60 268 1 PB

CORVINUS JOURNAL OF SOCIOLOGY AND SOCIAL POLICY Vol.
4 (2013) 1, 332
MODELING FREQUENCY AND TYPE OF

INTERACTION IN EVENT NETWORKS
Jrgen Lerner1
Margit Bussmann2
Tom A. B. Snijders3
Ulrik Brandes4
Abstract Longitudinal social networks are increasingly given by event data;
i.e., data coding the time and type of interaction between social actors. Examples
include networks stemming from computer-mediated communication, open
collaboration in wikis, phone call data and interaction among political actors. In
this paper, we propose a general model for networks of dyadic, typed events. We
decompose the probability of events into two components: the first modeling the
frequency of interaction and the second modeling the conditional event type, i. e.,
the quality of interaction, given that interaction takes place.
While our main contribution is methodological, for illustration we apply
our model to data about political cooperation and conficts collected with the
Kansas Event Data System. Special emphasis is given to the fact that some
explanatory variables affect the frequency of interaction while others rather
determine the level of cooperativeness vs. hostility, if interaction takes place.
Furthermore, we analyze if and how model components controlling for network
dependencies affect findings on the effects of more traditional predictors
such as geographic proximity or joint alliance membership. We argue that
modeling the conditional event type is a valuable and in some cases superior
alternative to previously proposed models for networks of typed events.
Keywords network analysis, statistical network models, event data, signed
networks, structural balance theory
1 Jrgen Lerner is postdoctoral researcher in Computer Science at University of Konstanz;
e-mail: lerner@inf.uni-konstanz.de
2 Margit Bussmann is professor of International Politics at University of Greifswald; e-mail:
margit.bussmann@uni-greifswald.de
3 Tom A. B. Snijders is professor of Statistics at University of Oxford; e-mail: tom.snijders@
nuffield.ox.ac.uk
4 Ulrik Brandes is professor of Computer Science at University of Konstanz; e-mail: Ulrik.
Brandes@uni-konstanz.de
CORVINUS JOURNAL OF SOCIOLOGY AND SOCIAL POLICY 1 (2013)
LERNERBUSSMANNSNIJDERSBRANDES
1 Introduction
More and more social network datasets encode interaction events (such as
sending an email or co-authoring a scientific article) rather than relational
states between actors (such as friendship or esteem). The increased availability
of event data is especially due to the advent of automated data collection
facilities. For instance, log-data of computer mediated communication (e.g.,
email, Usenet-groups, or social network services), open collaboration in
wikis, or phone-call data naturally gives rise to event networks. In this paper
we consider networks of political actors together with interaction events that
are routinely observed and reported in the news.
We consider networks of dyadic, typed events where the type is a real
number indicating the level of cooperativeness (if positive) or hostility (if
negative). From a modeling point of view, the general research questions
that we consider here are about the causes and effects of network interaction.
When analyzing the causes of network events, the network is seen as a
dependent variable and one seeks to answer questions like what makes actor
A interact more or less with actor B or what makes actor A engage in a
specific type of interaction towards actor B. When analyzing the effects
of events, the network is seen as an explanatory variable and one seeks to
answer what results from interaction (of a certain type). Here we consider
network events to be both the dependent and the explanatory variables;
more specifically, we want to find out how past events (and externally given
actor and dyad covariates) stochastically determine the frequency and type
of future events.
The occurrence of events of specific types can be modeled in at least two
distinct ways whose difference is crucial for this paper. For illustration, assume
that we want to model event networks with two types of interaction, positive
events encoding cooperation and negative events encoding hostilities, and
that we want to test hypotheses about the causes of both types of events. The
first way to do so is to adapt the model proposed by Butts (2008). In that
model there are two different rate functions for the two types of events and the
estimated parameters reveal which explanatory variables increase or decrease
the frequency of cooperative or hostile events, respectively. The second
possibility to model typed events is the one that we propose in this paper. In
our model there is one rate function modeling the frequency of events of any
type, and a type function modeling the conditional probability of cooperative
MODELING FREQUENCY AND TYPE OF INTERACTION IN EVENT NETWORKS
versus hostile interaction, given that interaction occurs.5 Thus, the estimated
parameters in our model reveal what triggers:
1. an increase/decrease in the frequency of interaction;
2. positive vs. negative interaction, given that interaction occurs.
There are at least two benefits resulting from this alternative model for typed
events. First, the conditional event type models are not restricted to a finite
number of event types (i. e., to categorical event types) but can also deal with
types characterized by continuous variables. Second, the results stemming
from the conditional event type models provide additional information about
the causes of events and may clarify seemingly counterintuitive findings that
result from modeling the frequency of typed events separately. For instance,
we demonstrate in the next section that a bivariate model for international
relations suggests that countries have an increased probability of engaging
in a militarized dispute with their alliance partners (compared to countries
with which they do not share an alliance membership). On the other hand, an
application of conditional event type models reveals that allies consistently
show a tendency to engage in cooperative rather than conflictive interaction
under the precondition that they do interact.
The remainder of this article is structured as follows. Section 2 introduces a
dataset on which we conduct an illustrative analysis, reports related previous
results in international relations research and develops the exemplary
hypotheses. Our newly-proposed model is described in Section 3 and results
of the illustrative application of the model are given in Section 4. Section 5
concludes and indicates future research.
2 Political Network Analysis

Scholars of international politics increasingly realize the advantages
of network analysis in various contexts (e. g., Maoz 2009; Hafner-Burton
and Montgomery 2006). One approach within social network analysis,
structural balance theory, is particularly well suited to addressing questions of
cooperations and conflict between states.6 A signed network (i. e., a network
5 Since we consider weighted events later in this paper, we model the conditional probability
density for event weights, rather than the conditional probability of positive/negative events;
the latter serves only for simplified illustration.
6 For a detailed description of structural balance theory see Heider (1946); Cartwright and
Harary (1956).
with positive and negative ties) is balanced if every semi-cycle has an even
number of negative ties.7 Structural balance theory (SBT) claims that actors
have a preference for balanced networks. Specifically, if two ties in a triplet
of actors are present and the third tie is to be created then its sign is predicted
by the following four rules resulting from SBT: the friend of a friend is a
friend, the friend of an enemy is an enemy, the enemy of a friend is an
enemy, and the enemy of an enemy is a friend.
The influence of common friends and enemies on a dyad in political
networks has been analyzed in Maoz et al. (2007) and Crescenzi (2007).
Maoz et al. (2007) compute the conditional probabilities of alliances and
militarized interstate disputes (MIDs) between two countries, given that these
satisfy the conditions of being (1) friends of enemies, (2) enemies of friends,
and (3) enemies of enemies (the relations friend and enemy are derived
from the alliance and MID relations, respectively). It turns out that all three
preconditions increase both the probability of alliances and the probability of
MIDs. Thus, the results simultaneously support and reject structural balance
theory. Seen from a different angle, actors that are indirectly related via a third
actor have a higher probability to interactboth positively and negatively. This
result can be refined by applying our newly-proposed network model: later
in this paper we show that actors that are (say) enemies of enemies have a
higher probability of interaction but, given that they do interact, their relation
has a tendency towards cooperationclearly supporting SBT. In related work,
Crescenzi (2007) defined a combined dyadic indicator that is positive if the
two actors evaluate most other actors consistently (both positive or both negative), negative if they evaluate most other actors inconsistently, and (close
to) zero if these effects cancel out. Crescenzi operationalized a test of SBT
by estimating the influence of this indicator on the time it takes until the next
MID in that dyad breaks out. Indeed, he found that dyads receiving a negative
score have shorter waiting times until the next conflict. This provides support
for the combined predictions of structural balance theory. In contrast to Maoz
et al. (2007) and Crescenzi (2007), we analyze the effect of indirect relations
on the conditional event type, rather than on the occurrence of ties. Thus, our
model estimates the sign of a tie (a, b) only if a does interact with b.8
7 A semi-cycle is a sequence of actors v1,v2,...,vk+1 = v1, k 3 where for all i =1,...,k there is a tie
from vi to vi+1 or vice versa.
8 Neither Maoz et al. (2007) nor Crescenzi (2007) use/s daily event data, but rather data coding
the yearly state of the world system on the country-level. Some researchers argue that yearly
data is too coarse-grained to capture quick responses to hostility as they occurred, e. g., in the
Israel/Palestine conflict (King and Lowe 2003, p. 617).
The analysis of event data (alternatively referred to as time-to-event

analysis, survival analysis, event history analysis, or lifetime analysis) is an
established research area; see Lawless (2003) for a general reference. Some
recent papers analyze network dependencies among events that happen in
dyads (e. g. Butts 2008; De Nooy 2008, 2011; Brandes et al. 2009; Stadtfeld
2010). Although event data analysis is common in political science (e. g.,
Box-Steffensmeier and Jones 1997), network dependencies are rarely
considered there. Exceptions include Goldstein et al. (2001) who applied
vectorautoregression to the dyadwise aggregated levels of cooperation/
conflict over short time-intervals and Hoff and Ward (2004) who estimate
dependencies in networks constructed from event data by aggregating over
the whole observation period. Our work differs from these references since we
do not aggregate events over time-intervals but rather model the probability
of each single event.
In this paper we propose a general model for networks of dyadic typed
events. With the increase in importance and availability of network event
data we hope that this model will be applied to a variety of data sets. As
an illustration, we apply it here to the publically-available data referred to
as Gulf data coded from full stories from the Kansas Event Data System
website (KEDS 2012).9 This data set consists of events related to the Persian
Gulf region for the period from April 15th, 1979 to March 31st, 1999. It
includes more than 304,000 events among 202 unique actors. For the analysis
done in this paper we exclude all non-state actors (such as ethnic groups or
international organizations) yielding 168 actors and more than 217,000 events
between them.
The KEDS (Schrodt et al. 1994) is a software tool that automatically
extracts daily events from news reports. Events encode who did when what
to whom and, thus, describe time-stamped, dyadic, typed interaction. Event
types are classified using the World Event/Interaction Survey (WEIS)
codes (McClelland 1976) and each event type is assigned a weight from the
interval [10, 10], where 10 stands for the most hostile and +10 for the most
cooperative type of interaction (Goldstein 1992). These codings are explained
in more detail in the following. Descriptive visualization, animation, and
clustering of this data set can be found, e. g., in Brandes et al. (2006) and
Brandes and Lerner (2008).
9 We have chosen the Gulf data set since mainly state actors are involved in this conflict; this
implies that a consistent set of established alternative explanatory variables (introduced later)
is available. Note that other data sets available from the KEDS website include many non-state
actors.
Roughly spoken, the KEDS software (Schrodt et al. 1994) extracts triples
of the form (subject, verb, object) from news wire reports. Each
triple encodes the information that the subject performs an action (specified
by the verb) towards the object. The subjects and objects are mapped to
actors defined by the analyst. The following excerpt illustrates the coding of
some of the actors included in the Gulf data:
AMERICA [USA]
CENTRAL INTELLIGENCE AGENCY [USA]
ISLAMIC_COUNTRIES [ARB]
ARAB_MONETARY_FUND [ARB]
GULF STATES [ARB]
Subjects and objects in news wire texts are interpreted as referring to

specific actors. For instance, the general term AMERICA as well as the more
specific CENTRAL INTELLIGENCE AGENCY are mapped to the same actor
labeled USA. As another example, the tokens ISLAMIC COUNTRIES, ARAB
MONETARY FUND, and GULF STATES (among others) are mapped to an actor
labeled ARB. Actor ARB is, thus, an example of a non-state actor which is
excluded from the network analyzed in this paper.
The following excerpt is an (incomplete) list of events that happened on
August 10th, 1990 in the Gulf region.
900810
900810
900810
900810
900810
900810
ARB
IRQ
IRQ
USA
USA
USA
IRQ
USA
ARB
IRQ
IRQ
IRQ
012
122
094
160
051
223
RETREAT
DENIGRATE
CALL FOR
WARN
PROMISE POLICY
MIL ENGAGEME
For instance, the last event (dated 900810; i. e., August 10th 1990) codes a
military action (WEIS event type 223), initiated by actor USA and directed to
the Iraq (IRQ). The text at the end of the line (MIL ENGAGEMENT) is a textual
description of the event type (which is not needed in the analysis, since it is
implied by the event type). In total there are more than 100 different types of
events.
The WEIS event types are mapped to an established scale whose entries
are referred to as Goldstein weights (Goldstein 1992) and indicate the level
of cooperativeness (if positive) or hostility (if negative). Examples of weights
associated with specific types are the following.
072
054
160
173
223
EXTEND MIL AID

ASSURE
WARN
SPECIF THREAT
MIL ENGAGEMENT
8.3
2.8
-3.0
-7.0
-10.0
Extending military aid is considered a highly cooperative action (weight

equal to 8.3), whereas warnings are mildly hostile (w = 3.0), specific threats
much more severe (w = 7.0), and military engagements are the most hostile
type of events (w = 10.0).
It must be kept in mind that using the KEDS data entails some problems
for the analysis. Since the data are generated on the basis of news reports
we do not, strictly speaking, estimate the tendency to interact but rather the
likelihood of interaction being reported in the news. We believe that, given
that the interpretation takes account of this bias, the results still are meaningful.
Interaction among political actors is not only influenced by previous
interaction but also by additional actor or dyad characteristics such as whether
they share a common border or are members in the same military alliance.
To control for different actor or dyad characteristics we rely on data from
a frequently-used model of international militarized disputes (Oneal and
Russett 2005). We will include several realist and liberal covariates, such
as geographic adjacency, capability distribution, the countries democracy
scores and trade flows; however, we will pay special attention to the effect
of military alliances. Research has not yet established whether military
alliances reduce or increase the likelihood that a militarized dispute breaks
out in a dyad (Bueno de Mesquita 1981; Bremer 1992; Oneal and Russett
2005; Kimball 2006). We will contribute to this debate by examining
whether two countries that share an alliance membership generally interact
more frequently, and if so, whether they behave more cooperatively or in
a more hostile way towards each other. Including alliance membership as
an explanatory variable is also quite illustrative from a methodological
point of view. As it turns out, the positive influence of joint alliances on the
conditional event type can be consistently validatedindependent of which
control variables we used. On the other hand, the positive relationship
between alliances and the frequency of dyadic interaction that can be
validated in a bivariate model diminishes, or even gets reversed, if we
control for network dependencies and other covariates. This illustrates that
the conditional event type is conceptually different from the absolute level
of (friendly or hostile) interaction and emphasizes the need to control for
network effects when testing associations among dyadic variables.
10
While the main contribution of this work is methodological, we present and

test several hypotheses to illustrate and exemplify how our newly-proposed
model can be applied in political science research and how it performs
on empirical network event data. We have chosen the below-mentioned
hypotheses since they illustrate different aspects of our model. Structural
balance theory explicitly predicts that dyads are dependent. More precisely,
interaction on a dyad (a, b) is claimed to depend on previous interaction on
(a, c) and (b, c), for any third actor c. On the other hand, the hypotheses about
the effect of alliances (H5 and H7) claim that interaction on (a, b) depends
on a binary indicator on the same dyad. Although H5 and H7, thus, make
no statement about dependencies among different dyads, we will see that
controlling for network effects leads to different findings for some of these
hypotheses. Thus, even if a particular research question is not about network
dependencies these should nevertheless be tested and, if present, be included
in the model.
Structural balance theory explains the type of events from a to b by the type
of indirect relations via a third actor. More detailed, SBT predicts that actors
behave:
H1 cooperatively towards the friends of their friends;
H2 hostile towards the friends of their enemies;
H3 hostile towards the enemies of their friends;
H4 cooperatively towards the enemies of their enemies.
Drawing on previous results on the effect of alliances, we hypothesize that
events among allies are rather cooperative than hostile. Thus:
H5 allies interact more cooperatively than non-allies. As hypotheses about
event frequencies, we test the following two:
H6 Transitivity of activity: the more actors a and b interacted (cooperatively
or hostile) with common others, the higher the event rate on the dyad (a, b).
Finally, we hypothesize that alliances are only established among countries
that, loosely speaking, have something to do with each other. Thus:
H7 if actor a and b are allies then the event rate on the dyad (a, b) is higher
than if a and b are not allies.
Note that the models we use later to test these hypotheses control for many
more network dependencies which are, however, not of central interest for
this paper and therefore not formulated as explicit hypotheses.
Specification of Conditional Network Event

