Multilevel Longitudinal Analysis of Social Networks

Johan Koskinen
University of Melbourne, Melbourne, Australia.
E-mail: [email protected]
Tom A.B. Snijders
University of Oxford, Oxford, United Kingdom; University of Groningen, Groningen, The
Netherlands.
arXiv:2201.12713v1 [stat.ME] 30 Jan 2022

E-mail: [email protected]

Summary. Stochastic actor-oriented models (SAOM) are a broadly applied modelling

framework for analysing network dynamics using network panel data. They have been
extended to address co-evolution of multiple networks as well as networks and behaviour.
This paper extends the SAOM to the analysis of multiple network panels through a random
coefficient multilevel model, estimated with a Bayesian approach. This is illustrated by a
study of the dynamic interdependence of friendship and minor delinquency, represented
by the combination of a one-mode and a two-mode network, using a sample of 81 school
classes in the first year of secondary school.

Keywords: Stochastic Actor-oriented Model; Random Coefficient Model; MCMC;

Social Influence; Delinquency; Two-mode network

1. Introduction

Social network research deals with analysing the dependencies among people or other social
units, dependencies induced by the relational ties that bind them together (Wasserman and Faust,
1994; Brandes et al., 2013; Robins, 2015). These dependencies can best be studied in a dynamic
approach, where the existence of a given configuration of ties leads to the creation, or supports
the maintenance, of other ties. While many of the endogenous network dependencies, like
triadic closure and balance, are of interest in their own right, there is a growing interest in the
dynamic interdependence of networks with other structures, such as actor variables (Veenstra
et al., 2013), other networks for the same actor set (Huitsing et al., 2014; Elmer et al., 2017), or
two-mode networks (Lomi and Stadtfeld, 2014).
Dynamic network data can be of various kinds. A frequently followed design is the collection
of network panel data, i.e., the observation of all relational ties (in one or more networks) and
other relevant variables, within a given group of social actors (such as individuals, firms, coun-
tries, etc.), at two or more moments in time, the ‘panel waves’. For modelling panel data for a
single network, represented by a digraph, the Stochastic Actor-oriented Model (‘SAOM’) was
proposed by Snijders (2001). This was extended to a joint model for changing actor variables
2 Johan Koskinen and Tom A.B. Snijders
(vertex attributes) and tie-variables by Steglich et al. (2010) and to a model for the interde-
pendent dynamics of multiple networks, potentially combinations of one-mode and two-mode
networks, by Snijders et al. (2013). These joint dynamic models can be combined under the
heading of ‘co-evolution’, as summarized in Snijders (2017).
Collecting longitudinal network data is very time-intensive and demands great care, but data
sets of longitudinal networks in many ‘parallel’ groups are becoming increasingly common; ex-
amples (among many others) are the study ‘Networks and actor attributes in early adolescence’
executed by Chris Baerveldt and Andrea Knecht which will be used in this paper (Knecht, 2006;
Knecht et al., 2010), and CILS4EU (Kalter et al., 2013).
While the SAOM has proved useful in analysing networks in single groups, the methodology has
been limited in studying the extent to which network dynamics generalise to different contexts
and what might differ systematically across groups of actors. The investigation of heterogeneity
across groups more generally, in the way multilevel models have proven useful (e.g., Goldstein,
2011; Snijders and Bosker, 2012), has not been possible. However, to find scientific regularities
it is more attractive to study multiple groups that may be regarded as a sample from a population
and to generalize to populations of networks (Snijders and Baerveldt, 2003; Entwisle et al.,
2007). For the Exponential Random Graph Model a multilevel methodology was proposed by
Slaughter and Koehly (2016) (see also Schweinberger et al., 2020).
This paper proposes a multilevel extension of the SAOM for data sets composed of disjoint
groups of actors, for which only networks within each group are considered. The actors are
nested within the groups. Since ties combine pairs of actors, the combined structure of actors
and ties cannot be regarded as being nested. This extension employs random coefficients like
the multilevel models mentioned above and draws on the likelihood-based estimation frame-
works of Koskinen and Snijders (2007) and Snijders et al. (2010). It also permits the investi-
gation of observable group-level variables, such as compositional and contextual factors, like
in standard multilevel modelling. Our example is a co-evolution of friendship networks and
delinquent behaviour represented by two-mode networks, therefore the elaboration focuses on
the co-evolution model of Snijders et al. (2013).
Combinations of networks are occasionally refered to as ‘multilayer networks’ (Kivelä et al.,
2014; Magnani and Wasserman, 2017) or ‘multilevel networks’ (Snijders, 2016), but in this pa-
per we use the term ‘multilevel networks’ to express the link to the random coefficient multilevel
models in the sense mentioned above.

2. Friendship and delinquency

As the motivating example, we consider the dynamic relation between friendship and delinquent
behaviour, using the study ‘Networks and actor attributes in early adolescence’. The data set
was collected by Andrea Knecht, supervised by Chris Baerveldt (Knecht, 2006). The data was
collected in 126 first-grade classrooms in 14 secondary schools in The Netherlands in 2003-
2004, using written questionnaires. The entire data set contains four waves with about three
months in between.
We focus on the friendship network and on the four questions about delinquency: stealing, van-
dalism, graffiti, and fighting, for each of which self-reported frequencies were given with five
 Multilevel Longitudinal Social Networks 3
categories. Written self-reports provide reliable measurements of delinquency for adolescents
(Köllisch and Oberwittler, 2004). The dynamic relation between a network such as friendship
and a changing actor variable such as the tendency to commit delinquent behaviour has two
sides: selection, changes of friendships as a function of the delinquent behaviour of the two in-
dividuals concerned; and influence, changes in delinquent behaviour of an actor as a function of
the network position of this actor and the delinquent behaviour of the others, especially those to
whom this actor has a friendship tie. A methodology to distinguish between selection and influ-
ence, using network and behaviour panel data, based on the SAOM, was proposed by Steglich
et al. (2010). The conclusions are not causal in the counterfactual sense, as demonstrated by
Shalizi and Thomas (2011), but in a temporal sense: does a change in behaviour follow on
some network configuration (‘influence’), or does a change in friendship follow on a behaviour
configuration (‘selection’). A further discussion of causality in network-behaviour systems was
given by Lomi et al. (2011).
The association between friendship and the tendency to delinquent behaviour was studied by
Knecht et al. (2010). This publication used the data set mentioned above, constructing an actor
variable representing delinquent behaviour as a sum score of the four items for the frequencies
of stealing, vandalism, graffiti, and fighting. It used the two-step multilevel method of Snijders
and Baerveldt (2003), in which first the SAOM is estimated for each classroom separately, after
which the results for the classrooms are combined. Since most of the classrooms were too small
for the satisfactory application of this — rather complicated — model, only 21 classrooms could
be used.
In the current paper we present an extension of this study, replacing the simplistic two-step
multilevel approach by an integrated random coefficient approach, which does not depend on
the condition of a convergent estimation algorithm for each classroom separately and therefore
can use a much larger part of the data set. Furthermore, we replace the model where delinquent
behaviour is represented by an actor variable with a model representing the four delinquency
items by a two-mode network. This allows a more detailed study of social influence. The ac-
tors are supposed to be influenced by their friends, which are those they mention as a friend
(friendship ties from the actor to the friends). In the former study, the tendency toward delin-
quent behaviour was regarded as a one-dimensional trait, measured by the sum score of the four
delinquent items; social influence was represented by the effect of the average of this trait over
the actor’s friends. The current study considers this together with another type of influence: the
effect of the friends’ behaviour for some specific delinquent behaviour on the same behaviour
of the actor.

3. Multilevel stochastic actor-oriented model

The stochastic actor-oriented model (Snijders, 2017) is a family of longitudinal network models
for network panel data. While networks are only observed at discrete time points, the model
assumes that the networks evolve in continuous time. This is necessary for representing the
feedback between the tie variables that can occur in the time elapsing between the observation
moments. Some history of continuous-time models for social network panel data is presented
in Snijders (2001). Continuous-time models for discrete-time panel data are well known (e.g.,
4 Johan Koskinen and Tom A.B. Snijders
Bergstrom, 1988; Hamerle et al., 1993; Singer, 1996). Their use for network panel data in
sociology is argued also by Block et al. (2018).

