Sei sulla pagina 1di 22

European Journal of Operational Research 177 (2007) 1253–1274

www.elsevier.com/locate/ejor

Interfaces with Other Disciplines

Modelling and estimation of social interaction effects


in new product diffusion
a,* b
Christos J. Emmanouilides , Richard B. Davies
a
Department of Economics, Aristotle University of Thessaloniki, P.O. Box 170, 546 21 Thessaloniki, Greece
b
University of Wales Swansea, Singleton Park, Swansea SA2 88P, Wales, UK

Received 1 February 2005; accepted 1 November 2005


Available online 13 February 2006

Abstract

We present a new product choice model in which individual agents are assumed to interact with each other across spa-
tial hierarchies under random Gaussian laws. A restricted form that allows estimation of the model and inference on the
random interaction structure on available data is then derived. This is essentially a random coefficients model which we
estimate on repeated cross-sectional data for the adoption of video cassette recorders in Britain. It is argued that the ran-
dom coefficient approach, based on economic and social-theoretic arguments allows for more accurate inference on the
structure of social interactions, and thus may provide a useful way to help alleviate the inferential problem caused by
the identification issues stated in the literature. This very same problem has for long been at the core of the debate between
alternative economic and marketing theories of new product diffusion.
 2006 Elsevier B.V. All rights reserved.

Keywords: Marketing; Multi-agent systems; Social interactions; Innovation diffusion; Choice models

1. Introduction

Innovation diffusion theory [46] and the relevant marketing research tradition (e.g. [4,21,36,37]) suggest
that the key diffusion mechanisms are the innovative and imitative behavior of individual adopters. Early
adopters differ from late adopters in that they have a high intrinsic propensity to adopt. This propensity relies
mainly on external information sources (e.g. mass media) and is not influenced by the adoption behavior of
other individuals. The reverse argument is hypothesized for later adopters. Economic theory has followed two
main approaches in explaining the adoption paths of new products in consumer markets (for a survey see [49]).
The first approach (e.g. [13,14]), suggests that diffusion is the outcome of individual heterogeneity in the pro-
pensity to adopt. This is determined mainly by economic factors such as income. The second approach, similar
to the marketing tradition, relies on awareness and information spreading mechanisms (e.g. [38,39]). For a
discussion of the debate between the two approaches in a marketing perspective we refer to [5].

*
Corresponding author. Tel.: +30 2310 996433.
E-mail address: cemman@econ.auth.gr (C.J. Emmanouilides).

0377-2217/$ - see front matter  2006 Elsevier B.V. All rights reserved.
doi:10.1016/j.ejor.2005.11.021
1254 C.J. Emmanouilides, R.B. Davies / European Journal of Operational Research 177 (2007) 1253–1274

It appears that the key distinguishing feature of the alternative theoretical approaches is the presence or
absence of social interactions. When present, social interactions usually subsume the form of direct communi-
cation links between agents on the demand side. Alternatively they reflect an expectation formation process in
which adoption levels signal quality and value (e.g. [31]). In an empirical setting, social interactions typically
manifest themselves as a dependence of the agents’ behavior on the behavior of other agents in some reference
group. For a thorough review of social interactions in economics we refer to [7].
In marketing and economics, most empirical applications have been based on aggregate growth-curve
models (e.g. Bass-type models, [36]). These models are not sufficient for hypothesis testing about the process
that drives adoption behavior; on the basis of aggregate fit, models based on different theoretical assump-
tions are often indistinguishable. Mahajan et al. [36,37] advocated an individual-level modelling approach
for providing a micro-level link to the aggregate diffusion models, and also to ‘‘. . . study the actual pattern
of social communication, and its impact on product perceptions, preferences and ultimate adoption.’’ [36, p.
388].
Adoption models based on individual decisions have been developed in the marketing and economics lit-
erature. These can be classified into analytical and empirically driven models. Analytical models (see [36,45]
for a review) usually assume a Bayesian learning scheme (e.g. [10]). Information flows about the product from
external (mass media) or internal sources (‘‘word of mouth’’) drive adoption probability and timing. While
these models explicitly account for the role of social interactions through information flows between actual
and potential adopters, they have not been thoroughly tested on empirical grounds.
Empirically driven models include the models of Weerahandi and Dalal [50] and Sinha and Chandrashek-
aran [9,47]. In the first model, social interactions do not affect the utility from adoption, but rather regulate the
arrival rate of purchase occasions. The other two empirical models are duration models of innovation adop-
tion, and they have not been used to test the presence of social interactions. In economics, Karshenas and
Stoneman [30] assumed that a firm’s expected utility from the adoption of computer controlled machine tools
is directly dependent on the number of the adopter firms in the UK engineering industry. In this way, industry-
level interactions are included in the formation of utility.
In this paper we specify and operationalize an empirical choice model that enables testing of the competing
theoretical hypotheses about the nature of the adoption process. Specifically, we want to examine if the pres-
ence of social interactions is empirically supported, while we control for observed individual heterogeneity in
socio-economic drivers of adoption and for unobserved heterogeneity in the process of interactions. As Man-
ski [40] argues about the more general problem of identification and inference of social interactions, this is not
a trivial econometric task. He suggests that identification is a core difficulty in testing competing behavioral
theories; social interaction effects cannot be easily distinguished from the effects of the average exogenous
characteristics of a group of decision agents (henceforth exogenous effects), and from the effects of shared
observed or unobserved characteristics of individuals or environmental influences in the group (henceforth
correlated effects). This problem is severe for linear processes in equilibrium. However, for non-linear far-
from-equilibrium processes, the identification and inferential prospects are much better, provided that inter-
action groups are more or less correctly specified. Dynamic models under which non-social forces are assumed
to act contemporaneously and social forces have a lagged effect on individuals, help overcoming the identifi-
cation problem, provided that the lagged social effects are not misspecified. Furthermore, it is important to
account for unobserved individual heterogeneity that otherwise may be masked by a ‘‘spurious’’ social effect
[1, p. 1007].
We believe that our approach can lead to a better understanding of the diffusion process. As such it can be
of managerial importance; marketing of new products is often strategically guided by assumptions regarding
the adoption process in which social interactions (e.g. ‘‘word of mouth’’) play a central role. The proposed
methodology, in an increasingly data-rich environment, may provide a useful way for analysts to identify
the extent of social interactions effects and position campains accordingly. Our modelling framework can also
be applied to forecasting the diffusion process as we present elsewhere. However, this issue is beyond the
agenda of the current paper.
The paper is structured as follows: in Section 2 we briefly present a dynamic binary choice model with a
social interactions component. The model is a dynamic non-equilibrium model that satisfies the technical con-
ditions for identification and inference we briefly discussed above. Building upon this model, in Section 3 we
C.J. Emmanouilides, R.B. Davies / European Journal of Operational Research 177 (2007) 1253–1274 1255

propose a richer structure for social interactions. Social interactions may arise directly through communica-
tion links (or ‘‘word of mouth’’) or indirectly, through an expectation formation process in which the per-
ceived benefit from adoption depends on the distribution of adopters in the social space of the individual.
We define this social space as an immediate spatial locality (i.e. a neighborhood or a small spatial unit)
and a wider space in which the individual has direct access through observation. Such a space can be a more
aggregate spatial hierarchy that transcends the immediate locality; the individual observes or forms an opinion
for the adoption levels in these hierarchies through sparse direct communication links or through communi-
cation media. We note at the outset that distinguishing between direct and indirect interaction effects requires
directional socio-metric data that are not usually available.
In Section 4 we present the resulting generic model for the choice probabilities. The interactions component
imposes a spatial neighborhood structure on the interaction process, and as such the choice model is a Markov
random field [32]. The spatial element of interactions is supported by evidence from the innovation diffusion
literature [25,8]. Rogers [46, pp. 267–268] stresses the importance of space and ‘‘neighborhood effects’’ on dif-
fusion and argues that ‘‘. . . space is probably one of the least studied variables in the diffusion process’’. Similarly,
Mahajan et al. [35, p. 21] advocate in a marketing context the integration of the temporal and spatial dimen-
sions of diffusion.
In Section 5 we operationalize the choice model with simplifications imposed by the nature of the available
data on the adoption of video cassette recorders in Britain. This is done through restrictions on the social
interactions component. The resulting model is a random coefficients model that allows us to control for
unobserved heterogeneity and thus alleviate the inferential problem stated earlier. It is known that failure
to control for unobserved heterogeneity in binary choice models leads typically to biased towards zero covar-
iate effects (e.g. [15, Chapter 9]). Through the formal analogy of Markov random fields to temporal Markov
chains, we also expect failure to control for unobserved spatial heterogeneity to lead to spurious social inter-
action effects, that is to an upwards bias of the relevant parameters; in the same way that uncontrolled tem-
porally correlated unobserved components can lead to spurious state dependence (i.e. dependence on previous
period own choice outcomes, [26, pp. 138–143], [27]), uncontrolled spatial correlations will lead to inflated
social interaction effects. Our empirical analysis confirms this conjecture. However, under the model assump-
tions, our results suggest that social interaction effects persist after controlling for observed heterogeneity and
misspecification due to omitted effects. Finally, in Section 6 we conclude with a discussion of the limitations of
our approach and propositions for future research.