MODELING
FREQUENCY
AND
TYPE
OF INTERACTION
IN EVENT
NETWORKS
Type
Models
11
Note
that
models
later
to
hypotheses
control
Note
that the
the
models we
we use
use
later
to test
test these
these
hypotheses
control for
for many
many
more
more network
network dependencies
dependencies which
which are,
are, however,
however, not
not of
of central
central interest
interest for
for
this
and
therefore
not
as
hypotheses.
this3paper
paper
andthe
therefore
not
formulated
as explicit
explicit
hypotheses.
We assume
that
occurrence
of events
and
the
type of events
are dependent
Specification
offormulated
Conditional
Network
Event
Typeevents
Models
on previous
on the same or on other dyads. It is the goal of the analyst
to test
and/or
control
for such
network
dependencies
and thereby
to establish
33 We
Specification
of
Network
Event
Specification
of Conditional
Conditional
Network
Event
assume that the occurrence of events and the type of events are
rules that govern the behavior of actors. As an example, if actors a and b
dependent
previous events on the same or on other dyads. It is the goal of
Type
TypeonModels
Models
both had
frequent
hostile
with
common
third actors
(i. e., if they
the analyst
to test
and/or interaction
control for such
network
dependencies
and thereby
assume
that
the
occurrence
of
events
and
the
type
of
events
are
dependent
Weto
assume
that
the
occurrence
of
events
and
the
type
of
events
are
dependent
are We
enemies
of
enemies)
then
this
may
increase
the
probability
that
establish rules that govern the behavior of actors. As an example, if actors a and b
on
previous
events
same
or
other
It
goal
on
previous
events
onthe
the
same
oron
on
otherdyads.
dyads.
Itisisthe
the
goalof
ofthe
theanalyst
analyst
a and
b both
had on
frequent
hostile
interaction
withmodel
common
third
actors
(i.
if section
interact
cooperatively
with
each
other.
The
introduced
ine.,this
to
and/or
control
for
such
network
dependencies
and
thereby
to
establish
totest
test
and/or
control
for
such
network
dependencies
and
thereby
to
establish
are enemies
of enemies)
then this may
increase
the probability
that a and
can bethey
applied
to perform
statistical
tests
for such
hypotheses.
Note that a
rules
that
the
of
As
an
ifif actors
and
rules
that govern
govern
the behavior
behavioreach
of actors.
actors.The
Asmodel
an example,
example,
actors
and bb
b interact
cooperatively
introduced
in thisaasection
preliminary
version
of
thiswith
model other.
has
been
proposed
in Brandes
et
al. (2009).
both
interaction
with
common
third
(i.
both had
had frequent
frequent hostile
hostile
interaction
with
common
third actors
actors
(i.e.,
e., ifif they
they
can be applied to perform statistical tests for such hypotheses. Note that a
are
enemies
enemies)
this
increase
the
arepreliminary
enemies of
ofversion
enemies)
thenmodel
this may
may
increase
the probability
probability
that
a and
and bb
of then
this
has been
proposed
in Brandes that
et
al. a(2009).
interact
interact cooperatively
cooperatively with
with each
each other.
other. The
The model
model introduced
introduced in
in this
this section
section
3.1can
Model
Overview
can be
be applied
applied to
to perform
perform statistical
statistical tests
tests for
for such
such hypotheses.
hypotheses. Note
Note that
that aa
preliminary
version
of
this
model
has
been
proposed
in
Brandes
et
al.
(2009).
preliminary
version
of
this
model
has
been
proposed
in
Brandes
et
al.
(2009).
3.1 Model
Overview of an observed sequence of events E = (e , . . . , e
To model
the probability
1
N ),
we assume that each event ei is only dependent on events that happened

ToModel
model theOverview
probability of an observed sequence of events E = (e1,...,eN),
3.1 To
Model
Overview
3.1
earlier.
obtain
furtheronassume
thathappened
this dependence
we assume
thata tractable
each event emodel,
is onlywe
dependent
events that
i
Toearlier.
modelthe
the
probability
an
observed
sequence
ofencoding
events
=the
(e
To
model
probability
observed
of
events
=
(e
......,,eeNN),), aspects
is completely
captured
byofofaan
dynamic
network
11,,essential
To
obtain
a tractable
model,
wesequence
further assume
thatEEthis
dependence
we
assume
that
each
event
e
is
only
dependent
on
events
that
happened
we
assume
that
each
event
e
is
only
dependent
on
events
that
happened
i
i
completely captured
by aactors.
dynamic network
encoding
the essential
aspects
of pastis interaction
among
The past
events
(i. e., the
events that
earlier.
Tointeraction
obtainaatractable
tractable
model,
we
further
assume
that
this dependence
dependence
earlier.
To
obtain
model,
we
further
assume
that
this
of before
past
among actors.
past
events
(i. e.,G
the
events
that
happen
happen
ei ) determine
theThe
event
network
and,
given
the
state of
ei
completely
capturedthe
byevent
dynamic
network
encoding
the
essential
aspects
isis completely
captured
by
aa dynamic
network
essential
aspects
before ei) determine
network
Gei and,encoding
given thethe
state
of Gei, the
next
Gei ,ofofthe
event among
ei is actors.
assumed
topast
be events
conditionally
of all
pastnext
interaction
among
actors. The
The
past
events
(i.all
e.,other
the independent
events that
that
past
interaction
e.,
the
events
event
ei is assumed
to be conditionally
independent
of(i.
events.
The
other
events.
The
of
ei given
GG
isand,
modeled
parametrically
so
happen
before
determine
the
event
network
given
the
state of
of
happen
before
determine
event
network
given
state
eG
ieeii and,
given Gei isthe
modeled
parametrically
so that
thethe
parameter
probability
ofeeii)e)iprobability
the
next
event
e
is
assumed
to
be
conditionally
independent
of
all
,i , the
next
event
e
is
assumed
to
be
conditionally
independent
of
all
thatGGthe
parameter
estimates
give
the
information
which
properties
of
G
i
eei estimates
i
give the information which properties of Gei increase/decrease
ei
other
events.
The
probability
of of
given
Geand
isofmodeled
modeled
parametrically
so network
other
The
of
eeii events
given
G
parametrically
so
eii is
the events.
frequency
of probability
events
and which
properties
the network
influence
increase/decrease
the
frequency
which
properties
of the
the
that
the parameter
parameter
estimates give
give the
the information
information which
which properties
properties of
of G
Geeii
that
the
estimates
conditional
event type.
influence
the conditional
eventoftype.
increase/decrease
the
frequency
events
and
which
properties
of
the
network
increase/decrease
the
frequency
of
events
and
which
properties
of
the
network
More formally, let E =(e1,...,eN) be a sequence of events and let
More
formally,
let E =
(e1type.
,type.
. . . , eN ) be a sequence of events and let
influence
the conditional
conditional
event
influence
the
event
More formally,
formally, let
let EE =
= (e
(e11,,......,,eeNN)) be
be aa sequence
sequence of
of events
events and
and let
let
More
()
()
()
()
= (() ; () ) = ()
(
, . .()
.(), k()
; 1 ,()
. . . , k )
()1
()
()
() ()
= (
(()
= (
(11 ,,......,,kk ;;11 ,,......,,kk ))
=
;;())) =
be the
parameters
ofofthe
model,
where
the
rate
parameters
() stochastically
()
bethe
thethe
parameters
the
model,
where
the
rate parameters
parameters ()
stochastically
be
parameters
the
stochastically
be
parametersof
of
themodel,
model,where
wherethe
the rate
rate
stochastically
()
determine
the
event
frequency
andthe
the
type
parameters
() stochastically
()
determine
the
event
frequency
and
type
parameters
stochastically
determine
the
event
stochastically
determine
the
eventfrequency
frequencyand
andthe
thetype
typeparameters
parameters stochastically
thetype,
event type,
as shall
we shall
later. The
ensity
function
determinedetermine
the event
as we
seeseelater.
Theprobability
probability
density
for an event sequence E = (e1,...,eN) is
function for an event sequence E = (e1 , . . . , eN ) is
99
f (E; ) = f (e1 |Ge1 ; ) f (e2 |Ge2 ; ) . . . f (eN |GeN ; ) .
(1)
Here f (ei |Gei ; ) denotes the conditional probability density for the event ei ,
JOURNAL OF SOCIOLOGY AND SOCIAL POLICY 1 (2013)
given the network of past events GeiCORVINUS
.
For a given observed sequence of events E = (e1 , . . . , eN ) the function
f (E; ) is the likelihood function when considered as a function of and hy-
determine the event type, as we shall see later. The probability density
function for an event sequence E = (e1 , . . . , eN ) is
12
f (E; ) = f (e1 |Ge1 ; ) f (e2 |Ge2 ; ) . . . f (eN |GeN ; ) .
(1)
theconditional
conditional
probability
density
the event ei ,
Here f (ei |Gei ; ) denotes
Here
denotes the
probability
density
for thefor
event
given
the
network
of
past
events
G
.
ei, given the network of past events Gei. ei
a given
observed
sequence
E )=the(efunction
) the
1 , . . . , eN
For For
a given
observed
sequence
of eventsofE events
= (e1,...,e
f(E;
q) function
N
(E;
) is thefunction
likelihood
function
when
a function
of and hyis fthe
likelihood
when
considered
as aconsidered
function ofas
q and
hypothesis
pothesis
testing is operationalized
by the
maximum
likelihood
i. e.,
testing
is operationalized
by the maximum
likelihood
estimates,
i. e.,estimates,
the
the values
that maximize
f (E;
) (see
Smith
2005). The followvalues
q that maximize
f(E; q) (see
Young
andYoung
Smith and
2005).
The following
ing sections
details
the
different components
sections
provideprovide
details about
the about
different
components
of our model. of our model.
3.2
Input Data
3.2 Input Data

The input data we consider consists of sequences of dyadic, typed events
The
consider
consiststyped)
of sequences
events
E =input
(e1 ,data
. . . ,we
eN ).
A (dyadic,
event ofe dyadic,
E istyped
defined
to Ebe a tuple
=(e
,...,e
).
A
(dyadic,
typed)
event
e
e
E
is
defined
to
be
a
tuple
e
=(a
,b
,w
,t ),
e e
e e
e 1= (aNe , be , we , te ), where:
where:

ae isis the
of e;of e;
thesource
source(initiator)
(initiator)
be is the target (addressee) of e;
target
of e; of the event e; and
bwee eisRthe
is the
type,(addressee)
coding the quality
t is the time when e happens.
we e R is the type, coding the quality of the event e; and
The source and the target of events are termed actors. Actors are, e. g.,
te is the time when e happens.
people, groups of people, organizations, or countries.
TimeThe
is given
on and
somethe
scale,
e. g., of
byevents
second,are
minute,
hour,actors.
day, month,
or are, e. g.,
source
target
termed
Actors
year.
In
the
KEDS
data
time
is
given
by
the
day.
Several
events
may
happen
people, groups of people, organizations, or countries.
duringTime
the same
time on
unit.
Thescale,
event e.sequence
is assumed
to behour,
in nonis given
some
g., by second,
minute,
day, month,
decreasing
order
with
respect
to
time.
The
order
of
events
that
happen
within
or year. In the KEDS data time is given by the day. Several events may
the same time unit is considered as undefined. We note that for our analysis
happen during the same time unit. The event sequence is assumed to be in
we do not need the absolute time t but rather the time difference Dt between
non-decreasing order with respect to time. The order of events that happen
events.
within
same
time eunit
considered
We note the
that for our
The
typethe
we of
an event
(alsoisreferred
to as as
its undefined.
weight) characterizes
analysis
we
do
not
need
the
absolute
time
t
but
rather
the
time
quality of e. In the exemplary application of this paper the weight we of an difference
t between
event
e is a realevents.
number from the interval [1, 1] obtained by dividing the
The weights
type weofofKEDS
an event
e (also
to as
its weight)
Goldstein
events
by ten.referred
A positive
weight
indicatescharacterizes
a
the quality
of e.
In the exemplary
application
this
paper value
the weight we
cooperative
event,
a negative
weight a hostile
event, andofthe
absolute
an event
e ismeasures
a real number
from the
[1, 1] obtained
by dividing
of ofevent
weights
the magnitude
of interval
cooperativeness
or hostility,
the Goldstein
weights
of KEDS
by ten.zero
A positive
respectively
(so that
this scale
has a events
non-arbitrary
indicatingweight
neutralindicates a
events). In other applications, events may have other types, e. g., binary,
multinomial, ordered multinomial, or event types
10 might be multidimensional.
While our model could be extended to these more general types of events this
is not considered in this paper.
3.3
Explanatory Variable: The Network of Past Interaction

