other. We assume that this structure is relatively unlikely
to form in our network, following the premise that ‘the enemy of my enemy is my friend’
(Cranmer and Desmarais, 2011; Maoz et al., 2007). However, as triangle causes degeneracy
by overestimating the propensity of actors to form triangles, geometric terms are used to
alleviate these concerns (Goodreau, Handcock, et al., 2008, p. 111).
To control for transitivity effects, we employ the geometrically weighted edge-wise shared
partner distribution statistic (gwesp) with the decay parameter α= 0 (Hunter, 2007). The
gwesp statistic captures the propensity of groups with pre-existing conflict ties to be more
likely to have multiple shared partners. Network sparsity means we do not include the
edgewise statistic (gwdsp) as gwesp effectively captures higher-order structures (Cranmer,
Leifeld, et al., 2017). Given few observed triangles, we expect to see a negative coefficient
resulting from estimation.
13

Empirical Analysis
Exploring the Conflict Network
Visualizing the conflict network and examining descriptive statistics provides us with insight
into structural features of the network. We use this to identify patterns of tie formation
between groups with certain attributes, and clustering effects.
The biggest connected component of the network in 2a shows some clustering along
identity cleavages; majority of high intensity conflict ties are formed between armed groups
with two or more salient cleavages. In contrast, groups varying along one identity dimension
such as tribal/communal groups have fewer and lower frequency ties with other groups, and
are predominantly located in the network periphery. The complete network in 2b provides
a fuller but noisier picture; over six years of conflict, most tribal groups have concentrated
on fighting similar groups, avoiding been drawn into the wider internationalised civil war.
This conflict environment provides less powerful groups with the windows of opportunity or
vulnerability to engage in resource competition to satisfy material greed, exact vengeance
for tribal feuds or resolve communal grievances (Pischedda, 2015).
Further examination of endogenous structures is included in the Appendix.
14

Al Humayqani Tribal Militia
Al Islah Party
Giants Brigade
Militia Abu al Abbas
Militia Hamoud Saeed al Mikhlafi
National Resistance Forces
Popular Resistance
Security Belt Forces
Shabwani Elite Forces
Southern Resistance
Supreme Political Council
Supreme Revolutionary Committee
Tihama Resistance
cleavage
1
2
3
4
powproj
100
200
300
400
500
battles
500
1000
(a) Main component of conflict network
Al Humayqani Tribal Militia
Al Islah Party
Giants Brigade
Militia Abu al Abbas
Militia Hamoud Saeed al Mikhlafi
National Resistance Forces
Popular Resistance
Security Belt Forces
Shabwani Elite Forces
Southern Resistance
Supreme Political Council
Supreme Revolutionary Committee
Tihama Resistance
cleavage
1
2
3
4
powproj
100
200
300
400
500
battles
500
1000
(b) Observed conflict network excluding isolates
Figure 2: Observed conflict network and main component for non-state armed groups in Yemen: node colour indicates identity
clevages, node size indicates power projection, names are included for key actors and thickness of ties indicates intensity of conflict
within dyads. Stress ma jorization layout for (a) and Fruchterman-Reingold layout for (b).
15

Mixing matrices help identify the presence of clustering in the observed network in Fig-
ure 3; these matrices provide the number of connected dyads for possible combinations of
identity cleavages to investigate the presence of patterns of homophily (assortative mixing)
or heterophily (disassortative mixing) (Goodreau, Handcock, et al., 2008; Newman, 2003).
Mixing matrix (Conflict)
1 2 3 4
1
2
3
4
22 27 19 3
27 20 34 1
19 34 16 7
3171
Mixing matrix (Random)
1 2 3 4
1
2
3
4
23 40 19 6
40 17 25 8
19 25 4 4
6 8 40
Figure 3: Mixing patterns for shared identity in observed and random networks.
Disassortative mixing is present in the observed network between groups active across
both one, two and three cleavages. Single cleavage groups interact with two clevage groups
(27 dyads), while two cleavage groups interact with three cleavage groups (34 dyads) and vice
versa. As the number of salient cleavages per actor increase, groups operating within more
cleavages engage in confrontations with moderately similar groups more than extremely sim-
ilar groups. Most conflict dyads are one degree apart in identity, usually made up of groups
on either side of the pro-government cleavage. Local tribal and communal rivalries repeat-
edly erupt into fighting in the network, and these are frequently exploited by major players
through alliances, exacerbating violence between similar groups. This disassortativity is
greatly amplified in the simulated network, which predicts most ties between groups with
one and two cleavages. One reason for this inflated heterophily prediction is the preponder-
ance of single-focus groups in the network: 144 of 189 nodes are only active across one or
two cleavages. Note that mixing matrices do not provide information about shared cleavages
within dyads, only the active cleavages for each actor. It provides context for the clash
of cross-cutting identities in Yemen, but to appropriately investigate the effect on violence
between groups, we must turn to inferential network models.
Exponential Random Graph Models
The ERGM is the most commonly used family of statistical models for inference about
patterns of tie formation in social networks given the existence of other endogenous network
structures (Desmarais and Cranmer, 2016; Snijders et al., 2006). Different assumptions of
dependence form the theoretical foundation of each different class of ERGMs, driven by
informed claims about ‘the [processes that influence the] type, extent and accumulation of
patterning that builds the network’ (Robins, 2014, p. 485). Independence of observations
16

cannot be presumed in civil war, it is likely that the presence of some conflict ties will
encourage, sustain or discourage ties between other dyads.
We provide a brief intuition about ERGMs; an abundant body of literature with a long
tradition exists on statistical models for detailed reference (Robins et al., 2007; Holland and
Leinhardt, 1981; Wasserman and Pattison, 1996). Given an observed network Xcomprised
of a set of nnodes and ties between them Xij , the model estimates the probability of a
tie existing Xij = 1 accounting for endogenous dependencies by considering the observed
network as the most likely single draw from an underlying multivariate probability distribu-
tion of possible networks, thus removing the requirement to assume independence (Cranmer
and Desmarais, 2011, p. 69). Estimation is carried out using Markov chain Monte Carlo
maximum likelihood estimation (MCMC-MLE) (Snijders, 2002).
We employ complex ERGMs to investigate probability of tie formation between groups
in the conflict network. Models were estimated using forward selection, step-wise variable
inclusion was guided by MCMC sample statistics and GOF diagnostics. Starting with a
simple ERGM capturing network density, we added relevant covariates to assess main effects
before including terms for second- and third-order effects. Geometric terms were added to
control for endogenous dependencies, resulting in dyad-dependent models. The final models
reproduced in Table 1 optimally test proposed hypotheses.
Prior to interpretation, we investigate model degeneracy and goodness-of-fit by examining
MCMC sample statistics and GOF diagnostics. The full specification in Model 5 models the
network reasonably well, and adjusting the model to improve prediction of medium-degree
nodes would likely lead to overfitting given the low density (Cranmer, Leifeld, et al., 2017).
Additional criteria used were likelihood-based measures: attributes and decay parameters
minimizing AIC, BIC, residual deviance and log-likelihood were retained during the fitting
process (Goodreau, Handcock, et al., 2008).
17