2. A simple model of interactive choice

Following a statistical physics approach to modelling interactions in many-particle systems (e.g. [19]),
[16,17] introduced a new product adoption model with a social interactions structure that is briefly reviewed
in Appendix A. In that model, the probability of binary action of individual i at time t conditional on a private
utility component, hi(t), and random interactions between decision units has the logistic form
" #!
X
pi ðsi ½tjJ ij ; hi ðtÞ; b; gðkÞÞ ¼ S b  ð2si ðtÞ  1Þ  J ij sj ðt  1Þ þ hi ðtÞ ; ð1Þ
j2gðkÞ

where
expðxÞ
SðxÞ ¼ ; ð2Þ
1 þ expðxÞ
and

• b1 is the ‘‘intensity of choice’’ or equivalently the inverse of the scale parameter of the logistic distribution
for the difference in the stochastic disturbances for the two alternative actions,

e1  e2  Lð0; p2 b2 =3Þ. ð3Þ


For details see [2]. For a recent survey of random utility models we refer to [3].
1256 C.J. Emmanouilides, R.B. Davies / European Journal of Operational Research 177 (2007) 1253–1274

• sj(t  1) is the choice outcome in {0, 1} at time period t  1 of individual j belonging to the same interaction
set, g(k), k = 1, . . . , K, with individual i.
• Jij are the interaction strengths between agents i, j, and are non-zero if and only if i and j belong to the same
influence set g(k). The strengths can be functions of the distance between agents, Jij = J(dij), dij = ji  jj. The
distance can be defined over any adequate spatial or social space. Interactions are frozen (time-invariant)
throughout P the development of the decision process for every interaction set. The sum
ðIÞ
ui ðtÞ ¼ j2gðkÞ J ij sj ðt  1Þ constitutes the interactive utility part, and represents the expected utility of a
decision arising from social interactions (e.g. word-of-mouth).
• hi(t) is the private utility part associated with individual i, assumed to be a linear combination of a set of
structural individual specific covariates, xi(t). Thus hi(t) = aTxi(t). In the innovation diffusion context it
quantifies the effect of exogenous individual characteristics on adoption decisions. It therefore accounts
for the intrinsic individual propensity to adopt a new product and the heterogenous response of agents
to the external influences from mass communication media.

In the terminology of Manski [40], this model can be sensibly interpreted as a non-linear dynamic ‘‘pure
endogenous-effects model’’; individuals located in a spatial cluster tend to share common socio-economic char-
acteristics that are included in the private utility part. Hence, the two utility components are moderately
related in the population, and the identification conditions [40, pp. 537–541] are satisfied. Also, the dynamic
non-linear specification further enhances the possibilities for identification.
The analytical properties of the model (1) were derived in [17], under the assumptions of a single interaction
set, i.e. K = 1, and of interaction strengths drawn from a normal distribution J ij  N ðJ0 ; r2J Þ; J0 > 0. It was
shown that the choice model admits up to three equilibrium solutions when bJ0 is greater than a critical value.
The critical value depends on the distributional properties of the private utility part in the population. For
large values of population heterogeneity in private utility, the model admits practically only one equilibrium
solution. On the other hand, the number of equilibrium solutions is not very sensitive to the distribution of the
interaction strengths. In [18] these results were extended to the case of temporally delayed transmission of
interactions.

3. A multiple network model for social interactions

In this section we extend the model (1) by allowing individuals to belong to more than one interaction set.
We are mainly interested in estimating a restricted version of the general model, and we do not consider at this
stage the analytical equilibrium properties of the proposed full model. This is the subject of future research.
We now discuss the assumptions underlying model formulation.
We assume that there are two independent interaction processes taking place at the micro-level of the dif-
fusion phenomenon. Past research in innovation diffusion (e.g. [46]) provides evidence that individuals form
direct pairwise communication links with other members of the social system and exchange information which
in turn affects adoption decisions. We shall term this a direct social interaction process. On the other hand,
individuals form expectations about the quality and value of an innovation by simply observing the average
adoption levels of the innovation in a reference population (e.g. [31]). These expectations are based on imper-
fect information received mostly through mass media of communication (e.g. newspapers, magazines, radio,
TV). We call this expectation formation mechanism an indirect social interaction process.
Evidence from geographical analyses of innovation diffusion suggests that the first interaction process oper-
ates mainly in local networks (e.g. [25,8]). These are usually defined as residential and travel to work areas.
Most of the individual social exchange activities tend to take place in such localities. The intensity of this pro-
cess is assumed to decay with the spatial distance between individuals, usually under a Newtonian-type inverse
power law or a negative exponential function (e.g. [42, Chapter 5]). However, the proliferation of long-dis-
tance means of two-way communications such as the telephone, and more recently the Internet, increases
the probability of long-range, albeit sparse, direct social interactions.
The second interaction process is also spatially distributed; mass media of communication tend to span
across localities, regions, national, even international levels. The limited empirical evidence for the extent
of indirect social interactions mainly derives from multinational diffusion research (e.g. [20,29]). The way indi-
C.J. Emmanouilides, R.B. Davies / European Journal of Operational Research 177 (2007) 1253–1274 1257

viduals form reference populations or groups and use them as a basis of their expectation formation process is
yet unclear. There is no fine grain law for the reference group formation, but there are theoretical arguments
that this is done on the basis of social proximity, which has a strong spatial dimension ([33], [1, p. 1010]).
In the absence of strong theory on the extent of social interactions, we proceed with the following
simplifications:

(a) Individuals belong to elementary influence sets g(k), k = 1, . . . , K, which we call elementary influence net-
works or simply networks. These are defined as collections of agents who are located in an adequately
small common spatial unit. Individuals in a network can interact with each other, according to the pro-
cesses defined above.
(b) Individuals are allowed to interact with agents who are located in different networks. We expect that
such interactions to be mostly of the indirect type.