13
Given a sequence of events E = (e1 , . . . , eN ) and a specific timepoint t (deVariable:

The network
Network of
noting 3.3
theExplanatory
current time),
the event
at Past
timeInteraction
t (referred to as network
of past events if t is implied) is a weighted graph Gt = (A; Wt ) defined as
Given a sequence of events E =(e1,...,eN) and a specific timepoint t (denoting
a function
of the set of past events E<t =t (referred
{e Eto; aste network
< t}; of
i. e.,
the set of
the current time), the event network at time
past
events events
that happen
before
Furthermore,
the Wevent
network
might encode
if t is implied)
is a t.weighted
graph Gt =(A;
) defined
as a function
t
(potentially
time-dependent)
actor,
dyad,
or
network
covariates
that are not
of the set of past events E<t = {e e E ; te <t}; i. e., the set of events that
happen
t. Furthermore,
the event
might encode
(potentially For ina function
of before
previous
events but
that network
give additional
information.
time-dependent)
actor,
dyad,
or
network
covariates
that
are
not
a
functionbeof the gross
stance, in the case of political networks, such covariates might
previous events but that give additional information. For instance, in the case
domestic
product of a country (as an example of an actor covariate) or the
of political networks, such covariates might be the gross domestic product of
geographical
distance
between
(as an
example
of a dyad
covariate).
a country
(as an example
of ancountries
actor covariate)
or the
geographical
distance
In our between
application
these
covariates
are
given
as
yearly
data.
countries (as an example of a dyad covariate). In our application
covariates are
yearly
Thethese
components
ofgiven
Gt =as(A;
Wtdata.
) are explained in the following. The set
The
components
of
G
=(A;
W
)
are
explained
in event
the following.
Thekeep
set A the set of
t
t
A consists of the actors that are involved
in any
(thus we
consists of the actors that are involved in any event (thus we keep the set of
actors fixed
over time) and W is a vector-valued function mapping each dyad
actors fixed over time) and Wt t is a vector-valued function mapping each dyad
(a, b) to(a,ab)value
thatthat
characterizes
theessential
essential
aspects
a interacted
to a value
characterizes the
aspects
of howofa how
interacted
with b in
i. e.,
before
Moreformally,
formally,
j A, i =
withthe
b inpast,
the past,
i. e.,
beforet.t. More
let Dlet
=D
{(i, =
j); {(i,
i, j ej)
A,; i i,
j}
be the
all dyads.
Wt isW
a tfunction
j} be the
setsetofofall
dyads.Then
Then
is a function
Wt : D Rd ; (a, b) (Wt,1 (a, b), . . . , Wt,d (a, b)) ,
where Wwhere
b)
R edenotes
valuein in
ith dimension,
t,i (a, b)
Wt,i(a,
R denotesthe
the real
real value
the the
ith dimension,
for i = for i =
1, . . . , d.
1,...,d.
In our concrete application, the network of past events Gt = (A; wt+ , wt )
=(A; wt +W
,wt =
) is(w + , w ),
In ournetwork
concrete application,
the network of pastweight
events G
t
is a weighted
with a two-dimensional
function
t
t
t
a weighted network with a two-dimensional weight function Wt =(wt+ ,wt), 10
encoding
past
cooperative
and
past
hostile
interaction,
respectively.
The
encoding past cooperative and past hostile interaction, respectively.10 The
10
value abuse
of cooperative/hostile
interaction
a particular
dyadby(a,the
b) same
increases
By a slight
of notation we denote
the of
weight
on a dyad
letter, namely
whenever
a
initiates
a
cooperative/hostile
event
e
targeted
at
b.
When
the
w, as the weight of an event; this should not cause any confusion.
difference between the current time t and the event time te increases, the
influence of e diminishes. The latter property is motivated by the assumption
11 and hostile actions. Assuming
that actors forget (or forgive) cooperative
that the rate of forgetfulness or forgiveness is only dependent on the current
weight, we obtain an exponentially decreasing influence of each event when
time increases. More precisely, let T1/2 e R>0 be a given positive number
10 By a slight abuse of notation we denote the weight on a dyad by the same letter, namely w, as
the weight of an event; this should not cause any confusion.
influence
influence of
of ee diminishes.
diminishes. The
The latter
latter property
property is
is motivated
motivated by
by the
the assumption
assumption
that
actors
forget
(or
forgive)
cooperative
and
hostile
actions.
that actors forget (or forgive) cooperative and hostile actions. Assuming
Assuming
that
that the
the rate
rate of
of forgetfulness
forgetfulness or
or forgiveness
forgiveness is
is only
only dependent
dependent on
on the
the current
current
14
weight,
influence
weight, we
we obtain
obtain an
an exponentially
exponentially decreasing
decreasing
influence of
of each
each event
event when
when
time
R>0 be
time increases.
increases. More
More precisely,
precisely, let
let T
T1/2
be aa given
given positive
positive number
number
1/2 R>0
+
+
denoting
the
halflife
of
the
influence
tt ::+D
denoting the
thehalflife
halflife of
of the
the influence
influence of
of
events.
Then,
the
function
ww
D
denoting
ofevents.
events.Then,
Then,the
thefunction
functionw
: D
t
R
is
defined
by
0
definedby
by
RR00is isdefined

ln(2) ln(2)
+
ln(2)
+ (i, j) =
ln(2)
|w
|
exp
(t
t
)
w
e
e
wtt (i,
j) = of a particular|w
exp
(t te ) T
e | (a,
value of cooperative/hostile
interaction
dyad
b) increases
T
1/2
T
T1/2
e:aevent
=i, b =j,
1/2
1/2
whenever a initiates a cooperative/hostile
targeted at b. When the
e:ae
e =i, be
e =j,
w >0, t <t
we
>0, event
te
ethe
e <t time te increases, the
difference between the current time t and
influence of e diminishes. The latter property is motivated by the assumption
11
that actors
forgive) cooperative
Assuming
and
the
w
R
is
by
0 actions.
andforget
the(orfunction
function
wtt ::D
D and
hostile
R0
is defined
defined
by1111
that the rate of forgetfulness or forgiveness
on the current
t is only dependent
0

weight, we obtain an exponentially decreasing
influence of each eventwhen
ln(2)

ln(2)
time increases. More precisely,
let T1/2 R
ln(2)
ln(2)
>0 be a given positive number
w
(i,
|w
te ))
ee || exp
T
(i, j)
j) =
=
|wfunction
expwt+: D(t
(t
denoting the halflife of w
thett influence
of events. Then, the
te
T
1/2
T1/2
T
e:ae =i, be =j,
1/2
1/2
R0 is defined by
e:ae =i, be =j,
w <0, t <t

wee <0, tee <t

ln(2)
ln(2)
|we | exp
wt+ (i, j) =
+ (t te ) T1/2 T1/2
e:ae=i,value
be=j,
Thus,
w
Thus, the
the
wtt+ (i,
(i, j)
j) is
is defined
defined as
as aa function
function of
of all
all weights
weights of
of events
events
we>0,value
te<t
+ involve
e
=
(a
,
b
,
w
,
t
)
that
i
as
source
(a
=
i)
and
j
as
target
(b
e
e
e
e
e
11
e
=
(a
,
b
,
w
,
t
)
that
involve
i
as
source
(a
=
i)
and
j
as
target
(bee =
= j),
j),
e
and the function wet : De Re0 ise defined
t by
that
happen
before the
time
tt (t
t),
and
that
have
positive
weight

current
e <
that
happen
current
time
(t
<
t),
and
that
have
positive
weight
before the
e
ln(2)
ln(2)
e =e
e e |w | exp (t
e
e
wt (i,>
j)
te ) j) is the
e
(w
Similarly,
w
ee > 0).
T1/2 sum
(w
0).
Similarly,
wtt(i,
(i,
j)T1/2is the
sum over
over events
events with
with negative
negative weight
weight
e:ae=i, be=j,
e
e
we<0, te<t
(w
How
strongly
an
event
ee is
counted
at
time
tt depends
on
the
time
(wee <
< 0).
0).
How
strongly
an
event
is
counted
at
time
depends
on
the
time+
t
e
Thus, the
value
w
(i,
j)
is
defined
as
a
function
of
all
weights
of
events
t
difference
ttee .. Each
Each time this
this difference increases
increases by
by T
T1/2
the factor for w is
tt
1/2 the factor for wee is
e = (ae , difference
be , we , te ) that involve
i as sourcetime
(ae = i) and jdifference
as target (be = j),
halved.
The
choice
of
T1/2
dependent
on
whether
the
analyst
is
that happen
before the
current
time t (t
andis
that
have positive weight
1/2
e <Tt),
halved.
The
choice
of
is
dependent
on
whether
the
analyst
is interested
interested
(we > 0).
Similarly,
wt (i, j) or
is the
sum over events
with negative
weight
e
1/2
e
in
short-term
long-term
responses
to
previous
events.
Estimation
of
short-term
or elong-term
responses
events. Estimation
of T
T1/2
1/2
(we < 0).inHow
strongly an event
is counted at time
t depends onto
the previous
timefrom
empirical
data
is
possible,
but
not
considered
in
this
paper.
The
last
difference
tt
.
Each
time
this
difference
increases
by
T
the
factor
for
w
is
1/2
e
e
1/2
from empirical
data
is
possible,
but
not
considered
in
this
paper.
The
last
analyst is interested
halved. The choice ln(2)
of T1/2 is dependent on whether the
ln(2)
factor
is
used
give
w
j)
of
1/2
factor
isresponses
used toto
toprevious
give events.
wtt (i,
(i,Estimation
j) the
the ofinterpretation
interpretation
of aggregate
aggregate weight
weight
T
in short-term
or long-term
T1/2
T1/2
1/2
from empirical
data is unit.
possible, but
not considered
in this paper.
The last accurate if time units are small
per
time
This
interpretation
is
quite
per
time
unit.
This
interpretation
is
quite
accurate
if
time
units
are
small
factor ln(2)
is used to give wt (i, j) the interpretation
of aggregate weight
T1/2
and
only
an
if
is
on a coarse scale.
t
an approximation
approximation
if iftime
time
is given
given
per timeand
unit.only
This interpretation
is quite
accurate
time units
are smallon a coarse scale.
and only an approximation
if time is given on
a coarse functions
scale.
Higher
weight
might
Higher dimensional
dimensional
weight
functions
might additionally
additionally encode,
encode, e.
e. g.,
g., the
the
Higher dimensional weight functions might additionally encode, e. g., the
minus
conflict
increase
defined
as
the
difference
of
the
current
value
of
w
conflict
increase
defined
ascurrent
the value
difference
of the current value of wtt minus
conflict increase
defined
as the difference
of the
of wt minus
its valueits
t time
unitst
in the
past, units
i. e., wt
value
time
inwtt
the. Whether
past, i.
i.thee.,
e.,levelw
woftt
. Whether the level of
its
value
t
time
units
in
the
past,
wtt
tt . Whether the level of
conflict is currently increasing or decreasing might have a significant effect
conflict
is
currently
increasing
or
decreasing
might
have
aa significant
effect
conflict
is
currently
increasing
or
decreasing
might
have
significant
effect
11
Note that w also maps to the non-negative numbers, since the absolute value |we | is
t
taken. 11
11 Note that w also maps to the non-negative

numbers,
since
t
t Dt
Note that w also maps to the non-negative
numbers,
since the
the absolute
absolute value
value |w
|wee || is
is
taken.
12
taken.
and the function w : D R is defined by
Thus, the value w (i, j) is defined as a function of all weights of events

e =(a ,b ,w ,t ) that involve i as source (a = i) and j as target (b = j), that
happen before the current time t (t <t), and that have positive weight (w >
0). Similarly, w (i, j) is the sum over events with negative weight (w < 0).
How strongly an event e is counted at time t depends on the time-difference
tt . Each time this difference increases by T the factor for w is halved.
The choice of T is dependent on whether the analyst is interested in shortterm or long-term responses to previous events. Estimation of T from
empirical data is possible, but not considered in this paper. The last factor
is used to give w (i, j) the interpretation of aggregate weight per time
unit. This interpretation is quite accurate if time units are small and only an
approximation if time is given on a coarse scale.
Higher dimensional weight functions might additionally encode, e. g., the
conflict increase defined as the difference of the current value of w minus its
value Dt time units in the past, i. e., w w
. Whether the level of conflict
is currently increasing or decreasing might have a significant effect on future
eventsadditional to the absolute level of conflict. The increase of conflict is,
12
however, not used as an explanatory 12
variable in this paper.
If e is an event in E, we sometimes write Ge for Gte. Note that Ge is only
dependent on events that happen earlier than e (and not on events that happen
in the same time unit as e).
3.4 Dependent Variable: The Next Event