Results
In Table 1, Models 1 and 2 present the main effects of informal cooperation and shared
identity between armed groups. Model 3 is the fully specified model including controls; all
specifications are dyad-dependent and control for network density, isolates, degree distribu-
tion and transitivity induced by triadic closure. For easier interpretation, we compare the
transformed odds ratios and 95% confidence intervals for the restricted and general model
in Figure 4 (see also transformed Model 3 in Table 2).
The probability of forming a tie is conditional on the rest of the network: estimated
coefficients are log-odds of establishing a conflict tie between armed groups ceteris paribus.
Edges is similar to the intercept in dyad-independent logistic regression models, indicating
that the homogeneous baseline likelihood of forming a conflict tie is lower than not forming
a tie, if controlling for group characteristics, other relationships and network dependencies.
We first evaluate the primary hypothesis H1 that shared identity influences conflict ties:
coefficients for the dyadic covariate indicate that odds are higher for armed groups which
share more salient cleavages. For groups sharing one more salient cleavage, the odds of
forming a tie increase 1.5 times more than other groups. Highly similar groups are more
likely to fight, but the effect sizes and significance vary across specifications while this effect
is consistent.
In the restricted model, groups are similarly likely to fight moderately similar others
with great significance, and some uncertainty surrounds this estimate (1.54; 95% CI : 1.19 −
1.99). In the general model, this high significance and certainty in the effect size is retained
(95% CI : 1.16 −1.94). The inclusion of informal cooperation absorbs some of the direct
effect of shared identity while increasing model fitness and improving log-likelihood.
Informal cooperation has significant positive coefficients across all specifications; odds of
forming a conflict tie are 2.20 times higher if armed groups have cooperated at least once
before. Although this correlation persists even when controlling for shared identity, the
estimated coefficients have high uncertainty and weak significance (95% CI : 1.11 −4.37),
weakly supporting H2. Recall that relational observations are pooled across nearly six years
to construct the dependent and independent networks without accounting for the temporal
dimension of tie formation or dissolution. Informal cooperation may occur in years following
conflict incidents or pre-date conflict, and the effect of time-varying mechanisms are pertinent
to draw accurate inferences about the effect of temporary alliances on dyadic conflict.
18

Model 2
Model 3
0 1 2 3 4 0 1 2 3 4
Coalition group
Foreign sponsor
Power projection
Geographic spread
Years active
Shared identity
Territorial control
Cooperation
Odds ratio
Model terms
Figure 4: Coefficients plot with odds ratios and 95% confidence intervals for relevant predictors. Restricted model 2; general model
3. Significant terms (blue); insignificant (red).
19

Figure 5 illustrates that few groups share more than three cleavages, leading to greater
uncertainty surrounding the predicted probability of conflict (in this case, going to battle).
Sharing two identity cleavages seems to increase the probability by 2%, conditional on the
rest of the network, while <1.5% for three cleavages drops to <0.5% for four cleavages.
This finding supports the hypothesis that moderately similar groups have greater incentives
to engage in violent competition than mostly dissimilar or similar groups. It further sug-
gests that examining the salience and commonality of cleavages between actors is equally as
important as understanding which cleavages provoke entrenched episodes of violence in this
context.
0
2500
5000
7500
10000
0 1 2 3 4
Shared identity
Frequency
Frequency of shared identity between armed actors
0.00
0.01
0.02
0.03
0 1 2 3 4
Shared identity
Predicted probability
Predicted probabilities for shared identity
Figure 5: Frequency of salient cleavages shared between armed groups and mean predicted probability of conflict.
20

We further explore the average marginal effect of shared identity on the probability of
conflict given tactical cooperation between dyads, compared to coalition membership in
Figure 6. The probability of conflict substantially increases for cooperative dyads with more
cleavages in common, from 19% to 31%, while only hovering between 0.4% to 2.1% for
non-cooperative dyads. Greater uncertainty surrounds this effect for higher cleavages as the
network contains fewer battles between extremely similar actors.
It is empirically helpful to compare these results with the differential effect for coalition
membership. This alliance implies long-term strategic, operational, or militaristic collabo-
ration between armed groups positioned along the master cleavage and supported by foreign
powers, Saudi Arabia and UAE, in this context. We would expect that shared identity would
not drive fighting between formally allied groups mobilised to participate in the civil war
directly. As we would expect in dyads where one group is part of a coalition, the probability
increases up to 17% with shared identity with considerable certainty. This effect would seem
to be largely driven by the fighting between the Saudi-led coalition members and the Houthi
and tribal militias. In contrast, dyads of two coalition members experience a decrease of 6.4%
in the probability of fighting if they are moderately similar instead of dissimilar groups. For-
eign influence has comparatively restrained infighting within the Saudi-led coalition and
between Saudi and UAE camps. For non-coalition dyads, shared cleavages do not seem to
matter for the probability of conflict.
21

0.00
0.25
0.50
0.75
1.00
0 1 2 3 4
Shared Identity
Probability of Conflict Tie
Cooperation no yes
Marginal Effect of Identity by Tactical Cooperation
0.00
0.25
0.50
0.75
1.00
0 1 2 3 4
Shared Identity
Membership none one both
Marginal Effect of Identity by Coalition Membership
Figure 6: Average marginal effects of tactical cooperation and coalition membership on probability of conflict in Model 3.
22

Controls The main effect of power projection has a significant positive coefficient, indicat-
ing that a unit increase in fighting capacity increases the odds of a conflict tie (1.0; 95% CI :
1.00 −1.01). We also evaluated the relative power hypothesis in the fully specified model
as an additional robustness test (see Appendix). Still, we find no significant effect of power
asymmetry on conflict and comparative reduction in goodness-of-fit.
Armed groups receiving foreign sponsorship are not significantly more likely to form
a conflict tie than non-sponsored groups across all specifications. The significant negative
coefficient for coalition groups suggests a decrease in the likelihood of tie formation compared
to independent groups. Few actors in the network are coalition groups, and substantial
variation in effect size indicates that this effect is heavily influenced by the inclusion of
covariates of interest, particularly power projection.
Whether groups in control of territory are significantly more likely to form a conflict
tie is consistently insignificant. In particular, the inclusion of shared identity and power
projection absorbs territorial control‘s effect size and significance. As expected, years active
and geographic range increases the likelihood of conflict. These effects are highly statistically
significant and have particularly relevant effect sizes in the full model: an additional year
of activity increases the odds by 1.32, and established military presence in one additional
administrative district increases odds by 1.12 for armed groups.
Endogenous effects While we are only controlling for the endogenous effects of pop-
ularity and third-party transitive effects, a brief global interpretation is provided to aid
understanding of the observed network structure.
GWDegree tests for network centralisation by measuring the anti-preferential attachment
tendency of ties: inverse of ties attaching to a few popular nodes. A positive coefficient
indicates that the observed network is more decentralised than a random network. The
likelihood of a tie attaching to a lower-degree node is greater than higher-degree nodes; in
other words, lower-degree nodes are repelled from connecting to popular nodes.
GWESP is significant across all models. Note that log-odds of a tie increase mono-
tonically with the number of common enemies (Goodreau, Handcock, et al., 2008; Hunter,
2007). This seemingly supports the conventional wisdom that ‘the enemy of my enemy is
my friend’ since significance in the general model suggests that transitivity effects persist
despite controlling for first and second-order effects of attributes and cooperative relations.
23

Table 1: ERGM results with log odds, standard errors and p-values predicting conflict ties
Model 1 Model 2 Model 3
Edges −7.787 (0.605)∗∗∗
−8.375 (0.631)∗∗∗
−8.364 (0.617)∗∗∗
Covariates
Cooperation network 0.916 (0.355)∗∗ 0.790 (0.349)∗
Shared identity 0.430 (0.131)∗∗ 0.404 (0.131)∗∗
Controls
Foreign sponsorship −0.099 (0.227) −0.122 (0.228) −0.111 (0.230)
Coalition group −1.537 (0.365)∗∗∗
−1.302 (0.350)∗∗∗
−1.449 (0.364)∗∗∗
Territorial control 0.352 (0.272) 0.527 (0.273) 0.492 (0.271)
Years active 0.294 (0.052)∗∗∗ 0.280 (0.052)∗∗∗ 0.278 (0.051)∗∗∗
Geographic spread 0.104 (0.018)∗∗∗ 0.109 (0.018)∗∗∗ 0.109 (0.017)∗∗∗
Power projection 0.005 (0.001)∗∗∗ 0.004 (0.001)∗∗∗ 0.005 (0.001)∗∗∗
Endogenous dependencies
Isolates −11.480 (0.245)∗∗∗
−11.021 (0.251)∗∗∗
−10.768 (0.245)∗∗∗
GW Degree (0.2) −9.738 (0.118)∗∗∗
−9.188 (0.118)∗∗∗
−8.920 (0.117)∗∗∗
GWESP (0) 0.789 (0.225)∗∗∗ 0.659 (0.230)∗∗ 0.645 (0.226)∗∗
AIC 1186.449 1182.199 1179.579
BIC 1272.084 1267.835 1273.000
Log Likelihood −582.224 −580.100 −577.790
∗∗∗p < 0.001; ∗∗ p < 0.01; ∗p < 0.05
Table 2: Transformed ERGM results for Model 3 with odds ratios and 95% confidence intervals
Odds ratio 95% CI
Edges 0.0002 [0.0001; 0.0008]∗
Covariates
Cooperation network 2.2043 [1.1129; 4.3662]∗
Shared identity 1.4983 [1.1581; 1.9386]∗
Controls
Foreign sponsorship 0.8950 [0.5698; 1.4057]
Coalition group 0.2349 [0.1151; 0.4793]∗
Territorial control 1.6359 [0.9622; 2.7812]
Years active 1.3203 [1.1946; 1.4593]∗
Geographic spread 1.1155 [1.0779; 1.1544]∗
Power projection 1.0046 [1.0031; 1.0061]∗
Endogenous dependencies
Isolates 0.0000 [0.0000; 0.0001]∗
GW Degree (0.2) 0.0001 [0.0001; 0.0002]∗
GWESP (0) 1.9069 [1.2246; 2.9694]∗
AIC 1179.579
BIC 1273.000
Log Likelihood −577.790
∗1 outside the confidence interval.
24

Discussion and Conclusions
This paper sought to explain intergroup conflict as motivated by shared identity cleavages
and tactical cooperation between armed groups while accounting for relational dependencies
arising from interactions in the wider environment. Despite power through group strength
being advanced as the dominant explanation for fratricide, we find that the salience of
identity cleavages to actors significantly influences whether and whom groups fight. Results
from dyad-dependent models accurately capturing endogenous dependencies raise doubts
about the certainty and validity of one-dimensional explanations of interactions between
groups. This paper finds evidence to hold that the most common form of cooperation
between groups is unlikely to deter conflict. The temporary, transient nature of tactical
cooperation implies shared geographic location. It presents opportunities to unveil private
information about the military strength of actors who join forces, exposing vulnerabilities
ripe for future exploitation.
The influence of power on conflict may be partially attributed to cleavages shared between
groups; in fact, battles are more likely to occur between groups whose identities align more
closely than military strengths, providing strong evidence favouring group identity based on
salient cleavages. Shared identity may particularly drive behaviours of less powerful groups
relying on communal loyalties to recruit fighters and access resources while distinguishing
themselves from each other. In contrast, consolidating political opportunity and maintaining
territorial control may motivate more powerful groups to fight. Identity and power likely
operate in conjunction for certain armed groups, based on their organisational needs and local
competition. Uncertainty around causal mechanisms and the reinforcing effect of cooperation
on identity requires greater clarity in scholarship.
Methodologically, we innovate by using inferential network analysis to examine armed
group interactions and EGRMs to model these networks. This method provides novel av-
enues for non-governmental and humanitarian policy specialists to explore the factors driving
hostilities, focus on removing central brokers and highly connected actors, and identify en-
dogenous dependencies that may be exploited to terminate the conflict early, broker durable
peaceful settlements and reduce the rate of fatalities.
Empirically, we analyse the oft-ignored case of the Yemeni civil war and examine a wider
set of armed groups actively involved to varying degrees in the national conflict then ad-
dressed in previous studies. In doing so, we contribute to developing layered explanations
of micro and meso-level violence that accounts for the trickle-down effect of fighting at the
macro-level along the master cleavage. We lay the groundwork for consideration of different
types of interactions (e.g., tactical or strategic alliances, coalition membership, tribal or com-
munal rivalries) when empirically testing explanations for violence between armed groups.
Our findings also contribute to addressing the broader question in the literature that whom
actors fight with helps explain who they fight.
Limitations and Future Work
We suggest several avenues for future research. Theoretically, while the highly fragmented
nature of the Yemeni civil war can be traced to its dense tribal/ethnic, social fabric and
troubled history, internal conflicts are increasingly divided along ideological lines, indicating
25

the need for further investigation of the differential impact of ideology on cooperation in
intergroup violence in this and similar cases.
Empirical studies would benefit from gathering fine-grained data on actor attributes and
behaviour through fieldwork or access to reliable sources. Our reliance on event data and
qualitative sources is not without its limitations, particularly concerns of bias and reliability
(Eck, 2012). Researchers could undertake text-mining of news reports and social media
channels to gather data while cross-checking the validity of this information. Survey data and
text-mining are frequently used together for network data collection in parallel disciplines.
Prospective network studies should account for temporal dependencies between conflict
years. Statistical formulations of dyadic interactions are likely distinctive for the formation
and dissolution of ties. At the same time, homophily based on identity, power or territorial
control could strongly influence conflict ties between groups; the inclusion of these covariates
may be misguided in models evaluating the dissolution of conflict ties. Actor or event-based
modelling may help explain drivers of tie formation over time. Each approach makes certain
(sometimes strict/unrealistic) assumptions about the data-generating process, which must
be carefully considered before selection.
26