To proceed with model development we require a law for the intensity and structure of interactions
within and between the influence sets defined above. These need to be separate for the two parallel social inter-
action processes. Due to the absence of detailed quantitative background theory to guide us, we have to
strongly rely on stochastic formulations. These formulations are kept fairly simple for mathematical tractabil-
ity and parsimony. Let i and j index agents who belong to elementary networks g(k) and g(l) respectively. g(k)
can be the same as g(l). Specifically, we assume that

(c) For the direct pairwise interaction process, the contribution to individual i’s utility through direct inter-
action with individual j is
ðk;lÞ ðk;lÞ ðk;lÞ
J ij  sj ðt  1Þ ¼ pij  I ij  sj ðt  1Þ; ð4Þ
ðk;lÞ
where is the probability of having a direct link between i and j. For simplicity, it is assumed that the
pij
ðk;lÞ
direct connection probability is a constant for each pair of networks, that is pij ¼ pðk;lÞ , for each i 2 g(k),
j 2 g(l). This parameter, which can be made distance dependent, quantifies the degree by which influence
sets are sparsely connected (or ‘‘diluted’’), and is expected to be small, indicating a large degree of dilution
ðk;lÞ
of direct social interactions. I ij is the direct interaction strength between individuals i and j, given that a
ðk;lÞ ðk;lÞ ðk;lÞ ðk;lÞ ðdÞ
link exists. It holds that J ij ¼ pij I ij ¼ pðk;lÞ I ij . Let N l ¼ pðk;lÞ N l be the number of active connec-
tions between agents in network g(k) with agents in network g(l). In the absence of evidence about the
true distribution of pairwise interactions we assume a convenient stationary random Gaussian law
!
2
ðk;lÞ J ðk;lÞ ½J ðk;lÞ 
J ij  N ðdÞ
; ðdÞ . ð5Þ
Nl Nl
The mean and the variance of the interaction strengths are scaled by the adjusted destination network
size so that the sum of interaction strengths is bounded and normalized to be of order one, O(1), inde-
pendently of the network size.
(d) For the process of indirect interactions, the contribution to the utility function of agent i is
ðk;lÞ ðk;lÞ ðlÞ ðk;lÞ ðk;lÞ ðk;lÞ
pi  Ji b i ðt  1Þ ¼ pi
m  Ji  ai  mðlÞ ðt  1Þ; ð6Þ
ðlÞ ðk;lÞ ðlÞ
where b i ðt
m 1Þ ¼  m ðt  1Þ is the decision level at network g(l) at time t  1 as perceived by
ai
agent i, and m(l)(t  1) is the true decision level at network g(l) at time t  1,
1 X
Nl
mðlÞ ðt  1Þ ¼ sj ðt  1Þ. ð7Þ
N l j¼1
ðk;lÞ
The term pi is the probability of individual i from elementary network g(k) taking into account the
network g(l) in his or her expectation formation process. For tractability, this
h probability
 is assumed
ðk;lÞ ðk;lÞ
constant for each pair of networks, that is pi ¼ pðk;lÞ for each i 2 g(k). ai 1
2 0; mðlÞ ðt1Þ is a parameter
indicating the deviation of the true from the perceived decision level in network g(l) for individual i. It is
1258 C.J. Emmanouilides, R.B. Davies / European Journal of Operational Research 177 (2007) 1253–1274

ðk;lÞ
assumed time-invariant, and constant for each pair of networks, ai ¼ aðk;lÞ for each i 2 g(k). Finally,
ðk;lÞ
Ji is the strength of indirect interactions for individual i. It is assumed to obey a time-invariant Gauss-
ian law that is independent of the size of the destination network
  2 
ðk;lÞ
Ji  N Jðk;lÞ ; Jðk;lÞ ; ð8Þ
with finite mean and variance.

Assumptions (a) and (b) partition the population into spatial groups where interactions may take place.
This is a form of quantization of the spatial dimension, and can be made as dense as required. In the limit,
a continuous spatial dimension is achieved. Assumptions (c) and (d) provide local laws for the distributions
and strengths of the interactions within and between the spatial units. The laws can take into account explicitly
the distance between the spatial units, for example by making the moments of the distributions distance-
dependent. At the expense of simplicity, the above assumptions can be modified to account for composite
social interaction spaces in which the spatial dimension is one component, others being, for example, certain
socio-economic characteristics. This would make the interaction laws dependent on the location of individuals
in this broader multidimensional social space.

4. Modelling choice probabilities—A general model

Under the assumptions of Section 3, the interactive utility component for individual i who belongs in net-
work k = 1, . . . , K, is
!
ðIÞ
X K X ðk;lÞ ðk;lÞ ðk;lÞ ðlÞ
ui ðtÞ ¼ J ij sj ðt  1Þ þ Ji a m ðt  1Þ . ð9Þ
l¼1 j2gðlÞ

The first part in the parenthesis is the contribution from pairwise interactions with agents in network l, while
the second is the contribution from indirect interactions with network l. Using (7) it can be shown that the two
components due to the parallel interaction processes are generally not identifiable, unless the interaction
strengths are explicitly dependent on individual characteristics. Indeed, inserting (7) in Eq. (9) we get
!
ðIÞ
X K X ðk;lÞ ðk;lÞ ðk;lÞ 1
X
ui ðtÞ ¼ J ij sj ðt  1Þ þ Ji a sj ðt  1Þ
l¼1 j2gðlÞ
N l j2gðlÞ
! !
X K X  ðk;lÞ ðk;lÞ ðk;lÞ 1
 X K X ðk;lÞ
¼ J ij þ Ji a sj ðt  1Þ ¼ Jij sj ðt  1Þ ; ð10Þ
l¼1 j2gðlÞ
Nl l¼1 j2gðlÞ

where
ðk;lÞ ðk;lÞ 1 ðk;lÞ
Jij ¼ J ij þ aðk;lÞ Ji
. ð11Þ
Nl
Since the two interaction processes are by definition statistically independent, from (5) and (8), and as
ðdÞ
N l ¼ pðk;lÞ N l ; it can be seen that
!
ðk;lÞ Jk;l J2k;l
Jij  N ; ; ð12Þ
Nl Nl
where

Jðk;lÞ
Jk;l ¼ þ aðk;lÞ Jðk;lÞ ; and ð13aÞ
pðk;lÞ
½J ðk;lÞ 2 ðk;lÞ 2 ½J
ðk;lÞ 2

J2k;l ¼ þ ½a  . ð13bÞ
pðk;lÞ Nl
Eqs. (12) and (13b) indicate that the variance component due to indirect interactions approaches zero faster
than the component due to the direct interactions as the size of the destination network increases. This implies
C.J. Emmanouilides, R.B. Davies / European Journal of Operational Research 177 (2007) 1253–1274 1259

that for large network sizes, most of the variance in the combined interaction strength comes from the direct
ðk;lÞ
interactions component. We term the random variables Jij interaction strengths, and they provide under
(12) and (13) a compact representation for the combined effect of the two interaction processes.
In the more general case these variables can be correlated with each other, and with the private utility com-
ponent. Without loss of generality, we may assume that the private utility component, hi(t), is decomposed
into two elements, #i(t) that is independent of the network and fixed for individual i at time t, and ai,k that
is shared between individuals in the elementary network (i.e. depends on the elementary network)
hi ðtÞ ¼ #i ðtÞ þ ai;k . ð14Þ
We shall attribute the dependence between the private utility and the interactions to membership in the par-
ticular network, that is to the second component in (14). To allow for these most generic possibilities, we as-
sume a joint multivariate normal distribution
   
ðk;lÞ
Jij ; ai;k  N lJ;a ; RJ;a ; ð15Þ

with mean vector of length K2 + K


( ! )K
 Jk;l K
lJ;a ¼ J; a ; J ¼ ; a ¼ fðak Þgk¼1 ; ð16Þ
Nl
k;l¼1

where brackets {Æ} indicate vectors, and ak is the mean of ai,k in network k. RJ;a is a (K2 + K) · (K2 + K)
covariance matrix given in partitioned form by
!
RJ RJ;a
RJ;a ¼ ; ð17Þ
Ra;J Ra

where
 
Ra ¼ ðrak ;al ÞKK ; Ra;J ¼ rak ;Jðl;mÞ ; RJ ¼ ðrklmn ÞK 2 K 2 ; ð18Þ
KK 2

k, l, m, n = 1, . . . , K. In particular,
  1
ðk;lÞ ðm;nÞ
rklmn ¼ Cov Jij ; Jpq ¼ pffiffiffiffiffiffiffiffiffiffiffi  Jk;l  Jm;n  qðk;l;m;nÞ ; i 2 gðkÞ; j 2 gðlÞ; p 2 gðmÞ; q 2 gðnÞ.
N lN n
ð19Þ
(k,l,m,n)
The term q is the correlation between the blocks (k, l) and (m, n) of the interaction matrix with K · K
partitions
0 1
Jcð1;1Þ Jcð1;2Þ  Jð1;KÞ
c
B ð2;1Þ C
 ðk;lÞ B
B Jc Jcð2;2Þ  Jð2;KÞ
c
C
C
J ¼ Jc ¼B . .. .. .. C ð20Þ
B . C
@ . . . . A
JðK;1Þ
c
ðK;2Þ
Jc  JðK;KÞ
c
PK
that describes the structure of interactions in a system of N ¼ k¼1 N k agents defined by assumptions (a)–(d)
ðk;lÞ
of Section 3. This matrix consists of blocks Jðk;lÞc ¼ ðJij Þ; i 2 gðkÞ; j 2 gðlÞ; k; l ¼ 1; . . . ; K, which are
themselves random matrices
0 1
ðk;lÞ ðk;lÞ ðk;lÞ
J 11 J 12    J 1N l
B C
B J ðk;lÞ J ðk;lÞ    J ðk;lÞ C
B 21 22 2N C
Jðk;lÞ ¼B C.
l
c B . C ð21Þ
B .. .
.. . .. .
.. C
@ A
ðk;lÞ ðk;lÞ ðk;lÞ
J Nk1 J Nk2    J NkNl
1260 C.J. Emmanouilides, R.B. Davies / European Journal of Operational Research 177 (2007) 1253–1274