In this section we model the probability density f(e|Ge;); i. e., the probability
density of an event e dependent on the event network Ge, compare Eq. (1).
11 Note that w also maps to the non-negative numbers, since the absolute value |we | is taken.
ability
density
of
event
on
event
3.4
Variable:
The
Event
ability
density
of an
an
event ee dependent
dependent
on the
the
event network
network G
Gee,, compare
compare
3.4Dependent
Dependent
Variable:
TheNext
Next
Event
Eq. (1).
(1).
InEq.
this
section
we we
model
thetheprobability
density
ff(e|G
;;);
probInThe
this
section
model
probability
density
(e|GeeNETWORKS
); i.i. e.,
e., the
the probMODELING
FREQUENCY
AND TYPE
OF INTERACTION
INprobability
EVENT
15
first
step
is
the
decomposition
of
the
density
of
into
The
first
step
is
the
decomposition
of
the
probability
density
of events
events
into
ability
density
of
an
event
e
dependent
on
the
event
network
G
,
compare
ability density of an event e dependent on the event network Gee , compare
a
rate
component
and
a
conditional
type
component.
Let
e
=
(a
,
b
,
w
e
e
e
Eq.
(1).(1).
a Eq.
rate
component and a conditional type component. Let e = (ae , be , we,, ttee))
Thein
first
stepobserved
is the decomposition
of theThe
probability density of
events into
be
an
event
sequence
density
of
e,
first
stepstep
isthe
the
decomposition
ofofthe
probability
density
into
first
is the
decomposition
theE.
probability
density of events
events
into
beThe
anThe
event
in
the
observed
sequence
E.
The probability
probability
density
of
e, given
given
a
rate
component
and
a
conditional
type
component.
Let
e
=(a
,b
,w
,t
)
be
e
e
e
e
network
of
past
events
G
,
is
decomposed
into
two
factors:
a
rate
component
and
a
conditional
type
component.
Let
e
=
b
,
w
,
t
)
a the
rate
component
and
a
conditional
type
component.
Let
e
=
(a
,
w
,
t
)
e
e
e
e
e
e
e
the network
of in
past
events G
intodensity
two factors:
e , is decomposed
an event
the observed
sequence
E. The probability
of e, given the
be event
an event
in the
observed
sequenceE.E.The
Theprobability
probability density
density of e, given
be an
in the
sequence
given
network
ofobserved
past events
Ge, is decomposed
into
two
factors:
()
()
ffof
(e|G
;
)
=
f
(a
,
b
,
t
|G
;
f
(w
|a
,
b
,
t
;
G
;
(2)
())into
network
of
past
events
G
,
is
decomposed
into
two
factors:
e
e
e
e
e
e
e
e
e
e
e
the the
network
past
events
G
,
is
decomposed
two
factors:
e e , be , te |Ge ;
(e|Ge ; ) = f (a
) f (we |ae , be , te ; Ge ; ())) ..
(2)
()
()
f (e|G,et;e=
) f=e;(a
f ()
(a
, b e, |G
te |G
;()
) ) frate
f(w
, b , t ; G ; is
() )) .. probability
(2)
(w
e |a
Here
(a
den, be),
|a
(2)
e , tecalled
e ; e the
ecomponent,
e ,e bee, tee; Gee;
Here fff(e|G
(aee,, ebb;ee)
, te|G
|Ge ; e()
),
called
the rate
component,
is the
the probability
density
that
the
next
event
happens
at
time
t
and
involves
a
as
source
()
e
e
Here
f
(a
,
b
,
t
|G
;
),
called
the
rate
component,
is
the
probability
den()
sityfthat
happens
time te()and
involves
ae as
source and
and
e e next
e
e event
Here
(ae, bthe
), called
the rateatcomponent,
is the
probability
dene , te |Ge ;
()
Here
fLikewise,
(ae,bevent
,t |Ge;happens
f),(w
called
iscalled
the
probability
density
bbesity
as
target.
bbe ,rate
tte t;;ecomponent,
G
),
the
conditional
type
()
that
the
next
at
time
and
involves
a
as
source
and
e |a
e ,,the
e ;;
Likewise,
e e
e
as
target.
f
(w
|a
,
G
),
called
the
conditional
type
sitye that the
event
happens
and
eat etime
e tee ()
e involves ae as source and
that next
the
next
event
happens
time
te and
a asthe
source
and b as target.
b as target.
Likewise,
f (we |aeat
, bprobability
involves
),density
called
conditional
is
the
conditional
that
event
has
type
e , te ; G
e ; ()
becomponent,
component,
is
the
conditional
probability
density
that
evente ee type
has
type w
wee,,
ase target.
Likewise,
f
(w
|a
,
b
,
t
;
G
;
),
callede the
conditional
type
()
e
e
e
e
e
Likewise,
f
(w
|a
,b
,t
;
G
;
),
called
the
conditional
type
component,
isethe
component,
is the
probability
density
that
event
e
has
type
w
,
conditional
e event
e e e involves
e
given
that
the
next
a
as
source,
b
as
target,
and
happens
at
e as target,
component,
is the
the next
conditional
probability
density
that
e hasthat
type
given
that
event
involves
aesource,
as source,
bevent
and
happens
at
e,
conditional
probability
density
that
e bhas
type
thewat
next
eevent
e we, given
given
that
the
next
event
involves
a
as
as
target,
and
happens
e
e
time
tte .. event
Defining
given
that
the
next
event
involves
a
as
source,
b
as
target,
and
happens
at
e
e
involves
a
as
source,
b
as
target,
and
happens
at
time
t
.
Defining
time
Defining
e
e
e
time ete . Defining

time te . Defining
()
()
)) =
bbee,,; tte()
|G
(E;
())()
=
, teee,,|G
) ee;; ()))
fff(E;
= f (aeff,b(a
e(a
(E;
e ()
e |G
()
f (E;
and
and
and
and
and
)=
feE
(ae , be , te |Ge ;
eE
eE
eE

() ()
|ae , be , te ; Ge ; () )
()
ff
)
=

(E;
(w
()) = f (wfe
())
(E;
e |a
e ,, b
e ,, tt()
e ;; G
e ;;
()
f
(E;
)
=
f
(w
|a
b
G
e
e
e
e
e
f (E; ) = eE
f (we |ae , be , te ; Ge ; )
eE
eE eE density into
we can decompose the joint probability
can
decompose
probability
into
wewe
decompose
the the
jointjoint
probability
densitydensity
into() into
wecan
can
decompose
the
joint
probability
density
into
()
we can decompose
the joint
probability
density
f (E; ) = f (E;
) f (E;
) ,
(3)
()
()
(E;
=
ff()
(E;
)) ff()
f (E;ff)
= )
f (E;
) f()
(E;
) , ())) ,,
(3)
(E;
(E;
)
=
(E;
(E;
where f is the rate component of the joint probability density, modeling the
(3)
(3)
occurrence
ofrate
events,
and f of
is the
theofjoint
conditional
type density,
component,
modeling
where
f fis the
component
probability
modeling
the
where
the
component
joint
probability
density,
modeling
the
where
f is
is
thef rate
rate
component
ofofthe
the
joint
probability
density,
modeling
the
where
is
the
rate
component
the
joint
probability
density,
modeling
the
the
distribution
of
event
types.
occurrence
of events,
and and
f isfthe
conditional
type
component,
modeling
occurrence
of
events,
is
the
conditional
type
component,
modeling
occurrence
of
events,
and
f
is
the
conditional
type
component,
modeling
occurrence
of events,Eq.
and(2)
f is
conditional
component,
Despite
itsofsimplicity,
is the
a key
concept type
in the
definitionmodeling
of condi-the
the
distribution
event
types.
the
distribution
of
types.
distribution
ofevent
event The
types.
tional
event
type
models.
main
decision
here
is
that
a
set
of
parameters
the
distribution
of
event
types.
Despite
its its
simplicity,
Eq. (2) is(2)
a key
in the definition
of condi- condiDespite
simplicity,
aaconcept
key
concept
in
definition
Despite
its simplicity,Eq.
Eq. (2)
is is
a key
concept
in the definition
of conditionalof
itsmodels.
simplicity,
Eq.
(2)
is
key
concept
insetthe
the
definition
of conditionalDespite
event
type
The
main
decision
here
is
that
a
of
parameters
event type models. The main decision here is that a set of parameters ()
tional
type
13 decision
tional event
event
type models.
models. The
The main
main
decision here
here is
is that
that aa set
set of
of parameters
parameters
stochastically determines the occurrence of events (of any type) and a disjoint
13
set of parameters () stochastically
determines the conditional type of an
13
event, given that an event happens on
13a particular dyad. The decomposition
in Eq. (2) implies that the rate parameters () and the type parameters ()
can be estimated separately and, more importantly, the maximum likelihood
estimates of () are independent of the specification of the event frequency f
and, vice versa, the maximum likelihood estimates of () are independent of
the specification of the distribution of the conditional event type f.
The argument for this modeling decision has been given in the introduction:
results about the determinants of the conditional event type give additional
()
disjoint set
of parameters
determines
the conditional
type
position
in Eq.
(2) implies thatstochastically
the rate parameters
() and
the type paramof
an
event,
given
that
an
event
happens
a particular
dyad.the
The
decom()
eters can be estimated separately
and,onmore
importantly,
maximum
position inestimates
Eq. (2) implies
rate parameters
() and the type
likelihood
of ()that
are the
independent
of the specification
of theparamevent
()
16
eters
can
be
estimated
separately
and,
more
importantly,
the
maximum
()
frequency f and, vice versa,
the
maximum
likelihood
estimates
of
are
likelihood
estimates
of () are independent
of the event
independent
of the distribution
of the conditional
event
frequency
f and,
vice versa,
the maximum
likelihood
estimates
of ()inare
type f . into
insights
the causes
of typed
events. Indeed,
it will
be shown
this
independent
of
the
specification
of
the
distribution
of
the
conditional
event
The
argument
for this modeling
been given
the intropaper
that
for our example,
the effectdecision
of jointhas
alliances
on theinfrequency
of
type f . results about the determinants of the conditional event type give
duction:
dyadic
events is very sensitive to the inclusion or exclusion of specific control
The argument
for this
modeling
decision
has Indeed,
been given
inbethe
introadditional
insights
into
the causes
of typed
events.
it will
shown
in
variables.
On
the about
other
hand,
sharing
an alliance
membership
consistently
has
duction:
results
the determinants
ofof the
conditional
event
type give
this
paper
that
for
our
example,
the
effect
joint
alliances
on
the
frequency
aadditional
positive effect
on
the
conditional
event
type
meaning
that
this
association
insightsisinto
thesensitive
causes oftotyped
Indeed,
it will beofshown
in
dyadic events
very
the events.
inclusion
or exclusion
specific
isof
robust
tothat
the inclusion
or exclusion
of control
variables
and
is even
robust
this
paper
for
our
example,
the
effect
of
joint
alliances
on
the
frequency
control variables. On the other hand, sharing an alliance membership consiswhether
or not weisconsider
any network
at all.
of dyadic
to thedependencies
inclusion
or exclusion
ofthat
specific
tently
has events
a positive very
effectsensitive
on the conditional
event typemeaning
this
Next,
we
clarify
the
functional
form
of
the
density
for
the conditional
event
control variables.
On to
thethe
other
hand, or
sharing
an alliance
membership
association
is robust
inclusion
exclusion
of control
variables consisand
is
tently
has
a
positive
effect
on
the
conditional
event
typemeaning
that
this
type
(event
quality)
and
the
event
rate
(event
frequency).
We
assume
that
even robust whether or not we consider any network dependencies at all. the
association
robust
theactor
inclusion
or
offor
control
variables
and
is
type
we ofweanisclarify
event
e to
from
aform
to actor
is dependent
the current
state
Next
the
functional
of exclusion
thebdensity
theon
conditional
event
()
even
robust
whether
or
not
we
consider
any
network
dependencies
at
all.
of
the
network
G
and
the
type
parameters
.
type (event quality)
and the event rate (event frequency). We assume that
e
() density()
Next
we
the functional
form
of
for ,),
the conditional
event
For
parameters
()
= a(
,...,
the
conditional
the
typegiven
we clarify
of an
event
e from
actor
to1 the
actor
b iskdependent
on the
current
type (event
quality)
and
the the
event
rate
(event
frequency).
We
assume that
()
state
of the network
and
type
parameters
.
distribution
of the G
weight
w
of
event
e
=(a,
b,
w,
t)
is
modeled
as a
e
()
()
the For
typegiven
we ofparameters
an event efrom
b is
dependent
ondistribution
the current
()
= actor
(
.to
. , actor
,
),
the
conditional
normal
distribution,
leading
to1 a, a.likelihood
which
is
the
same
as
that
k
state
ofweight
the network
Ge and
type
parameters
()
. a normal distribution,
of a
the
w of event
e()
= the
(a, b,
w,
t)
is
modeled
as
of
linear
regression
model:
()
()
For given
parameters
=
(1 same
, . . . , as
kthat
, ),of
the
conditional
distribution
leading
to a likelihood
which
is the
a linear
regression
model:
of the weight w of event e = (a, b, w, t) is modeled as
a normal distribution,

2
[w ab ]
leadingfto(w|a,
a likelihood
same as 1that exp
of a linear
regression
.model:
(4)
b; Ge ; ()which
) = isabthe
, 2 (w) =

2 2 2
2 2
1
[w
]
ab
f (w|a,
(4)
b; Ge ;event
() ) =weight
ab ,2(w)== (G ; exp
()
Here, the
expected
) ispostulated
to . be deab
ab2
e2
2 2
()
()
()
pendent
on the
parameters
statistics
Here, the
expected
event(weight
ab) and
= the
(Gvalues
; )ofisvarious
postulated
to be
1 ,...,
Here, the expected event weight
ab =k()ab (Ge ;ab()e) is postulated to be de()
sh (Ge ; a, b),onh the
= 1,parameters
. . . , k that(1
characterize
the
network
around
a
and
b
(see
dependent
,...,()
)
and
the
values
of
various
statistics
()
pendent
on the parameters
(1expected
, . . . , kk
) andweight
the values
of various
statistics
event
Sect.
the
is modeled
function
sh(G3.5).
; a, More
b), h precisely,
=1,...,k that
characterize
the network
aroundasaaand
b (see
e
sh (Ge ; a, b), h = 1, . . . , k that characterize the network around a and b (see
Sect. 3.5). More precisely, the expected
event
weight
is
modeled
as
a
function
k
event
Sect. 3.5). More precisely, the()expected
weight is modeled as a function
()
ab (Ge ; ) =
h sh (Ge ; a, b) .
(5)
k

h=1
()
ab (Ge ; () ) =
h sh (Ge ; a, b) .
(5)
() reveal
The maximum likelihood estimates
h=1 of the weight parameters h
dependencies between characteristics of the network and later observed
() event
The maximum likelihood estimates of the weight parameters h reveal
The maximum
likelihood
estimates
weight
parameters
h()event
reveal
dependencies
between
characteristics
of of
thethe
network
and
later observed
14
dependencies between characteristics of the network and later observed event