3.1. Data structure

We assume that we have panel network data for G independent groups. The groups are a col-
lection of mutually exclusive fixed sets of nodes N1 , . . . , NG , with time-dependent one-mode
networks for each of them. In our example, these nodes represent individuals and the network
represents the friendships among them. We assume that there may only be network ties between
nodes in the same node sets, and at any point in time t, the network in group Ng is represented
[g] [g]
by a binary adjacency matrix X [g] (t) = (Xij (t))(i,j)∈Ng ×Ng , where Xij (t) = 1 if there is a
tie from i to j at time t, and zero otherwise. Self-ties are excluded. In addition, we have two-
mode networks with a common second-mode node set H, which here is the set of the H = 4
[g]
delinquency behaviours. The delinquency behaviours are dichotomized, and Zih (t) indicates
whether individual i in group g engages in behaviour h at time t. These two-mode tie variables
are collected in a matrix Z [g] (t). Jointly we denote the one-mode and two-mode network by
Y [g] (t) = (X [g] (t), Z [g] (t)), for g = 1, . . . , G. The supports of X [g] and Z [g] are denoted Xg
and Zg , respectively, with joint support Yg = Xg × Zg .
For the data, we assume that Y [g] (t) is observed at discrete points in time, t0 , t1 , . . . , tM , where
M can be as small as 2. The inferential target is to model how Y [g] (tm−1 ) changed into Y [g] (tm )
for m = 1, . . . , M − 1.

3.2. Model specification for a single group

The model for the SAOM in a single group can be described without the notational dependence
on the group membership, as one-mode ties are only defined within groups. Therefore we drop
the superscript [g]. The process is actor-oriented in the sense that transitions in the process are
modelled as choices by actors i ∈ N to change outgoing tie variables Xij or Zih . It is assumed
that Y (t), t1 ≤ t ≤ tM , given the available covariates, is a Markov process in continuous time.
We present the SAOM for the case of co-evolution of a one-mode and a two-mode network;
this can be generalized to more networks and to co-evolution with behavioural variables, see
Snijders (2017).
At any moment t in continuous time, at most one actor i may make a change in at most one tie
variable Xij or Zih ; this can be creation of the tie (0 to 1) or termination (1 to 0). This restriction
was proposed already by Holland and Leinhardt (1977), and it implies that the dynamic model is
decomposed in the smallest possible changes; these changes are called mini-steps. Basic ingre-
dients of the model are rate functions λX Z
i (θ, y) and λi (θ, y) which indicate the rates at which
actor i gets an opportunity, respectively, to change some one-mode tie Xij (j ∈ N , j 6= i) or
to change some two-mode tie Zij (j ∈ H); and evaluation functions fiX (θ, y) and fiZ (θ, y)
indicating the value, as it were, that actor i attaches to state y of the combined networks when
making, respectively, a change in network X or in network Z . The rate functions define the
expected frequency of the mini-steps and the evaluation functions define the probability distri-
bution of their results. For simple models the number of opportunities has a Poisson distribution.
 Multilevel Longitudinal Social Networks 5
Since the choice situations with respect to the one-mode network (friendship) and the two-mode
network (delinquency behaviour) are different, different considerations for the actors may apply,
and the evaluation functions fiX (θ, y) and fiZ (θ, y) will not be the same.
By the properties of the exponential distribution, the time until the first opportunity for change
of any kind by any actor is exponentially distributed with rate
X
λ+ λX Z


+ (θ, y) = i (θ, y) + λi (θ, y) ,
i∈N

and the probability that actor i ∈ N is selected for changing a tie variable in V ∈ {X, Z} is

λVi (θ, y)
.
λ++ (θ, y)

Given that i is selected for making a change in network V , the option set consists of all outgoing
tie variables in network V , together with the option ‘no change’. The set of outcomes reachable
in a mini-step by actor i in network V is denoted AVi (y), with
0 0 0
AX
i (x, z) ⊆ {(x , z) ∈ Y : ||x − x || ≤ 1, xkj = xkj , ∀j and ∀k 6= i}

and
0 0 0
AZ
i (x, z) ⊆ {(x, z ) ∈ Y : ||z − z || ≤ 1, zkh = zkh , ∀h and ∀k 6= i} .
Here ||B − C|| denotes the Hamming distance between adjacency matrices B and C . Usually
the subset “⊆” will be implemented as equality “=”, but the subset symbol is used because
there could be constraints on the state space, such as in the case of changing composition or
absorbing states.
Conditionally on y , and on i being selected to make a change in network V , the probability that
the outcome of the choice is y 0 is

exp fiV (θ, y 0 ) exp fiV (θ, y 0 ) − fiV (θ, y)

V 0
pi (θ, y, y ) = P
= P
(1)
ỹ∈AV
i (y)
exp fiV (θ, ỹ) ỹ∈AV
i (y)
exp fiV (θ, ỹ) − fiV (θ, y)

if y 0 ∈ AVi (y), and 0 if y 0 6∈ AVi (y). Note that since y ∈ AVi (y), the probability of no change,
i.e., y 0 = y , is positive.

3.2.1. Interpretation of process

For notational convenience, we further use the symbol y instead of y 0 in the role of outcome of
the mini-step. Typically, the evaluation functions fiV (θ, y) are modelled as weighted functions
of statistics calculated on y , X
fiV (θ, y) = θkV sVki (y) .
k

The statistics sVki (y)

are briefly called effects, and will be functions pertaining to actor i and
the network neigbourhood of i. Usual effects sVki (x, z) are counts of subgraphs (configurations)
6 Johan Koskinen and Tom A.B. Snijders
that include ties originating with actor i. Since no information is available on the timing of
the mini-steps, the focus of modeling is on the evaluation functions and not on the rate func-
tions (an exception is the diffusion model of Greenan, 2015). Often the rate functions λXi (θ, y)
and λZ i (θ, y) are chosen to be constant between observation moments, and that is what will
be assumed
further on. If the evaluation function fiX θ, (x, z) does not depend on z and

fiZ θ, (x, z) does not depend on x, the dynamics of the one-mode and two-mode networks
are independent. In our example the interest is in the interdependence between friendship and
delinquent behaviour, which is reflected by statistics that depend on both networks jointly.
The model can be interpreted as a sequential discrete-choice model where actors make choices
about their outgoing ties, using random utilities (Maddala, 1983), under the restriction that they
can change no more than one outgoing tie variable. From that perspective the model can be
interpreted as a process whereby actors chose to change their network ties or their behaviour to
what they deem most preferable, allowing for a random element in their decisions. The model
does not strictly require this interpretation and Snijders (2017) treats a wide variety of different
model specifications, including differential treatments of creating and terminating ties, more
elaborate specifications of the rate functions, and options for non-directed networks.
Of particular importance are cross-network effects sX Z
ki (x, z) and ski (x, z) depending on x as
well as z , reflecting the mutual dependence between the one-mode and the two-mode network.
In our application, where the networks are friendship and delinquent behaviours, the following
cross-network effects are used. As mnemonic indicators, we use ‘o’ for outgoing friendship ties,
’i’ for incoming friendship ties, and ’d’ for ties in the delinquency network. The subgraphs used
are illustrated in the pictograms, where nodes of the first mode are denoted by circles, nodes
of the second mode by squares, one-mode ties by straight arrows, and two-mode ties by curly
arrows. The superscript V indicates that the effect applies to V = X as well as V = Z . Note
that effects sVki refers to actors i, who consider changing some outgoing tie in network V . In the
pictograms, the parts with a tie i → j have the role of dependent variables for friendship, and
the parts with a tie i h have the role of dependent variables for delinquency.

(a) od: the product of the number of outgoing friendships and the number of delinquent be-
haviours of i,
X X
sVod,i (x, z) = xij zih . h i j
j h

(b) id: the product of the number of incoming friendships and the number of delinquent be-
haviours (note the exchange of i and j ),
X X
sX
id,i (x, z) = xij zjh , h j i
j h
X X
sZ
id,i (x, z) = xji zih . h i j
j h

(c) odd: a mixed triadic effect: the number of friendships of i weighted by the number of
 Multilevel Longitudinal Social Networks 7
delinquent behaviours i and j have in common,

h
X
sVodd,i (x, z) = xij zih zjh .
j,h i j

(d) od_av: a mixed four-node effect that is not a subgraph count: the total number of delin-
quent behaviours reported by i multiplied by the average number of delinquent behaviours,
centered, reported by all i’s friends,
P  P

j x ij z j` − z̄
P `
X
sZ
od_av,i (x, z) = zih ,
h j xij

h i j `

where z̄ is the average observed outdegree for Z in the group. Here 0/0 is defined as 0.

Effect ‘od_av’ is used only for explaining the dynamics of the Z network, the other three are
used for explaining the dynamics of both networks. Brief interpretations of these effects, for
positive parameter values, are the following.
For explaining the friendship dynamics (‘selection’):

(a) The ‘od’ effect indicates that those who engage in more delinquent behaviours will be
more active in nominating friends.

(b) The ‘id’ effect indicates that those who engage in more delinquent behaviours will be more
popular as friends.

(c) The ‘odd’ effect indicates that actors will tend to be friends with those who engage in the
same delinquent behaviours.

And for explaining the delinquency dynamics (‘influence’):

(a) The ‘od’ effect indicates that those who nominate more friends will tend to engage in more
delinquent behaviours.

(b) The ‘id’ effect indicates that those who are more popular as friends will tend to engage in
more delinquent behaviours.

(c) The ‘odd’ effect indicates that actors will tend to engage in the same delinquent behaviours
as their friends.

(d) The ‘od_av’ effect indicates that those whose friends on average are more delinquent will
also themselves tend to engage in more delinquent behaviours.
8 Johan Koskinen and Tom A.B. Snijders
The last two effects (‘odd’ and ‘od_av’) are the most clear expressions of the idea of social
influence, both implying that the probability distribution of changes in delinquent behaviour of
the actor is a function of the delinquent behaviour of the actor’s friends. Effect ‘odd’ is social
influence operating for specific acts of delinquent behaviour, while ‘od_av’ is a generalized
influence at the level of the sum scores of delinquency.

3.3. Data augmentation

The SAOM with rates (λVi (θ, y)) and one-step jump probabilities (pVi (θ, y, y 0 )) defines a dis-
crete Markov chain in continuous time with intensity matrix defined for y 6= y 0 , and V ∈
{X, Z}, by
λi (θ, y) pVi (θ, y, y 0 ) if y 0 ∈ AVi (y)
 V
0
q(y, y ) = (2)
0 otherwise.
The process can be defined as a marked point process. Only in trivial cases, such as the random
walk on a |Y|-cube (Aldous, 1983), is Bayesian inference for such models tractable (Koskinen
and Snijders, 2007). For two waves of observations y(tm ) and y(tm+1 ), the likelihood is a |Y|
times |Y| matrix
P T = eT Q ,
for T = tm+1 − tm , which is huge. The model is doubly intractable given that both the likeli-
hood and the posterior involve intractable normalising constants. Index the mini-steps by r =
1, . . . , R (where R is random), and denote the results of the mini-steps by v r = (ir , V r , y r ) and
the holding times by (sr ). Koskinen and Snijders (2007) propose to augment data by performing
joint inference over the model parameters θ as well as the unobserved sequences (v r ) and (sr ).
The sequence (ir , V r , y r ) must be such that if y r differs from y r−1 it is only in variable V r and
row ir of the adjacency matrix. The augmented data likelihood, conditional on y 0 =
y(tm ), for
a sequence of holding times (sr ) and results of mini-steps v = (v r ) = (ir , V r , y r ) , is given
by
R
( )
X
p∗AUG (v r ), (sr ) | y 0 , θ = exp − sr λ+ r−1


+ (θ, y )
r=1
R
Y
× λVirr (θ, y r−1 ) pVirr (θ, y r−1 , y r ) .
r=1

It is more efficient to work with the marginal model pAUG v | y(tm ) which is p∗AUG v, s | y(tm )

marginalised over holding times s. In the sequel we will assume constant rates λVi = λV for
both networks V = X, Z , in which case the augmented likelihood is

pAUG (v r ) | y 0 , θ = exp − λ+



+ (tm+1 − tm ) (3)
R R
λ+ (tm+1 − tm ) Y  λVr  Vr


× + p (θ, y r−1 , y r ) ;
R! λX + λZ ir
r=1

see Snijders et al. (2010) where also an approximation for non-constant rates is given.
 Multilevel Longitudinal Social Networks 9
The Markov assumption
implies that the likelihood for a sequence
of augmented data v =
v(t1 ), . . . , v(tM ) , given observation y = y(t0 ), y(t1 ), . . . , y(tm ) is
M
Y −1


pAUG (v | y, θ) = pAUG v(tm+1 ) | y(tm ), θ . (4)
m=1

The model pAUG v|y(tm−1 ), θ when marginalised over all paths v that start in y 0 = y(tm−1 )

and end in y R = y(tm ) is the data likelihood per wave

pSAOM y(tm )), θ
with the obvious extension for multiple waves.

4. Hierarchical model

We assume that each group g follows the same specification, i.e., has the same expressions for
the rate and evaluation functions, although the number of actors ng = |Ng | may be different.
Each group g has associated with it a group-specific parameter θ[g] . Heterogeneity across groups
typically takes the form of contextual and compositional effects.
While comparing structure across networks is a natural thing to do and has attracted some
attention (e.g., Faust and Skvoretz, 2002), it is clear that comparing structure across different-
sized networks is non-trivial (Anderson et al., 1999). One key problem is the way the average
degree scales with network size, something that has been studied for cross-sectional networks
(Erdős and Rényi, 1960; Krivitsky et al., 2011; Shalizi and Rinaldo, 2013). We assume that the
variation in group sizes ng as well as in average degrees is limited. Based on a combination
of Krivitsky
P et al. (2011) and Snijders (2005, p. 243) we suggest that including an effect of
log(ng ) j xij will make the other parameters comparable, and that for this effect a parameter
of −1/2 would be expected if none of the other parameters reflects differential group sizes.
Components of θ[g] that are variable across g are similar to random slopes in regular multilevel
modeling (Goldstein, 2011; Snijders and Bosker, 2012). The question of whether to allow all
group-level parameters to vary across groups needs to be guided by specific case considerations
as well as computational aspects just as in multilevel models in general. We partition the pa-
rameter vector θ[g] for group g into subvectors γ [g] , of dimension p1 , containing the variable
parameters, and η , of dimension p2 , containing the constant parameters. We write the group-
wise parameters as the partitioned vector
 [g] 
[g] γ
θ = .
η

When p1 = 0 we have the so-called multi-group model (Ripley et al., 2021, Section 11.2). In
classical multilevel modeling it is usual to apply models with only a few random slopes. How-
ever, it seems that Bayesian estimation allows entertaining models with more random slopes
(Eager and Roy, 2017). For group-level covariates, such as interventions or indicators of group
composition, it is natural that their effects are fixed.
10 Johan Koskinen and Tom A.B. Snijders
We draw on standard hierarchical modelling approaches and assume that the group-level pa-
iid
rameters have a multivariate normal distribution γ [g] ∼ Np1 (µ, Σ). We assume that (µ, Σ) and
η are a priori independent with priors (µ, Σ) ∼ π(µ, Σ | Γ) and η ∼ π(η | µ0,η , Σ0,η ).
An exception to this should be made for the rate parameters λ, which are necessarily positive.
They reflect particular circumstances of groups and issues of study design, and will always be
included among the variable parameters γ [g] . The multivariate normal distribution is assumed
to be truncated to positive values for these parameters. The values of µ and Σ will in practice
be such that the non-truncated distribution has an extremely small probability for negative rate
parameters. An alternative is to employ a transformed normal or a Gamma distribution, which
is conjugate for the Poisson counts (Koskinen and Snijders, 2007). However, the multivariate
normal gives a simple unified treatment for all varying parameters.
With this hierarchical specification, denoting the multivariate normal density by φ, the joint
probability density function for data y [1] , . . . , y [G] , parameters γ [1] , . . . , γ [G] , and µ, Σ, η is given
by
YG
φ(γ [g] | µ, Σ) pSAOM (y [g] | γ [g] , η) .


π µ, Σ | Γ π(η|µ0,η , Σ0,η ) (5)
g=1

5. Prior specifications

We present the inference scheme for a specific choice of priors. Other prior specifications may
be considered (see Appendix B) but the MCMC scheme largely remains unchanged.

5.1. Varying parameters: conjugate prior

For multivariate normal distributions with unknown expected value µ and covariance matrix Σ,
the conjugate prior distribution is the inverse Wishart distribution for Σ, and conditional on Σ
for µ a multivariate normal distribution:

• Σ ∼ InvWishartp (Λ0 , ν0 ), and conditionally on Σ

• µ | Σ ∼ Np (µ0 , Σ/κ0 ) .

This is treated, e.g., in Gelman et al. (2014), Section 3.6, and O’Hagan and Forster (2004),
Chapter 14. Thus, the hyper-parameters of the prior are Λ0 , ν0 , κ0 . The expected value for the
inverse Wishart(Λ, ν ) distribution is
 1
E Σ = Λ
ν−p−1
provided ν > p + 1, and the mode is (ν + p + 1)−1 Λ (O’Hagan and Forster, 2004). Thus,
the central tendency of the inverse Wishart(Λ, ν ) distribution may be taken to be about ν −1 Λ.
Parameter Λ is on the scale of the sum of squares of a sample of size ν from a distribution with
variance-covariance matrix Σ. The number of degrees of freedom ν0 can be regarded as the
 Multilevel Longitudinal Social Networks 11

Λ0 ν0
µ0 κ0

Σ µ Σ0,η µ0,η

η
γ [1] γ [G]

[1] ... [1] [G] ... [G]

vt1 vtM vt1 vtM

[1] ... [1] [G] ... [G]

Yt1 YtM Yt1 YtM

Fig. 1. Dependence structure of hierarchical SAOM, representing only the first and last groups
g = 1, G.

effective sample size that has led to the prior information. The value of κ0 can be interpreted
as the proportionality between Σ, the uncertainty about the groupwise parameters γ [g] given the
average population value µ, and the prior uncertainty about µ. Having the same proportionality
of this kind for all parameters is rather restrictive, but as a first approach we prefer to use a
conjugate prior which leads to relatively simple procedures for this already complicated model.

5.2. Constant parameters

For most components of the group-constant parameter η we assume an improper prior with
constant density π(η) ∝ c. This is justified because for the estimation of η the information
from all groups is combined, leading for η to a quite weak dependence on the prior. However,
for effects of group-level covariates the situation is different, and for those components of η a
multivariate normal prior distribution will be assumed.

6. Estimation

The dependence structure amongst all variables is given in Figure 1. Parameters can be es-
timated by an MCMC procedure, sampling the random variables indicated by the circles in
Figure 1, going up in the figure. The parameters in rectangular boxes are given hyperparame-
ters.
12 Johan Koskinen and Tom A.B. Snijders
Mini-steps
For all groups g independently, sequences v [g] of outcomes of mini-steps (ir , V r , y r ) are sam-
pled by an extension of the Metropolis-Hastings procedures of Koskinen and Snijders (2007)
and Snijders et al. (2010). The extension consists of the insertion of the determination of V r .
The target probability function is (4) for given y = y [g] and θ = (γ [g] , η).
 Multilevel Longitudinal Social Networks 13
Groupwise varying parameters
Groupwise varying parameters γ [g] are sampled for given v [g] and η, µ, Σ, again for all groups
g independently, by Metropolis Hastings steps with target density

φ(γ [g] | µ, Σ) pAUG (v [g] | γ [g] , η) .

Here φ is the multivariate normal density and pAUG was given in (4). A random walk proposal
distribution is used, like in Schweinberger (2007, Ch. 5.4) and Koskinen and Snijders (2007,
Section 4.4). The covariance matrix for the proposals is C [g] as defined below in the section on
initial values, scaled to obtain approximately 25 % acceptance rates (Gelman et al., 1996).
Constant group-level parameters

The constant parameter η with prior density π(η | µ0,η , Σ0,η ) can be sampled in two ways, both
using Metropolis Hastings steps analogous to the sampling of the groupwise varying parameters.
The first way draws random walk proposals for η with additive perturbations from the multivari-
[0]
ate normal distribution with mean 0 and covariance matrix Cη given below, scaled to obtain
approximately 25 % acceptance rates. The target distribution is
G
Y
π(η | µ0,η , Σ0,η ) pAUG (v [g] | y [g] , γ [g] , η) .
g=1

The second way draws random walk proposals for additive changes in the entire vectors γ [g] , η ,


excluding the basic rate parameters. Now the perturbations come from the multivariate normal
distribution with mean 0 and covariance matrix C [0] given below, again scaled to obtain approx-
imately
25 % acceptance rates. The proposal is to add this perturbation identically to the vectors
θ[g] , η for all j . The target distribution is
G
Y
π(η | µ0,η , Σ0,η ) φ(γ [g] | µ, Σ) pAUG (v [g] | y [g] , γ [g] , η) .
g=1

Global parameters

Given realisations of the varying group-level parameters γ [1] . . . , γ [G] , global parameters µ and
Σ can be updated using Gibbs-sampling steps from the full conditional posteriors, as explained
in Gelman et al. (2014), Section 3.6, and O’Hagan and Forster (2004), Chapter 14. The condi-
tional distribution of µ given γ [1] . . . , γ [G] , Σ is given by
 
[1] [G] G κ0 1
µ | Σ, γ , . . . , γ ∼ Np γ̄ + µ0 , Σ
κ0 + G κ0 + G κ0 + G

with γ̄ = (1/G) g γ [g] , in which we recognize the posterior mean as a weighted sum of the
P
group-level parameters and the prior mean.
14 Johan Koskinen and Tom A.B. Snijders
For the posterior variance-covariance matrix of γ [g] we have

Σ | γ [1] , . . . , γ [G] ∼ InvWishartpγ (Λ1 , ν0 + G),

where
κ0 G
Λ1 = Λ0 + Q + (γ̄ − µ0 )(γ̄ − µ0 )0 ,
κ0 + G
G
X
Q = (γ [g] − γ̄)(γ [g] − γ̄)0 .
g=1

The influence of the prior is mainly carried by Λ0 and the last term of Λ1 , which involves κ0
and µ0 . Since the central tendency of the inverse Wishart(Λ, ν ) distribution is about ν −1 Λ, this
shows that the posterior distribution of Σ for large values of G will be close to the variance-
covariance matrix of the γ [g] .
Combining the updates

Sequentially the within-group ministeps v , the group-level parameters γ , and the global param-
eters η, µ, Σ are updated. To achieve good mixing, more updates are required for v than for the
other parameters.
Initial values

Initial values are obtained in a procedure consisting of two stages. First, parameters are esti-
mated for the model where all parameters in θ[g] that are coefficients in the linear predictor are
assumed to be constant across groups, but the basic rate parameters are allowed to be group-
dependent, i.e., a multi-group model. This estimation uses the Robbins-Monro algorithm pro-
posed for obtaining method-of-moments estimates in Snijders (2001), in a brief version because
great precision is not necessary here. This yields an estimated value θ̂(0) , with estimated covari-
ance matrix C [0] . The components of this vector and matrix corresponding to η are denoted η̂ (0)
[0]
and Cη .
Second, for each of the groups g separately, starting from the provisional estimate θ̂(0) , and
keeping the components η̂ (0) constant, a small number of Robbins-Monro steps again following
Snijders (2001) are taken to improve the estimate of θ[g] . The result is used as initial value for
(1)
θ̂[g] . The covariance matrix Cg for the proposal distribution for θ[g] is a weighted combination
of the covariance matrix for this estimate and C [0] .

7. Data and model definition

Data were collected in the first year of secondary school in 14 schools in the Netherlands in
2003-2004, with students being on average slightly older than 12 years at the first wave. There
were four waves, with three months in between. Allowing for the social processes to be unstable
 Multilevel Longitudinal Social Networks 15
at the very start of the school year, we used the last three waves. These will be called waves 1-3
from now on, which yields period 1 as the period from wave 1 to wave 2 and period 2 as the
period from wave 2 to 3. Network X was the friendship network, Z the two-mode network of
delinquent behaviours with four second-mode nodes: stealing, vandalism, graffiti, and fighting.
Covariates used were sex (female=1, male=2), language spoken at home, and advice. The Dutch
secondary school system is tiered and ‘advice’ here is defined as the recommended secondary
school level according to the advice given in the last grade of primary school. It is ordered from
low to high with range 1–9.
A basic measure for network stability in a period is the Jaccard coefficient (Batagelj and Bren,
1995), defined for network X and period m as
P
min{xij (tm ), xij (tm+1 )}
P ij
ij max{xij (tm ), xij (tm+1 )}

and for Z similarly.

Delinquency was dichotomized to construct the two-mode network. The coding was zih = 1 if
individual i answered having done the behaviour at least once in the past three months, but for
fighting the threshold was ‘at least twice’ because apparently this was rather common.
Of the original set of 126 classrooms, the criteria for including the classrooms were having less
than 20% missing data in the first two waves for both networks, but less than 10% in the first
wave for the delinquency network; having at least 10 persons with non-missing advice; and
having Jaccard coefficients higher than 0.2 for both networks and both periods. Furthermore,
one group was excluded because it was considered an outlier with a density for the delinquency
network of more than 0.50 for all waves. This leaves 81 groups.

7.1. Model specification

The mutual dependence between friendship and delinquent behaviour was represented by the
effects discussed in Section 3.2.1. Here we discuss the effects operating only on the friendship
and those operating only on the delinquency network. For the mathematical definition of the
effects we refer to Appendix A.
The structural part of the model for friendship dynamics was defined in accordance with what
is usual for friendship networks. The outdegree is a necessary effect, representing the balance
between creation and termination of ties. Reciprocity and transitive triplets effects were in-
cluded together with their interaction following Block (2015). As degree effects were included
outdegree-activity, indegree-popularity, and reciprocal degree-activity; for the latter a negative
parameter is expected, reflecting that actors with more reciprocated friendship ties will tend to
create fewer new ties. For the covariates we included homophily effects with respect to sex,
language, and advice, expecting positive parameters. Furthermore, the logarithm of group size
was included to account for group size differences, where a parameter in the neighbourhood of
−0.5 was expected.
For the delinquency network effects included were the outdegree effect, outdegree-activity and
indegree-popularity reflecting, respectively, differences between students and between delin-
quent activities, and effects of sex, advice, and classroom mean advice.
16 Johan Koskinen and Tom A.B. Snijders
For this multilevel network model with Bayesian estimation, it was mentioned above that it
is possible to specify fairly many parameters as randomly varying between groups, but not
too many. In any case, the rate parameters must vary randomly between groups. A moderate
number of random effects was chosen. Random effects were given to outdegree, reciprocity,
indegree-popularity, reciprocal degree-activity, transitivity, same language, and similar advice
for the friendship network; and to outdegree and outdegree-activity for the delinquency network.