The distributional properties of each block are given by (12) and (13). Random matrices have been studied in
several branches of physics and multicomponent dynamical systems (e.g. [12]). More recently, interest on such
matrices has grown in statistics, in the context of Markov random fields (alternatively known as Gibbs mea-
sures) (e.g. [11]) in spatial statistics (e.g. [11]) and in Markov Chain Monte Carlo (MCMC) methods and
Bayesian image analysis (e.g. [22]). This interest is mainly due to the following fact: Coupled multi-component
systems with evolution parameters the elements of a random matrix have dynamics determined by the matrix’
eigenvalue spectra (e.g. [12]).
Inserting the interactive utility (10) in (1), the choice probabilities for the multiple network model
become
" #!
  XK X ðk;lÞ
ðk;lÞ
pi si ½t ¼ 1 Jij ; hi ðtÞ; b; gðkÞ ¼ S b Jij sj ðt  1Þ þ hi ðtÞ ; ð22Þ
l¼1 j2gðlÞ

where g(k) denotes the assumed collection of K interaction sets, and S(.) is the logit function (2). For N
randomly interacting agents, (22) defines a system of N coupled equations of motion for the adoption
probabilities. The degree of coupling is regulated by the form of the random interaction matrix (20). Ana-
lytical and/or numerical study can reveal the dynamical properties of the choice system defined by such
equations.
Let
" #
X
K X ðk;lÞ
ui ½si ðtÞ ¼ ð2si ðtÞ  1Þ Jij sj ðt  1Þ þ hi ðtÞ ð23Þ
l¼1 j2gðlÞ

denote the utility for choice si[t] 2 {0, 1}. For the choice model to be complete and estimable, the joint distri-
bution of the population choice outcomes at any time period has to be specified. Because we condition the
individual choice outcomes at period t on realized choice outcomes at the previous time period, t  1, the joint
distribution can be written as a product of independent individual choice probabilities
Y
N
expðbui ½sðtÞÞ expðU ðsðt  1ÞÞÞ
PrðsðtÞÞ ¼ ¼P 0
; ð24Þ
i¼1
1 þ expðbui ½sðtÞÞ s 2S expðU ðs ðt  1ÞÞÞ
0

PN
where U ðsðt  1ÞÞ ¼ i¼1 ½bui ½si ðtÞ, S is the set of all possible combinations of the choice outcomes in the
population, and s(t  1) = (s1(t  1), . . . , sN(t  1)) is the realized vector of choices at time t  1. The joint dis-
tribution (24) is a Boltzmann–Gibbs distribution, and is known by the Gibbs–Markov equivalence theorem
[44, p. 7] to be a markov random field [32]. Markov random fields are markov chains defined in the spatial
dimension, with the time index of a temporal markov chain now transformed to index the neighborhood of
an agent in space.
ðk;lÞ
The joint distribution establishes the likelihood function, and involves a large number of parameters, Jij
and ai,k. When N ! 1 the sum in the denominator (known as the partition function) becomes infinite, and has
no closed-form thus rendering exact maximum likelihood estimation difficult. Modified likelihood methods
(e.g. pseudo-likelihood, [11]) have been proposed for the binary response case, but estimation of standard
errors of estimates is cumbersome and requires knowledge of the joint probability structure [48, p. 510]. Model
simplification is necessary for empirical progress to be made. The degree of simplification depends on the
available data. In the next Section we reduce the complexity of the general model by imposing further assump-
tions and deriving a restricted model form estimable on available data.

5. Operationalizing the choice model

5.1. The data

The data for the empirical study of the adoption of video cassette recorders (VCRs) in Britain are extracted
from the General Household Survey (GHS). This is a continuous cross-sectional survey of households, initi-
C.J. Emmanouilides, R.B. Davies / European Journal of Operational Research 177 (2007) 1253–1274 1261

ated at 1971. The sampling follows a two-stage design with postcode sectors as the primary sampling units.
Initially, postcode sectors are allocated to major strata. These are created by subdividing the standard British
regions and further distinguishing between Metropolitan and non-Metropolitan counties. In such a way 12
regional divisions arise. Major strata are divided into minor strata of equal size, with the number of minor
strata per major stratum proportional to the size of the major stratum. Since 1983 the sampling frame has been
divided into 576 minor strata, and one primary sampling unit is selected from each minor stratum per year.
Each year about 11,000–13,000 households are contacted, with the response rate consistently high at each sur-
vey wave. For example in 1993–1994 the minimum response rate (including only completely co-operating
households) was 72% and the maximum response rate (including partially co-operating households) was
82%. Households are sampled only one time in the history of the survey. Among the information collected
is household ownership of key consumer durables and socio-demographic characteristics of households. More
details can be found in [23].
Under this sampling scheme, each year a different primary sampling unit is selected from the same minor
strata. We take the 576 minor strata as the elementary spatial units (the networks) of our analysis. All primary
sampling units in a minor stratum are spatially close enough to justify the assumption of the presence of some
direct social interactions transmitted from one time period to the other within the strata. However, due to the
small sample sizes (N a ¼ 172 across all 10 survey waves) and measurement error arising from the fact that dif-
ferent primary sampling units are sampled each year in each major stratum, we expect the effect of the local
direct interactions to be contaminated. We would speculate this will result in an inefficient and downwards
biased systematic effect of local direct interactions, and will make the tests of significance more conservative.
The adoption units are households.
Information on VCR ownership was first collected in 1983. For consistency with the spatial sampling
scheme, we analyze household adoption data from 1984 to 1993. The data across survey waves have been
pooled in order to inject temporal variation into explanatory variables and thus to assess time varying char-
acteristics of the adoption process. The possibility of non-stationary parameters has been allowed for through
inclusion of interaction terms between polynomials of calendar time and individual covariates. For a discus-
sion of the benefits of pooling cross-sectional survey data and the issues involved we refer to [41]. The time
interval chosen for the analysis is 12 months, to ensure that all the minor strata are sampled at each time per-
iod. The pooled data cover ten annual periods with total sample size N = 89,741. The sample characteristics
are summarized in Table 1. The diffusion path of VCRs is depicted in Fig. 1.

Table 1
Sample characteristics—geographical regions and small areas in sample
Launch year of VCRs in Britain t0 1975
Observed time window ½tobs obs
1 ; t 11  [1983, 1993]
Total pooled sample size N 98,790
Effective sample size Neff (tobs P 1984) 89,741
Penetration range ½mobs obs
1 ; m11  [0.16, 0.74]
Geographical regions: Regional samples (Nr)
1. North England 5841
2. Yorkshire and Humberside 9167
3. North West England 11,459
4. East Midlands 7170
5. West Midlands 9258
6. East Anglia 3733
7. Greater London 10,836
8. Outer Metropolitan 10,093
9. South East England 8568
10. South West England 8095
11. Wales 5117
12. Scotland 9453
Total: N = 98,790
Primary sampling units: 576 per year Average size: N a ¼ 172
1262 C.J. Emmanouilides, R.B. Davies / European Journal of Operational Research 177 (2007) 1253–1274

Fig. 1. Adoption path of VCRs, six-monthly data.