14
weights. For instance, if a particular statistic
sh(Ge; a, b) encodes how much b
attacked a in the past, then a (significantly) negative value for the associated
parameter h() would imply that actors show a tendency to initiate hostile
events towards attackers.
The modeling of the event frequency is slightly more complicated, since not
the frequency itself but rather the waiting time between events is observed.
However, estimating event frequencies from observed time-to-event data is
a common task in lifetime analysis (Lawless 2003) from where we get the
necessary methodology.
parameter
hinstance,
would imply
that actors
showsha(Gtendency
to initiate
hostileb
weights. For
if a particular
statistic
how much
e ; a, b) encodes
events towards
attacked
a in theattackers.
past, then a (significantly) negative value for the associated
()
The modeling
of the
event
frequency
is slightly
more complicated,
since
parameter
h would
imply
that
actors show
a tendency
to initiate hostile
not
the
frequency
itself
but
rather
the
waiting
time
between
events
is
ob- 17
MODELING
FREQUENCY
AND TYPE OF INTERACTION IN EVENT NETWORKS
events towards
attackers.
served.
However, of
estimating
frequencies
frommore
observed
time-to-event
The modeling
the eventevent
frequency
is slightly
complicated,
since
datathe
is afrequency
common task
lifetime
analysis
(Lawless
2003)
from where
get
not
itselfinbut
rather
the waiting
time
between
events we
is obA necessary
key concept
in modeling event times is the so-called hazard function (also
the
methodology.
served.
However,
estimating event frequencies from observed time-to-event
called
intensity
function,
or rate
function).
notation
weweassume
concept
in modeling
event
times(Lawless
isTo
thesimplify
so-called
hazard
dataAiskey
a common
task
in lifetime
analysis
2003) from
wherefunction
get
here
that
event
times
are
known
exactly
and
that
at
any
point
in
time
there
(alsonecessary
called intensity
function, or rate function). To simplify notation
wecan
the
methodology.
happen
at
most
one
event;
the
likelihood
function
for
the
occurrence
of
events
assume
here
that event
times are
known
that hazard
at any function
point in
A key
concept
in modeling
event
timesexactly
is the and
so-called
can
happen
at most
likelihood
for the
istime
laterthere
derived
without
this
assumption
(i. e.,the
where
isfunction
only
known
up to
(also
called
intensity
function,
orone
rateevent;
function).
Totime
simplify
notation
we
occurrence
of
events
is
later
derived
without
this
assumption
(i.
e.,
where
some
fixed
precision
and
where
more
than
one
event
may
happen
during
assume here that event times are known exactly and that at any point inthe
time time
is only
known
up at
to most
some one
fixedevent;
precision
and wherefunction
more than
time
there
can
happen
the likelihood
for one
the
same
interval).
event
mayb)happen
the in
same
interval).
occurrence
of
events
is dyad
later
derived
without
this
assumption
(i. in
e., time,
whereand
Let (a,
eD
beduring
any
the time
network,
let
t denote
a point
(a,
b)
D be up
the network,
t denote
pointthan
in time,
time
is
only
known
somein fixed
precisionletand
where amore
one
let
NLet
(t),
defined
by anytodyad
ab
and
let
N
(t),
defined
by
ab
event may happen during the same time interval).
Let (a, b) DNbe(t)
any
= dyad
|{e in
E ;the
ae network,
= a, be =let
b, tte denote
t}|, a point in time,
ab
and let Nab (t), defined by
denote the number of events that happen on the dyad (a, b) before or at time
E ; apoint
ab (t) = |{e atime
e = a,t bto
e = b, te t}|,
t. denote
The function
N
ab mapping
the number
of events that happen
on the dyad (a, b) before or at time
the number
of events athat
happen
on
dyad
E[N
(t + tt)
N
| Gb)
t.denote
The function
ab mapping
time
point
tothe
ab
ab (t)(a,
t ] before or at time
(6)
(t) = lima time point t to
t. The function ababmapping
t0
t
E[Nthe
+ t)(a,
b).
Nab (t)
| Gt ] the function E[]
ab (tdyad
is called the hazard function for
(Here,
(6)
ab (t) = lim
denotes the expectation oft0
the argument.) t
Intuitively, the hazard function
cancalled
be interpreted
the expected
of b).
events
in athe
time
intervalE[]
of
is
the hazardasfunction
for thenumber
dyad (a,
(Here,
function
length
one.
Thus,
(t)
is
also
referred
to
as
the
event
rate
on
the
dyad
ab
denotes the expectation
of the argument.) Intuitively, the hazard function
is called
thet.hazard
function
theone
dyad (a, can
b). (Here, the
E[]
(a,
time
Note
if atfor
most
on function
(a, b), this
canb)beatinterpreted
as that,
the expected
numberevent
of eventshappen
in a time
interval
of
denotes
the
expectation
of
the
argument.)
Intuitively,
the
hazard
function
definition
is equivalent
to the
morereferred
usual definition
the hazard
ascan
length
one.
Thus, ab (t)
is also
to as theofevent
rate onfunction
the dyad
be
interpreted
as
the
expected
number
of
events
in
a
time
interval
of
length
being
the
conditional
probability
density
that
the
event
happens
at
time
t,
(a, b) at time t. Note that, if at most one event can happen on (a, b), this
given
that
it
did
not
happen
before
(Lawless
2003).
The
definition
given
in
one.
Thus,
(t)
is
also
referred
to
as
the
event
rate
on
the
dyad
(a,
b)
definition is equivalent
to the more usual definition of the hazard function as at
ab
Eq.
6t.is
preferable
case,one
since
it generalizes
repeated
time
Note
that, ifinatour
most
event
can
happen
on
(a,
b),events.
this
being
the
conditional
probability
density
that
the to
event
happens
at definition
time t,
Wethat
assume
the
hazard
is a definition
function
of of
the
current
of the network
it did
not
happen
before
(Lawless
2003).
Thestate
definition
given
in
isgiven
equivalent
tothat
the
more
usual
the
hazard
function
as being
12
but
that
this function
is time-invariant.
Similar
totothe
event type,
the t,
event
Eq. conditional
6 is preferable
in our
case,
since it
generalizes
repeated
events.
the
probability
density
that
the event
happens
at
time
given
12 We assume that the hazard is a function of the current state of the network
Foritmore
hazard before
functions
that have 2003).
an explicit
dependency
see Lawless
that
did general
not happen
(Lawless
Thetime
definition
given
in Eq. 6
12
that this function
is time-invariant.
Similar
the event
type, We
the event
(2003).
isbut
preferable
in our case,
since it generalizes
to to
repeated
events.
assume
12
Forthe
more
generalishazard
functions
an explicit
time
that
hazard
a function
of that
the have
current
state of
thedependency
network see
butLawless
that this
(2003). is time-invariant.12 Similar 15
function
to
the
event
type,
the
event
rate
ab()(Gt;
()
()
(Gt ; () ) on
is dependent
on the rate
(1the
, . .values
. , k )of
()
rate
) isab
dependent
the rate parameters
()parameters
=(1 () ,...,k ())=and
and the values of various statistics sh (G
a, b), h = 1, . . . , k that characterize
15 t ; that
various
statistics sh(Gt; a, b), h =1,...,k
characterize the network around
the network around

a and b. More precisely,
the rate is specified to be a
a and b. More precisely, the rate is specified to be a function
function
ab (t) = ab (Gt ;
()
) = exp
k

h=1
()
h
sh (Gt ; a, b)
(7)
Theexponential
exponentiallink
link
function
form
ensures
a positive
event
rate.
The
function
form
ensures
a positive
event
rate.
()
The maximum likelihood estimates for the rate parameters h reveal dependencies
between
the
network
past
eventstime
anddependency
the frequency
of future
12
For more general
hazard
functions
thatof
have
an explicit
see Lawless
(2003).
events. For instance, if a particular statistic sh (Gt ; a, b) encodes how much b
interacted with a in the past, then a significantly positive associated param()
eter h would indicate that actors
reciprocate activity. The rate parameters
do not reveal whether responses are positive (more cooperation), neutral, or
negative (more hostility).
()
()
()
rate
ab (Gt ; around
) is dependent
theprecisely,
rate parameters
= (1 , . to
. . , be
k a)
the network
a and b. on
More
the rate()
is specified
and
the
values
of
various
statistics
s
(G
;
a,
b),
h
= 1, . . . , k that characterize
function
h
t
the network around a and b. More precisely,
the rate is specified
to be a
k

18
function
()
ab (t) = ab (Gt ; () ) = exp
sh (Gt ; a, b) .
(7)
k h

h=1
()
()
()
The maximum
deab (t) likelihood
= ab (Gt ; estimates
) = expfor thehrate
sparameters
. h reveal(7)
h (Gt ; a, b)
The exponential
link the
function
formofensures
a positive
event
rate.
pendencies
between
network
pasth=1
events
and the
frequency
of
future
()
TheFor
maximum
likelihood
estimates
for thes rate
h how
reveal
de- b
events.
instance,
if a particular
statistic
(Gt; parameters
a, b) encodes
much
h
The exponential
linkthe
function
form
a positive
event
rate. of future
pendencies
between
network
of ensures
past events
and the
frequency
() parameter
interacted
with a inlikelihood
the past, then
a significantly
positive associated
The For
maximum
estimates
for theshrate
hhow
reveal
de-b
events.
instance, if a particular
statistic
(Gt ;parameters
a, b) encodes
much
()
would
indicate
that
actors
reciprocate
activity.
The
rate
parameters
h
pendencies
between
network
events and
the frequency
futuredo
interacted
with
a in the past,
thenofapast
significantly
positive
associatedofparam()
not
whether
responses
are reciprocate
positive s(more
cooperation),
neutral,
events.
instance,
if athat
particular
statistic
encodes
much bor
eter reveal
h For
would
indicate
actors
activity.
ratehow
parameters
h (G
t ; a, b) The
negative
(more
hostility).
interacted
with
a in theresponses
past, then
significantly
associated
paramdo
not reveal
whether
area positive
(morepositive
cooperation),
neutral,
or
Theh()hazard
already
determines
likelihood
function
for the
eter
would
indicate
that
actors
reciprocatethe
activity.
The rate
parameters
negative
(morefunction
hostility).
occurrence
of whether
events
which
we derive
next. We
to takefunction
into neutral,
account
that
do not
responses
are
positive
(more
cooperation),
or
Thereveal
hazard
function
already
determines
thehave
likelihood
for the
negative
hostility).
we
have a(more
fixed
timewhich
precision
(e. g.,next.
a dayWe
in our
application)
and
occurrence
of events
we derive
haveillustrative
to take into
account that
The
hazard
function
determines
thedyads
likelihood
functionduring
for and
the
we have
a fixed
time
(e.or
g.,
a day
in our
illustrative
application)
that
several
events
onprecision
thealready
same
on
different
may happen
the
occurrence
ofevents
eventsonwhich
we
next.
We dyads
have tomay
into
account
that
that several
the
or on
different
happen
duringlet
the
interval
of timepoints
thatsame
getderive
the
same
timestamp.
Intake
the
following,
the
we
have of
a fixed
precision
g.,same
a to
day
ininterval
our illustrative
application)
interval
timepoints
that
get(e.
the
timestamp.
the following,
let and
theof
expression
time time
interval
always
refer
an
ofInlength
one consisting
that
severaltime
events
on
same
or
on different
may
happen
during
expression
interval
always
refer
to an
of
length
one
consisting
the
timepoints
with
thethe
same
timestamp,
e. interval
g.,dyads
all timepoints
within
onethe
day.
interval
of timepoints
that
thetimestamp,
same timestamp.
the following,
let one
the
of
the timepoints
the get
same
e. g.,that
allIntimepoints
within
Henceforward,
wewith
assume
for
sake of simplicity
the event network
does
expression
time interval
alwaysfor
refer
to of
ansimplicity
interval ofthat
length
day.
Henceforward,
we assume
sake
the one
eventconsisting
network
not
get timepoints
updated during
athe
time
interval.
of
thenot
withduring
same
timestamp,
does
get updated
a time
interval. e. g., all timepoints within one
Let
t denote
a particular
time
interval
and
let,let,
forfor
any
dyad
(i,
j) e D,
day.
Henceforward,
we assume
for
sake of
simplicity
that
event
Let
t denote a particular
time
interval
and
any the
dyad
(i, j)network
D,
does not get updated during a time interval.
E ;interval
ae = i, and
be =let,
j, tfor
t}|, dyad (i, j) D,
Let t denote anparticular
ij (t) = |{etime
e =any
denote the number

of events
int}|,
the interval t. The
nij (t)
= |{e that
E ; happen
ae = i, beon= (i,
j, j)
te =
use
of time
thatthat
we go
from aoncontinuous-time
model t.toThe
a
denote
theintervals
numberimplies
of events
happen
(i, j) in the interval
denote
the number
events
that
happen
on (i,
j) interval
in the interval
t. The
discrete-time
model, ofwhere
thethat
number
events
in
t has amodel
Poisson
use
of time
intervals
implies
we
gooffrom
a continuous-time
to a
use of time intervals
implies
that
we
a continuous-time
model
to a
distribution
with
parameter
thisgoisfrom
also
the
expected
number
ofa events
ij (t);
discrete-time
model,
where the
number
of
events
in interval
t has
Poisson
discrete-time
model,
where
the
number
of
events
in
interval
t
has
a
Poisson
in
this
time
interval.
For
any
dyad
(i,
j)
the
probability
that
exactly
n
(t)in
ij
distribution with parameter ij(t); this is also the expected number of events
distribution
with
parameter
ijinterval
(t); thistisisalso the expected number of events
events
happen
on
(i,
j)
in
the
this time interval. For any dyad (i, j) the probability that exactly nij(t) events
in this time interval. For any dyad (i, j) the probability that exactly nij (t)
happen on (i, j) in the interval
tnis
ij (t)
exp(
events happen on (i, j) in the
interval
t is ij (t))
ij (t)
,
nij (t)!
nij (t)
ij (t)
exp(ij (t))
,
nij (t)!
16
see, e. g., Lawless (2003). If t denotes the first and tN the last timestamp, the
see, e. g., Lawless (2003). If t1 1denotes
the first and
t the last timestamp,
16 density
rate component of the joint probability
is13 13N
the rate component of the joint probability density is
f (E; () ) =
tN

ij (t)nij (t) exp(ij (t))
.
nij (t)!
t=t ijD
(8)
Taking into account that in a given time interval there may be many
dyads on which no event happens, and letting Dact (t) denote the set of active
dyads
at time
t, i.one
e., event
dyads
on which
least
one Eq.
event
we can
13
Note that,
if at most
happened
in anyat
time
interval,
(8) ishappens,
indeed identical
to Eq.
rewrite
this (2008,
as p.163), although it is arranged in a different way.
(2) in Butts

tN

ij (t)nij (t)
()
CORVINUS JOURNAL
OF SOCIOLOGY
AND SOCIAL POLICY 1 (2013)
f (E; ) =
exp
ij (t) .
(9)
nij (t)!
t=t
ijD
1
ijDact (t)
Note that the second product is over all dyads that are active at time t, while
the rate component of the joint probability density is13

see, e. g., Lawless (2003). If t1 denotes the first and tN the last timestamp,
13
tN density
the rate component of the joint probability

ij is
(t)nij (t) exp(ij (t))
()
.
(8)19
f
(E;
)
=
MODELING FREQUENCY
AND TYPE OF INTERACTION IN EVENT NETWORKS
nij (t)!
tN
nij1(t)
t=t

ijD
ij (t)
exp(ij (t))
.
(8)
f (E; () ) =
Takinginto
intoaccount
account
that
ingiven
aij (t)!
given
interval
be many
n
Taking
that
in
a
timetime
interval
therethere
maymay
be many
dyads
t=t1 ijD
dyads on which no event happens, and letting Dact (t) denote the set of active
on
which
no
event
happens,
and
letting
D
(t)
denote
the
set
of
active
dyads
at
Taking into
account
that
interval
there one
mayevent
be many
dyads
at time
t, in
i. e.,a given
dyads time
on which
atactleast
happens, we can
time
t,
i.
e.,
dyads
on
which
at
least
one
event
happens,
we
can
rewrite
this
as
dyads on which
no event
and letting Dact (t) denote the set of active
rewrite
this happens,
as
dyads at time t, i. e., dyads on which
we can
at least one event
happens,

tN
rewrite this as

ij (t)nij (t)