7.2. Prior specification

For the rate parameters a data-dependent normal prior was used, with means and covariance
matrices given by the robust mean and 0.5 times the covariance matrix of the rate parameter
estimates in the multi-group estimation.
For the parameters of the evaluation function, the determination of the prior distribution was
based on existing experience with modeling friendship networks, together with the desire not
to influence the results too strongly, while still obtaining convergence of the MCMC process.
It should be noted that non-zero prior means chosen for structural parameters below have little
influence on the results. The evidence for reciprocity, for example, is typically strong enough to
overwhelm the prior (see Appendix B for a brief illustration) and the performance of the MCMC
is typically not contingent on a strong prior. Naturally, it would be unwise to choose a strong
informative prior for any parameter that is the main target of inference.
The effects used in this example all are scaled in such a way that their parameters have sizes
usually between −1 and +1, except for the outdegree parameter which is negative, reflecting the
sparsity of the networks, and reciprocity, which often has a parameter between +1 and +3. This
implies a prior uncertainty of the global means µ with a standard deviation of approximately 1;
for the outdegree parameters the prior uncertainty is larger. Furthermore, the groups will tend
to be similar to each other, which we express by the prior expectation that the between-group
standard deviations are 10 times smaller than the prior standard deviations for the elements of
µ. This is reflected by the value κ0 = 0.01.
These considerations led to prior means of −2 for the outdegree parameters, +1 for recipro-
city, +0.2 for transitive triplets, and 0 for all other coefficients of random effects. For the
13-dimensional prior Wishart distribution, ν0−1 Λ0 was chosen as a diagonal matrix with diago-
nal values 0.01 except for the two outdegree parameters, which had value 0.1; and number of
degrees of freedom ν0 = 15.
For the fixed effects, improper constant prior distributions were used for all except the effects of
the group-level variables which are log group size and group mean of advice; for these parame-
ters the prior distribution was normal with mean 0 and variance 0.04.

8. Results

8.1. Descriptive statistics

Table 1 gives the overall means and Jaccard similarity coefficients of the four delinquent acts
for the pooled data. They are positively associated.
 Multilevel Longitudinal Social Networks 17
Table 1. Overall means and similarity coefficients of delinquent acts
mean Jaccard similarity
stealing vandalism graffiti fighting
stealing 0.13 – 0.29 0.22 0.27
vandalism 0.20 0.29 – 0.27 0.33
graffiti 0.17 0.22 0.27 – 0.23
fighting 0.21 0.27 0.33 0.23 –

Table 2. Descriptives for actor variables

mean σ̂ τ̂ icc
sex M 0.53 0.50 0.03 0.00
advice 6.71 0.88 1.44 0.73
language Dutch 0.91 0.27 0.07 0.06
delinquency wave 1 0.76 1.03 0.21 0.04
delinquency wave 2 0.90 1.10 0.25 0.05
delinquency wave 3 0.90 1.14 0.27 0.05

A measure for delinquency is the outdegree in the two-mode network, i.e., the number of delin-
quent behaviours reported by a student. For this variable and for the covariates the means,
within-classroom and between-classroom standard deviations (σ̂ and τ̂ ), and the intraclass cor-
relation coefficients (icc) (calculated according to Snijders and Bosker, 2012, Chapter 3) are
reported in Table 2. From the intraclass correlation coefficients we see that the classrooms are
quite homogeneous with respect to advice, not assortative with respect to sex, while for the level
of delinquency and whether the Dutch language is spoken at home assortativity is positive but
low.
Some descriptive statistics for the set of 81 friendship networks are presented in Table 3. These
include reciprocity, defined as the proportion of ties i → j that is reciprocated by j → i,
and transitivity, defined as the proportion of two-paths i → j → h that is closed by i → h.
Average degrees are about 4, average reciprocity is about 0.60, and average transitivity is about
0.56. These are quite usual figures for friendship networks. The Jaccard measure for network
stability ranges for friendship from 0.28 to 0.75, with a mean of 0.51. This indicates that a good
proportion of ties remains in place from one wave to the next.
Some descriptive statistics for the two-mode delinquency networks are given in Table 4. The

Table 3. Descriptives for friendship networks: means and standard devia-

tions across the 81 groups.
wave 1 wave 2 wave 3
mean (s.d.) mean (s.d.) mean (s.d.)
mean outdegree 4.01 (0.66) 4.17 (0.60) 4.03 (0.68)
s.d. outdegree 2.62 (0.88) 2.72 (0.73) 2.55 (0.73)
s.d. indegree 1.96 (0.54) 1.97 (0.57) 1.99 (0.50)
reciprocity 0.59 (0.08) 0.60 (0.09) 0.60 (0.09)
transitivity 0.55 (0.09) 0.56 (0.09) 0.56 (0.09)
Jaccard with next wave 0.50 (0.09) 0.52 (0.08)
proportion missings 0.03 (0.03) 0.07 (0.06) 0.06 (0.04)
18 Johan Koskinen and Tom A.B. Snijders
Table 4. Descriptives for delinquency networks: means and standard devi-
ations across the 81 groups.
wave 1 wave 2 wave 3
mean (s.d.) mean (s.d.) mean (s.d.)
mean outdegree 0.76 (0.29) 0.91 (0.34) 0.92 (0.35)
s.d. outdegree 1.01 (0.23) 1.09 (0.20) 1.12 (0.24)
s.d. indegree 2.02 (0.83) 2.14 (1.01) 2.01 (1.00)
Jaccard with next wave 0.39 (0.09) 0.43 (0.10)
proportion missings 0.01 (0.02) 0.06 (0.05) 0.06 (0.05)

students report on average slightly less than one out of the four delinquent acts. The Jaccard
measure for stability ranges from 0.21 to 0.70, with a mean of 0.41. Here also there is some
change from one wave to the next, but not too much.

8.2. Modelling results

For the MCMC procedure, three parallel chains were used, each of 70,000 steps; each step
consisted of 200-800 updates of v in the 81 groups (with a total of 66,200), five updates of η ,
and one update of each of γ , µ, and Σ. Of the 70,000, the first 10,000 were the warming phase.
The homogeneity of the three chains was good according to the R̂ measure of Gelman et al.
(2014), which was less than 1.05 for all global parameters.
Posterior means, standard deviations, and credibility intervals of the parameters are given in
Table 5. Appendix C illustrates the posterior distributions of some parameters. The estimated
model for friendship dynamics is usual and has the usual interpretation (e.g., Fujimoto et al.,
2018; Ripley et al., 2021). We focus the interpretation on the mutual dependency of delinquency
and friendship, using the mnemonic indicators given above in the list of cross-network effects.
Dependent variable: friendship
The delinquency degree, i.e., the number of delinquent acts practised, is a measure for delin-
quent behaviour. Effects of delinquent behaviour on friendship dynamics are minor. The table
shows that delinquency has virtually no effect on the popularity as a friend (‘id’) and a negative
effect on the activity in nominating friends (‘od’): those who report more delinquent behaviour
tend to nominate slightly fewer friends. Practising the same delinquent acts (‘odd’) has no
appreciable effect on friendship formation.
Dependent variable: delinquent acts
We start with discussing the five effects not related to friendship. The parameter for the outde-
gree effect (µ̂k = −2.250) indicates a reluctance to practising delinquency, stronger for girls
than for boys (η̂k = 0.207). There is hardly a differentiation between the four delinquent acts
(indegree popularity, η̂k = 0.012) but quite a strong differentiation between students (outde-
gree activity, µ̂k = 0.406), expressing that those currently practising more delinquency have a
stronger tendency to add new delinquent acts. The evidence is inconclusive for effects of school
advice, and of its classroom mean (the latter is negative with 0.91 posterior probability).
Social influence is represented by the four mixed effects of friendship and delinquency on the
dynamics of delinquent behaviour. The effects of indegrees (‘id’) and outdegrees (‘od’) show
 Multilevel Longitudinal Social Networks 19

Table 5. Posterior summaries for delinquency networks

Effect par. (psd) CI betw. sd
friendship 0.025 0.975
outdegree (density) –2.325 (0.068) –2.46 –2.19 0.386
reciprocity 2.044 (0.061) 1.93 2.16 0.332
transitive triplets 0.457 (0.015) 0.43 0.49 0.100
transitive recipr. triplets –0.149 (0.016) –0.18 –0.12
indegree - popularity –0.074 (0.012) –0.10 –0.05 0.092
outdegree - activity 0.038 (0.004) 0.03 0.05
reciprocal degree - activity –0.186 (0.014) –0.21 –0.16 0.085
same sex 0.660 (0.024) 0.61 0.71
log class size –0.127 (0.166) –0.46 0.20
advice similarity 0.105 (0.084) –0.06 0.27 0.250
same language 0.172 (0.033) 0.11 0.24 0.172
delinq. degree popularity ‘id’ –0.001 (0.013) –0.03 0.02
delinq. degree activity ‘od’ –0.043 (0.013) –0.07 –0.018
same delinquent acts ‘odd’ 0.040 (0.030) –0.02 0.10
delinquency
outdegree (density) –2.250 (0.087) –2.41 –2.08 0.333
indegree - popularity 0.012 (0.011) –0.01 0.03
outdegree - activity 0.406 (0.018) 0.37 0.44 0.098
sex (M) 0.207 (0.042) 0.13 0.29
advice 0.018 (0.022) –0.02 0.06
classroom mean advice –0.042 (0.031) –0.10 0.02
friendship indegree activity ‘id’ –0.004 (0.012) –0.03 0.02
friendship outdegree activity ‘od’ –0.073 (0.015) –0.10 –0.04
same delinq. acts as friends ‘odd’ 0.267 (0.037) 0.19 0.34
av. number of delinq. acts of friends ‘od_av’ –0.125 (0.051) –0.22 –0.02
par = posterior mean µ̂, η̂; psd = posterior standard deviation of µ, η;
betw. sd = posterior between-groups standard deviation σ̂.
20 Johan Koskinen and Tom A.B. Snijders
a similar pattern to what was found for friendship dynamics: there is a negative effect of the
outdegree for friendship on the number of delinquent acts reported. There is a rather strong
tendency to practise the same delinquent acts as one’s friends (‘odd’, η̂k = 0.267) but a negative
effect of the average delinquency of friends (‘od_av’, η̂k = −0.125).

The combination means that, for a given delinquent act h, if more of i’s friend practise it then i
will have a higher probability of also starting to practise it and a lower probability of stopping
with it; by contrast, if i’s friends are more delinquent on average but none of them practises act h
then i will have a lower probability of starting to practise h, and a higher probability of stopping.
However, the former, positive influence effect is stronger than the latter, negative influence
effect, because its parameter is higher in absolute value and the range of the explanatory variable
corresponding to ‘odd’ — which is the number of friends practising act h — is equal to 9, which
is larger than the within-group range of the explanatory variable corresponding to ‘od_av’, equal
to 4.

Concluding, there is a weak social selection effect, where those who are more delinquent tend
to nominate fewer friends, and a rather strong social influence effect in the sense of practising
the same delinquent behaviours as one’s friends, but a weak effect of avoiding the delinquent
behaviours not practised by one’s friends if the friends are more delinquent otherwise. This
contrasts with the results of Knecht et al. (2010), who used the same data set but found no evi-
dence of social influence. That publication used a simpler two-stage multilevel network method
which allowed the inclusion of only 21 classrooms of this data set. Another difference is that the
earlier publication did not distinguish the four separate delinquent acts in a two-mode network,
but used an aggregate measure of delinquency. This leads to incomparability between the data
used in the analysis, because the two-mode network representation required dichotomization
of the four delinquency variables, while they were added, without dichotomization, in Knecht
et al. (2010).

9. Conclusions

Network analysis has typically been concerned with describing and modelling network pro-
cesses for individual networks only. We have proposed a modelling framework for generalising
network inference beyond the specifics of individual groups to a population of networks. The
model is a hierarchical extension of the Stochastic Actor-oriented Model (Snijders, 2017) for
longitudinal network panel data, using random coefficients to represent differences between
groups. This allows taking into consideration group-level effects, e.g., interventions or compo-
sitional characteristics, and their cross-level interactions with within-group effects. A further
possibility is to investigate the network dynamics in many small groups, e.g., of size 5 to 10, for
which an analysis per group does not give meaningful results; an example is Dolgova (2019).

The methods are implemented in the R package RSiena (Ripley et al., 2021). They have been
available in beta versions since a few years, which already led to applications, e.g., Boda (2018).
The MCMC algorithm proposed in this paper is a straightforward procedure, and future work
will be devoted to making it more efficient.
 Multilevel Longitudinal Social Networks 21
Acknowledgements

This work was supported in part by award R01HD052887 from the US Eunice Kennedy Shriver
National Institute of Child Health and Human Development (John M. Light, Principal Investi-
gator). We are grateful to Ruth Ripley and her programming and support in the foundational
stages of this project at Nuffield College and the Department of Statistics at Oxford.

10. References
Aldous, D. (1983) Minimization algorithms and random walk on the d-cube. The Annals of
Probability, 11, 403–413.
Anderson, B. S., Butts, C. and Carley, K. (1999) The interaction of size and density with graph-
level indices. Social Networks, 21, 239–267.
Batagelj, V. and Bren, M. (1995) Comparing resemblance measures. Journal of Classification,
12, 73–90.
Bergstrom, A. R. (1988) The history of continuous-time econometric models. Econometric
Theory, 4, 365–383.
Block, P. (2015) Reciprocity, transitivity, and the mysterious three-cycle. Social Networks, 40,
163–173.
Block, P., Koskinen, J., Hollway, J., Steglich, C. and Stadtfeld, C. (2018) Change we can believe
in: comparing longitudinal network models on consistency, interpretability and predictive
power. Social Networks, 52, 180–191.
Boda, Z. (2018) Social influence on observed race. Sociological Science, 5, 29–57.
Brandes, U., Robins, G., McCranie, A. and Wasserman, S. (2013) What is network science?
Network Science, 1, 1–15.
Dolgova, E. (2019) On Getting Along and Getting Ahead: How Personality Affects Social Net-
work Dynamics. Ph.D. thesis, Erasmus University, Rotterdam. URL: https://repub.
eur.nl/pub/119150/Dissertation_Dolgova_final_printer2.pdf.
Eager, C. and Roy, J. (2017) Mixed effects models are sometimes terrible.
Elmer, T., Boda, Z. and Stadtfeld, C. (2017) The co-evolution of emotional well-being with
weak and strong friendship ties. Network Science, 5, 278–307.
Entwisle, B., Faust, K., Rindfuss, R. R. and Kaneda, T. (2007) Networks and contexts: Variation
in the structure of social ties. American Journal of Sociology, 112, 1495–1533.
Erdős, P. and Rényi, A. (1960) On the evolution of random graphs. A Matematikai Kutató
Intézet Kőzleményei, 5, 17–61.
Faust, K. and Skvoretz, J. (2002) Comparing networks across space and time, size and species.
Sociological Methodology, 32, 267–299.
Fujimoto, K., Snijders, T. and Valente, T. W. (2018) Multivariate dynamics of one-mode and
two-mode networks: Explaining similarity in sports participation among friends. Network
Science, 6, 370–395.
Gelman, A., Carlin, J. B., Stern, H. S., Dunson, D. B., Vehtari, A. and Rubin, D. B. (2014)
Bayesian Data Analysis. Boca Raton, FL: Chapman & Hall / CRC, 3d edn.
Gelman, A., Roberts, G. O. and Gilks, W. R. (1996) Efficient Metropolis jumping rules. In
22 Johan Koskinen and Tom A.B. Snijders
Bayesian Statistics 5 (eds. J. Berardo, J. Berger, A. Dawid and A. Smith), 599–607. Oxford:
Clarendon Press.
Goldstein, H. (2011) Multilevel Statistical Models. London: Edward Arnold, 4th edn.
Greenan, C. C. (2015) Diffusion of innovations in dynamic networks. Journal of the Royal
Statistical Society, Series A, 178, 147–166.
Hamerle, A., Singer, H. and Nagl, W. (1993) Identification and estimation of continuous time
dynamic systems with exogenous variables using panel data. Econometric Theory, 9, 283–
295.
Holland, P. W. and Leinhardt, S. (1977) A dynamic model for social networks. Journal of
Mathematical Sociology, 5, 5–20.
Huitsing, G., Snijders, T. A., Van Duijn, M. A. and Veenstra, R. (2014) Victims, bullies, and
their defenders: A longitudinal study of the coevolution of positive and negative networks.
Development and Psychopathology, 26, 645–659.
Jeffreys, H. (1998) The theory of probability. OUP Oxford.
Kalter, F., Heath, A. F., Hewstone, M., Jonsson, J. O., Kalmijn, M., Kogan, I. and van Tubergen,
F. (2013) Children of immigrants longitudinal survey in four European countries (CILS4EU):
Motivation, aims, and design. Tech. rep., GESIS: GESIS Data Archive, Cologne.
Kivelä, M., Arenas, A., Barthelemy, M., Gleeson, J. P., Moreno, Y. and Porter, M. A. (2014)
Multilayer networks. Journal of Complex Networks, 2, 203–271.
Knecht, A. (2006) Networks and actor attributes in early adolescence [2003/04]. Ics-codebook
no. 61, The Netherlands Research School ICS, Department of Sociology, University of
Utrecht, Utrecht. Persistent data set identifier urn:nbn:nl:ui:13-ehzl-c6.
Knecht, A., Snijders, T. A. B., Baerveldt, C., Steglich, C. and Raub, W. (2010) Friendship and
delinquency: Selection and influence processes in early adolescence. Social Development,
19, 494–514.
Köllisch, T. and Oberwittler, D. (2004) Wie ehrlich berichten männliche jugendliche über ihr
delinquentes verhalten? KZfSS Kölner Zeitschrift für Soziologie und Sozialpsychologie, 56,
708–735.
Koskinen, J. H. and Snijders, T. A. B. (2007) Bayesian inference for dynamic social network
data. Journal of Statistical Planning and Inference, 13, 3930–3938.
Krivitsky, P. N., Handcock, M. S. and Morris, M. (2011) Adjusting for network size and compo-
sition effects in exponential-family random graph models. Statistical Methodology, 8, 319–
339.
Lomi, A., Snijders, T. A. B., Steglich, C. and Torlò, V. J. (2011) Why are some more peer
than others? Evidence from a longitudinal study of social networks and individual academic
performance. Social Science Research, 40, 1506–1520.
Lomi, A. and Stadtfeld, C. (2014) Social networks and social settings: Developing a coevolu-
tionary view. KZfSS Kölner Zeitschrift für Soziologie und Sozialpsychologie, 66, 395–415.
Maddala, G. (1983) Limited-dependent and Qualitative Variables in Econometrics. Cambridge:
Cambridge University Press, 3rd edn.
Magnani, M. and Wasserman, S. (2017) Introduction to the special issue on multilayer networks.
Network Science, 5, 141–143.
O’Hagan, A. and Forster, J. (2004) Bayesian Inference, vol. 2B of Kendall’s Advanced Theory
of Statistics. London: Arnold.
 Multilevel Longitudinal Social Networks 23
Ripley, R. M., Snijders, T. A. B., Bóda, Z., Vörös, A. and Preciado, P. (2021) Manual for Siena
version 4.0. Tech. rep., Oxford: University of Oxford, Department of Statistics; Nuffield
College. URL: http://www.stats.ox.ac.uk/siena/.
Robins, G. (2015) Doing social network research: Network-based research design for social
scientists. London etc.: Sage.
Schweinberger, M. (2007) Statistical Methods for Studying the Evolution of Networks and Be-
havior. Ph.D. thesis, University of Groningen, Groningen.
Schweinberger, M., Krivitsky, P. N., Butts, C. T. and Stewart, J. (2020) Foundations of finite-,
super-, and infinite-population random graph inference. Statistical Science, 35, 627–662.
Shalizi, C. R. and Rinaldo, A. (2013) Consistency under sampling of exponential random graph
models. Annals of Statistics, 41, 508–535.
Shalizi, C. R. and Thomas, A. C. (2011) Homophily and contagion are generically confounded
in observational social network studies. Sociological Methods & Research, 40, 211–239.
Singer, H. (1996) Continuous-time dynamic models for panel data. In Analysis of Change (eds.
U. Engel and J. Reinecke), chap. 6, 113–134. De Gruyter.
Slaughter, A. J. and Koehly, L. M. (2016) Multilevel models for social networks: Hierarchical
Bayesian approaches to exponential random graph modeling. Social Networks, 44, 334–345.
Snijders, T. A. B. (2001) The statistical evaluation of social network dynamics. In Sociological
Methodology – 2001 (eds. M. E. Sobel and M. P. Becker), vol. 31, 361–395. Boston and
London: Basil Blackwell.
— (2005) Models for longitudinal network data. In Models and Methods in Social Network
Analysis (eds. P. Carrington, J. Scott and S. Wasserman), chap. 11, 215–247. New York:
Cambridge University Press.
— (2016) The multiple flavours of multilevel issues for networks. In Multilevel Network Anal-
ysis for the Social Sciences (eds. E. Lazega and T. A. B. Snijders), 15–46. Cham: Springer.
— (2017) Stochastic actor-oriented models for network dynamics. Annual Review of Statistics
and Its Application, 4, 343–363.
Snijders, T. A. B. and Baerveldt, C. (2003) A multilevel network study of the effects of delin-
quent behavior on friendship evolution. Journal of Mathematical Sociology, 27, 123–151.
Snijders, T. A. B. and Bosker, R. J. (2012) Multilevel Analysis: An Introduction to Basic and
Advanced Multilevel Modeling. London: Sage, 2nd edn.
Snijders, T. A. B., Koskinen, J. H. and Schweinberger, M. (2010) Maximum likelihood estima-
tion for social network dynamics. Annals of Applied Statistics, 4, 567–588.
Snijders, T. A. B., Lomi, A. and Torlò, V. (2013) A model for the multiplex dynamics of two-
mode and one-mode networks, with an application to employment preference, friendship, and
advice. Social Networks, 35, 265–276.
Steglich, C. E. G., Snijders, T. A. B. and Pearson, M. A. (2010) Dynamic networks and behavior:
Separating selection from influence. Sociological Methodology, 40, 329–393.
Veenstra, R., Dijkstra, J. K., Steglich, C. and Van Zalk, M. H. (2013) Network–behavior dy-
namics. Journal of Research on Adolescence, 23, 399–412.
Wasserman, S. and Faust, K. (1994) Social Network Analysis: Methods and Applications. New
York and Cambridge: Cambridge University Press.
24 Johan Koskinen and Tom A.B. Snijders
Appendix

A. Statistics

A comprehensive list and definition of all effects currently employed in SAOMs is provided in
Ripley et al. (2021, Chapter 12). The effects used in this paper are defined as follows.

(a) OutdegreeP (density)

sX
d,i (x) = j xij
(b) ReciprocityP
sX
rec,i (x) = j xij xji
(c) transitive triplets
sX
P
tt,i (x) = j,h xij xih xhj
(d) transitive reciprocated triplets effect
X
P
strt,i (x) = j,h xij xji xih xhj
(e) indegree-popularity
sX
P
idp,i (x) = j,h xij xhi
(f) outdegree-activity
sX
P
oda,i (x) = j,h xij xih
(g) reciprocal degree
P activity
sX
rda,i (x) = j,h xij xih xhi
(h) covariate B P ego
sX
ego,i (x) = j xij Bi
(i) same covariate PB
sX
same,i (x) = j xij I{Bi = Bj }
(j) covariate B P similarity
sX


sim,i (x) = j xij c − |Bi − Bj |/range(B) ,
where c is a centering constant

The effects for network Z are similar. The cross-network effects were defined in Section 3.2.1.
The effect of covariates in Table 5 are ego effects, unless indicated as ‘same’ or ‘similarity’.

B. Prior sensitivity

B.1. Prior variance

As outlined in Section 6, the influence of the prior is mainly from Λ0 and will affect the infer-
ence both for µ and the group-wise parameters γ [g] , through Σ. As an illustrative example of
the prior scale, we consider here the subset of 21 classrooms used in Knecht et al. (2010) for a
simplified model for the network only, Y [g] = X [g] (g = 1, . . . , 21) and M = 2 periods. In the
structural part, we have omitted indegree popularity, outdegree activity, and reciprocity activity
but added 3-cycles. To further simplify the model, delinquency is treated as a nodal covariate.
All parameters are variable and p2 = 0. Figure 2 provides the credibility intervals for µk using
 Multilevel Longitudinal Social Networks 25

basic rate parameter friends outdegree (density) reciprocity transitive triplets transitive recipr. triplets

4
3
7

4
0

3
2

2
6

1
−2

1
µ

µ
0
5

0
0

−1
−4

−1
4

−2
−2

−2
−6

−3

−3
3

−1 0 1 2 3 4 −1 0 1 2 3 4 −1 0 1 2 3 4 −1 0 1 2 3 4 −1 0 1 2 3 4

logσ02 logσ02 logσ02 logσ02 logσ02

3−cycles sex similarity 3 delinq alter delinq ego delinq ego x delinq alter
2

2
4

2
1
0

0
2

0
µ

µ
−1
−2

−2
0

−2

−2
−4

−3

−4
−2

−4
−4

−1 0 1 2 3 4 −1 0 1 2 3 4 −1 0 1 2 3 4 −1 0 1 2 3 4 −1 0 1 2 3 4

logσ02 logσ02 logσ02 logσ02 logσ02

Fig. 2. Credibility intervals (95%, dark; 99% light) for µ with default prior and Λ0 = σ02 I for
different values of log(σ02 )

the Normal - Inverse Wishart prior with µ0 = 0, ν0 = 12, κ0 = 1, Λ0 = σ02 I , for different
values of σ02 (a very small number of draws, 300, have been used here).
For small values of σ02 , the credibility intervals are noticeably tighter than for increasingly large
values σ02 , when the prior variance overwhelms the data. The central tendencies (posterior
means) are remarkably constant as a function of σ02 , and are hardly pulled towards the prior
mean of zero, even for values of σ02 as small as 0.25 (the smallest value in the plots).
[g]
The influence on the group parameters γk of the same set of priors is illustrated in Figure 3.
Note the difference in vertical scale. The inference on these parameters is remarkably robust
to the prior variance. Only for extreme values of σ02 and for a few classrooms do we see a big
change in group-level parameters. The very wide intervals are due to two specific classrooms.
More specifically, in one classroom (number 20) ‘transitive reciprocated triples’ and ‘3-cycles’
were collinear, which manifests itself in extremely large intervals for these parameters when
large σ02 prevents this classroom from borrowing information from the other classrooms. An-
other classroom (number 11) had a similar issue with structural parameters and in addition a
‘sex similarity’ effect that is difficult to estimate because of the very skewed sex distribution in
this classroom. The issues with these two classrooms also manifests themselves in increasingly
poor mixing for γ [g] for large values of σ02 (results available upon request from the authors).
26 Johan Koskinen and Tom A.B. Snijders

basic rate parameter friends outdegree (density) reciprocity transitive triplets transitive recipr. triplets

15
6

2.0
10

●
●●
●●●
−2

5
●●●
●●
●●●●
●●

1.5
●
8

10
●

4
● ●
●
−3

●
●

1.0
●
●
6
θ

θ
● ●
●●●
●
● ●
−4

● ●●●● ●
2
●●●
● ●●

5
●
●●● ●
●
●
●●

0.5
● ●●
●●
●● ●●
●● ●●
4

● ●
●●
1

●●
●●●
● ●
−5

● ●
●●

0.0
0

●
●●●
2

●●●●

0
●●●●●●●
●●●●●
●
−1
−6

5 10 15 20 5 10 15 20 5 10 15 20 5 10 15 20 5 10 15 20

3−cycles sex similarity delinq alter delinq ego delinq ego x delinq alter
3

●●●●●●
3
0

●●●●●
●●●●●●●

2
●●●
15

●
●●
●●●●●
2

●●●
●●●●●

0
2

●●●●
●
−5

−2
1
10

1
−10
θ

θ
●●●

−4
●●●●
●●● ●●●
0

●●●● ●●●
●●●●●●● ●●
0

● ●●
●●
● ●●●●
● −6
5
−15

−1

● ●
−1

−8

●●●
−2

●●●
●●●●●
●●●●●●●
−2
−20

●●
0

●
−10

5 10 15 20 5 10 15 20 5 10 15 20 5 10 15 20 5 10 15 20

Fig. 3. Equal 95% tail prediction intervals for γ [g] for different values of σ02 , from σ02 = 0.25 (light
grey) to σ02 = 113 (dark grey). Groups ordered according to posterior predictive mean
 Multilevel Longitudinal Social Networks 27
B.2. Reference prior
We may consider the influence of the prior for µ and Σ on the predictive distributions for γ [g]
by comparing these to posteriors from group-level parameters estimated independently. The
a priori dependence of µ on Σ can be decoupled by setting π(µ, Σ) = π(µ)π(Σ). When
G > p1 + 1, we may chose an improper prior for (µ, Σ) for reference.
Jeffreys rule (Jeffreys, 1998) is a principled choice for a reference prior. For the multivariate
normal distribution, it is given by

p(µ, Σ) ∝ |Σ|−(p1 +2)/2 ,

or (for the independence-Jeffreys prior)

p(µ, Σ) ∝ |Σ|−(p1 +1)/2 .

For the conjugate model this corresponds to κ0 → 0, ν0 → 0, and letting the determinant of Λ0
tend to 0. Jeffreys prior is still conjugate for µ and Σ, and as such does not alter the updating
scheme outlined in Section 6.
Figure 4 illustrates the inference obtained from (horizontal) fitting the model separately to each
classroom, assuming a constant prior, and (vertical) the predictive distributions obtained from
the hierarchical SAOM with Jeffreys prior. The two previously mentioned classrooms 11 and 20
are omitted for reasons mentioned above. Both analyses are based on the other 19 classrooms.

Figure 4 demonstrates a negligible influence of the prior on the distributions for γ [g] . Fig-
ure 5 presents the posterior densities for µk and shows that these are also centered on the raw,
[g]
un-weighted means γ̄k of γk from the separate estimations. This shows that imposing the mul-
tivariate normal model for the group-level parameters does not yield results that differ strongly
from the individual group-level inference. In order to make use of all classrooms we would
however require more additional classrooms to borrow strength across groups, and impose a
more informative prior for µ and Σ.

C. Posteriors

For the estimated model of Table 5, Figure 6 presents the posterior distributions of the pop-
ulation mean µk for the delinquency outdegree - activity effect and of the constant parameter
ηk for the effect (‘odd’) of same delinquency acts as friends. For both parameters it is evident
that they are positive with a high posterior probability. There is less posterior uncertainty about
[g]
µk than ηk but the former is a population mean of varying group-wise parameters γk whereas
the latter is a constant parameter. This variability across groups is illustrated in Figure 7 (right
[g]
panel), which shows boxplots (without outliers) of the posterior distributions of γk for groups
g ordered according to the posterior means, with horizontal credibility bands in grey for µk .
The variability in group-level means is greater than the variability in µk but γ [g] is clearly pos-
itive with high posterior probability for all groups g . The length of the 95%CI for delinquency
outdegree - activity is 0.072 and the length of the 95%CI for same language is 0.132 but both
28 Johan Koskinen and Tom A.B. Snijders

10
−1

5
15

4
1.0
8
−2

3
6
10

0.5
−3

2
4

1
−4

0.0

0
5

2
−5

−0.5
0

−2
5 10 15 −5 −4 −3 −2 −1 0 2 4 6 8 10 −0.5 0.0 0.5 1.0 −2 0 1 2 3 4 5

basic rate parameter friends outdegree (density) reciprocity transitive triplets transitive recipr. triplets
2

1
4

4
3

0
0

1
2

2
−1

0
−2

−2

−1

0
−4

−2
−3
−1

−2
−6

−2

−6 −4 −2 0 2 −2 −1 0 1 2 3 4 −3 −2 −1 0 1 −2 −1 0 1 2 −2 0 2 4

3−cycles sex similarity delinq alter delinq ego delinq ego x delinq alter

Fig. 4. Prediction-intervals for γ [g] fitted independently (horizontal axis) against predictions from
Hierarchical SAOM using Jeffreys prior (vertical) (excluding classrooms g = 11, 20)
1.5

2.5
1.2
0.4

1.5
2.0
1.0
0.3

0.8

1.5

1.0
0.2

1.0
0.5

0.4

0.5
0.1

0.5
0.0

0.0

0.0

0.0

0.0

2 4 6 8 10 −3.5 −2.5 −1.5 0.5 1.5 2.5 0.0 0.4 0.8 −1.0 0.0 0.5

basic rate parameter friends outdegree (density) reciprocity transitive triplets transitive recipr. triplets
1.5
2.0

1.5
1.5

1.5
1.5

1.0

1.0
1.0

1.0
1.0

0.5

0.5
0.5

0.5
0.5
0.0

0.0

0.0

0.0

0.0

−1.0 0.0 0.5 −0.5 0.5 1.5 −1.0 0.0 0.5 −1.0 0.0 1.0 −0.5 0.5 1.5

3−cycles sex similarity delinq alter delinq ego delinq ego x delinq alter

Fig. 5. Posteriors for µ fitted using Jeffreys prior with vertical line representing γ̄k and the
[g]
horizontal bar γ̄k ± 2sd(γk ) from independently fitting each group (excluding rogue g = 11, 20)
 Multilevel Longitudinal Social Networks 29

^
µ k

^
η k

0.15 0.20 0.25 0.30 0.35 0.40 0.45

Parameter

Fig. 6. Posterior densities for µk for delinquency outdegree - activity and ηk for same delin-
quency acts as friends (‘odd’) with 95% credibility intervals in dark and light grey, respectively.

parameters are positive with high posterior probability. There is a greater variation in the group-
level parameters for same language, however (Figure 7, left panel), and Pr(γ [g] > 0 | y) ranges
from 0.30 to 1.00 with a median of 0.88.
30 Johan Koskinen and Tom A.B. Snijders

friends: same language del: outdegree − activity

0.8

0.5
0.6
parameter

parameter

0.4

0.0

0.2

−0.5 0.0

group group

[g]
Fig. 7. Boxplots (without outliers) for posteriors of γk for same language and delinquency
outdegree - activity, ordered by posterior means. The horizontal dark and light grey bands
indicate the 90% and 99% credibility intervals for µk .