5.2. Specification of the restricted model

5.2.1. The interactive utility component


Full spatio-temporal analysis of the detailed interaction structures between the minor strata is not possible
due to lack of information regarding the relative locations or distances of the elementary interaction units, and
due to the complexity and size of the interaction matrix. We impose additional structure on this matrix by
assuming a simple partitioning of the interactions for each household into three spatially hierarchical sources:
(a) interactions originating from the network it belongs to, (b) interactions originating from networks that
belong to the same region, and (c) to interactions originating from networks that belong to other regions. This
structure is depicted schematically in Fig. 2.
Let k = 1, . . . , K, and let m(k) = 1, . . . , M, the region in which area k belongs. In our data K = 576, and
M = 12. Let k 0 , k00 denote areas such that k 0 2 m(k), and k00 62 m(k). Assume households i, j, p, q such that
i, j 2 g(k), p 2 g(k 0 ), and q 2 g(k00 ). Then, we assume:

(1) The interaction strengths between households in area k are homogeneous and equal to the mean of the
interaction strengths in (16)

ðk;kÞ Jk;k
Jij ¼ Jðk;kÞ ¼ ; ð25Þ
Nk
0 0
and vary by area, Jðk;kÞ 6¼ Jðk ;k Þ ; k 6¼ k 0 . Eq. (25) implies that these interactions are also symmetric,
ðk;kÞ ðk;kÞ
Jij ¼ Jji ¼ Jðk;kÞ .
(2) Households in area k interact homogeneously with households in all areas k 0 that belong to same region
m(k),

ðk;k 0 Þ Jk;R
Jij ¼ JðkÞ;R ¼ ; ð26Þ
N mðkÞ  N k
where Nm(k) indicates the size of the region. The denominator scales the strengths so that their sum is
bounded and normalized between zero and one. The interaction strengths vary by area,
0 ðk;k 0 Þ ðk 0 ;kÞ 0
JðkÞ;R 6¼ Jðk Þ;R , and are in general asymmetric, Jij ð¼ JðkÞ;R Þ 6¼ Jji ð¼ Jðk Þ;R Þ. The index R indi-
cates that the interactions refer to cross-network interactions in the same region.
C.J. Emmanouilides, R.B. Davies / European Journal of Operational Research 177 (2007) 1253–1274 1263

Region m 1

(k ),R
Area k'

(k ),C (k ' ),R

(k ),R
Area k
Region m 2

(k ),C

ij
( k ,k )
= ji
( k ,k )
= (k,k)

i j

Region m 3 Area k

Fig. 2. Network diagram of the assumed restricted structure of interactions.

(3) Households in area k interact homogeneously with agents in all areas k00 that do not belong to the same
region m(k),

ðk;k 00 Þ Jk;C
Jij ¼ JðkÞ;C ¼ ; ð27Þ
N  N mðkÞ
00
where N is the total population size. The interaction strengths vary by area, JðkÞ;C 6¼ Jðk Þ;C , and are in
ðk;k 00 Þ ðkÞ;C ðk 00 ;kÞ ðk 00 Þ;C
general asymmetric, Jij ð¼ J Þ 6¼ Jji ð¼ J Þ. The index C indicates that the interactions
refer to cross-network interactions in different regions.
(4) The utility component that is shared between individuals in an elementary network (14) is homogeneous
in the network and equal to the mean in (16)
ai;k ¼ ak . ð28Þ

Under these assumptions, the interaction matrix (20) reduces to the K · K partitioned matrix with homo-
geneous partitions rearranged to reflect the groupings of areas into regions

J ¼ Jrðm;nÞ ; m; n ¼ 1; . . . ; M;

Jðm;nÞ
r ¼ Jðk;lÞ
c ; k 2 m; l 2 n; ð29Þ
 
ðk;lÞ ðk;lÞ
Jc ¼ Jij ; i 2 gðkÞ; j 2 gðlÞ.

The subscripts r and c indicate that the corresponding matrices are defined at the regional and area levels
respectively. The matrix J has K · K partitions. Under the imposed restrictions, each partition is a matrix with
identical elements. In total, there are 3 Æ K distinct element values in J. These follow a restricted form of the
multivariate normal distribution of the interaction strengths, Eqs. (15)–(19), which reduces to
1264 C.J. Emmanouilides, R.B. Davies / European Journal of Operational Research 177 (2007) 1253–1274

ðJk ; ak Þ  N ðlk ; Rk Þ; ð30Þ


where Jk is the vector of the three random interaction variates,
!
 ðk;kÞ ðkÞ;R ðkÞ;C Jk;k Jk;R Jk;C
Jk ¼ J ; J ; J ¼ ; ; ; ð31Þ
N k N mðkÞ  N k N  N mðkÞ

lk is the mean vector



lk ¼ JA ; JR ; JC ; a0;k ; ð32Þ
K
where a0;k is the mean of fak gk¼1 . JA ; JR , and JC are the means of the random variates (31) over areas
k = 1, . . . , K, and
0 2 1
rJA rJA ;JR rJA ;JC rJA ;a0
B C
B rJR ;JA r2JR rJR ;JC rJR ;a0 C
Rk ¼ B Br
C ð33Þ
@ JC ;JA rJC ;JR r2JC rJR ;a0 C
A
ra0 ;JA ra0 ;JR ra0 ;JC r2a0
is the covariance matrix. The coarse-graining of the interaction matrix greatly reduces the dimensionality of
the mean vector (16) and the complexity of the covariance matrix (17). The latter is redefined to reflect the
imposed restrictions.
It can be shown that the interactive utility (10) for an individual i in area k and region m reduces to
!
ðIÞ
X
K X ðk;lÞ
ui ðtÞ ¼ Jij sj ðt  1Þ
l¼1 j2gðlÞ
0 1
B X ðk;kÞ X X X X C
¼ bB
@ J sj ðt  1Þ þ JðkÞ;R sj0 ðt  1Þ þ JðkÞ;C sj00 ðt  1ÞC
A;
j2gðkÞ l2mðkÞ; j0 2gðlÞ l0 62mðkÞ j00 2gðl0 Þ
l6¼k

and using the definition (31),


ðIÞ
ui ðtÞ ¼ Jk;k mk ðt  1Þ þ Jk;R mk;RðkÞ ðt  1Þ þ Jk;C mk;CðkÞ ðt  1Þ; ð34Þ
where
1 X
mk ðt  1Þ ¼ sj ðt  1Þ; ð35aÞ
N k j2gðkÞ
1 X
mk;RðkÞ ðt  1Þ ¼ sj0 ðt  1Þ; and ð35bÞ
N mðkÞ  N k j0 2mðkÞ;j2gðlÞ

1 X
mk;CðkÞ ðt  1Þ ¼ sj00 ðt  1Þ. ð35cÞ
N  N mðkÞ j00 62mðkÞ

Relationships (34) and (35) reflect an implicit first order mean-field approximation for each of the three spa-
tial components of the interactive utility. The mean field approximation enabled by the homogeneity assump-
tion for the interaction strengths (Eqs. (25)–(27)), replaces the sum of the interaction terms in each of the three
components, with their corresponding average over the units in the spatial component. Such methods are
often employed in the statistical mechanics of spin systems, under the name of Weiss or Bragg–Williams field
approximations (e.g. [6, Chapter 6]). Even though it ignores the correlations between the state variables,
Corr(si(t), sj(t  1)), it has been proved to describe quite well the behavior of statistical multi-component sys-
tems under a wide range of conditions, and is asymptotically (in the limit of large system size) exact (e.g. [51,
Chapter 4]). The above restricted form of the interactive utility is a mathematically convenient expression for
C.J. Emmanouilides, R.B. Davies / European Journal of Operational Research 177 (2007) 1253–1274 1265

the social interactions. It may be untenable in reality, but it provides a way to disaggregate the interaction
process and to relax the commonly made assumption of spatially homogeneous random population mixing
that is made in most aggregate and individual-level diffusion models.

5.2.2. The private utility component


In (14) we decomposed the private utility part, hi(t), in two components. We parameterize the first compo-
nent, #i(t), as a linear combination of household specific variables, #i(t) = aTxi(t). Under the assumption of
(28), the second component is common between households in network k, and equal to ak . This random
parameter, distributed according to (30), accounts for unobserved, time-invariant, local spatial effects in the
private utility of households. Such effects could arise from unobserved time-invariant factors, such as back-
ground socio-economic, supply side, or environmental variables. Failure to account for these effects may result
to biases in the parameter estimates of the observed effects. Thus, the private utility becomes

hi ðtÞ ¼ aT xi ðtÞ þ ak . ð36Þ

We selected the following variables xi(t) that reflect household characteristics that may be relevant to the
adoption phenomenon:

(1) Age of household head in years (AGE). Rogers [46] suggests that there is no consistent evidence that age is
a predictor of innovativeness.
(2) Calendar time in years (TIME). The effects of unobserved variables with a temporal trend are captured
by the effects of calendar time.
(3) Annual net household income (CNI, log CNI), deflated to 1990 pounds sterling. Income is consistently
associated with early adoption in most studies of innovation diffusion [46]. Since income is lognormally
distributed, we have used its natural logarithm transformation to capture the median effects.
(4) Number of adult household members (NAD). NAD is a measure of household composition and is
expected to act as a multiplier in adoption decisions.
(5) Number of children in household (NCH). This is another measure of household composition, that may
have an effect on adoption probability.
(6) Economic status of household head (ECST). This binary variable {0, 1} indicates whether the household
head is in or out of employment. It is a measure of wealth related to income, but it may also act as a
proxy for unobserved household characteristics.
(7) Educational level of household head (EDL). A five point ordinal/categorical variable,
{1 = High, 2 = Middle, 3 = Lower, 4 = None, 5 = Missing}. The last category is due to the large number
of missing values in the sample. It is suggested by Rogers [46] that high educational level is associated
with early adoption.
(8) Marital status of household head (MAR). A three point categorical variable {1 = Married or cohabiting
couples, 2 = Single, 3 = Widowed or divorced}.
(9) Gender of household head (SEX), a binary variable {1 = Male, 2 = Female}.
(10) Home ownership (OWN). A binary variable, {1 = Owns or buying house, 2 = Rents house}, which is
another measure of wealth complementary to income and economic status.
(11) Socio-economic segment of household head (SEG). This is a standard three point ordinal measure of
social status based on occupation: {1 = High, 2 = Middle, 3 = Low}. The first category includes employ-
ers, managers, and professionals in self-employment or ‘‘white collar’’ employees. The second category
includes intermediate non-manual and skilled manual workers. The latter category includes students,
agricultural workers, and those never worked.

5.2.3. The restricted form of the choice probabilities


The choice probabilities (22) become
 h i
ðIÞ
pi ðtÞ ¼ S b ui ðtÞ þ hi ðtÞ ; ð37Þ
1266 C.J. Emmanouilides, R.B. Davies / European Journal of Operational Research 177 (2007) 1253–1274

ðIÞ
where ui ðtÞ and hi(t) are given by (34) and (14) respectively. The random parameters of the utility function
can be decomposed as
g
Jk;k ¼ JA þ J k;k ;
g
Jk;R ¼ JR þ J k;R ;
ð38Þ
g
Jk;C ¼ JC þ J k;C ; ak ¼ a0;k þ ag
0;k ;

where J g g g
k;k ; Jk;R ; Jk;C , and g
a0;k are zero mean multivariate normal variates that preserve the covariance ma-
trix (33). The term a0;k is then absorbed into the systematic regressor part, aTxi(t).

5.3. Model estimation

In this subsection we describe the estimation procedure for the specified logistic choice model with random
coefficients (37). We employ marginal maximum likelihood, assuming a joint multivariate normal distribution
for the random coefficients as specified in the previous subsection. Integration over this distribution is per-
formed with Gaussian quadrature. For details on the approach we refer to [34].
Let k = 1, . . . , K denote geographical areas with Ik the total number of observations across all survey waves
in area k. Note that for the pooled-cross sectional data in hand we have one observation per individual house-
hold. In line with the model specification, assume the random regression parameters for the intercept and the
interaction effects are drawn from a joint multivariate normal distribution
e
b ðkÞ  N ð0; RÞ; ð39Þ

where e g;J
b ðkÞ ¼ ð J k;k
g;J
k;R
g ;g
k;C a0;k Þ is the vector of the random parameters that vary across areas k and R is
the 4 · 4 variance–covariance matrix of the random coefficients given by (33). This matrix describes the be-
tween-areas variation of these parameters. Then the logistic choice probability model with random coefficients
for each area k becomes

expðnik ½tÞ
pðsik ½t ¼ 1jxik ½t; b; R; dk Þ ¼ ; ð40Þ
1 þ expðnik ½tÞ
where

nik ½t ¼ xTik ½tb þ zTik ½tR1=2 dk ð41Þ

is the linear predictor consisting of a fixed and a random part, with

• zik[t] = (mk(t  1), mk,R(k)(t  1), mk,C(k)(t  1), 1), where mk(t  1) is the adoption level at period t  1 in
area k where household i resides, and mk,R(k)(t  1) is the average adoption level at period t  1 in the areas
other than k located within the region where area k belongs. mk,C(k)(t  1) is the average adoption level in
all other regions,
• dk  N(0, I4) IID random vectors, where I4 is the 4 · 4 unit matrix,
• b is the 1 · p vector of fixed effects, corresponding to the 1 · p covariate vector xik[t].

The observations si[t],sj[t], i 2 g(k), j 2 g(l) are conditionally independent given the random terms dk. There-
fore, the negative marginal log-likelihood for the observed responses s = {sik}, i = 1, . . . , N, k = 1, . . . , K under
the model (40) is
X K Z Z Z Z
 log Lðs; b; RÞ ¼  log P k ðdk ÞUðdk Þ ddk ; ð42Þ
k¼1
d

where

• U(dk) is the density of the 4-variate normal distribution for the random vector dk,
• Pk(dk) is the likelihood for area k, conditional on the random vector, dk,
C.J. Emmanouilides, R.B. Davies / European Journal of Operational Research 177 (2007) 1253–1274 1267

YIk s ½t
expðnik ½tÞ ik
P k ðdk Þ ¼ . ð43Þ
i¼1
1 þ expðnik ½tÞ

In order to maximize the above negative log-likelihood, the multiple integral in (42) is approximated by
Gauss-Hermite quadrature; Each integral is calculated as a sum over a number of R (in our case R = 5) quad-
rature points with locations lr, and probability masses pr, r = 1, . . . , R. To facilitate the computations we
employ a Cholesky decomposition CCT = R of the random part’s covariance matrix R (for details, see [28]) with
the lower triangular Cholesky factor
C ¼ ðcij Þ; i; j ¼ 1; . . . ; 4;
whose elements are related to the covariance matrix entries through
X
i
r2i ¼ c2ij ; i ¼ 1; . . . ; 4; j ¼ 1; . . . ; i;
j¼1
ð44Þ
X
j
rij ¼ cil  cjl ; l ¼ 1; . . . ; j; i 6¼ j.
l¼1

Under this decomposition, the approximate marginal negative log-likelihood becomes


" " ! # #
X
K X
R X
R X
Ik
 log Lðs; b; RÞ ¼  log ... . . . exp ðsik ½tnik ½t  logð1 þ expðnik ½tÞÞÞ pq4 . . . pq1 ;
k¼1 q1 ¼1 q4 ¼4 i¼1

ð45Þ
with nik ½t ¼ xTik ½tb þ zTik ½t  C  Lr , where Lr is the 1 · 4 vector of quadrature points locations.
Estimation was carried out with PMIX, a software package developed by the first author for the analysis of
ordinal and categorical data, using probability models and neural networks. In its current version, it can han-
dle up to five Gaussian random coefficients, and can perform non-parametric estimation of binary random
effects models. The software is written in FORTRAN 90 and optimization is performed using the IMSL rou-
tines DBCON and DLCON. The first routine uses a quasi-Newton method and an active set strategy to solve
optimization problems subject to simple bounds on the variables e.g. [24]. The second routine optimizes a
general objective function subject to linear equality/inequality constraints, using Powell’s TOLMIN method
[43].

6. Results

The estimation results are given in Table 2 for the spatially homogeneous and the spatially heterogenous
model. There was no empirical support for interaction effects that transcend the regional level, so we report
results that include social interaction effects up to the regional level.
We observe that in the heterogenous model, most of the estimated parameters are slightly higher in absolute
value than in the homogeneous model, as expected; this is an innocuous scaling effect. Also, the direction of
effects in the two models is the same for all variables. As we conjectured earlier in Section 1, the social inter-
action effects decrease when we take into account unobserved heterogeneity at the spatial level. This is partic-
ularly so for the local interaction effects, where the estimated strengths are more than halved in the
heterogenous model. In the following we discuss the estimated covariate effects.

6.1. Age of household head

AGE is significantly associated with adoption. Its effects are strong in magnitude, non-linear, and they vary
over time (Fig. 3). Maximum adoption probability occurs at about the age of 35. The location of the maxi-
mum on the AGE axis remains constant, and the probability of adoption increases over time. As the age
of household head decreases below 38 years, the decline in adoption probability is slow. As the age increases
1268 C.J. Emmanouilides, R.B. Davies / European Journal of Operational Research 177 (2007) 1253–1274

Table 2
Parameter estimates (standard errors in parentheses) and fit statistics
Terms Spatial homogeneous model Spatial inhomogeneous model
Fixed regression part
Intercept a0 25.304 (1.719) 25.665 (1.537)
NCH a1 0.034 (0.024) 0.035 (0.025)
NAD a2 0.074 (0.032) 0.075 (0.034)
CNI/104 a3 0.335 (0.098) 0.339 (0.089)
log(CNI) a4 2.102 (0.188) 2.125 (0.182)
AGE a5 0.314 (0.039) 0.324 (0.031)
(AGE2)/102 a6 0.657 (0.080) 0.676 (0.065)
(AGE3)/104 a7 0.370 (0.052) 0.381 (0.043)
OWN2 a8 0.286 (0.021) 0.295 (0.022)
SEG2 a9 0.046 (0.023) 0.052 (0.023)
SEG3 a10 0.096 (0.031) 0.098 (0.032)
MAR2 a11 0.529 (0.042) 0.535 (0.042)
MAR3 a12 0.050 (0.038) 0.046 (0.038)
SEX2 a13 0.227 (0.033) 0.226 (0.033)
EDL2 a14 0.454 (0.040) 0.453 (0.040)
EDL3 a15 0.055 (0.037) 0.053 (0.037)
EDL4 a16 0.122 (0.042) 0.121 (0.043)
EDL5 a17 0.123 (0.034) 0.122 (0.034)
ECST2 a18 0.056 (0.026) 0.060 (0.027)
TIME a19 0.722 (0.110) 0.757 (0.095)
TIME2/10 a20 0.006 (0.004) 0.009 (0.004)
ARE JA 0.298 (0.066) 0.145 (0.080)
REG JR 2.248 (0.232) 2.169 (0.273)
NCH:TIME a21 0.010 (0.002) 0.010 (0.002)
NAD:TIME a22 0.028 (0.003) 0.029 (0.003)
(CNI:TIME)/105 a23 0.138 (0.066) 0.140 (0.058)
log(CNI):TIME a24 0.054 (0.012) 0.055 (0.011)
AGE:TIME a25 0.013 (0.003) 0.014 (0.002)
(AGE2:TIME)/103 a26 0.310 (0.057) 0.322 (0.047)
(AGE3:TIME)/105 a27 0.221 (0.037) 0.228 (0.031)
Random regression parta for the spatial inhomogeneous model
0 1 0 1
b
c 11 0:606ð0:139Þ
B
b ¼ @b C B C
C c 21 bc 22 A ¼ @ 0:393ð0:228Þ 0:469ð0:126Þ A;
b
c 31 bc 32 b c 33 0:112ð0:079Þ 0:297ð0:053Þ 0:078ð0:039Þ
0 1 0 1
r 2JA
b 0:367
B C
b ¼Bb r 2JR
b C¼B C
R @ r JA ;JR A @ 0:238 0:374 A
b
r a0 ;JA b r a0 ;JR b 2
ra 0:068 0:095 0:107
0

Spatial homogeneous model Spatial inhomogeneous model


Fit statistics
Sample size 89,741 89,741
Fitted parameters 30 36
Model deviance (2 log L) 81,144.830 81,097.454
Null model deviance 124,026.214 124,026.214
LR test statistic (p-value) 42,881.384 in 29 df (0) 42,928.760 in 35 df (0)
AIC 0.9049 0.9045
a b of the random coefficients variance–
bc ij ; i; j ¼ 1; 2; 3 denote the elements of the estimated Cholesky decomposition factor, C,
covariance matrix.

above 35, there is again a decline in probability, but a faster one. The estimated temporal variation indicates
that there is a relationship between age and earlier adoption for the case of VCRs, in contrast to most diffusion
studies that did not find any significant association.
C.J. Emmanouilides, R.B. Davies / European Journal of Operational Research 177 (2007) 1253–1274 1269

16 23 30 37 44 51 58 65 72 79 86
93.5 93.5

92.5 92.5

91.5 0.7 91.5


0.8
0.6 90.5
Time (calendar years)

90.5
0.9 0.54
89.5 0. 3 89.5
0. 2
88.5 0. 88.5

87.5 0.1 87.5

86.5 86.5

85.5 85.5

84.5 84.5

16 23 30 37 44 51 58 65 72 79 86
Age (years)

Fig. 3. Estimated effects of age on adoption probability over time. Evaluation at the sample mean of the other covariates. The contour
labels indicate adoption probability levels.

6.2. Income

CNI effects are generally large, non-linear, and time varying (Fig. 4). Maximum adoption probability
occurs at high incomes, with the location of maximum shifting over time from about £60,000 in 1984 towards
higher income values. Adoption probability increases non-linearly and fast with income, up to the value that
corresponds to the annual maximum in probability. Then, it declines slowly. Over time, the adoption proba-
bility increases in a varying manner for each income group, and the rate of decrease in probability for higher
incomes becomes smaller.

10 45 80 115 150 185


93.5 93.5

92.5 92.5

91.5 91.5
Time (calendar years)

90.5 90.5
0.9
89.5 89.5
0.8
88.5 88.5
0.7
0.2

87.5 0.6 87.5


0.5
0.1

86.5 86.5
0.4
0.3

85.5 0.3 85.5


0.
4

84.5 0.2 84.5

10 45 80 115 150 185


Income (£ 000s)

Fig. 4. Estimated effects of income on adoption probability over time. Evaluation at the sample mean of the other covariates. The contour
labels indicate adoption probability levels.
1270 C.J. Emmanouilides, R.B. Davies / European Journal of Operational Research 177 (2007) 1253–1274

6.3. Number of adults

As expected, there is a positive effect of NAD on adoption probability. This increases linearly with time.
The more adults in the household the earlier adoption occurs.

6.4. Educational level

There is a non-linear variation of the adoption probability with education level. Low educational level is
associated with higher adoption probability, while for middle educational level the adoption probability is
the smallest among all groups. The effects are relatively large, and almost constant over time.

6.5. Calendar time

Adoption probability increases as an almost linear function of calendar time, however the effect is more
complex due to interactions with other covariates. It reflects among other factors, price decline, and other sup-
ply side effects such as increased availability and distribution over time.

6.6. Marital status

Households with single heads, are less likely to own a VCR than other households. There is no significant
difference in ownership between households with married/cohabiting couples and households with divorced/
widowed heads.

6.7. Regional penetration in previous time period

Here we uncover a constant over time, strong average social interaction effect of the indirect type. The effect
declines marginally when we account for unobserved spatial heterogeneity. The presence of indirect social
interactions is supported by our data.

6.8. Local network penetration in previous time period

Here the average effect is smaller than that of the regional penetration. Its value is halved and almost loses
its statistical significance when we control for unobserved spatial heterogeneity. However, these results cannot
be conclusive due to the error in the way we measure this variable, as we argue in Section 5.1.

6.9. Unobserved heterogeneity

The results from the heterogenous model reveal a substantial variation in the strengths of direct and indi-
rect social interactions across local networks or areas. The variation in the direct interactions is very large
r 2JA ¼ 0:367Þ as compared to its average value ðJA ¼ 0:145Þ. This result may be attributed to measurement
ðb
errors, but further examination is required to reach a conclusion. The negative covariance between the direct
and indirect social interaction random effects ðb r JA ;JR ¼ 0:238Þ indicates that in local networks where direct
social interactions are stronger, the indirect social interactions tend to be smaller and vice versa. This finding
may indicate that people who rely more on ‘‘external’’ influences (i.e. form expectations on the basis of more
aggregate measures of adoption mediated through media of mass communications) tend to rely less on inter-
nal ones (that is more direct communication links between agents) and vice versa.
r 2a0 ¼ 0:107Þ in comparison with
On the other hand, spatial heterogeneity in the private utility is negligible ðb
its average value ðb a 0 ¼ 25:655Þ, and uncorrelated with the random part of the social interactions
r a0 ;JA ¼ 0:068; b
ðb r a0 ;JR ¼ 0:095Þ. These results indicate that there is very little empirical support for the
presence of time-invariant unobserved factors operating at the local network level that may affect the propen-
sity of households in the network to adopt. Also, these factors, to the extent they exist do not seem to be cor-
related with the random social interactions component of the utility function.
C.J. Emmanouilides, R.B. Davies / European Journal of Operational Research 177 (2007) 1253–1274 1271

6.10. Other effects

As expected, adoption probability increases with home ownership, and socio-economic status. Finally, even
though we control for marital status and household structure, households with male heads are more likely to
own a VCR than those with a female head. This effect may reflect particular social conditions in such house-
holds. The number of children in the home increases the adoption probability and the chance of early
adoption.

7. Concluding comments

We proposed a methodology for modelling spatially disaggregated social interaction effects on the adoption
of innovations. In an illustrative application we estimated a restricted version of the full interactive choice
model and tested the presence of social interactions while we controlled for unobserved spatial heterogeneity
and observed household heterogeneity. We found empirical evidence that social interactions operate at the
regional spatial level, and possibly at the local network level. However, measurement errors preclude a definite
conclusion about the extent of the latter. These results indicate that direct and indirect interactions may play
an important role in the adoption process. Several socio-economic factors also proved to be important drivers
of adoption. Thus we are tempted to conclude that the diffusion process in the case of VCRs is the combined
outcome of adopter’s heterogeneity in the propensity to adopt and of social interactions.
However, there are several limitations to our approach that point to directions for future research; The
cross-sectional sampling mechanism does not allow us to account for unobserved household heterogeneity.
It is also responsible for the measurement errors in the local measures of spatial interactions. An assessment
of the degree of this error is required. The lack of knowledge of the relative distances between households
and the primary sampling units does not allow us to test for richer interaction structures. Individual time
series would permit testing of more structured interactions (possibly asymmetric), and also of temporal
variation in the interaction effects. Finally, more informed construction of interaction sets may be required,
allowing for interactions to depend on household characteristics other than spatial location. Such a task
points towards the collection of richer data, including direct measurement of the influence set formation
process.

Acknowledgements

An earlier version of this paper was written in 1998 as part of the first author’s Ph.D. thesis. The financial
support for the Ph.D. research was supplied by the Greek Section of Scholarships of the ‘‘Alexander S. Onas-
sis’’ Foundation and is gratefully acknowledged. We would like also to thank the Economic and Social Re-
search Council (ESRC) in the UK for providing us with the General Household Survey data.

Appendix A

In this appendix we expose briefly the derivation of model (1) in Section 2. The model belongs to the class of
the Ising lattice models of statistical physics (e.g. [19, Chapter 2]) and as such is similar to the basic interac-
tions-based model of Brock and Durlauf [7]. Assume the introduction of a new product in a market of N het-
erogenous individuals. For individual i the choice outcome (not adopt/adopt) at time period t is denoted by
yi(t) and takes values in the binary set {1, 1}. Individuals are assumed to form mutually exclusive interaction
sets (or ‘‘reference groups’’, ‘‘communities of interest’’, etc.) where interactions take place through the obser-
vation of the previous period choice outcomes of each member of the set. There are K such groups in the pop-
ulation, denoted by g(k), k = 1, . . . , K. Then, the discrete binary choice problem is specified as

max e fy i ðtÞ; xi ðtÞ; pi ðt; S i Þ; eðy i ðtÞÞg;


U ðA:1Þ
y i ðtÞ2f1;1g

where Ue is a stochastic indirect utility function, xi(t) is a vector of exogenous characteristics relevant to
the adoption decision, pi(t; Si) is a probability measure conditional on social interactions, and e(yi(t)) is the
1272 C.J. Emmanouilides, R.B. Davies / European Journal of Operational Research 177 (2007) 1253–1274

unobserved random utility associated with the choice outcome. The utility function is assumed additively
separable and is specified as
e ¼ uðPÞ ½y i ðtÞ; xi ðtÞ þ uðIÞ ½y i ðtÞ; pi ðt; S i Þ þ eðy i ðtÞÞ.
U ðA:2Þ

In this specification, u(P)[Æ] is the non-interactive or private utility component and u(I)[Æ] is the interactive utility
component. The private utility of agent i is assumed to be a linear combination of a set of structural individual
ðPÞ
specific covariates, xi(t). Thus, if we set #i(t) = bTxi(t), ui ½y i ðtÞ ¼ y i ðtÞ  #i ðtÞ. The interactive utility of agent i
is specified as
ðIÞ
X
ui ½y i ðtÞ ¼ y i ðtÞ Jij y j ðt  1Þ. ðA:3Þ
j2gðkÞ

This specification, similar to the Ising model of ferromagnetic interactions in statistical physics (e.g. [19, Chap-
ter 2]), is derived by the assumptions regarding the interaction process. Each agent i interacts with agent j
(j 5 i) in the same interaction set g(k) with a strength Jij . The interaction strength is a weight put by agent
i to the choice behavior of agent j at the previous time period, yj(t  1). It is positive when there is imitative
behavior and/or transmission of positive signals regarding the benefits from the specific choice action. It is
negative otherwise. Interactions are assumed to be random variables that are time-invariant throughout the
development of the decision process for every interaction set. The distribution of the interactions in set
ðkÞ
g(k) has a mean equal to J0 =N k . The mean is scaled by the set size Nk so that the sum of interaction strengths
is finite and of order one, O(1), independently of the set size.
Under the usual assumption of iid extreme value distributed unobserved random utilities, it can be shown
with some algebra that the conditional individual probability of binary action takes the logistic form
" #!
X
pi ðy i ðtÞjJij ; #i ðtÞ; b; gðkÞÞ ¼ S 2by i ðtÞ Jij y j ðt  1Þ þ #i ðtÞ ; ðA:4Þ
j2gðkÞ

where
expðxÞ
SðxÞ ¼ ; ðA:5Þ
1 þ expðxÞ
and
" #
ðIÞ ðPÞ
X
ui ½y i ðtÞ ¼ ui ½y i ðtÞ þ ui ½y i ðtÞ ¼ y i ðtÞ Jij y j ðt  1Þ þ #i ðtÞ ðA:6Þ
j2gðkÞ

are the deterministic utilities associated with each action and their difference respectively. b is the scale
parameter of the logistic distribution for the difference in the unobserved random utilities e(yi(t) = 1) 
e(yi(t) = 1)  L(0, p2b2/3).
To obtain the form (1) of Section 2 for the ordinary definition of the choice variable and choice set,
si(t) 2 {0, 1}, we apply the reparameterization

fy i ðtÞ ! si ðtÞ : si ðtