8

THE STATISTICAL EVALUATION OF

SOCIAL NETWORK DYNAMICS

Tom A. B. Snijders*

A class of statistical models is proposed for longitudinal network

data. The dependent variable is the changing (or evolving) rela-
tion network, represented by two or more observations of a directed
graph with a fixed set of actors. The network evolution is modeled
as the consequence of the actors making new choices, or withdraw-
ing existing choices, on the basis of functions, with fixed and ran-
dom components, that the actors try to maximize. Individual and
dyadic exogenous variables can be used as covariates. The change
in the network is modeled as the stochastic result of network effects
(reciprocity, transitivity, etc.) and these covariates. The existing
network structure is a dynamic constraint for the evolution of the
structure itself. The models are continuous-time Markov chain mod-
els that can be implemented as simulation models. The model
parameters are estimated from observed data. For estimating and
testing these models, statistical procedures are proposed that are
based on the method of moments. The statistical procedures are
implemented using a stochastic approximation algorithm based on
computer simulations of the network evolution process.

1. INTRODUCTION

Social networks represent relations (e.g., friendship, esteem, collabora-

tion, etc.) between actors (e.g., individuals, companies, etc.). This paper
is concerned with network data structures in which all relationships within
a given set of n actors are considered. Such a network can be represented
by an n 3 n matrix x 5 ~ x ij !, where x ij represents the relation directed

*University of Groningen, The Netherlands

361
362 SNIJDERS

from actor i to actor j ~i, j 5 1, + + + , n!. Only dichotomous relations are

considered here: the relation from i to j either is present, denoted x ij 5 1,
or absent, denoted x ij 5 0. Self-relations are not considered, so that the
diagonal values x ii are meaningless. They are formally defined as x ii 5 0.
This x is the adjacency matrix of the directed graph by which the network
can be represented; it is also called the sociomatrix.
More specifically, we consider longitudinal data on entire net-
works. It is supposed that the data available are a time series x~t !, t [
$t1 , + + + , tM % of social networks for a constant set $1, + + + , n% of actors. The
observation times are ordered—i.e., t1 , t2 , + + + , tM . The number M of
time points is at least 2. The purpose of the statistical analysis is to obtain
an insight in the evolution of the network, where the initial state x~t1 ! is
taken for granted.
Longitudinal social network data are a complex data structure,
requiring complex methods of data analysis for a satisfactory treatment.
Holland and Leinhardt (1977a, 1977b) and Wasserman (1977) already pro-
posed to use continuous-time Markov chains as a model for longitudinal
social networks. In a continuous-time model, time is assumed to flow on
continuously, although observations are available only at the discrete time
points t1 to tM , and between the observations the network is assumed to
change unobserved at random moments as time progresses. Continuous-
time models offer, in principle, greater flexibility than the discrete-time
Markov chain models elaborated—e.g., by Katz and Proctor (1959) and
Wasserman and Iacobucci (1988).
A basic continuous-time Markov chain model for dichotomous
social networks, the reciprocity model, was elaborated by Wasserman
(1977, 1979, 1980) and further investigated by Leenders (1995a, 1995b)
and Snijders (1999). This model is limited because it assumes dyad inde-
pendence. A dyad is defined as the pair ~ x ij , x ji ! of relations between two
actors i and j. Dyad independence means that the dyads ~X ij ~t !, X ji ~t !!
evolve as mutually independent Markov chains. This assumption effec-
tively allows one to change the analysis from the level of the network to
the level of the dyad. This is computationally attractive but does not leave
much room for realistic statistical modeling. Effects related to depen-
dence in the relations between sets of three or more actors—e.g., transi-
tivity (“a friend of my friend is my friend”)—cannot be represented by
models with dyad independence. Other continuous-time models for social
network evolution were proposed by Wasserman (1980) and Mayer (1984),
but these models were also very restrictive in order to allow parameter
estimation.
STATISTICAL EVALUATION OF NETWORK DYNAMICS 363

Markov chain Monte Carlo (MCMC) methods can be used to devel-

op statistical procedures for quite general probability models for the evo-
lution of social networks, provided that these models can be implemented
as stochastic simulation models. This was proposed by Snijders (1996)
for data defined by sociometric rankings. Snijders and Van Duijn (1997)
sketched how this approach can be used for dichotomous social net-
work data. They also indicated how such an actor-oriented model must be
specified in order to obtain the dyad-independent models of Wasserman
and Leenders. Empirical applications of these stochastic actor-oriented
models were presented in Van de Bunt (1999) and Van de Bunt, Van Duijn,
and Snijders (1999). The present paper extends this method to data
observed at more than two time points, specifies a more efficient and sim-
pler stochastic approximation algorithm, and presents a wider array of
effects that can be included in the model.
The basic idea for our model for social network evolution is that the
actors in the network may evaluate the network structure and try to obtain
a “pleasant” (more neutrally stated, “positively evaluated”) configuration
of relations. The actors base their choices in the network evolution on the
present state of the network, without using a memory of earlier states. How-
ever, they are assumed to have full knowledge of the present network. This
represents the idea that actors pursue their own goals under the constraints
of their environment, while they themselves constitute each others’ chang-
ing environment (cf. Zeggelink 1994). It is immaterial whether this “net-
work optimization” is the actors’ intentional behavior; the only assumption
is that the network can be modeled as if each actor strives after such a
positively evaluated configuration. This evaluation is defined as a func-
tion of the network as regarded from the perspective of the focal actor, and
depends on parameters that are to be estimated from the data. This approach
to network evolution is in line with the theoretical sociological principle
of methodological individualism and was referred to by Snijders (1996)
as a stochastic actor-oriented model. The evaluation includes a random
element to account for the deviation between theoretical expectation and
observed reality, which leads to a kind of random utility model (cf. ran-
dom utility models commonly used in econometrics and treated, e.g., in
Maddala [1983] ). The models can be implemented as stochastic sim-
ulation models, which is the basis for the MCMC procedure for parameter
estimation. This is a frequentist procedure, using the method of moments.
The MCMC implementation of the method of moments uses a stochastic
approximation algorithm that is a descendant of the Robbins-Monro (1951)
algorithm.
364 SNIJDERS

2. CONTINUOUS-TIME MARKOV CHAINS

This section gives a brief introduction to continuous-time Markov chains.

Karlin and Taylor (1975) and Norris (1997) give general treatments of
this kind of stochastic process model. More elaborate introductions to
continuous-time Markov chain models for social networks are given by
Leenders (1995b) and Wasserman (1979, 1980).
The available data are assumed to be two or more observations of
social networks; but the present section is phrased, more generally, in terms
of an arbitrary finite outcome space Y. The finitely many observation times
t1 to tM are embedded in a continuous set of time points T 5 @t1 , tM # 5
$t [ R 6t1 # t # tM %. Thus it is assumed that changes can take place unob-
served between the observation moments. This is not unrealistic and allows
a more versatile and natural mathematical treatment.
Suppose that $Y~t !6t [ T % is a stochastic process where the Y~t !
have a finite outcome space Y and the time parameter t assumes values in
a bounded or unbounded interval T , R. Such a stochastic process is a
Markov process or Markov chain if for any time ta [ T, the conditional
distribution of the future, $Y~t !6t . ta % given the present and the past,
$Y~t !6t # ta %, is a function only of the present, Y~ta !. This implies that for
any possible outcome x [ Y, and for any pair of time points ta , tb ,
P$Y~tb ! 5 x6Y~t ! 5 y~t ! for all t # ta %
(1)
5 P$Y~tb ! 5 x6Y~ta ! 5 y~ta !%+

The Markov chain is said to have a stationary transition distribution if the

probability (1) depends on the time points ta and tb only as a function of
the elapsed time in between, tb 2 ta . It can be proved that if $Y~t !6t [ T %
is a continuous-time Markov chain with stationary transition distribution,
then there exists a function q : Y 2 r R such that
P$Y~t 1 dt ! 5 y6Y~t ! 5 x%
q~ x, y! 5 lim for y Þ x
dtf0 dt
(2)
1 2 P$Y~t 1 dt ! 5 x6Y~t ! 5 x%
q~ x, x! 5 lim +
dtf0 dt
This function q is called the intensity matrix or the infinitesimal genera-
tor. The element q~ x, y! is referred to as the rate at which x tends to change
into y. More generally, an event is said to happen at a rate r, if the proba-
STATISTICAL EVALUATION OF NETWORK DYNAMICS 365

bility that it happens in a very short time interval ~t, t 1 dt ! is approxi-

mately equal to r dt.
The simultaneous distribution of the Markov chain $Y~t !6t $ ta %
with stationary transition distribution is determined completely by the prob-
ability distribution of the initial value Y~ta ! together with the intensity
matrix. Specifically, the transition matrix

P~tb 2 ta ! 5 ~ P$Y~tb ! 5 y6Y~ta ! 5 x%!x, y[Y

is defined by

P~t ! 5 e Qt,

where Q is the matrix with elements q~ x, y! and the matrix exponential is

defined by
`
Q ht h
e Qt 5 (
h50 h!
+

The reasons for specializing the model to Markov processes with

stationary transition distributions are that such models often are quite
natural, and that they lend themselves well to computer simulation. The
resulting dynamic computer simulation models can be regarded as a type
of discrete event simulation model as discussed by Fararo and Hum-
mon (1994).

3. STOCHASTIC ACTOR-ORIENTED MODELS FOR

NETWORK EVOLUTION: SIMPLE SPECIFICATION

The specification of the model developed in this paper has three ingredi-
ents: the rate function, the objective function, and the gratification func-
tion. A simple specification is determined by only the objective function,
with a constant rate function and a gratification function equal to zero.
The model is explained first for this simple specification. The rate and
gratification functions are treated in a later section.

3+1+ Basic Model Ingredients

The class of all sociomatrices—i.e., of all n 3 n matrices of 0-1 elements

with a zero diagonal—is denoted by X. Note that X has 2 n~n21! elements,
366 SNIJDERS

a number that is so huge that analytical calculations based on the intensity

matrix will be out of the question for most purposes.
It is assumed that all actors “control” their outgoing relations, which
are collected in the row vector ~Xi1~t !, + + + , Xin ~t !! of the sociomatrix. Actors
have the opportunity to change their outgoing relations at stochastic times;
in the interval between the observation moments tm and tm11 these oppor-
tunities occur at a rate rm . When actors change their outgoing relations,
they are assumed to strive after a rewarding configuration for themselves
in the network. This goal is modeled in the so-called objective function f
discussed below, to which a random component is added, representing the
actors’ drives that are not explicitly modeled. The actors are assumed to
have all information required to calculate their own objective function.
This information can be extensive or limited, depending on the model.
At any single time point, at most one actor may change his outgo-
ing relations. Furthermore, he may change only one relation at the time.
Of course, many small changes between two observation times can result
in a big difference between the two observed networks. The fact that the
model specification focuses on changes of single relations is the major
reason why continuous time modeling is relatively straightforward. (An
example of a continuous-time model for social networks where more than
one relation can change at one time point is given by Mayer [1984].) It
should be noted that the fact that the actors take into account the present
network structure that is common to them all, introduces a high degree of
interdependence between them (when one marginalizes out, rather than
conditions upon, the current network structure).

3+2+ Objective Function

The objective function for actor i is denoted by

fi ~ b, x!, x [ X, (3)

and indicates the degree of satisfaction for actor i inherent in the rela-
tional situation represented by x. This function depends on a parameter
vector b. In the simple model specification of this section, the parameter
of the statistical model is u 5 ~ r, b!, where r 5 ~ r1 , + + + , rM21 ! is the
vector of change rates during the time periods from tm to tm11 ~m 5
1, + + + , M 2 1!.
STATISTICAL EVALUATION OF NETWORK DYNAMICS 367

Suppose that at some moment t, actor i has the opportunity to change

her outgoing relations. At this moment, actor i determines the other actor
j with whom she will change her relation x ij . If immediately before time t
actor i does have a relation to actor j, then a change implies withdrawing
the relation; if immediately before time t actor i does not have a relation
to actor j, then a change implies initiating the relation. Given the present
state x of the network, the network that results when the single element x ij
is changed into 1 2 x ij (i.e., from 0 to 1 or from 1 to 0), is denoted by
x~i V j !. Note that x~i V j ! refers to an entire adjacency matrix. When
the current network is x, actor i has the choice between x~i V j ! for all
possible j 5 1, + + + , n, j Þ i. It is assumed that actor i chooses the j that
maximizes the value of her objective function fi ~ b, x~i V j !! plus a ran-
dom element,

fi ~ b, x~i V j !! 1 Ui ~t, x, j !+ (4)

The term Ui ~t, x, j ! is a random variable, indicating the part of the actor’s
preference that is not represented by the systematic component fi . It is
assumed that these random variables are independent and identically dis-
tributed for all i, t, x, j. The assumption that the actor tries to maximize
(4), which refers to the state obtained immediately after making this sin-
gle choice, can be regarded as an assumption of myopia: the actor does
not consider the longer-term, or indirect, effects of her choices.

3+3+ Markov Chain with Random Utility Component

These functions are used in the following way to define a continuous-time

Markov chain X~t ! with the finite outcome space X.
Events—i.e., changes of the network structure—take place at dis-
crete time points; in between these points, the network structure remains
constant. The process is modeled as being right-continuous: If a change
takes place from state x 0 to state x 1 at time t0 , then there is an e . 0 such
that X~t ! 5 x 0 for t0 2 e , t , t0 , while X~t ! 5 x 1 for t0 # t , t0 1 e.
The actions of the n actors depend only on the current state of the
network, not on the history of how this network came into being. All actors
change their relations one at a time at stochastic moments at a rate rm .
This means that at each time point t [ ~tm , tm11 !, the time until the next
change by any actor has the negative exponential distribution with param-
eter nrm and the expected waiting time until the next change by any actor
368 SNIJDERS

is 10~nrm !. When an event occurs, all actors have the same probability
10n to be the one to change one of their outgoing relations. Given that
actor i may change an outgoing relation, she chooses to change her rela-
tion to that actor j ~ j Þ i ! for whom the value of (4) is highest.
It is convenient to let the Ui ~t, x, j ! have the type 1 extreme value
distribution (or Gumbel distribution) with mean 0 and scale parameter 1
(Maddala 1983). This assumption is commonly made in random utility
modeling in econometrics. When this distribution is used, the probability
that the given actor i chooses the other actor j for changing the relation x ij
is the multinomial logit expression (cf. Maddala [1983, p. 60] ),
exp~ fi ~ b, x~i V j !!!
pij ~u, x! 5 ~ j Þ i !+ (5)
n

(
h51, hÞi
exp~ fi ~ b, x~i V h!!!

3+4+ Intensity Matrix

It was mentioned in Section 2 that stationary transition distributions of

continuous-time Markov chains are characterized by their intensity matrix.
In our case, where relations are allowed to change only one at a time, the
intensity matrix can be represented by functions qij ~ x!, indicating the
change rates of x to x~i V j ! for j Þ i. All other change rates are 0. These
functions are defined for i, j 5 1, + + + , n, i Þ j as
P$X~t 1 dt ! 5 x~i V j !6 X~t ! 5 x%
qij ~ x! 5 lim + (6)
dtf0 dt
The intensity matrix q~ x, y! defined in (2) is related to qij ~ x! by

5
qij ~ x! if y 5 x~i V j !
0 if x and y differ in more than one element (7)
q~ x, y! 5
2 ( q ~x!
iÞj
ij if x 5 y+

Note that directed graphs x and y differ in exactly one element ~i, j ! if and
only if y 5 x~i V j ! and x 5 y~i V j !.
For the Markov chain in the simple model specification of the
present section, qij ~ x! is given for time period ~tm , tm11 ! by
qij ~ x! 5 rm pij ~u, x!+ (8)
STATISTICAL EVALUATION OF NETWORK DYNAMICS 369

3+5+ Specification of the Model

The objective function must contain the substantive ingredients of the

model, including, for example, actor attributes and structural properties
of the directed graph. Since actors have direct control only of their outgo-
ing relations, only the dependence of fi on row i of the adjacency matrix
has an influence on the behavior of the model.
A convenient choice for the objective function is to define it as
a sum

L
fi ~ b, x! 5 (b s
k51
k ik ~ x!, (9)

where the weights bk are statistical parameters indicating the strength of

the corresponding effect sik ~x!, controlling for all other effects in the model,
and the sik ~ x! are relevant functions of the digraph that are supposed to
play a role in its evolution. All formulas given below for possible compo-
nents sik refer to a contribution to the objective function of actor i, while
the other actors to whom i could be related are indicated by j.
Effects can be distinguished according to whether they depend only
on the network x—in which case they can be regarded as endogenous
network effects—or also on covariates, which are supposed to be deter-
mined exogenously. Covariates can be of two kinds: (1) actor-dependent
covariates V with values vi for actor i, or (2) pair-dependent (dyadic)
covariates W with values wij for the ordered pair ~i, j !. Only constant (i.e.,
time-independent) covariates are considered.
The following list is a collection of network effects, as possibilities
for the functions sik in (9).