()
f (E; ) =
exp
ij (t) .
(9)

nij (t)!
tN
nij (t)

t=t1ij (t)ijD
ijD
act (t)
exp
f (E; () ) =
ij (t) .
(9)
nij (t)!
t=t
ijD
Note that 1theijD
second
that are active at time t, while
act (t) product is over all dyads
the
sumthat
is over
all dyads,
including
inactive
ones.
If are
it isactive
knownatthat
Note
the second
product
is over
all dyads
that
timeduring
t, while
Note that thea second
productt isthere
over cannot
all dyads
that
are active
at time
t, while
time
interval
be
an
event
on
some
dyads,
then
theseduring
have a
the
sum
is
over
all
dyads,
including
inactive
ones.
If
it
is
known
that
the sum is over
all left
dyads,
ones. Thereby
If it is known
thatfor
during
to be
outincluding
from the inactive
summation.
one can,
instance, address
timet interval
t therebecannot
be an
some dyads,
then
these have to be
a time interval
there where
cannot
an of
event
on event
some on
dyads,
have
situations
the set
actors
changes
over then
time: these
if actor
a is not in the
left
out
from
the
summation.
Thereby
one
can,
for
instance,
address
situations
to be left outnetwork
from the
Thereby
can,a for
instance,
at summation.
time t, then all
dyads one
having
as source
or address
as target
have to be

where
the
set
of
actors
changes
over
time:
if
actor
a
is
not
in
the
network
at
the
(t)
ij
situations where
the
set
of
actors
changes
over
time:
if
actor
a
is
not
in
left out from the summation in the normalizing constant e ijD
.
time t,t, then
then all
all dyads
dyads having
havingaaas
assource
sourceororasastarget
target
have
to
be
left
out
from
network at time
have
to
be

left out fromthe
thesummation
summationininthe
thenormalizing
normalizingconstant
constant e ijD ij (t) .
3.5
3.5
Network Statistics
The general
model outlined so far can be applied to test many hypotheses
Network
Statistics
concerning
theStatistics
interplay between network structure and the frequency and
3.5
Network
The general quality

model outlined
far can
bespecialization
applied to test
manybyhypotheses
of dyadicso
events.
The
is done
using various statistics
concerning the
interplay
between
network
structure
and that
the frequency
and are similar
inThe
Eqs.
(5)
and
(7).
The
particular
statistics
we define
general model outlined so far can be applied
to testbelow
many hypotheses
quality of dyadic
events.
The specialization
is done by
using
various statistics
to
those
of
previous
statistical
models
for
cross-sectional
(Robins
et al. 2007)
concerning
the
interplay
between
network
structure
and
the
frequency
and
in Eqs. (5) and
The particular
statistics
that2005).
we define
below
are similar
or (7).
longitudinal
networks
(Snijders
The
statistics
are illustrated in
quality
of
dyadic
events.
The
specialization
is
done
by
using
various
statistics
to those of previous
Table 1.statistical models for cross-sectional (Robins et al. 2007)
in Eqs.
(5) and(Snijders
(7).here.]
The 2005).
particular
statistics
thatare
weillustrated
define below
or longitudinal
networks
The
statistics
in are similar to
[Table 1 about
those
of
previous
statistical
models
for
cross-sectional
(Robins
et al. 2007)
Table 1.
Note that some of these statistics are used to test the hypotheses
onor
networks
(Snijders
2005).
The
statistics
are
illustrated
in
Table
[Table 1 longitudinal
about
here.]
structural balance theory, while others mostly serve to control for certain 1.
Note thattrivial
some regularities
of these
arestatistics
used A
to control
test used
thestatistic
hypotheses
Note
that
somestatistics
of g.,
these
are
to testthat
theon
hypotheses
(e.
inertia).
always
has toon
structural balance
theory,
whiletheory,
others while
mostlyothers
serve mostly
to control
for to
certain
structural
balance
serve
control
for
certain
13
Note that,
if at
most one event
happened
any time
interval,
Eq.has
istoindeed identical
trivial regularities
(e. g.,
inertia).
A inertia).
control
statistic
that
always
trivial
regularities
(e. g.,
Aincontrol
statistic
that(8)always
has to be
13
to Eq. (2) in Butts (2008, p.163), although it is arranged in a different way.

meaningful
results
is Eq.
the (8)constant
statistic, defined by
Note that, iftaken
at mostto
oneobtain
event happened
in any time
interval,
is indeed identical
to Eq. (2) inconstant
Butts (2008, p.163),
although
it isThe
arranged
in a different
(G ; a,
b) = 1.
constant
statisticway.
just controls for possible
t
deviation from zero of the average 17

event weight (similar argument for the
rate). The function constant
(G
;
a,
b), as well as the ones whose definition
t
17
follows, correspond to the statistics sh (Gt; a, b) in Eqs. (5) as well as (7). First
we propose some further statistics for use in Eq. (5) to specify the distribution
of the conditional event type; subsequently statistics for use in Eq. (7) are
proposed, specifying the event frequency.
20
Table 1: Illustration of some statistics explaining the tie from actor a towards
actor b at time t. A sign indicates that there is a version for positive and
for negative weights. The symmetric positve/negative weight on a dyad i, j
is defined by wt,sy (i, j) = wt (i, j) + wt (j, i). Dashed lines indicate negative
ties.
name
inertia
reciprocity
friendOfFriend
friendOfEnemy
enemyOfFriend
enemyOfEnemy
formula
wt (a, b)
wt (b, a)

iA
wt,+sy (a, i) wt,+sy (i, b)
iA wt,sy (a, i)
+
iA wt,sy (a, i)
iA wt,sy (a, i)
b depends on
a
b
a
b
i2
i1
b
wt,+sy (i, b)
i2
i1
wt,sy (i, b)
i2
i1
wt,sy (i, b)
i2
i1
i1
a
i2
activitySource
iA
i3
wt (a, i)
i1
a
popularityTarget
iA
wt (i, b)
i2
b
i3
The most simple model would assume that actors just continue to act in
the way they did in the past. For instance, if actor a initiated many hostilities
targeted at actor b, the dyad (a, b) is likely to be a hostile one in the future.
This effect is controlled for by the two statistics capturing the inertia of
1
positive respectively negative events, defined by
inertia+(Gt; a, b) = wt+(a, b) inertia(Gt; a, b) = wt (a, b) .
The type parameter associated with inertia+ is expected to be positive
(events from a to b are more cooperative if a cooperated with b in the past)
21
and the parameter associated with inertia is expected to be negative

(events from a to b are more hostile if a fought b in the past).
A non-trivial, but very reasonable, network effect is that actors reciprocate,
i. e., actor a adapts its events towards actor b in accordance to how b treated
a in the past. This is captured for positive and negative events by the two
statistics
reciprocity+(Gt; a, b)= wt +(b, a) reciprocity(Gt; a, b)= wt(b, a).
A positive estimate for the type parameter associated with reciprocity+

would imply that actors reward cooperation; a negative estimate for the type
parameter associated with reciprocity would imply that actors retaliate
when receiving hostilities.
Structural balance theory predicts that the relation of two actors a and b is
dependent on whether they have common friends or foes. In the following we
take it as an indication of friendship if two actors cooperate (in either direction)
and as an indicator that they are enemies if they exchange hostilities. Let
wt+,sy(i, j)= wt+(i, j)+ wt+(j, i) denote the symmetrized positive weight on adyad
(i, j) and let wt,sy(i, j)= wt(i, j)+ wt(j, i) denote the symmetrized negative
weight. To test the four hypotheses of structural balance theory, we define
four statistics measuring to what extent b is with respect to a a friend of a
friend,
a friend of an enemy, an enemy of a friend, and an enemy of an enemy,
respect to a a friend of a friend, a friend of an enemy, an enemy of a friend,
respectively.
and an enemy of an enemy, respectively.
friendOfFriend(Gt ; a, b) =

iA
friendOfEnemy(Gt ; a, b) =

iA
enemyOfFriend(Gt ; a, b) =

iA
enemyOfEnemy(Gt ; a, b) =

iA
wt,+sy (a, i) wt,+sy (i, b)

wt,sy (a, i) wt,+sy (i, b)
wt,+sy (a, i) wt,sy (i, b)
wt,sy (a, i) wt,sy (i, b)
The
thethe
assumption
that
(third,
fourth,
etc.)
Thesquare
squareroot
rootexpresses
expresses
assumption
thata second
a second
(third,
fourth,
etc.)
actor
who
is
an
enemy
of
a
and
a
friend
of
b
has
a
decreasing
marginal
effect
actor who is an enemy of a and a friend of b has a decreasing marginal effect
on how strongly a perceives b as a friend of an enemy (compare Snijders
on
how strongly a perceives b as a friend of an enemy (compare Snijders et
et al. 2010). According to the hypotheses developed before, the type parameters associated with friendOfFriend and enemyOfEnemy are predicted to
be positive and those associated with friendOfEnemy and enemyOfFriend
are predicted to be negative.14CORVINUS JOURNAL OF SOCIOLOGY AND SOCIAL POLICY 1 (2013)
As a matter of fact, some actors are more active than others, some do
rather initiate hostile events (aggressive actors), and others are more cooperative. Likewise, some actors are typical targets of hostilities, while others
22
al. 2010). According to the hypotheses developed before, the type parameters
associated with friendOfFriend and enemyOfEnemy are predicted to be
positive and those associated with friendOfEnemy and enemyOfFriend
are predicted to be negative.14
As a matter of fact, some actors are more active than others, some do rather
initiate hostile events (aggressive actors), and others are more cooperative.
Likewise, some actors are typical targets of hostilities, while others tend to
experience cooperation. To control for such differences in actors network
positions or roles, we introduce a set of statistics dependent on the degree
of actors. These statistics vary in three dimensions: (1) outdegree (activity)
vs. indegree (popularity); (2) positive vs. negative weight; and, (3) whether
we want to analyze the influence of these degree statistics on the initiator of
events (source) or on the addressee of events (target). Together we obtain
we obtain
eight
differenttwo
statistics;
them are
defined
below,are
theimplied
otheight
different
statistics;
of themtwo
areof
defined
below,
the others
ersanalogy.
are implied
analogy.
Thesource
activity
of the
actor
with respect
to is
by
Theby
activity
of the
actor
withsource
respect
to positive
events
positive events is defined to be
defined to be

activitySource+ (Gt ; a, b) =
wt+ (a, i) ,
iAare defined below, the othwe obtain eight different statistics; two of them
ers
are
implied
by
analogy.
The
activity
of
the
source
with respect
to is
whilethethepopularity
popularityofofthethetarget
targetactor
actorwith
withrespect
respectactor
negative
events
while
totonegative
events
is
positive events is defined to be
defined
defined to
to be
be

w+ (a, i) ,
activitySource+
(G
t ; a, b) =
t (i, b) .
popularityTarget
(G
w
t ; a, b) =
t
iA
iA
For instance,
a positive
estimate
for the
type parameter
associated
with
while
the popularity
of the
target for
actor
to negative
events
is
For instance,
a positive
estimate
thewith
typerespect
parameter
associated
with
+
activitySource
would
imply
that
actors
a
who
initiate
a
lot
of
cooperation
defined
to
be
+
activitySource would imply that actors a who initiate a lot of cooperation
with (rather than fight) the
(towards
i),
totocooperate
(towards any
any actor
actor
i),are
aremore
morelikely
likely
cooperate
(rather than fight)
popularityTarget
(Gthe
b) =event.
wwith
t ; a,next
t (i, b) .
particular
actor
b
who
is
the
target
of
the particular actor b who is the target of the next
event.
iA
Constant
or or
slowly
changing
actor
andand
dyaddyad
characteristics
may may
additionally
Constant
slowly
changing
actor
characteristics
addior
alternatively
explain
the
behavior
of
political
actors.
The
following
statistics
For
instance,
a
positive
estimate
for
the
type
parameter
associated
with
tionally or alternatively explain the behavior of political actors. The
fol+
activitySource
would
that actors
a who
initiate
a lot are
of cooperation
characterize
a dyad
(a, b) imply
by various
covariates.
The
statistics
takenstatisfrom a
lowing statistics
characterize
a dyad
(a, b) by
various
covariates.
The
(towards
any actor
i),
are moredyad
likelyanalysis
toforcooperate
(rather
than fight)
tics
are model
taken
from
a standard
model
non-directed
dyad
analysis
used
standard
for non-directed
used inwith
the study
of
international
thethe
particular
who is the
target
ofbinary
the next
event.
in
study
ofactor
international
relations
and
Russett
2005).(GThe
relations
(Oneal
andbRussett
2005).
The(Oneal
variable
allies
;t a, bib) is
actor
dyad
may
naryConstant
; a, b)atisleast
one one
ifand
and
only characteristics
if
a and
b have
at addileast
thave
one
if variable
and onlyorallies(G
if slowly
a and bchanging
common
joint
alliance
membership.
tionally
or alternatively
explain
the behavior
of political
actors. that
The folone common
alliance
membership.
Hypothesis
H5 predicts
the
Hypothesis
H5joint
predicts
that
the conditional
type
parameter
associated
with
lowing statistics
a dyad (a, b)
by various
statisconditional
type characterize
parameter associated
with
allies covariates.
is positive.The
Statistic
tics are taken tfrom
model forratio
non-directed
dyad analysis
lnCapRatio(G
; a, b) aisstandard
the logarithmized
of the capability
score ofused
the
14
Cthe
learly,
by symmetrizing
the friend
and enemy
relations
weRussett
lose some
information,
since
15
in
study
ofactor
international
relations
(Oneal
2005).
bimore
powerful
divided
by
the
score
of
theand
less
powerful
actor.The
Prethe direction of ties might cause different behavior. For instance, it might be possible that
nary
variable
allies(G
;
a,
b)
is
one
if
and
only
if
a
and
b
have
at
least
vious
have
thatfriends
a preponderance
national
capabilities
is
actorsresults
fight those
whoshown
attackt their
but are indiferentof
to those
who are
attacked by their
one
common
joint
alliance
membership.
Hypothesis
predicts
that
the
5the
related
to
lessdistinguish
conflict
within
the dyad
(Hegre
2008). H
binary
statistic
friends.
If we
all
combinations
of signs
and directions
ofThe
two
ties that
indirectly
conditional
parameter
associated
with
allies
isrefinement
Statistic
minorPowers(G
a,
b) is16one
if
and
if neither
aThis
nor
bpositive.
is a major
power.not
relate a with type
b, we
statistics
for only
structural
balance.
is, however,
t ; obtain
lnCapRatio(G
is the
ratioinofthe
thedyad
capability
score
of the
considered
this
paper.
t ; a,
The
polity in
score
ofb)the
less logarithmized
democratic actor
gives the
value
of
15
more
powerful
actor
divided
by
the
score
of
the
less
powerful
actor.
Prethe statistic polityWeakLink(Gt ; a, b). Since previous results have reported
vious
results
haveshow
shown
that a to
preponderance
national
capabilities
is
that democracies
a tendency
not fight
eachofother
(Russett
and Oneal
CORVINUS
JOURNAL OF SOCIOLOGY
AND SOCIAL
POLICY
1 (2013)
related
to expect
less conflict
within
dyad (Hegre
2008).with
The
binary statistic
2001), we
a positive
typethe
parameter
associated
polityWeakLink.
minorPowers(G
is one if and tonly
noronly
b is ifa amajor
t ; a, b)
The binary statistic
contiguity(G
; a, b)ifisneither
one if aand
and bpower.
share
The
polity
scoreorofa the
democratic
actor
the dyad
the value
of
a land
border
sealess
border
less than
400inmiles
long gives
and the
variable
23
allies is positive. Statistic lnCapRatio(Gt; a, b) is the logarithmized ratio
of the capability score of the more powerful actor divided by the score of the
less powerful actor.15 Previous results have shown that a preponderance of
national capabilities is related to less conflict within the dyad (Hegre 2008).
The binary statistic minorPowers(Gt; a, b) is one if and only if neither a nor
b is a major power. The polity score of the less democratic actor in the dyad
gives the value of the statistic polityWeakLink(Gt; a, b). Since previous
results have reported that democracies show a tendency to not fight each other
(Russett and Oneal 2001), we expect a positive type parameter associated
with polityWeakLink. The binary statistic contiguity(Gt; a, b) is one if
and only if a and b share a land border or a sea border less than 400 miles long
and the variable lnDistance(Gt; a, b) is the logarithmized distance between
the capitals of a and b. The statistic lnTrade(Gt; a, b) is the logarithm of the
average of trade going from a to b and from b to a and lnJointIGO(Gt; a,
b) is the logarithmized number of joint memberships in intergovernmental
organizations of a and b.
In principle, the statistics defined above can also be taken for the specification
of the event rate in Eq. (7). However, structural balance theory makes no
predictions about whether actors interact more or less frequently with, say, the
friends of their enemies; it is just predicted that the tie is likely to be a hostile
one. For this reason we argue that the rate is better specified by statistics that
ignore the sign of previous interaction as defined in the following, leaving out
the argument (Gt; a, b).
inertia
=

reciprocity
=

triangle
=

activitySource
=

inertia+ + inertia

reciprocity+ + reciprocity

friendOfFriend + friendOfEnemy
+enemyOfFriend + enemyOfEnemy

activitySource+ + activitySource
The definition of the statistics activityTarget, popularitySource,

and popularityTarget is analogous to activitySource. The rate
parameters associated with these statistics reveal dependencies between
particular aspects of the network of past events and future event frequencies.
For instance, a positive rate parameter associated with reciprocity would
imply that if b interacted a lot with a (positively or negatively) then the
current event frequency on the dyad (a, b) is typically higher than without
this precondition; a negative rate parameter points to a decreased event rate.
15 The capability score of a country is a composite measure taking into account, among other
things, demographic, economic, and military strength.
24
3.6 Parameter Estimation

The maximization of the log-likelihood of the type parameters, see Eq. (4),
is very efficient and can be done by simply solving the usual OLS system of
equations. The maximization of the log-likelihood of the rate parameters (see
Eq. (8)) can be done numerically, e. g., by using the established NewtonRaphson algorithm. The computation of the rate parameters is much more
time-consuming since the normalization constant in Eq. (9) contains a sum
over all pairs of actors. The computation can be speeded up by approximating
this sum via sampling over a sffiucient number of pairs; compare Butts
(2008). The results in this paper, however, are computed without such an
approximation.
4 Results and Discussion

In this section we report and discuss the estimated parameters on the Gulf
network, restricted to state actors. The half life parameter T1/2 is set to 30 days.
4.1 Conditional Type Parameters

Table 2 shows the estimated conditional type parameters for three models,
the first built from the 16 signed network effects, the second built from the
covariate statistics, and finally a combined model which includes network and
covariate effects. The log-likelihood of the null model M0 (no effects except
the constant offset) is 143,964, its BIC is 287,940, its AIC is 287,930. The
maximized likelihood and information criteria for the other models are listed
in Table 2. With respect to the information criteria the network model is better
than the covariate model, but the combination of networks and covariates
yields a strong further model improvement.16
16 Note that the models including covariates are estimated using a slightly smaller set of events
due to dropping events with missing values; thus the network-only model can only be
compared with the null model and the covariate model only with the joint model, according
to the information criteria.
25
Table 2: Estimated conditional type parameters and
Table 2 Estimated
conditional
type
prameters and standard errors (in brackets)
standard
errors (in
brackets).
statistic
inertia+
inertia
reciprocity+
reciprocity
friendOfFriend
enemyOfFriend
friendOfEnemy
enemyOfEnemy
activitySource+
activitySource
activityTarget+
activityTarget
popularitySource+
popularitySource
popularityTarget+
popularityTarget
lnCapRatio
allies
polityWeakLink
minorPowers
lnTrade
contiguity
lnDistance
lnJointIGO
constant
#events
log-likelihood
ll-ratio to M0
#params
BIC
AIC
event network model covariate model

0.214 (0.012)*
-0.085 (0.003)*
0.124 (0.014)*
-0.082 (0.004)*
0.246 (0.027)*
-0.206 (0.014)*
-0.224 (0.015)*
0.113 (0.008)*
0.051 (0.003)*
-0.008 (0.001)*
0.040 (0.004)*
0.002 (0.002)
-0.008 (0.005)
0.003 (0.001)*
-0.020 (0.005)*
0.004 (0.001)*
0.002 (0.001)*
0.118 (0.003)*
3.2E4 (1.8E4 )
0.097 (0.003)*
0.028 (0.001)*
-0.093 (0.003)*
0.011 (0.001)*
-0.097 (0.003)*
-0.082 (0.001)*
-0.017 (0.008)*
217 479
200 886
-136 669
-137 844
7 295
6 120
17
9
273 547
275 798
273 372
275 706
combined model
0.192 (0.012)*
-0.071 (0.003)*
0.075 (0.014)*
-0.052 (0.004)*
0.138 (0.027)*
-0.119 (0.015)*
-0.137 (0.015)*
0.057 (0.008)*
0.009 (0.004)*
0.001 (0.002)
-0.006 (0.004)
0.013 (0.002)*
0.023 (0.005)*
-0.007 (0.002)*
0.005 (0.005)
-0.005 (0.002)*
-0.007 (0.001)*
0.106 (0.003)*
-0.001 (1.9E4 )*
0.042 (0.004)*
0.017 (0.001)*
-0.060 (0.003)*
0.013 (0.001)*
-0.076 (0.003)*
-0.002 (0.008)
200 886
-134 441
9 523
25
269 187
268 932
The estimated parameters for the network-only model are reported in the
first column in Table 2. The hypotheses
derived from structural balance theory
1
(H1 to H4) are fully supported by our analysis. The parameter associated with
friendOfFriend is significantly positive, supporting H1. This implies that
actors have a tendency to cooperate with the friends of their friends: the more
b is connected by past cooperative events to a third actor c who, in turn, is
connected by past cooperative events to a, the higher (i. e., more cooperative)
is the average weight of future events from a to b. The parameter associated
26
with friendOfEnemy and with enemyOfFriend are significantly negative,

supporting H2 and H3 respectively. This implies that actors have a tendency
to fight the friends of their enemies as well as the enemies of their friends.
Finally, the parameter associated with enemyOfEnemy is significantly
positive as predicted by hypothesis H4. This implies that actors in the Gulf
conflict have a tendency to cooperate with the enemies of their enemies.
The second column of Table 2 shows the parameters estimated for a
covariates-only model. The main purpose of the politically relevant
covariates is checking whether the network effects reported above are merely
side-effects of certain actor or dyad characteristics. We do not wish to debate
here the extent to which the conclusions obtained from this data analysis
may be regarded as generalizable tests of IR theories: by restricting the data
to a focused region, the Gulf, we have selected actors which have specific
characteristics with respect to democracy, trade, geographic closeness, etc.
The parameter associated with allies (binary variable encoding whether
the two actors have at least one alliance or not) is significantly positive,
supporting hypothesis H5. This implies that events tend to be more cooperative
among allied actors. As we will see in the following, this result is robust to the
inclusion of network effects and also to the exclusion of all other covariate
statistics. Other parameters, which are not related to our hypotheses, are not
discussed here.
Finally, we estimate a model in which network effects and covariate effects
are combined (third column in Table 2). It is remarkable that controlling
for covariates does not change the signs of any parameter associated with
inertia (positive and negative), reciprocity (positive and negative), and
the four structural balance effects friendOfFriend, friendOfEnemy,
enemyOfFriend, and enemyOfEnemy. In particular, the support for of
structural balance theory (hypotheses H1 to H4) is robust and not just the
side-effect of actor characteristics. There are some changes of parameters
associated with covariates that are not related to our hypotheses. The effect
of alliances on the conditional event type, however, remains significantly
positive; i. e., allies interact more cooperatively if they interact. This gives
further evidence that the validation of hypothesis H5 is robust and not just the
result of uncontrolled network dependencies.
To further test the robustness of the effect of alliances on the conditional
event type we fit a bivariate model that contains only the alliance statistic
as predictor (and a constant offset). In this model we obtain a significantly
positive type parameter equal to 0.112 (0.002) for allies and a constant of
0.132 (0.001). In a different model built from the 16 network statistics (as
above), the constant, and the alliance statistic (without any other covariate)
27
we estimate a significantly positive type parameter 0.077 (0.002) for

allies. Thus, the effect of alliances on the conditional event type is indeed
consistently positive.
4.2 Event Frequency

Table 3 shows the estimated rate parameters for three models, the first built
from the seven unsigned network effects, the second built from the covariate
statistics, and finally the joint model including both network and covariate
effects. The log-likelihood of the null model M0 (no effects except the
constant offset) is 1,629,098, its BIC is 3,258,208 and its AIC is 3,258,198.
The maximized likelihood and information criteria for the other models
are listed in Table 3. The information criteria show that here the covariate
model is better than the network model, but the combination of networks and
covariates yields a strong further model improvement.
Table 3 Event Table
rate parameters
and parameters
standard errors
3: Event rate
and standard errors.
statistic
event network model
inertia
-0.114 (0.002)
reciprocity
-0.090 (0.003)
triangle
0.506 (0.002)
activitySource
0.202 (0.001)
activityTarget
0.168 (0.001)
popularitySource
0.094 (0.001)
popularityTarget
0.131 (0.001)
lnCapRatio
allies
polityWeakLink
minorPowers
lnTrade
contiguity
lnDistance
lnJointIGO
constant
-6.774 (0.002)
#events
217, 479
log-likelihood
-1 302 411
ll-ratio to M0
326 687
#params
8
BIC
2 604 920
AIC
2 604 838
covariate model
-0.289 (0.002)
0.064 (0.006)
-0.137 (0.001)
-2.726 (0.007)
0.062 (0.001)
1.362 (0.006)
-0.287 (0.002)
1.344 (0.005)
-6.964 (0.017)
200, 886
-1 271 416
357 682
9
2 542 941
2 542 850
combined model
1.9E4 (0.002)
0.042 (0.003)
0.348 (0.003)
0.161 (0.001)
0.118 (0.001)
0.073 (0.001)
0.119 (0.001)
-0.225 (0.002)
-0.223 (0.006)
-0.122 (0.001)
-1.970 (0.007)
0.142 (0.001)
1.310 (0.007)
-0.343 (0.002)
1.313 (0.005)
-7.530 (0.016)
200, 886
-1 029 255
599 843
16
2 058 705
2 058 542
28
The rate parameters estimated for the network-only model are reported
in the first column in Table 3. The parameter associated with triangle is
significantly positive, supporting hypothesis H6. This implies that activity is
transitively closed: the more b is connected by past (cooperative or hostile)
events to a third actor c who, in turn, is connected by past events to a, the
higher is the frequency of future events from a to b.
The rate parameters resulting from the covariate-only model are reported
in the second column in Table 3. The rate parameter associated with allies
is significantly positive. Thus, allied actors interact more frequently than nonallied actors, giving support to hypothesis H7. It is interesting to consider
this result together with the conditional type parameter associated with the
allies statistic (see Table 2): allies interact more, and if they interact, their
interaction is rather friendly. This might nevertheless lead to the observation
that the absolute (rather than conditional) probability of conflictive interaction
is higher among allies (compare Bremer 1992).
Finally, we report the event frequency parameters in the combined model
(third column in Table 3). The sign of the rate parameter associated with the
triangle statistic did not change when covariates are included. In particular,
the validation of hypothesis H6 (transitivity of activity) is robust to the
inclusion of covariates. There is only one change in the signs of the covariate
parameters when the model is augmented by network effects: the parameter
associated with allies changes from significantly positive to significantly
negative. Thus, the finding allies interact more, related to hypothesis H7, is
sensitive to whether we control for past interaction on the dyad (inertia and
reciprocity statistics), activity and popularity of the actors, and indirect
relations (triangle statistic). Controlling for these network dependencies,
allies do not interact with higher frequency.
For comparison, we estimate a bivariate event frequency model that
includes only the allies statistic and, as always, the constant. The rate
parameter associated with the allies statistic in this model is 1.600 (0.005)
and for constant we get a parameter of 6.700 (0.002). Thus, while allies
consistently interact more cooperatively if they interact, the finding that allies
interact more frequently is only a side-effect of ignored variables. Indeed,
without any control variables we obtain the result that allies interact about
e1.6 =4.95 times as much as non-allies. Controlling for the effects of other
covariates, this factor is reduced to (approximately) e0.064 =1.07. Finally,
controlling for covariates and network effects, a joint alliance membership
even diminishes the rate of interaction by a factor of e0.223 =0.8.
It is noteworthy that some covariates influence the frequency of interaction
while others rather have an effect on the conditional event type. For instance,
29
in contrast to alliances, geographic adjacency (operationalized by the