Appendix
Attribute data and independent covariates
We provide descriptive statistics for attribute data gathered on all 189 armed groups included
in the sample.
Table 3: Summary Statistics for Armed Group Attribute Data
Statistic N Mean Median St. Dev. Min Pctl(25) Pctl(75) Max
coalition 189 0.026 0 0.161 0 0 0 1
sponsorship 189 0.169 0 0.376 0 0 0 1
territory 189 0.153 0 0.361 0 0 0 1
geographic range 189 1.741 1 2.480 1 1 1 17
years active 189 1.725 1 1.406 1 1 2 6
power projection 189 17.315 1 63.830 1 1 4 581
identity 189 1.899 2 0.872 1 1 2 4
ethnic 189 0.042 0 0.202 0 0 0 1
tribal/communal 189 0.767 1 0.424 0 1 1 1
religious 189 0.460 0 0.500 0 0 1 1
ideological 189 0.466 0 0.500 0 0 1 1
pro-government 189 0.164 0 0.371 0 0 0 1
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
tribalco
ideologic
yrsactive
sponsor
territory
ethnic
geosprd
coalition
powrproj
progov
identity
religious
tribalco
ideologic
yrsactive
sponsor
territory
ethnic
geosprd
coalition
powrproj
progov
identity
religious
Figure 7: Correlations between independent covariates and controls.
27

Further exploration of the conflict network
We examine the underlying endogenous structure through comparison of degree distribution,
edge-wise and dyad-wise shared partners of the observed and random simulated networks,
reproduced in Figure 8. The observed network consists of many low-degree and few high-
degree nodes. Shared partner distributions differ significantly as our network has more nodes
with multiple edge-wise and dyad-wise shared partners while the random network mostly
consists of nodes with one shared partner and isolates.
Degrees (Conflict)
Frequency
0 2 4 6 8 10
0 50 100 150 200
Degrees (Random)
Frequency
0 2 4 6 8 10
0 50 100 150 200
esp1 esp2 esp3 esp4 esp5
Edge-wise Shared Partners (Conflict)
Frequency
0 5 10 15 20 25
esp1 esp2 esp3 esp4 esp5
Edge-wise Shared Partners (Random)
Frequency
0 5 10 15 20 25
dsp1 dsp2 dsp3 dsp4 dsp5
Dyad-wise Shared Partners (Conflict)
Frequency
0 500 1000 1500
dsp1 dsp2 dsp3 dsp4 dsp5
Dyad-wise Shared Partners (Random)
0 500 1000 1500
Figure 8: Comparison of random and observed network structures.
28

Assessing convergence and model fit
Model convergence Degeneracy can occur if model specification does not represent the
observed network well, or if unstable endogenous terms are used, and produces simulated
networks which are nearly full or empty. In the worst case, non-convergence leads to complete
failure of parameter estimation (Hunter, Goodreau, and Handcock, 2008; Snijders et al.,
2006). To improve model convergence, we increased MCMC burn-in to 10000 and sample
size to 20000 for all specifications.
Figure 9 includes MCMC diagnostic graphics for fully specified Model 3. Trace plots
from last round of estimation show the Markov chain is mixing well as sample statistics vary
stochastically around the observed mean at each step with limited serial correlation between
them. Deviations of sample statistics from observed values are normally distributed around
zero for most statistics, except categorical covariates with few observations: the wave pattern
is partly an artifact of these statistics and infrequently-occurring value categories are retained
for hypotheses testing purposes.
Goodness-of-Fit We assess goodness-of-fit prior to analysing substantive results as poorly
fitting models generate results of little interpretative or practical value in understanding the
data-generating process. (Cranmer, Leifeld, et al., 2017, p.245) point out that neglecting to
capture endogenous dependencies may result in omitted variable bias and poor parameter
estimates.
The gof() command simulates networks using fixed model parameters to compute mea-
sures for structural network features, i.e. degree distribution, edge-wise and dyad-wise shared
partners and geodesic distances (Hunter, Goodreau, and Handcock, 2008; Hunter, Goodreau,
and Handcock, 2013). The R() output includes statistics for observed and simulated endoge-
nous features to facilitate comparison, and provides estimated p-values to gauge differences
between both statistics; note that the auxiliary statistic does not significantly differ between
the observed network and simulations.
In Figure 10, key structural features of the observed network are plotted against simulated
networks for the fully specified model. The black line represents values for the observed
network while the boxplots indicate distributions of structural features and grey lines indicate
confidence intervals for 1000 simulated networks. The line generally crosses through median
of boxplots for most counts of endogenous statistics, fitting dyad-wise shared partners and
minimum geodesic distances particularly well; the model under-predicts the proportion of
nodes with medium degrees and edge-wise shared partners, underestimating dyads with
shortest paths of length 3.
Additionally, we evaluate goodness-of-fit through the comparison of simulated and ob-
served networks to ensure that endogenous structures have been appropriately modelled.
Figure 11 illustrates that the full model closely reproduces network structure improving
confidence in our specification. Figure 12 and Figure 13 illustrate the autocorrelation and
crosscorrelation between model terms in the final specification. Geweke diagnostics are pre-
sented in Figure 14. This is a convergence diagnostic for Markov chains which produces
z-scores for a test of equality of means between first and last parts of the chain. If the
Markov chain is in the stationary distribution, the means are equal and Geweke’s statistic
has an asymptotically standard normal distribution.
29

5000000 10000000 15000000 20000000
-20 0 20 40
Iterations
Trace of edges Density of edges
0.00 0.01 0.02 0.03 0.04
-34 -25 -15 -5 0 5 10 15 20 25 30 35 40
5000000 10000000 15000000 20000000
-40 -20 0 20 40
Iterations
Trace of nodefactor.sponsor.1 Density of nodefactor.sponsor.1
0.000 0.010 0.020 0.030
-48 -40 -30 -20 -10 0 5 15 25 35 45
5000000 10000000 15000000 20000000
-30 -20 -10 0 10 20
Iterations
Trace of nodefactor.coalition.1 Density of nodefactor.coalition.1
0.00 0.01 0.02 0.03 0.04 0.05
-31 -25 -20 -15 -10 -5 0 5 10 15 20 25
5000000 10000000 15000000 20000000
-40 -20 0 20 40
Iterations
Trace of nodefactor.territory.1 Density of nodefactor.territory.1
0.000 0.010 0.020
-54 -40 -30 -20 -10 0 10 20 30 40 50
5000000 10000000 15000000 20000000
-200 0 100 200 300
Iterations
Trace of nodecov.years Density of nodecov.years
0.000 0.001 0.002 0.003 0.004
-270 -200 -100 0 50 100 200 300
5000000 10000000 15000000 20000000
-600 -200 0 200 400 600
Iterations
Trace of nodecov.geopres Density of nodecov.geopres
0.0000 0.0010 0.0020
-614 -400 -200 0 100 300 500 633
Figure 9: MCMC diagnostic graphics for Model 3.
30

5000000 10000000 15000000 20000000
-15000 -5000 0 5000 15000
Iterations
Trace of nodecov.powproj
-15000 -10000 -5000 0 5000 10000 15000
0.00000 0.00004 0.00008
Density of nodecov.powproj
N = 15009 Bandwidth = 572.6
5000000 10000000 15000000 20000000
-10 -5 0 5 10
Iterations
Trace of edgecov.alliancenet Density of edgecov.alliancenet
0.00 0.04 0.08 0.12
-12 -10 -8 -6 -4 -2 0 2 4 6 8 10 12
5000000 10000000 15000000 20000000
-40 -20 0 20 40 60
Iterations
Trace of edgecov.simidmat Density of edgecov.simidmat
0.000 0.010 0.020
-51 -40 -30 -20 -10 0 10 20 30 40 50 60
5000000 10000000 15000000 20000000
-20 -10 0 10 20
Iterations
Trace of isolates Density of isolates
0.00 0.02 0.04 0.06
-21 -16 -12 -8 -4 0 2 4 6 8 12 16 20
5000000 10000000 15000000 20000000
-20 -10 0 10 20
Iterations
Trace of gwdeg.fixed.0.2
-20 -10 0 10 20
0.00 0.02 0.04 0.06
Density of gwdeg.fixed.0.2
N = 15009 Bandwidth = 0.89
5000000 10000000 15000000 20000000
-30 -10 0 10 20 30 40
Iterations
Trace of gwesp.fixed.0 Density of gwesp.fixed.0
0.00 0.01 0.02 0.03 0.04
-30 -20 -10 -5 0 5 10 15 20 25 30 35 40
31

Figure 10: Comparison of simulated (red) and observed (blue) networks.
0 5 11 18 25 32 39 46 53 60
0.0 0.2 0.4 0.6
Degree
Frequency
0 2 4 6 8 10 12 14 16
0.0 0.2 0.4 0.6 0.8
Edge-wise shared partners
Frequency
0 2 4 6 8 10 12 14 16
0.0 0.2 0.4 0.6 0.8 1.0
Dyad-wise shared partners
Frequency
123456789 11
0.0 0.2 0.4 0.6 0.8
Geodesic distances
Frequency
Figure 11: Goodness-of-fit assessment for full specification.
32

0 5000 10000 15000 20000 25000 30000
-1.0 -0.5 0.0 0.5 1.0
Lag
Autocorrelation
edges:chain1
0 5000 10000 15000 20000 25000 30000
-1.0 -0.5 0.0 0.5 1.0
Lag
Autocorrelation
nodefactor.sponsor.1:chain1
0 5000 10000 15000 20000 25000 30000
-1.0 -0.5 0.0 0.5 1.0
Lag
Autocorrelation
nodefactor.coalition.1:chain1
0 5000 10000 15000 20000 25000 30000
-1.0 -0.5 0.0 0.5 1.0
Lag
Autocorrelation
nodefactor.territory.1:chain1
0 5000 10000 15000 20000 25000 30000
-1.0 -0.5 0.0 0.5 1.0
Lag
Autocorrelation
nodecov.years:chain1
0 5000 10000 15000 20000 25000 30000
-1.0 -0.5 0.0 0.5 1.0
Lag
Autocorrelation
nodecov.geopres:chain1
0 5000 10000 15000 20000 25000 30000
-1.0 -0.5 0.0 0.5 1.0
Lag
Autocorrelation
nodecov.powproj:chain1
0 5000 10000 15000 20000 25000 30000
-1.0 -0.5 0.0 0.5 1.0
Lag
Autocorrelation
edgecov.alliancenet:chain1
0 5000 10000 15000 20000 25000 30000
-1.0 -0.5 0.0 0.5 1.0
Lag
Autocorrelation
edgecov.simidmat:chain1
0 5000 10000 15000 20000 25000 30000
-1.0 -0.5 0.0 0.5 1.0
Lag
Autocorrelation
isolates:chain1
0 5000 10000 15000 20000 25000 30000
-1.0 -0.5 0.0 0.5 1.0
Lag
Autocorrelation
gwdeg.fixed.0.2:chain1
0 5000 10000 15000 20000 25000 30000
-1.0 -0.5 0.0 0.5 1.0
Lag
Autocorrelation
gwesp.fixed.0:chain1
Figure 12: Autocorrelations for Model 3.
33

edges ndfctr.c.1 ndcv.yr ndcv.pw edgcv.s gw..0.2
gwsp..0 isolats edgcv.l ndcv.gp ndfctr.t.1 ndfctr.s.1
1
-1
0
Figure 13: Crosscorrelations for Model 3.
34

6000000 8000000 10000000 12000000
-2 -1 0 1 2
First iteration in segment
Z-score
edges
6000000 8000000 10000000 12000000
-2 -1 0 1 2
First iteration in segment
Z-score
nodefactor.sponsor.1
6000000 8000000 10000000 12000000
-2 -1 0 1 2
First iteration in segment
Z-score
nodefactor.coalition.1
6000000 8000000 10000000 12000000
-2 -1 0 1 2
First iteration in segment
Z-score
nodefactor.territory.1
6000000 8000000 10000000 12000000
-2 -1 0 1 2
First iteration in segment
Z-score
nodecov.years
6000000 8000000 10000000 12000000
-2 -1 0 1 2
First iteration in segment
Z-score
nodecov.geopres
6000000 8000000 10000000 12000000
-2 -1 0 1 2
First iteration in segment
Z-score
nodecov.powproj
6000000 8000000 10000000 12000000
-3 -2 -1 0 1 2 3
First iteration in segment
Z-score
edgecov.alliancenet
6000000 8000000 10000000 12000000
-2 -1 0 1 2
First iteration in segment
Z-score
edgecov.simidmat
6000000 8000000 10000000 12000000
-2 -1 0 1 2
First iteration in segment
Z-score
isolates
6000000 8000000 10000000 12000000
-2 -1 0 1 2
First iteration in segment
Z-score
gwdeg.fixed.0.2
6000000 8000000 10000000 12000000
-2 -1 0 1 2
First iteration in segment
Z-score
gwesp.fixed.0
Figure 14: Geweke statistics for Model 3.
35

Additional models for power
We also tested the proposition that relative power influences intergroup conflict. Power
asymmetry may have both first-order and second-order effects, e.g., stronger groups form
more conflict ties, and they are especially attractive opponents for equally matched or weaker
groups (Fjelde and Nilsson, 2012; Nyg˚ard and Weintraub, 2015; Krause, 2017; Gade, Hafez,
and Gabbay, 2019).
Power difference is conveniently equivalent to a pairwise matrix of differences in power
projection included as an edge covariate in the model. This hypothesis asserts that the
likelihood of forming a conflict tie is based on the relative difference in fighting capacity. It
is unexpected but not surprising that the negative coefficient is insignificant, indicating that
armed groups do not tend to form conflict ties with groups of similar strength, even when
controlling for shared identity.
Table 4: ERGM results with log odds, standard errors and p-values predicting conflict ties
Model 1 Model 2
Edges −7.830 (0.600)∗∗∗
−8.376 (0.644)∗∗∗
Covariates
Cooperation network 0.757 (0.351)∗
Shared identity 0.394 (0.135)∗∗
Power projection 0.007 (0.002)∗∗∗ 0.006 (0.002)∗∗∗
Power difference −0.002 (0.002) −0.001 (0.002)
Controls
Foreign sponsorship −0.184 (0.240) −0.130 (0.244)
Coalition group −1.439 (0.361)∗∗∗
−1.460 (0.379)∗∗∗
Territorial control 0.328 (0.281) 0.461 (0.291)
Years active 0.308 (0.052)∗∗∗ 0.282 (0.054)∗∗∗
Geographic spread 0.109 (0.018)∗∗∗ 0.112 (0.018)∗∗∗
Endogenous dependencies
Isolates −11.262 (0.242)∗∗∗
−10.649 (0.258)∗∗∗
GW Degree (0.2) −9.510 (0.120)∗∗∗
−8.793 (0.112)∗∗∗
GWESP (0) 0.738 (0.230)∗∗ 0.622 (0.234)∗∗
AIC 1191.448 1181.158
BIC 1277.084 1282.363
Log Likelihood −584.724 −577.579
∗∗∗p < 0.001; ∗∗ p < 0.01; ∗p < 0.05
36

Data and secondary sources
Table of Sources
Type Name of Source Details
Data
Sources
Armed Conflict Location & Event Data Project Event data (2015-2020)
Yemen Data Project Event data (2015-2020)
UCDP Non-State Conflict Dataset (v.21.1) Dyad-year (1989-2020)
EPR Ethnic Dimensions (EPR-ED) Group-level cleavages
(2021)
Secondary
Sources
Saba News Agency Yemen news agency
Yemen Press Agency Yemen news agency
Al Masdar Online Yemen news agency
Hodhod Yemen News Agency Yemen news agency
Al Arabiya Saudi news agency
Middle East Monitor UK media institution
BBC World News UK/Global news agency
Reuters International Global news agency
Sana’a Center for Strategic Studies Yemen think-tank
Council on Foreign Relations US think-tank
UN Security Council on Yemen Res., SG & GA reports
Main
Texts
Paul Dresch, A Modern History of Yemen Cambridge University,
2000
Paul Dresch, Tribes, Government, History in Yemen Clarendon Press, 1989
Shelagh Weir, A Tribal Order: Politics & Law in
Yemen
University of Texas, 2007
Ahmed Ibrahim Abushouk (eds.), Hadhrami Dias-
pora in Southeast Asia: Identity Maintenance Or
Assimilation?
Brill Publishing, 2009
Helen Lackner, Why Yemen Matters: A Society in
Transition
Saqi Publishing, 2014
Marieke Brandt, Tribes & Politics in Yemen Oxford University, 2017
Marie Christine Heinze, Yemen and the Search for
Stability
Bloomsbury Publishing,
2018
Charles Schmidt & Robert D. Burrowes, Historical
Dictionary of Yemen (3rd ed.)
Rowman & Littlefield,
2018
37

References
Abdelal, Rawi et al. (2009). “Identity as a Variable”. In: Measuring Identity: A Guide
for Social Scientists. Ed. by Alastair Iain Johnston et al. Cambridge: Cambridge
University Press, pp. 17–32. doi:10.1017/CBO9780511810909.002 (cit. on p. 11).
Abduljaber, Malek (2018). “The Determinants of Political Cleavages in Jordan, Tunisia,
and Yemen: An Analysis of Political Attitudes Structure in the Arab World”. In:
Digest of Middle East Studies 27.1, pp. 97–120. doi:10.1111/dome.12132 (cit. on
p. 6).
Akcinaroglu, Seden (2012). “Rebel Interdependencies and Civil War Outcomes:” in:
Journal of Conflict Resolution 56.5 (cit. on pp. 3,7–9).
Asal, Victor et al. (Aug. 1, 2019). “Carrots, Sticks, and Insurgent Targeting of Civil-
ians”. In: Journal of Conflict Resolution 63.7, pp. 1710–1735. doi:10.1177/0022
002718789748 (cit. on p. 6).
Bacon, Tricia (Apr. 4, 2018). Why Terrorist Groups Form International Alliances.
University of Pennsylvania Press. 352 pp. isbn: 978-0-8122-9502-3 (cit. on p. 4).
Bapat, Navin A. and Kanisha D. Bond (2012). “Alliances between Militant Groups”.
In: British Journal of Political Science 42.4, pp. 793–824 (cit. on p. 3).
Borgatti, Stephen P., Martin G. Everett, and Jeffrey C. Johnson (2018). Analyzing
Social Networks. Second edition. London: SAGE Publications (cit. on p. 12).
Buhaug, Halvard, Scott Gates, and P¨aivi Lujala (2009). “Geography, Rebel Capability,
and the Duration of Civil Conflict:” in: Journal of Conflict Resolution 53.4 (cit. on
p. 8).
Bultmann, Daniel (2018). “The Social Structure of Armed Groups. Reproduction and
Change during and after Conflict”. In: Small Wars & Insurgencies 29.53 (cit. on
pp. 5,6).
Burton, Lucy (2015). “Rebel Group Interactions in Multi-Party Civil Wars”. Ph.D.
UCL (University College London). url:https://discovery.ucl.ac.uk/id/epr
int/1471553/ (visited on 09/05/2020) (cit. on pp. 3,6).
Carboni, Andrea (2018). Yemen – The Battle of Sana’a.url:https://acleddata.c
om/2018/02/16/yemen-the-battle-of-sanaa/ (visited on 09/04/2020) (cit. on
p. 8).
Carboni, Andrea and Luca Nevola (2019). Inside Ibb: A Hotbed of Infighting in Houthi-
Controlled Yemen.url:https://acleddata.com/2019/10/03/inside-ibb-a-h
otbed-of-infighting-in-houthi-controlled-yemen/ (visited on 07/06/2020)
(cit. on p. 7).
Cederman, Lars-Erik and Yannick Pengl (2019). Global Conflict Trends and Their
Consequences. United Nations, p. 55. url:https://www.un.org/development/d
esa/dpad/wp-content/uploads/sites/45/publication/SDO%5C_BP%5C_Ceder
man%5C_Pengl.pdf (visited on 01/08/2020) (cit. on p. 2).
38

Cederman, Lars-Erik, Andreas Wimmer, and Brian Min (Jan. 2010). “Why Do Ethnic
Groups Rebel? New Data and Analysis”. In: World Politics 62.1, pp. 87–119. doi:
10.1017/S0043887109990219 (cit. on p. 5).
Christia, Fotini (2008). “Following the Money: Muslim versus Muslim in Bosnia’s Civil
War”. In: Comparative Politics 40.4, pp. 461–480 (cit. on pp. 3,6).
— (2012). Alliance Formation in Civil Wars. Cambridge: Cambridge University Press
(cit. on p. 3).
Conrad, Justin, Kevin T. Greene, et al. (July 10, 2021). “Competition from Within:
Ethnicity, Power, and Militant Group Rivalry”. In: Defence and Peace Economics
0, pp. 1–16. doi:10.1080/10242694.2021.1951595 (cit. on p. 3).
Conrad, Justin and William Spaniel (2021). “Outbidding as Deterrence: Endogenous
Demands in the Shadow of Group Competition”. In: Militant Competition: How
Terrorists and Insurgents Advertise with Violence and How They Can Be Stopped.
Cambridge: Cambridge University Press, pp. 125–161. doi:10.1017/9781108992
275.005 (cit. on p. 5).
Cranmer, Skyler J. and Bruce A. Desmarais (2011). “Inferential Network Analysis
with Exponential Random Graph Models”. In: Political Analysis 19.1, pp. 66–86
(cit. on pp. 13,17).
Cranmer, Skyler J., Philip Leifeld, et al. (2017). “Navigating the Range of Statistical
Tools for Inferential Network Analysis”. In: American Journal of Political Science
61.1, pp. 237–251 (cit. on pp. 12,13,17,29).
Crawford, James (2002). “The ILC’s Articles on Responsibility of States for Interna-
tionally Wrongful Acts: A Retrospect”. In: American Journal of International Law
96.4, pp. 874–890 (cit. on p. 9).
Cremaschi, Simone et al. (2020). “Do Non-State Armed Groups Influence Each Other
in Attack Timing and Frequency? Generating, Analyzing, and Comparing Empir-
ical Data and Simulation”. In: Computational Conflict Research. Ed. by Emanuel
Deutschmann et al. Computational Social Sciences. Cham: Springer International
Publishing, pp. 167–182 (cit. on p. 4).
Cunningham, Kathleen Gallagher, Kristin M. Bakke, and Lee J. M. Seymour (2012).
“Shirts Today, Skins Tomorrow: Dual Contests and the Effects of Fragmentation
in Self-Determination Disputes”. In: Journal of Conflict Resolution 56.1, pp. 67–93
(cit. on p. 5).
Curiel, Rafael Prieto, Olivier Walther, and Neave O’Clery (2020). “Uncovering the
Internal Structure of Boko Haram through Its Mobility Patterns”. In: Applied
Network Science 5.1, pp. 1–23 (cit. on p. 9).
De la Calle, Luis and Ignacio S´anchez-Cuenca (2015). “How Armed Groups Fight:
Territorial Control and Violent Tactics”. In: Studies in Conflict & Terrorism 38.10,
pp. 795–813 (cit. on p. 8).
39

Derpanopoulos, George (2018). “Essays on Networks, Dictatorships, and Political Vi-
olence”. UCLA. url:https: / /escholarship.org/uc/item/33c9t14j (visited
on 09/06/2020) (cit. on p. 2).
Desmarais, Bruce A. and Skyler J. Cranmer (2016). “Statistical Inference in Politi-
cal Networks Research”. In: The Oxford Handbook of Political Networks. Ed. by
Jennifer Nicoll Victor, Alexander H. Montgomery, and Mark Lubell. New York:
Oxford University Press, p. 982 (cit. on p. 16).
Dorff, Cassy, Max Gallop, and Shahryar Minhas (2019). “Networks of Violence: Pre-
dicting Conflict in Nigeria”. In: The Journal of Politics 82.2, pp. 476–493 (cit. on
pp. 4,8).
Eck, Kristine (2009). “From Armed Conflict to War: Ethnic Mobilization and Conflict
Intensification”. In: International Studies Quarterly 53.2, pp. 369–388 (cit. on p. 3).
— (2012). “In Data We Trust? A Comparison of UCDP GED and ACLED Conflict
Events Datasets:” in: Cooperation and Conflict 47.1, pp. 124–141 (cit. on p. 26).
Findley, Michael G. and Joseph K. Young (Nov. 1, 2012). “More Combatant Groups,
More Terror?: Empirical Tests of an Outbidding Logic”. In: Terrorism and Political
Violence 24, pp. 706–721. doi:10.1080/09546553.2011.639415 (cit. on p. 5).
Fjelde, Hanne and Desir´ee Nilsson (2012). “Rebels against Rebels: Explaining Violence
between Rebel Groups”. In: Journal of Conflict Resolution 56.4, pp. 604–628 (cit.
on pp. 3,36).
Gade, Emily Kalah, Michael Gabbay, et al. (2019). “Networks of Cooperation: Rebel
Alliances in Fragmented Civil Wars:” in: Journal of Conflict Resolution 63.9,
pp. 2071–2097 (cit. on pp. 3,4).
Gade, Emily Kalah, Mohammed M. Hafez, and Michael Gabbay (2019). “Fratricide in
Rebel Movements: A Network Analysis of Syrian Militant Infighting:” in: Journal
of Peace Research 56.3, pp. 321–335 (cit. on pp. 3,36).
Ghanem, Abdulkareem (Feb. 28, 2019). Addressing Social Fragmentation in Yemen.
Sana’a Center For Strategic Studies. url:https://sanaacenter.org/publicat
ions/analysis/7137 (visited on 09/02/2021) (cit. on p. 6).
Gleditsch, Nils Petter et al. (2002). “Armed Conflict 1946-2001: A New Dataset”. In:
Journal of peace research 39.5, pp. 615–637 (cit. on pp. 9,12).
Goodreau, Steven M., Mark S. Handcock, et al. (2008). “A Statnet Tutorial”. In:
Journal of Statistical Software 24.9, pp. 1–27 (cit. on pp. 13,16,17,23).
Goodreau, Steven M., James A. Kitts, and Martina Morris (2009). “Birds of a Feather,
or Friend of a Friend? Using Exponential Random Graph Models to Investigate
Adolescent Social Networks*”. In: Demography 46.1, pp. 103–125 (cit. on p. 6).
Hafez, Mohammed M. (2017). “Fratricidal Rebels: Ideological Extremity and Warring
Factionalism in Civil Wars”. In: Terrorism and Political Violence.issn: 0954-6553
(cit. on pp. 3,6).
40

Hafner-Burton, Emilie M., Miles Kahler, and Alexander H. Montgomery (2009). “Net-
work Analysis for International Relations”. In: International Organization 63.3,
pp. 559–592 (cit. on pp. 4,12).
Handcock, Mark S. et al. (2019). Ergm: Fit, Simulate and Diagnose Exponential-Family
Models for Networks. The Statnet Project (https://statnet.org). url:https://C
RAN.R-project.org/package=ergm (cit. on p. 12).
Holland, Paul W. and Samuel Leinhardt (1981). “An Exponential Family of Probabil-
ity Distributions for Directed Graphs”. In: 76.373, pp. 33–50 (cit. on p. 17).
Hunter, David R. (2007). “Curved Exponential Family Models for Social Networks”.
In: Social Networks 29.2, pp. 216–230 (cit. on pp. 13,23).
Hunter, David R., Steven M. Goodreau, and Mark S. Handcock (2008). “Goodness of
Fit of Social Network Models”. In: Journal of the American Statistical Association
103.481, pp. 248–258 (cit. on p. 29).
— (2013). “Ergm.Userterms: A Template Package for Extending Statnet”. In: Journal
of Statistical Software 52.2 (cit. on p. 29).
Kalyvas, Stathis N. (2003). “The Ontology of “Political Violence”: Action and Identity
in Civil Wars”. In: Perspectives on Politics 1, pp. 475–494. doi:10.1017/S15375
92703000355 (cit. on pp. 5,7).
Kalyvas, Stathis N. and Laia Balcells (2010). “International System and Technologies
of Rebellion: How the End of the Cold War Shaped Internal Conflict”. In: American
Political Science Review 104, pp. 415–429. doi:10.1017/S0003055410000286 (cit.
on p. 3).
Kolaczyk, Eric D. and G´abor Cs´ardi (2014). Statistical Analysis of Network Data with
R. Use R! New York: Springer New York.
Krause, Peter (2017). Rebel Power: Why National Movements Compete, Fight, and
Win. Ithaca: Cornell University Press (cit. on p. 36).
Levy, Michael A. (2016). “Gwdegree: Improving Interpretation of Geometrically-Weighted
Degree Estimates in Exponential Random Graph Models”. In: Journal of Open
Source Software 1.3, p. 36 (cit. on p. 13).
Lilja, Jannie and Lisa Hultman (2011). “Intra-Ethnic Dominance and Control : Vio-
lence against Co-Ethnics in the Early Sri Lankan Civil War”. In: Security Studies
20.2, pp. 171–197 (cit. on p. 3).
Maoz, Zeev (2012). “Preferential Attachment, Homophily, and the Structure of Inter-
national Networks, 1816–2003:” in: Conflict Management and Peace Science 29.3,
pp. 341–369 (cit. on p. 6).
Maoz, Zeev et al. (2007). “What Is the Enemy of My Enemy? Causes and Consequences
of Imbalanced International Relations, 1816–2001”. In: Journal of Politics 69.1,
pp. 100–115 (cit. on p. 13).
Mendelsohn, Barak (2021). “The Battle for Algeria: Explaining Fratricide among
Armed Nonstate Actors”. In: Studies in Conflict & Terrorism 44.9, pp. 776–798.
doi:10.1080/1057610X.2019.1580419 (cit. on p. 4).
41

Newman, Mark E. J. (2003). “Mixing Patterns in Networks”. In: Physical Review E
67.2 (cit. on p. 16).
Nyg˚ard, H˚avard Mokleiv and Michael Weintraub (2015). “Bargaining Between Rebel
Groups and the Outside Option of Violence”. In: Terrorism and Political Violence
27.3, pp. 557–580 (cit. on pp. 3,36).
Oakes, Penelope (2002). “Psychological Groups and Political Psychology: A Response
to Huddy’s ”Critical Examination of Social Identity Theory””. In: Political Psy-
chology 23.4, pp. 809–824. issn: 0162-895X (cit. on p. 5).
Pearlman, Wendy and Kathleen Gallagher Cunningham (2011). “Nonstate Actors,
Fragmentation, and Conflict Processes:” in: Journal of Conflict Resolution 56.1,
pp. 3–15 (cit. on p. 7).
Phillips, Brian J (Jan. 1, 2015). “Enemies with Benefits? Violent Rivalry and Terrorist
Group Longevity”. In: Journal of Peace Research 52.1, pp. 62–75. doi:10.1177/0
022343314550538 (cit. on p. 4).
Pischedda, Costantino (2015). “Wars Within Wars: Understanding Inter-Rebel Fight-
ing”. Columbia University (cit. on p. 14).
— (Nov. 2020). Conflict Among Rebels: Why Insurgent Groups Fight Each Other.
Columbia University Press, 264 Pages. isbn: 978-0-231-55274-5 (cit. on pp. 3–5).
Popovic, Milos (2018). “Inter-Rebel Alliances in the Shadow of Foreign Sponsors”. In:
International Interactions 44.4, pp. 749–776 (cit. on p. 8).
Quinn, Jason, Madhav Joshi, and Erik Melander (2019). “One Dyadic Peace Leads to
Another? Conflict Systems, Terminations, and Net Reduction in Fighting Groups”.
In: International Studies Quarterly 63.4, pp. 863–875 (cit. on p. 3).
Raleigh, Clionadh et al. (2010). “Introducing ACLED: An Armed Conflict Location
and Event Dataset”. In: Journal of Peace Research 47.5, pp. 651–660 (cit. on p. 9).
Robins, Garry (2014). “Exponential Random Graph Models for Social Networks”. In:
The SAGE Handbook of Social Network Analysis. In collab. with John Scott and
Peter Carrington. London: SAGE Publications, pp. 484–500 (cit. on p. 16).
Robins, Garry et al. (2007). “An Introduction to Exponential Random Graph (P*)
Models for Social Networks”. In: Social Networks 29.2, pp. 173–191 (cit. on p. 17).
Roy, Emile and Andrea Carboni (2019). Yemen’s Fractured South: Shabwah and Hadra-
mawt.url:https://acleddata.com/2019/05/09/yemens-fractured-south-s
habwah-and-hadramawt/ (visited on 08/08/2020) (cit. on p. 7).
Rubin, Donald B., ed. (1987). Multiple Imputation for Nonresponse in Surveys. Wiley
Series in Probability and Statistics. Hoboken: John Wiley & Sons, Inc. (cit. on
p. 12).
San´ın, Francisco Guti´errez and Elisabeth Jean Wood (2014). “Ideology in Civil War:
Instrumental Adoption and Beyond”. In: Journal of Peace Research 51.2, pp. 213–
226 (cit. on p. 3).
42

Snijders, Tom A. B. (2002). “Markov Chain Monte Carlo Estimation of Exponential
Random Graph Models”. In: Journal of Social Structure 3.2, pp. 1–40 (cit. on
p. 17).
Snijders, Tom A. B. et al. (2006). “New Specifications for Exponential Random Graph
Models”. In: Sociological Methodology 36.1, pp. 99–153 (cit. on pp. 16,29).
Steinwand, Martin C. and Nils W. Metternich (2020). Who Joins and Who Fights?
Explaining Tacit Coalition Behavior among Civil War Actors. Unpublished. So-
cArXiv. url:https://osf.io/ksx3m (visited on 09/06/2020) (cit. on p. 7).
Taraktas, Basak (2019). Studying Alliance Formation Using Case Studies and Network
Analysis. SAGE Research Methods Cases. Los Angeles: SAGE Publications (cit. on
p. 6).
Tokdemir, Efe et al. (Apr. 1, 2021). “Rebel Rivalry and the Strategic Nature of Rebel
Group Ideology and Demands”. In: Journal of Conflict Resolution 65, pp. 729–758.
doi:10.1177/0022002720967411 (cit. on p. 5).
Turner, John C. et al. (1987). Rediscovering the Social Group: A Self-Categorization
Theory. Rediscovering the Social Group: A Self-Categorization Theory. Cambridge,
MA, US: Basil Blackwell, pp. x, 239. isbn: 978-0-631-14806-7 (cit. on p. 5).
Van Buuren, Stef (2018). Flexible Imputation of Missing Data. Second edition. Chap-
man and Hall/CRC Interdisciplinary Statistics Series. Boca Raton: CRC Press,
Taylor & Francis Group (cit. on p. 12).
Wasserman, Stanley and Philippa Pattison (1996). “Logit Models and Logistic Re-
gressions for Social Networks: I. An Introduction to Markov Graphs and p”. In:
Psychometrika 61.3, pp. 401–425 (cit. on p. 17).
Weidmann, Nils B. (Nov. 1, 2016). “Micro-Cleavages and Violence in Civil Wars:
A Computational Assessment”. In: Conflict Management and Peace Science 33,
pp. 539–558. doi:10.1177/0738894215570433 (cit. on pp. 4,7).
Weinstein, Jeremy M. (2006). Inside Rebellion: The Politics of Insurgent Violence.
Cambridge Studies in Comparative Politics. Cambridge: Cambridge University
Press. doi:10.1017/CBO9780511808654 (cit. on pp. 3,5).
43