1. Density effect, defined by the out-degree

si1 ~ x! 5 x i1 5 (x j
ij ;

2. Reciprocity effect, defined by the number of reciprocated relations

si 2 ~ x! 5 (x
j
ij x ji ;
370 SNIJDERS

3. Popularity effect, defined by the sum of the in-degrees of the others

to whom i is related,

si3 ~ x! 5 (xj
ij x 1j 5 (x (x
j
ij
h
hj ;

4. Activity effect, defined by the sum of the out-degrees of the others to

whom i is related,

si 4 ~ x! 5 (xj
ij x j1 5 (x (x
j
ij
h
jh ;

5. Transitivity effect, defined by the number of transitive patterns in i ’s

relations (ordered pairs of actors ~ j, h! to both of whom i is related,
while also j is related to h!,

si5 ~ x! 5 (x
j, h
ij x ih x jh ;

6. Indirect relations effect, defined by the number of actors to whom i is

indirectly related (through one intermediary—i.e., at sociometric dis-
tance 2),

si6 ~ x! 5 #$ j 6 x ij 5 0, maxh ~ x ih x hj ! . 0%;

7. Balance, defined by the likeness between the out-relations of actor i

to the out-relations of the other actors j to whom i is related,
n n
si 7 ~ x! 5 ( ( ~b
j51
x ij
h51
0 2 6 x ih 2 x jh 6!, (10)
hÞi, j

where b0 is a constant included for convenience. If the density effect

is included in the model (which normally will be the case), the num-
ber b0 can be chosen so as to obtain the clearest interpretation with-
out essentially changing the model specification.
For example, to have a balance effect that is not too strongly
correlated with the density effect, the number b0 in (10) can be cho-
sen so that the average of the second sum in this equation over all
actors and over the first M 2 1 time points is 0. In other words,
STATISTICAL EVALUATION OF NETWORK DYNAMICS 371

M21 n n
1
b0 5
~M 2 1!n~n 2 1!~n 2 2! m51 i, j51( ( ( 6x h51
ih ~tm ! 2 x jh ~tm !6+
hÞi, j

(11)
This list can be extended, in principle, indefinitely. Potentially important
additional types of effect are nonlinear effects—i.e., nonlinear functions
of sik defined above, the out-degree x i1 being the primary candidate for
such a nonlinear transformation; and other subgraph counts in which actor
i is involved, of which the reciprocity and transitivity effects are examples.
In practically all applications it will be advisable to include the
density effect, because the other effects listed above should be controlled
for this effect. The reciprocity effect is so fundamental in social relations
that it is advisable also to include this effect in most applications.
The transitivity and balance effects, and the indirect relations effect
when it has a negative weight, all are different mathematical specifica-
tions of the intuitive idea that actor i has a “closed” or transitive personal
network—i.e., the others to whom i is related tend to have comparatively
many relations among themselves. Verbal theories will not often be detailed
enough to distinguish between these effects. It can be determined empir-
ically if one or some of these three effects succeed better than the others
in accounting for the observed degree of closure, or transitivity, in the
data.
For each actor-dependent covariate V there are the following three
basic potential effects. (The notation for the functions sik does not explic-
itly indicate their dependence on the covariate values vj .)
8. Covariate-related popularity, defined by the sum of the covariate
over all actors to whom i has a relation,

si 8 ~ x! 5 (x
j
ij vj ;

9. Covariate-related activity, defined by i ’s out-degree weighted by

his covariate value,
si 9 ~ x! 5 vi x i1 ;
10. Covariate-related dissimilarity, defined by the sum of absolute
covariate differences between i and the others to whom he is related,

si10 ~ x! 5 (x
j
ij 6vi 2 vj 6+
372 SNIJDERS

Positive covariate-related popularity or activity effects will lead to

associations between the covariate and the in-degrees and out-degrees,
respectively. A negative covariate-related dissimilarity effect will lead to
relations being formed especially between actors with similar values on
the covariate.
This list can be extended—for example, by including covariate val-
ues in the definitions of the network effects listed above. This represents
interactions between the covariate and the network effect.
The main effect for a pair-dependent covariate is as follows:

11. Covariate-related preference, defined by the sum of the values of

wij for all others to whom i is related,

si11 ~ x! 5 (x
j
ij wij +

Here also, the list can be extended by including covariate values in the
definition of network effects.
Theoretical insights into the relational process and experience with
modeling this type of data have to determine the effects that are included.

4. MOMENT ESTIMATORS

Let the objective function be given by (9), so that the parameter of the
statistical model is u 5 ~ r, b!. The dimensionality of b is denoted L and
the total number of dimensions for u is K 5 M 21 1 L. Analogous to what
was proposed for a similar model by Snijders (1996), this parameter can
be estimated by the method of moments (explained for general statistical
models—for example, by Bowman and Shenton 1985). This means that a
statistic Z 5 ~Z 1 , + + + , Z K ! is used, for which u is determined as the solution
of the K-dimensional moment equation

E u Z 5 z, (12)

where z is the observed outcome. This moment equation will be specified

further by certain ways of conditioning on the initial and intermediate
outcomes x~t1 ! to x~tm21 !.
First the choice of the statistic Z is discussed, and then a MCMC
algorithm that can be used to approximate the solution of the moment
equation.
STATISTICAL EVALUATION OF NETWORK DYNAMICS 373

For the estimation, no assumptions whatsoever are made about the

initial state x~t1 !. Therefore, the estimation is carried out conditional on
this initial state, and this state is not used to obtain any information about
the value of the parameter.
In the absence of a formal method such as a reduction to sufficient
statistics, the statistics Z k should be chosen so that they are relevant for
the components of the parameter u in the sense that the expected values of
Z k ~k 51, + + + , K ! are sensitive to changes in the components of u. One way
to specify this is to require that for all k
]E u Z k
. 0+
]uk

A more stringent specification is to require that this property hold not

only for all separate coordinates of the parameter vector, but also for all
linear combinations:

a' S D
]E u Z
]u
a . 0 for all a [ RK, a Þ 0, (13)

where ~]E u Z0]u! is the matrix of partial derivatives. This requirement is

far from implying the statistical efficiency of the resulting estimator, but
it confers a basic credibility to the moment estimator and it ensures the
convergence of the stochastic approximation algorithm mentioned below.
The components of u 5 ~ r, b! are the rates of change rm in the
time interval ~tm , tm11 ! and the weights bk in the objective function (9).
The motivation for the statistics Z i , at this moment, is of a heuristic nature,
based on their obvious connection to the parameters and supported by
sufficiency considerations in certain special cases.
For rm , a relevant statistic is the total amount of change in the mth
time period measured by the number of differences between two consec-
utive observation moments,
n
Cm 5 ( 6X
i, j51
ij ~tm11 ! 2 X ij ~tm !6+ (14)
iÞj

This choice for the statistic relevant for rm can be supported by noting
that if b 5 0, which reduces the model to the trivial situation where the
X ij ~t ! are randomly changing 0-1 variables, Cm is a sufficient statistic
for rm .
374 SNIJDERS

For bk , a relevant statistic is the sum over all actors i of the digraph
statistics sik , observed at time tm11 ,
n
Smk 5 (s
i51
ik ~X~tm11 !!+ (15)

This statistic has an immediate intuitive appeal: If bk is larger, then the

actors strive more strongly to have a high value of sik , so that it may be
expected that Smk will be higher for all m. The statistics Smk are combined
over the M 2 1 time intervals by an unweighted sum.
Combining all these proposals, the moment estimator for u is
defined as the solution of the system of equations
E u $Cm 6 X~tm ! 5 x~tm !% 5 cm ~m 5 1, + + + , M 2 1! (16)
M21 M21

(
m51
E u $Smk 6 X~tm ! 5 x~tm !% 5 (s
m51
mk ~k 5 1, + + + , L!, (17)

where cm and smk are the observed outcomes of the statistics Cm and Smk .
Although in our experience these equations mostly seem to have
exactly one solution, they do not always have a solution. This can be seen
as follows. For a fixed value of b, the left-hand side of (16) is an increas-
ing function of rm , tending to an asymptote which is lower than the max-
imum possible value of cm , this maximum being n~n 2 1!. This implies
that the method proposed here is not suitable for observations x~tm ! and
x~tm11 !, which are too far apart in the sense of the metric (14). For such
observations the dependence of x~tm11 ! on the initial situation x~tm ! is
practically extinguished, and it may be more relevant to estimate the param-
eters of the process generating x~tm11 ! without taking this initial situation
into account.
For the trivial submodel where all X ij ~t ! are independent, the exis-
tence of maximum likelihood and moment estimators is discussed in
Snijders and Van Duijn (1997).

4+1+ Covariance Matrix of the Estimator

The delta method (e.g., see Bishop, Fienberg, and Holland 1975, sec. 14.6)
can be used to derive an approximate covariance matrix for the moment
estimator u.Z (This holds generally for moment estimators; see Bowman
and Shenton 1985, formula 5.) For a homogeneous notation for the param-
STATISTICAL EVALUATION OF NETWORK DYNAMICS 375

eters rm and b, denote Cmm 5 Cm and formally define Cmk 5 0 for k Þ m,

and denote
Z m 5 ~Cm1 , + + + , Cm, M21 , Sm1 , + + + , SmL !+

Then the moment equations (16, 17) can be written as

M21 M21

(
m51
E u $Z m 6 X~tm ! 5 x~tm !% 5 (z
m51
m+ (18)

Further denote
M21

( u5 ( cov$Z
m51
m 6 X~tm ! 5 x~tm !% (19)

] M21
Du 5
]u ( E$Z
m51
m 6 X~tm ! 5 x~tm !%+ (20)

Then it follows from the delta method, combined with the implicit func-
tion theorem and the Markov property for the X~t ! process, that the approx-
imate covariance matrix of uZ is

Z ' Du21
cov~ u! (uD '21
u + (21)

It is plausible that these estimators have approximately normal dis-

tributions, although a proof is not yet available. Based on the assumption
of normally distributed estimates, the parameters can be tested using the
t-ratios defined as the parameter estimate divided by its standard error,
referred to a standard normal null distribution. (In other words, the test is
carried out as a t-test with infinite degrees of freedom; this test should be
regarded as a rough approximation, since no definite results are yet avail-
able on the distribution of this test statistic.)

4+2+ Conditional Moment Estimation

The method of moments can be modified by conditioning on the out-

comes cm of Cm ~m 5 1, + + + , M 2 1! rather than using moment equations
involving these statistics. This provides a more stable and efficient
algorithm and reduces the parameter estimated by the method of moments
to the L-dimensional b. This can be helpful especially for larger values
of M.
376 SNIJDERS

The modified method is based on the property that the distribution

of a continuous-time Markov chain X~t ! remains invariant when the time
parameter is divided by some constant value while the rate parameter is
multiplied by the same value. Specifically, when the rate parameter rm
obtains for all t $ tm , then the distribution of X~tm 1 t !, conditional on
X~tm ! and for t . 0, depends on rm and t only through their product, trm .
The modified method can be loosely described as follows. For each period
m independently, the Markov chain is started at time t 5 0 with the initial
value x @m# 5 x~tm ! and a rate parameter equal to 1. The process is stopped
@m#
at the first moment t when (ij 6 X ij ~t ! 2 x ij 6 5 cm . This value of t is
expected to be close to the product rm ~tm11 2 tm ! and the statistics observed
at this moment are compared with the statistics calculated from observa-
tion x~tm11 !.
To explain this more formally, denote by X ~1! ~t ! a Markov chain
evolving according to our model with a fixed and constant rate parameter
~1!
r 5 1 and a given value of b, and denote by Sk ~t ! the corresponding
statistics (15). Independent replications of this stochastic process, starting
at t 5 0 with X ~1! ~0! 5 x~tm !, are used as models for the M 2 1 periods.
Define the statistic
n
C ~1! ~t ! 5 ( 6X
i, j51
~1!
ij ~t !
~1!
2 X ij ~0!6 (22)
iÞj

and the stopping time

Tmfin 5 min$t $ 06C ~1! ~t ! $ cm %+ (23)

The conditional moment estimator for b is defined as the solution of

M21 M21

(
m51
~1!
E b $Sk ~Tmfin !6 X ~1! ~0! 5 x~tm !% 5 (s
m51
mk ~k 5 1, + + + , L! (24)

and, given the resulting estimate b,Z rm is estimated by

r[ m 5 ~tm11 2 tm !21 E bZ $Tmfin 6 X ~1! ~0! 5 x~tm !%+ (25)

It follows from the general theory of Markov chains that for all
possible values of cm the stopping time Tmfin is finite with probability 1,
and even has a finite expected value. Therefore the difficulties with the
definition of the estimator for large values of cm , as discussed for the uncon-
STATISTICAL EVALUATION OF NETWORK DYNAMICS 377

ditional moment estimator, do not arise here. However, this consolation is

only theoretical, because in practice, for large t the value of C ~1! ~t ! fluc-
tuates randomly about an asymptote lower than the maximum possible
value of n~n 2 1!, and the stopping time Tmfin is indeed finite but horribly
large. The simulation-based algorithm, explained below, is not practically
feasible for values of cm larger than this asymptote.

5. STOCHASTIC APPROXIMATION

The moment equations for the two estimation methods are defined by (18)
and (24), but the conditional expectations that are central in these equa-
tions cannot be calculated explicitly (except for some special and rather
trivial cases, as discussed in Snijders and Van Duijn [1997] ). However, it
is rather straightforward to simulate random digraphs with the desired dis-
tributions. Therefore, stochastic approximation methods—in particular,
versions of the Robbins-Monro (1951) procedure—can be used to approx-
imate the moment estimates. Introductions to stochastic approximation
and the Robbins-Monro algorithm are given, for example, by Ruppert
(1991) and Pflug (1996).
The algorithm to solve the equation (12) is based on a sequence uZ N
generated according to the iteration step

uZ N11 5 uZ N 2 a N D021 ~Z N 2 z!, (26)

where Z N is generated according to the probability distribution defined by

the parameter value uZ N . For a N , a sequence is used that converges slowly
to 0. D0 is a positive diagonal matrix. In principle, the optimal choice of
D0 might be nondiagonal. However, Polyak (1990), Ruppert (1988), and
Yin (1991) (as discussed also by Pflug [1996, sec. 5.1.3] and Kushner and
Yin, 1997) showed that if all eigenvalues of the matrix of partial deriva-
tives, ~]E u Z0]u!, have positive real parts and certain regularity condi-
tions are satisfied, then convergence at an optimal rate can be achieved
when D0 is the identity matrix, with a N a sequence of positive numbers
converging to 0 at the rate N 2c , where 0+5 , c , 1. To obtain this optimal
convergence rate, the solution of (12) must be estimated not by the last
value uZ N itself, but by the average of the consecutively generated uZ N val-
ues. This algorithm is a Markov chain Monte Carlo algorithm because the
iteration rule (26) indeed defines a Markov chain.
378 SNIJDERS

The convergence properties of this algorithm hold asymptotically

for N r `. To have good properties already for relatively low values of
N, it is important to specify the algorithm in such a way that it quickly
comes close to the target value. This can be achieved by applying a result
due to Pflug (1990), who showed that the limiting first order autocorrela-
tion of the sequence ~Z N 2 z! generated by (26) is negative. This means
that as long as the partial sums of successive values of the product ~Z N 2
z! ' ~Z N21 2 z! are positive, it must be assumed that the sequence uZ N still is
drifting toward the limit point rather than wandering around the limit point,
so that it is not desirable to decrease the step sizes a N . Therefore a N remains
constant as long as there still seems to be such a drift going on, except that
when N gets too large a N is decreased anyway, in order to retain the con-
vergence rate N 2c for the sequence a N .
These ideas are combined in the specification of the algorithm as
given in the appendix. The algorithm provides an arbitrarily accurate
approximation to the solution of (12) as well as an estimate of the covari-
ance matrix (21). It is available in the freeware PC program SIENA (see
the discussion in Section 10).

6. AN EVOLVING NETWORK OF UNIVERSITY

FRESHMEN

As an illustration, data are used from a study conducted by Van De Bunt

(1999), which were analyzed also by Van De Bunt, Van Duijn, and Snijders
(1999). For a more extensive description of this data set we refer to these
publications. In the present paper, this data set is used only as an illustra-
tion without paying much attention to the theoretical interpretations.
The actors in this network are a group of 32 university freshmen
who were following a common study program in a Dutch university. This
group comprised 24 female and 8 male students. The number of observa-
tions used here is M 5 3. The data used here are those for the time points
labeled t2 , t3 , and t4 in Van De Bunt, Van Duijn, and Snijders (1999).
There are 3 weeks between time points t2 and t3 , and also between t3 and
t4 . For the purpose of this illustration, the time points are relabeled t1 , t2 ,
and t3 . The relation studied is defined as “at least a friendly relationship,”
referred to here as a positive relation ~ x ij 5 1!. The absence of a positive
relation is referred to as a null relation ~ x ij 5 0!.
There is missing data due to nonresponse, increasing from 9% at t1
to 19% at t3 . This incompleteness of data is treated in the estimation pro-
STATISTICAL EVALUATION OF NETWORK DYNAMICS 379

cedure in the following ad hoc fashion. (It will be important to conduct

further studies to evaluate this way of dealing with incomplete data, and
compare it with potential alternatives.)
Missing data are treated in a simple way, trying to minimize their
influence on the estimation results. The simulations are carried out over
all n 5 32 actors. In the initial observation x~tm ! for each period, missing
entries x ij ~tm ! are set to 0. In the course of the simulations, however, these
values are allowed to become 1 like any other values x ij ~t !. For the calcu-
lation of the statistics Smk and Cm , the values of x ij ~tm ! as well as of
X ij ~tm11 ! are set to 0 whenever at least one of the two observations x ij ~tm !
and x ij ~tm11 ! is missing.
To get a basic impression of the data, it may be noted that densities
(calculated over the available data) at the three observation moments
increase from 0.15 via 0.18 to 0.22. The number of observed changes
between the observations at t1 and t2 was 60 (out of 744 directed pairs
~i, j ! for which the value of x ij was observed at observations t1 and t2 !;
between t2 and t3 this was 51 (out of 679 observations).
The first model estimated includes the basic effects of density and
reciprocity, together with the three basic triadic effects: transitivity, indi-
rect relations, and balance. The purpose of this stage in the analysis is to
investigate which of these triadic effects are empirically supported by these
network evolution data. The number b0 in (10) is defined by (11). The
conditional moment estimator was used and the algorithm was specified
as described in the appendix, except that to increase precision 5 sub-
phases were carried out in phase 2 and n3 5 1000 steps were made in
phase 3. The results are displayed as Model 1 in Table 1.
The estimated rate parameters, r[ 1 5 3+87 and r[ 2 5 3+10, indicate
that on average the actors made 3.87 changes of relationships between the
first two observations, and 3.10 changes between the last two observa-
tions. (This includes two-way changes between two observations that
remained unobserved because they canceled each other.)
As suggested in Section 4.1, the effects are tested by t-statistics
defined by the ratio of parameter estimate to standard error, referred to as
standard normal distribution. There is a strongly significant reciprocity
effect ~t 5 1+9800+31 5 6+39!. Of the three triadic effects, the indirect
relations effect is significant ~t 5 20+34700+074 5 24+69!, but the other
two are not significant at the 5 percent level, although the transitivity
effect comes close. When the balance effect was deleted from the model,
the t-value for the transitivity effect became 1.94 (results not shown
380 SNIJDERS

TABLE 1
Parameters and Standard Errors for Models Estimated Using Observations at t1 , t2 , t3

Model 1 Model 2 Model 3

Effect Par. (s.e.) Par. (s.e.) Par. (s.e.)

Rate (period 1) 3.87 3.78 3.91
Rate (period 2) 3.10 3.14 3.07
Density 21.48 (0.30) 21.05 (0.19) 21.13 (0.22)
Reciprocity 1.98 (0.31) 2.44 (0.40) 2.52 (0.37)
Transitivity 0.21 (0.11) — —
Balance 20.33 (0.66) — —
Indirect relations 20.347 (0.074) 20.557 (0.083) 20.502 (0.084)
Gender activity — — 20.60 (0.28)
Gender popularity — — 0.64 (0.24)
Gender dissimilarity — — 20.42 (0.24)

here), just short of significance at the 5 percent level. The results obtained
when deleting the two nonsignificant effects from the model are shown as
Model 2 in Table 1. The indirect relations effect becomes larger, and the
density and reciprocity effects change, because these effects now must
also represent the effects represented by transitivity and balance in Model
1. It can be concluded that there is evidence of a tendency to have closed
networks in the sense of a relatively low number of indirect relations;
controlling for this effect and for reciprocity, there is no significant ten-
dency toward a high number of transitive triplets or toward balanced rela-
tionships. No significant evidence was found for other structural network
effects (estimation results not shown here).
As a next step, the three basic effects of gender were included in
the model. In the original data set, gender was represented by a dummy
variable equal to 0 for women and 1 for men. The means were subtracted
from this variable as well as from the dissimilarity variable 6vi 2 vj 6. Given
that the proportion of women was 75 percent, this leads to the variable vi
being 20.25 for women and 10.75 for men, and the dissimilarity vari-
able being 20.387 for equal-gender pairs and 0.613 for unequal-gender
pairs. The results for the model including the structural effects of reciproc-
ity and indirect relations as well as the three covariate effects of gender
are presented in Table 1 as Model 3. It can be concluded that women are
more active in creating positive relations than men ~t 5 20+6000+28 5
22+14!, while men receive more positive choices ~t 5 0+6400+24 5 2+67!,
STATISTICAL EVALUATION OF NETWORK DYNAMICS 381

but there are no significant (dis)similarity effects associated with gender.

The control for gender does not have an important influence on the reci-
procity or indirect relations effects.
The results based on the observations at these three moments can
be compared with results based on only two of these observations. This
can be used to check the model assumption that the parameter values bk in
the time interval between t1 and t2 are the same as between t2 and t3 . Fur-
thermore, for the analysis of the evolution of the network from t1 to
t3 , this illustrates the greater precision obtainable by including the infor-
mation about the network at t2 . The comparison is made only for Model 3
and reported in Table 2.
None of the estimates are significantly different between the peri-
ods t1–t2 and t2–t3 . This supports the use of a common model for the entire
period t1–t3 .
To compare the Model 3 column of Table 1 with the t1 , t3 column
of Table 2, the estimates in the former column are called “three-
observation” and those in the latter column “two-observation” estimates.
It appears that the corresponding estimates differ at most by about one
“two-observation” standard error; for all parameters but one, the three-
observation estimates are closer than the two-observation estimates to the
mean of the separate estimates for the t1–t2 and t2–t3 periods. The three-
observation standard errors all are clearly smaller than the two-observation
standard errors. This provides some support for the expected greater reli-
ability of the three-observation as compared with the two-observation
estimates.

TABLE 2
Parameter Estimates and Standard Errors for Model 3, Estimated
from Two Observations

Observations t1 , t2 t2 , t3 t1 , t3

Effect Par. (s.e.) Par. (s.e.) Par. (s.e.)

Rate 3.64 3.21 5.29
Density 20.99 (0.32) 21.30 (0.28) 20.78 (0.31)
Reciprocity 2.36 (0.52) 2.89 (0.67) 2.40 (0.48)
Indirect relations 20.432 (0.113) 20.653 (0.140) 20.536 (0.146)
Gender activity 20.75 (0.40) 20.39 (0.42) 20.77 (0.36)
Gender popularity 0.40 (0.31) 1.03 (0.44) 0.36 (0.26)
Gender dissimilarity 20.35 (0.35) 20.58 (0.43) 20.22 (0.31)
382 SNIJDERS

7. EXTENDED MODEL SPECIFICATION

The general model specification contains, in addition to the objective func-

tion, two other elements: (1) the rate function, which shows that actors
may differ in the rate at which they change their relations; and (2) the
gratification function, which shows that various effects may operate dif-
ferently for the creation of a relation (where x ij goes from 0 to 1) than for
its dissolution ~ x ij changing from 1 to 0).

7+1+ Rate Function

The rate function for actor i is denoted

l i ~ r, a, x, m! for x [ X, (27)

and indicates the rate at which actor i is allowed to change something in

his outgoing relations in the time period tm # t , tm11 . In the simple
specification given above, this rate function depended only on m and not
on i or x, and was defined as l i ~ r, a, x, m! 5 rm . The roles of the statis-
tical parameters r and a are discussed below.
These rate functions and the conditional independence of the actors
imply that at each time point t, the time until the next change by any actor
has the negative exponential distribution with parameter
n
l 1 ~ r, a, x, m! 5 ( l ~ r, a, x, m!, for x 5 x~t !, t
i51
i m # t , tm11 (28)

(provided that this next change still is before time tm11 !. The parameter of
the negative exponential distribution is taken here as the reciprocal of the
expectation, so the expected waiting time until the next change after time
t is 10l 1 ~ r, a, x~t !, m! (where a possible change to the following time
interval is not taken into account). Given that a change occurs, the prob-
ability that it is actor i who may change his out-relations is

l i ~ r, a, x, m!
+ (29)
l 1 ~ r, a, x, m!

Nonconstant rate functions can depend, for example, on actor-

specific covariates or on network statistics expressing the degree to which
the actor is satisfied with the present network structure. Of course, the
STATISTICAL EVALUATION OF NETWORK DYNAMICS 383

rate function must be restricted to positive values. In order not to burden

the specification with too many complications, it is proposed that the rate
function be defined as a product
l i ~ r, a, x, m! 5 l i1 l i 2 l i3

of factors depending, respectively, on period m, actor covariates, and the

actor’s personal network. The corresponding factors in the rate function
are the following:

1. The dependence on the period can be represented by a simple factor

l i1 5 rm

for m 5 1, + + + , M 2 1.
2. The effect of actor covariates with values vhi can be represented by
the factor

l i 2 5 exp S ( a v D+
h
h hi (30)

3. The dependence on the network can be modeled, for example, as a

function of the actor’s out-degree, in-degree, and number of recipro-
cated relations. Define these by

x i1 5 (x
j
ij , x 1i 5 (x
j
ji , x i ~r! 5 (x
j
ij x ji

(recalling that x ii 5 0 for all i !.

Snijders and Van Duijn (1997) investigated how the rate func-
tion should be specified in order to obtain Wasserman’s (1979) reci-
procity model as a special case. Denoting the corresponding parameter
by a1 , for the dependence on the out-degree this led to the factor

l i3 5
x i1
n 21
exp~a1 ! 1 1 2 S
x i1
n 21
exp~2a1 !+ D (31)

This defines a linear function of the out-degree, parametrized in such

a way that it is necessarily positive.
For a general dependence on the out-degree, in-degree, and
number of reciprocated relations, one can use an average of such terms,
the second and third one depending on x 1i and x i~r! , respectively.
384 SNIJDERS

It would be interesting to explore other specifications of the rate function,

expressing in a theoretically more satisfactory way the circumstances and
characteristics upon which it depends how quickly actors change their
relations.

7+2+ Gratification Function

The basic motivation for the third model ingredient, the gratification
function, is that a given effect may operate more strongly, or less strongly,
for the creation than for the dissolution of relations. For example, it is
conceivable that although actors prefer to establish reciprocated rela-
tions, they are quite willing to initiate as yet unreciprocated relations;
but that, once they have a reciprocated relationship, they are very reluc-
tant to let it go—for example, because of the investments accumulated
in this relation (cf. Van De Bunt [1999] ). This would mean that the rec-
iprocity effect is greater for dissolution than for creation of ties. Such a
difference cannot be represented by the objective function alone. There-
fore the model includes also a gratification function
gi ~g, x, j !, defined for i, j 5 1, + + + , n, i Þ j, x [ X, (32)

which indicates the instantaneous gratification experienced by actor i when,

from the given network configuration x, element x ij is changed into its
opposite, 1 2 x ij .
When a gratification function is included in the model, expression
(4) for the momentary objective function maximized by i is replaced by
the sum of the actor’s preference for the new state, the gratification expe-
rienced as a result of the change, and a random element:
fi ~ b, x~i V j !! 1 gi ~g, x, j ! 1 Ui ~t, x, j !+ (33)

Using the same assumptions for the random term Ui ~t, x, j ! as above, the
probabilities of the various possible new states x~i V j ! are now given by
exp~r~u, i, j, x!!
pij ~u, x! 5 ~ j Þ i !, (34)
n

(
h51, hÞi
exp~r~u, i, h, x!!

where
r~u, i, j, x! 5 fi ~ b, x~i V j !! 1 gi ~g, x, j !+
STATISTICAL EVALUATION OF NETWORK DYNAMICS 385

These probabilities do not change when a term is added to r~u, i, j, x! that

does not depend on j. It is often more convenient to work with
r~u, i, j, x! 5 fi ~ b, x~i V j !! 2 fi ~ b, x! 1 gi ~g, x, j !+ (35)

The instantaneous effect gi is a more general model component than the

objective function fi , because the objective function depends only on the
new state x~i V j !, whereas the gratification function depends arbitrarily
on the new state as well as the old state x. The reason for not working with
just the gratification function is that the objective function, attaching a
value to each network configuration, often is conceptually more attractive
and better interpretable than the instantaneous gratification effect.
The gratification function can be specified by a weighted sum,
H
gi ~g, x, j ! 5 (g
h51
h rijh ~ x! (36)

for certain statistics rijh ~ x!, each containing either a factor x ij (if it reflects
the gratification involved in withdrawing a relation—i.e., changing x ij from
1 to 0) or a factor ~1 2 x ij ! (if the effect is about the gratification involved
in creating a relation). Some examples of such terms are the following:

1. g1 x ij x ji : Indicator of a reciprocated relation; a negative value of g1

reflects the costs associated with breaking off a reciprocated relation.
2. g2 ~1 2 x ij ! (h x ih x hj : The number of actors through whom i is indi-
rectly related to j; a positive value of g2 reflects that it is easier to
establish a new relation to another actor j if i has many indirect rela-
tions to j via others who can serve as an introduction.
3. g3 x ij wij : The value wij for another actor to whom i has a relation; e.g.,
a negative value of g3 reflects the costs for i associated with breaking
off an existing relation to other actors j with a high value for wij .

7+3+ Intensity Matrix and Simulation

The model that includes an arbitrary rate function l i ~ r, a, x, m!, an objec-

tive function, and a gratification function, still is a continuous-time Mar-
kov chain. The intensity matrix q~ x, y! is still given by (7), now with
qij ~ x! 5 l i ~ r, a, x, m!pij ~u, x!, (37)

where pij now is given by (34).

386 SNIJDERS

Note that it is straightforward to define an algorithm that simulates

this stochastic process. Schematically, this can be done as follows. Sup-
pose that the present time point is t [ @tm , tm11 !. The time until the next
change by any actor is generated by a negative exponential distribution
with parameter (28), provided that the moment so determined is before
time tm11 . The actor who is to change a relation (i.e., the row of the adja-
cency matrix in which a change will occur) is actor i with probability
(29). The other actor with whom actor i will change the relation (column
of the adjacency matrix) is j with probability (34). When j is chosen, ele-
ment x ij is changed into its opposite, 1 2 x ij .

7+4+ Choice of Statistics for Estimation

The use of the method of moments requires also the selection of statistics
that are relevant for the parameters included in the rate and gratification
functions.
A tentative choice for statistics to estimate the parameters ah in
(30) is provided by the total amounts of change weighted by vhi ,
M n
CM1h21 5 ( ( 6X
m51 i, j51
ij ~tm11 ! 2 x ij ~tm !6vhi + (38)
iÞj

To estimate the parameter a1 in (31) for the effect of out-degree on rate of

change, the statistic
M n
CM1H 5 ( ( 6X
m51 i, j51
ij ~tm11 ! 2 x ij ~tm !6 x i1 ~tm ! (39)
iÞj

can be used (where H is the total number of covariates used for modeling
the rate function), and similarly for the effects of the in-degree and the
number of reciprocated relations. These choices are intuitively plausible
and have led to reasonable estimates in some trial data sets, but more
research is required.
For the parameters gh included in the gratification function (36), a
relevant statistic is
M21 n
Rh 5 ( ( 6X
m51 i, j51
ij ~tm11 ! 2 x ij ~tm !6rijh ~ x~tm !!, (40)
iÞj
STATISTICAL EVALUATION OF NETWORK DYNAMICS 387

which is the sum of the rijh values of newly formed relations if rijh con-
tains a factor ~1 2 x ij !, and the sum of rijh values of disappeared relations
if rijh contains a factor x ij .
These statistics CM1h and R h are used in the method of moments in
the same way as (m Smk in (17) and (25).

8. CONTINUATION OF THE EXAMPLE

Continuing the example of the network of university freshmen, the effect

(31) of the out-degrees on the rate of change is included, and the gratifi-
cation function is defined as the sum of the effect of breaking recipro-
cated relations and the effect of gender difference on breaking a relation,

gi ~g, x, j ! 5 g1 x ij x ji 1 g2 x ij 6vi 2 vj 6

where vi indicates the gender of actor i.

The results are given as Model 4 in Table 3. It can be concluded
that the tendency of actors with higher out-degrees to change their rela-
tions more often is close to significance at the 5 percent level ~t 5 0+900
0+47 5 1+91!, and that relations with other actors of the opposite sex are
terminated more quickly than those with others of the same sex ~t 5 1+640

TABLE 3
Parameter Estimates and Standard Errors for Model with Rate
and Gratification Effects

Model 4

Effect par. (s.e.)

Rate (period 1) 5.05
Rate (period 2) 3.95
Out-degree effect on rate 0.90 (0.47)
Density 20.99 (0.20)
Reciprocity 2.82 (0.56)
Indirect relations 20.508 (0.091)
Gender activity 20.52 (0.31)
Gender popularity 0.55 (0.30)
Gender dissimilarity 0.08 (0.37)
Breaking reciprocated relation 20.58 (1.06)
Breaking relation with different-gender other 1.64 (0.62)
388 SNIJDERS

0+62 5 2+65!. The effect of reciprocity on breaking a relation is not differ-

ent from what may be expected from the main reciprocity effect ~t 5
20+5801+06 5 20+55!. Comparing these results with those for Model 3 in
Table 1, it can be concluded that the activity and popularity effect for
gender now are somewhat weaker (having lost their significance at the 5
percent level), and the main gender dissimilarity effect has vanished due
to the inclusion of the effect of gender dissimilarity on breaking a rela-
tion. Thus Model 4 suggests that friendly relations with actors of the oppo-
site sex are less stable, and that there is no evidence (as one might
erroneously conclude from Model 3) that friendly relations are initiated
less with actors of the opposite sex than with those of the same sex.