contiguity statistic) consistently increases the frequency of interaction,
independent of whether we control for network dependencies or not. The
effect of contiguity on the conditional event type is consistently negative
so thatconsidering their frequent interactionadjacent countries indeed are
dangerous dyads (Bremer 1992). A different behavior can be seen for the
lnCapRatio statistic (logarithmized capability ratio). While its effect on the
event rate is consistently negative (unequal capability decreases interaction
frequency), its effect on the conditional event type changes when we control
for network dependencies: ignoring these leads to a seemingly positive effect
(more cooperation when the capability ratio increases). In contrast, controlling
for network effects yields a negative association between capability ratio
and the conditional event type. The last-mentioned findings also emphasize
the need to control for statistical dependency among dyadic observations:
ignoring these might lead to spurious associations that vanish or even become
reversed if network effects are taken into account.
5 Conclusion
We propose a general model for the dynamics of networks that are given as
sequences of dyadic, typed events. Our model exploits the time information
of event data and can test and control for potentially complex dependencies
among dyadic observations. The most distinctive feature of our model is that
we decompose the joint probability of typed events into two components, the
first modeling the frequency of events of any type and the second modeling
the conditional type of events, given that events happen. Thus, our model is an
alternative to previously proposed models that estimate the rate of events of
different types separately. This distinction indeed turned out to be crucial when
tackling substantive research questions. For instance, previous work showed
that enemies of enemies have a higher probability of engaging in conflict and
in cooperation (Maoz et al. 2007), which simultaneously rejects and supports
structural balance theory. In contrast, we showed that the conditional event
type among enemies of enemies is pushed towards cooperation, which clearly
supports SBT. Thus, our main methodological conclusion is the following:
when we want to test a typical hypothesis in political network analysis, such
as does condition X lead to more or less conflict, we have to clarify first
whether we mean the absolute level of conflictive interaction or rather the
tendency of conflict vs. cooperation, given that interaction occurs.
The separation of the joint probability density into its rate component
30
and conditional type component provides deeper insight into how specific
predictors increase or decrease the probabilities of events of a given type. For
instance, alliance ties increase the probability of dyadic conflict in a bivariate
model without any control variables. This result is refined by our model:
alliances push the conditional event type towards cooperationindependent
of which control variables we use. Thus, given that two actors do interact,
their interaction is more friendly and less conflictual if they are allies. On
the other hand, the finding that allies interact more frequently is only due to
uncontrolled variables; if we control for network effects and other covariates,
alliances even decrease the event frequency. Thus, alliance ties co-occur with
frequent interaction but, apparently, they do not cause it.
We emphasize that the concrete empirical findings that have been included
for illustration in this paper should be treated with caution for at least two
reasons. First, the KEDS data includes dyadic events only if they have
been reported in the news. Thus, issues such as news bias or media fatigue
might influence the results in a systematic way. This is, strictly spoken, not
a problem with our model but rather of the particular data. When more and
more event data become available the same model can be applied to data sets
that (hopefully) suffer less from news bias. Second, the data we analyzed
included events that are restricted to the Gulf regionthereby selecting actors
that have specific characteristics with respect to democracy scores, geographic
proximity, trade relations, capability, etc. An issue for future work is to repeat
the analyses on global datasets that are not restricted to specific regions, e. g.,
King and Lowe (2003); OBrian (2010). Otherwise, it remains unclear which
of the observed patterns are special to the Gulf conflict and which ones are
universal in the dynamics of international relations.
The model itself leads also to several possibilities for extension and
refinement. Obviously, the decomposition of the probability density of
typed events into a rate component and a conditional type component, as in
Eq. (2), is not restricted to one-dimensional real-valued event types. More
general event types (such as binominal, multinomial, ordered multinomial, or
multidimensional types) would just require the adaptation of the specification
of the distribution for the conditional event type and the adaptation of the
estimation procedure for the type parameters. Developing and applying
models for more general event types is a promising area for future research.
31
References
Box-Steffensmeier, Janet M.-Bradford S. Jones (1997), Time is of the essence: Event
history models in political science, American Journal of Political Science Vol. 41,
No. 4, pp. 14141461.
Brandes, Ulrik -Daniel Fleischer -Jrgen Lerner (2006), Summarizing dynamic
bipolar conflict structures, IEEE Transactions on Visualization and Computer
Graphics Vol. 12, No. 6, pp. 14861499.
Brandes, Ulrik -Jrgen Lerner (2008), Visualization of conflict networks, in:
Kauffmann, Mayeul, ed., Building and Using Datasets on Armed Conflicts, IOS
Press, pp. 169188.
Brandes, Ulrik -Jrgen Lerner -Tom A. B. Snijders (2009), Networks evolving
step by step: Statistical analysis of dyadic event data, in Proc. 2009 Intl. Conf.
Advances in Social Network Analysis and Mining, IEEE Computer Society, pp.
200205.
Bremer, Stuart A. (1992), Dangerous dyads: Conditions affecting the likelihood of
interstate war, 18161965, Journal of Conflict Resolution Vol. 36, No. 2, pp.
309341.
Bueno de Mesquita, Bruce (1981), The War Trap, Yale University Press.
Butts, Carter T. (2008), A relational event framework for social action, Sociological
Methodology Vol. 38, No. 1, pp. 155200.
Cartwright, Dorwin -Frank Harary (1956), Structural balance: A generalization of
Heiders theory, The Psychological Review Vol. 63, No. 5, pp. 277293.
Crescenzi, Mark J. C. (2007), Reputation and interstate conflict, American Journal
of Political Science Vol. 51, No. 2, pp. 382396.
De Nooy, Wouter (2008), Signs over time: Statistical and visual analysis of a
longitudinal signed network, Journal of Social Structure Vol. 9, No. 1.
De Nooy, Wouter (2011), Networks of action and events over time. A multilevel
discrete-time event history model for longitudinal network data, Social Networks
Vol. 33, No. 1, pp. 3140.
Goldstein, Joshua S. (1992), A conflict-cooperation scale for WEIS international
events data, Journal of Conflict Resolution Vol. 36, No. 2, pp. 369385.
Goldstein, Joshua S. -Jon C. Pevehouse -Deborah J. Gerner -Shibley Telhami (2001),
Reciprocity, triangularity, and cooperation in the Middle East, 1979-97, Journal
of Conflict Resolution Vol. 45, No. 5, pp. 594620.
Hafner-Burton, Emilie M. -Alexander H. Montgomery (2006), Power positions:
International organizations, social networks, and conflict, Journal of Conflict
Resolution Vol. 50, No. 1, pp. 327.
Hegre, Havard (2008), Gravitating toward war: Preponderance may pacify, but power
kills, Journal of Conflict Resolution Vol. 52, No. 4, pp. 566589.
Heider, Fritz (1946), Attitudes and cognitive organization, The Journal of
Psychology Vo. 21, pp. 107112.
32
Hoff, Peter D. -Michael D. Ward (2004), Modeling dependencies in international

relations networks, Political Analysis Vol. 12, No. 2, pp. 160175.
KEDS (2012), http://eventdata.psu.edu/data.dir/gulf.html, Web-page, last accessed
on September 20th, 2012.
Kimball, Anessa L. (2006), Alliance formation and conflict initiation: The missing
link, Journal of Peace Research Vol. 43, No. 4, pp. 371389.
King, Gary Will Lowe (2003), An automated information extraction tool for
international conflict data with performance as good as human coders: A rare
events evaluation design, International Organizations Vol. 57, pp. 617642.
Lawless, Jerald F. (2003), Statistical Models and Methods for Lifetime Data, Wiley.
Maoz, Zeev (2009), The effects of strategic and economic interdependence on
international conflict across levels of analysis, American Journal of Political
Science Vol. 53, No. 1, pp. 223240.
Maoz, Zeev -Lesley G. Terris -Ranan D. Kuperman -Ilan Talmud (2007), What is
the enemy of my enemy? Causes and consequences of imbalanced international
relations, 18162001, Journal of Politics Vol. 69, No. 1, pp. 100115.
McClelland, Charles A. (1976), World Event/Interaction Survey Codebook (ICPSR
5211) .
OBrian, Sean P. (2010), Crisis early warning and decision support: Contemporary
approaches and thoughts on future research, International Studies Review Vol.
12, pp. 87104.
Oneal, John R. Bruce Russett (2005), Rule of three, let it be? When more really
is better, Conflict Management and Peace Science Vol. 22, No. 4, pp. 293310.
Robins, Garry -Tom A. B. Snijders -Peng Wang -Mark Handcock -Philippa Pattison
(2007), Recent developments in exponential random graph (p*) models for social
networks, Social Networks Vol. 29, No. 2, pp. 192215.
Russett, BruceJohn R. Oneal (2001), Triangulating peace -democracy,
interdependence, and international organizations, W.W. Norton & Company.
Schrodt, Phillip A. -Shannon G. Davis -Judith L. Weddle (1994), Political science:
KEDSa program for the machine coding of event data, Social Science Computer
Review Vol. 12, No. 3, pp. 561588.
Snijders, Tom A. B. (2005), Models for longitudinal network data, in: Carrington,
Peter J. -John Scott -Stanley Wasserman, eds., Models and Methods in Social
Network Analysis, Cambridge University Press.
Snijders, Tom A. B. Gerhard G. van de Bunt Christian E. G. Steglich (2010),
Introduction to actor-based models for network dynamics, Social Networks Vo.
32, pp. 4460.
Stadtfeld, Christoph (2010), Who Communicates With Whom? Measuring
Communication Choices on Social Media Sites, in Proc. IEEE 2nd Intl. Conf.
Social Computation (SocialCom), pp. 564569.
Young, G. A. and Smith, R. L. (2005), Essentials of Statistical Inference, Cambridge
University Press.

60 268 1 PB

Transféré par

Informations du document

Titre original

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

60 268 1 PB

Transféré par

Droits d'auteur :

Formats disponibles

CORVINUS JOURNAL OF SOCIOLOGY AND SOCIAL POLICY Vol.

MODELING FREQUENCY AND TYPE OF

CORVINUS JOURNAL OF SOCIOLOGY AND SOCIAL POLICY 1 (2013)

MODELING FREQUENCY AND TYPE OF INTERACTION IN EVENT NETWORKS

2 Political Network Analysis

MODELING FREQUENCY AND TYPE OF INTERACTION IN EVENT NETWORKS

The analysis of event data (alternatively referred to as time-to-event

Subjects and objects in news wire texts are interpreted as referring to

MODELING FREQUENCY AND TYPE OF INTERACTION IN EVENT NETWORKS

EXTEND MIL AID

Extending military aid is considered a highly cooperative action (weight

While the main contribution of this work is methodological, we present and

Specification of Conditional Network Event

we assume that each event ei is only dependent on events that happened

f (E; ) = f (e1 |Ge1 ; ) f (e2 |Ge2 ; ) . . . f (eN |GeN ; ) .

f (E; ) = f (e1 |Ge1 ; ) f (e2 |Ge2 ; ) . . . f (eN |GeN ; ) .

3.2 Input Data

we e R is the type, coding the quality of the event e; and

Explanatory Variable: The Network of Past Interaction

Given a sequence of events E = (e1 , . . . , eN ) and a specific timepoint t (deVariable:

11 Note that w also maps to the non-negative

and the function w : D R is defined by

Thus, the value w (i, j) is defined as a function of all weights of events

3.4 Dependent Variable: The Next Event

dependencies between characteristics of the network and later observed event

CORVINUS JOURNAL OF SOCIOLOGY AND SOCIAL POLICY 1 (2013)

the network around

denote the number

the rate component of the joint probability density is13

The general quality

to Eq. (2) in Butts (2008, p.163), although it is arranged in a different way.

deviation from zero of the average 17

CORVINUS JOURNAL OF SOCIOLOGY AND SOCIAL POLICY 1 (2013)

wt,+sy (a, i) wt,+sy (i, b)

MODELING FREQUENCY AND TYPE OF INTERACTION IN EVENT NETWORKS

and the parameter associated with inertia is expected to be negative

A positive estimate for the type parameter associated with reciprocity+

wt,+sy (a, i) wt,+sy (i, b)

MODELING FREQUENCY AND TYPE OF INTERACTION IN EVENT NETWORKS

allies is positive. Statistic lnCapRatio(Gt; a, b) is the logarithmized ratio

The definition of the statistics activityTarget, popularitySource,

3.6 Parameter Estimation

4 Results and Discussion

4.1 Conditional Type Parameters

MODELING FREQUENCY AND TYPE OF INTERACTION IN EVENT NETWORKS

Table 2: Estimated conditional type parameters and

event network model covariate model

CORVINUS JOURNAL OF SOCIOLOGY AND SOCIAL POLICY 1 (2013)

with friendOfEnemy and with enemyOfFriend are significantly negative,

MODELING FREQUENCY AND TYPE OF INTERACTION IN EVENT NETWORKS

we estimate a significantly positive type parameter 0.077 (0.002) for

4.2 Event Frequency

CORVINUS JOURNAL OF SOCIOLOGY AND SOCIAL POLICY 1 (2013)

MODELING FREQUENCY AND TYPE OF INTERACTION IN EVENT NETWORKS

in contrast to alliances, geographic adjacency (operationalized by the

CORVINUS JOURNAL OF SOCIOLOGY AND SOCIAL POLICY 1 (2013)

MODELING FREQUENCY AND TYPE OF INTERACTION IN EVENT NETWORKS

CORVINUS JOURNAL OF SOCIOLOGY AND SOCIAL POLICY 1 (2013)

Hoff, Peter D. -Michael D. Ward (2004), Modeling dependencies in international

CORVINUS JOURNAL OF SOCIOLOGY AND SOCIAL POLICY 1 (2013)

Vous aimerez peut-être aussi