9. ASYMPTOTIC DISTRIBUTION AND RELATION

WITH THE p * MODEL

If it is possible to reach every state from every given initial state in a finite
number of steps (as is the case here), the distribution of a Markov chain
with stationary intensity matrix on a finite outcome space tends to a unique
limiting distribution as t r `, independent of the initial distribution. For
a certain specification of our model, this limiting distribution is the p *
model for social networks proposed by Wasserman and Pattison (1996),
generalizing the Markov graph distribution proposed by Frank and Strauss
(1986). The p * model is a family of probability distributions for a single
observation x on a stochastic directed graph X. The probability distribu-
tion for the p * model is defined by
exp~ b ' z~ x!!
P$X 5 x% 5 (41)
k~ b!
where z~ x! is a vector of statistics of the digraph and k~ b! is a normaliza-
tion factor. The following proposition indicates a specification for the actor-
oriented model that yields the p * distribution as the limiting distribution.
Proposition 1. Define for all i the objective function by
fi ~ b, x! 5 b ' z~ x! (42)
and the gratification function by gi 5 0. Furthermore, define the rate func-
tion by
n
l i ~ x! 5 ( exp~ b z~ x~i V h!!!+
h51
'
(43)
hÞi
STATISTICAL EVALUATION OF NETWORK DYNAMICS 389

Then the limiting probability distribution of X~t ! for t r ` is the p * dis-

tribution with probability function (41).

Proof. It follows from (34), (37), and (43) that

qij ~ x! 5 exp~ b ' z~ x~i V j !!!+

Note that the symbol x~i V j ! can be understood as the result of taking
matrix x and applying the operation of changing x ij into 1 2 x ij . Applying
this operation twice returns the original matrix x, which can be repre-
sented as ~ x~i V j !!~i V j ! 5 x. Therefore,

qij ~ x~i V j !! 5 exp~ b ' z~ x!!

which implies

exp~ b ' z~ x!!qij ~ x! 5 exp~ b ' z~ x~i V j !!!qij ~ x~i V j !!

and, for Q defined by (7), that

exp~ b ' z~ x!!q~ x, y! 5 exp~ b ' z~ y!!q~ y, x!

for all x, y. In terms of the theory of Markov chains (e.g., Norris 1997,
pp. 124–25), this means that the intensity matrix Q and the distribution
(41) are in detailed balance, which implies that (41) is the stationary dis-
tribution for Q. Since all states communicate with one another, the station-
ary distribution is unique and (41) is also the limiting distribution. Q.E.D.
An interpretation of the rate function (43) is that actors for whom
changed relations have a higher value, will indeed change their rela-
tions more quickly.

10. DISCUSSION

The procedure proposed in this paper provides a method for the analysis
of two or more repeated observations on a social network, in which net-
work as well as covariate effects are taken into account. In view of pro-
cesses in the real-life evolution of social networks, in which endogenous
network effects cumulate continuously over time, the continuous-time
nature of this model will be attractive in many applications. The proce-
dure is available in SIENA (Simulation Investigation for Empirical Net-
work Analysis, available free of charge from http:00stat.gamma.rug.nl0
390 SNIJDERS

snijders0siena.html), which runs under Windows, and is contained in the

StOCNET package (http:00stat.gamma.rug.nl0stocnet).
The present article provides the basic procedure, but this method-
ology could benefit from further elaborations and improvements, for exam-
ple, along the following lines. The algorithm has been proved to work
well in various applications, but it is rather time-consuming and improve-
ments may be possible. A proof of the sufficient condition for its conver-
gence (see the appendix: the eigenvalues of @D021 ]E u Z0]u# should have
positive real parts) is still lacking. The frequency properties of the stan-
dard errors and the hypothesis tests are based on large sample approxima-
tions and should be investigated. The robustness of the proposed estimates
and tests to deviations from the model assumptions is an interesting point
for further study. The method of moments was chosen because of its fea-
sibility, but it may be possible to develop other estimation methods for
this model. As additions to the toolbox, it would be useful to have mea-
sures for goodness of fit and some kind of standardized effect sizes. The
present implementation contains an ad hoc way of dealing with missing
data, which merits further investigation.
Although the model is presented as an actor-oriented model, it uses
an extremely simple and myopic behavioral model for the actors. This
simplicity is a strength because more complicated models for the behav-
ior of actors in a relational network would be more restrictive and less
general in their domain of applicability. On the other hand, for specific
applications it could be interesting to develop statistical network evolu-
tion models incorporating a sociologically more interesting behavioral
model.
Further extensions are possible. An extension to relations with
ordered outcome categories would increase the scope of the model. One
could also think of extending the model to include unobserved heteroge-
neity by means of random effects, but this would lead the model outside
of the realm of complete observations of the state of a Markov process,
and therefore require more complex estimation methods.

APPENDIX: STOCHASTIC APPROXIMATION ALGORITHM

The purpose of the algorithm is to approximate the solution of the moment

equation (12). In this appendix, the solution is denoted by u0 . As men-
tioned in the text above, the algorithm uses the idea of Polyak (1990) and
Ruppert (1988) to employ a diagonal matrix D0 in the iteration step (26)
STATISTICAL EVALUATION OF NETWORK DYNAMICS 391

and estimate the solution by partial averages of uZ N rather than the last
value; and it uses the idea of Pflug (1990) to let the values of a N remain
constant if the average products of successive values ~Z N 2 z!~Z N21 2 z!
are positive, since this suggests that the process still is drifting toward its
limit value. However, the specification used here deviates from Pflug’s
proposal by requiring, for the premature decrease of a N , that for each
coordinate the partial sum of the product of successive values be nega-
tive, rather than requiring this only for the sum over the coordinates. Fur-
thermore, the number of steps for which a N is constant is bounded between
a lower and an upper limit to ensure that a N is of order N 2c .
A crucial condition for the Polyak-Ruppert result about the opti-
mal convergence rate of the partial sums of uZ N to the solution of (12), is
the assumption that all eigenvalues of the matrix of partial derivatives,
~D021 ]E u Z0]u!, have positive real parts; see Yin (1991), Pflug (1996), or
Kushner and Yin (1997). This condition is implied by condition (13) if D0
is the identity matrix. For our model and the proposed statistics used in
the moment estimators, we conjecture that this condition is satisfied, but
the proof is still a matter of further research. Whether the algorithm yields
an estimate that indeed solves the moment equation (12) to a satisfactory
degree of precision is checked in the “third phase” of the algorithm below.
The practical experience with the convergence of the algorithm is, for
most models applied to most data sets, quite favorable.
The reason for incorporating the matrix D0 is to achieve better com-
patibility between the scales of Z and of u. The diagonal elements of D0
are defined as the estimated values of the derivatives ]E u ~Z k !0]uk where u
is at its initial value. To see that this leads to compatibility of the scales of
Z and u, note that in the extreme case where var ~Z k ! 5 0 and the diagonal
elements of D0 are equal to ]E u ~Z k !0]uk , (26) for a N 5 1 is just the itera-
tion step of the Newton-Raphson algorithm applied to each coordinate of
Z separately. Thus, beginning the algorithm with a N in the order of mag-
nitude of 1 will imply that the initial steps have an approximately right
order of magnitude.
The algorithm consists of three phases, which can be sketched as
follows. The number of dimensions of u and of Z is denoted by p and the
initial value is denoted u1 .

• Phase 1: In this phase a small number n1 of steps are made to estimate

D~u1 ! 5 ~]E u ~Z!0]u!6u5u1 , using common random numbers; the diago-
nal elements of this estimate are used to define D0 .
392 SNIJDERS

This is described formally as follows. Denote by ej the j ' th unit

vector in p dimensions. In step N, generate Z N0 ; u1 and Z Nj ; u1 1
ej ej , where all the p 1 1 random vectors use a common random number
stream to make them strongly positively dependent and where ej are
suitable constants. For different N, the random vectors are generated
independently. Compute the difference quotients

dNj 5 ej21 ~Z Nj 2 Z N 0 !;

for small values of ej the expected value of the matrix dN 5 ~dN1 , + + + , dNp !
approximates D~u1 !. However, ej must be chosen not too small because
otherwise the variances of the dNj become too large.
At the end of this phase, estimate Eu1 Z and D~u1 ! by
n n
1 1 1 1
zS 5 (
n1 N51
Z N 0 and DZ 5
n1 N51 (
dN ,

respectively, make one estimated Newton-Raphson step,

uZ n1 5 u1 2 DZ 21 ~ zS 2 z!,

and use the diagonal matrix DE 5 diag~ D! Z in Phase 2.

• Phase 2: This is the main phase, consisting of several subphases. The
number of iteration steps per subphase is determined by a stopping
2
rule, but bounded for subphase k by a minimum value n2k and a max-
1
imum value n2k . In each subphase, a N is constant. The only differ-
ence between the subphases is the value of a N .
1
The subphase is ended after less than n2k steps as soon as the
2
number of steps in this subphase exceeds n2k while, for each coordinate
Z k , the sum within this subphase of successive products ~Z Nk 2 z k ! 3
1
~Z N21, k 2 z k ! is negative. If the upper bound n2k is reached, then the
subphase is terminated anyway.
In each iteration step within each subphase, Z N is generated
according to the current parameter value uZ N . After each step, this value
is updated according to the formula

uZ N11 5 uZ N 2 a N DE 21 ~Z N 2 z!+ (44)

At the end of each subphase, the average of uZ N over this subphase is

used as the new value for uZ N .
STATISTICAL EVALUATION OF NETWORK DYNAMICS 393

The value of a N is divided by 2 when a new subphase is entered.

2 1
The bounds n2k and n2k are determined so that N 304 a N tends to a finite
positive limit.
The average of uZ N over the last subphase is the eventual esti-
mate u.Z
• Phase 3: Phase 3 is used only for the estimation of D~u! and S~u!,
using common random numbers for the estimation of the derivatives;
and as a check for the (approximate) validity of (12). Therefore the value
of uZ N is left unchanged in this phase and is equal to the value obtained
after the last subphase of Phase 2. The simulations and the estimation of
E u Z and D~u! are as in Phase 1. The covariance matrix of Z, required
for the calculation of (21), is estimated in the usual way.

This algorithm contains various constants that can be adapted so as to

achieve favorable convergence properties. Experience with various data
sets led to the following values. The number of steps in Phase 1 is n1 5
7 1 3p. The values of ei are chosen at least 0.1, in most cases 1.0, because
the variability obtained by the use of small values of ei is more serious
than the bias obtained by the use of this large value. The minimum num-
2
ber of steps in subphase 2+k is n2k 5 2 4~k21!03 ~7 1 p! and the maximum
1 2
number is n2k 5 n2k 1 200. The initial value of a N in Phase 2 is 0.2. The
default number of subphases is 4, but fewer or more numerous subphases
can be used to obtain smaller or larger precision. The default number of
steps in Phase 3 is n3 5 500. Phase 3 takes much time because each step
requires p 1 1 simulations; however, the variance estimate is rather unsta-
ble if the number of steps is much smaller.

REFERENCES

Bishop, Yvonne M., Stephen E. Fienberg, and Paul Holland. 1975. Discrete Multivar-
iate Analysis. Cambridge, MA: MIT Press.
Bowman, K.O., and L.R. Shenton. 1985. “Method of Moments.” Pp. 467–73 in Ency-
clopedia of Statistical Sciences, vol. 5, edited by S. Kotz, N. L. Johnson, and C. B.
Read. New York: Wiley.
Doreian, Patrick, and Frans N. Stokman, eds. 1997. Evolution of Social Networks.
Amsterdam: Gordon and Breach.
Fararo, Thomas J., and Norman P. Hummon. 1994. “Discrete Event Simulation and
Theoretical Models in Sociology.” Advances in Group Processes 11:25– 66.
Frank, Ove, and David Strauss. 1986. “Markov Graphs.” Journal of the American
Statistical Association 81:832– 42.
394 SNIJDERS

Holland, Paul, and Samuel Leinhardt. 1977a. “A Dynamic Model for Social Net-
works.” Journal of Mathematical Sociology 5:5–20.
———. 1977b. “Social Structure as a Network Process.” Zeitschrift für Soziologie
6:386– 402.
Karlin, S., and H. M. Taylor. 1975. A First Course in Stochastic Processes. New York:
Academic Press.
Katz, Leo, and Charles H. Proctor. 1959. “The Configuration of Interpersonal Rela-
tions in a Group as a Time-dependent Stochastic Process.” Psychometrika
24:317–27.
Kushner, Herbert J. and George G. Yin. 1997. Stochastic Approximation: Algorithms
and Applications. New York: Springer.
Leenders, Roger Th. A. J. 1995a. “Models for Network Dynamics: A Markovian Frame-
work.” Journal of Mathematical Sociology 20:1–21.
———. 1995b. Structure and Influence. Statistical Models for the Dynamics of Actor
Attributes, Network Structure, and Their Interdependence. Amsterdam: Thesis
Publishers.
Maddala, G. S. 1983. Limited-dependent and Qualitative Variables in Econometrics.
Cambridge, England: Cambridge University Press.
Mayer, T. F. 1984. “ Parties and Networks: Stochastic Models for Relationship Net-
works.” Journal of Mathematical Sociology 10:51–103.
Norris, J. R. 1997. Markov Chains. Cambridge, England: Cambridge University Press.
Pflug, Georg Ch. 1990. “Non-asymptotic Confidence Bounds for Stochastic Approx-
imation Algorithms with Constant Step Size.” Monatshefte für Mathematik
110:297–314.
———. 1996. Optimization of Stochastic Models. Boston: Kluwer.
Polyak, B. T. 1990. “New Method of Stochastic Approximation Type.” Automation
and Remote Control 51:937– 46.
Robbins, H., and S. Monro. 1951. “A Stochastic Approximation Method.” Annals of
Mathematical Statistics 22:400– 407.
Ruppert, David. 1988. “Efficient Estimation from a Slowly Convergent Robbins-
Monro Process.” Technical Report No. 781, School of Operations Research and
Industrial Engineering, Cornell University.
———. 1991. “Stochastic Approximation.” In Handbook of Sequential Analysis, edited
by B. K. Gosh, and P. K. Sen. New York: Marcel Dekker.
Snijders, Tom A. B. 1996. “Stochastic Actor-oriented Models for Network Change.”
Journal of Mathematical Sociology 21:149–72. Also published in Doreian and Stok-
man (1997).
———. 1999. “The Transition Probabilities of the Reciprocity Model.” Journal of
Mathematical Sociology 23:241–53.
Snijders, Tom A. B., and Marijtje A. J. Van Duijn. 1997. “Simulation for Statistical
Inference in Dynamic Network Models.” Pp. 493–512 in Simulating Social Phe-
nomena, edited by R. Conte, R. Hegselmann, and P. Terna. Berlin: Springer.
Van de Bunt, Gerhard G. 1999. Friends by Choice. An Actor-oriented Statistical Net-
work Model for Friendship Networks Through Time. Amsterdam: Thesis Publish-
ers, 1999.
STATISTICAL EVALUATION OF NETWORK DYNAMICS 395

Van de Bunt, Gerhard G., Marijtje A. J. Van Duijn, and Tom A. B. Snijders. 1999.
“Friendship Networks Through Time: An Actor-oriented Statistical Network
Model.” Computational and Mathematical Organization Theory 5:167–192.
Wasserman, Stanley. 1977. “Stochastic Models for Directed Graphs.” Ph.D. disserta-
tion, Department of Statistics, Harvard University.
———. 1979. “A Stochastic Model for Directed Graphs with Transition Rates Deter-
mined by Reciprocity.” Pp. 392– 412 in Sociological Methodology 1980, edited by
K. F. Schuessler. San Francisco: Jossey-Bass.
———. 1980. “Analyzing Social Networks as Stochastic Processes.” Journal of the
American Statistical Association 75:280–94.
Wasserman, Stanley, and Katherine Faust. 1994. Social Network Analysis: Methods
and Applications. New York: Cambridge University Press.
Wasserman, Stanley, and D. Iacobucci. 1988. “Sequential Social Network Data.” Psy-
chometrika 53:261–82.
Wasserman, Stanley, and Philippa Pattison. 1996. “Logit Models and Logistic Regres-
sion for Social Networks: I. An Introduction to Markov Graphs and p * .” Psy-
chometrika 61:401–25.
Yin, George. 1991. “On Extensions of Polyak’s Averaging Approach to Stochastic
Approximation.” Stochastics 36:245– 64.
Zeggelink, Evelien P. H. 1994. “Dynamics of Structure: An Individual Oriented
Approach.” Social Networks 16:295–333.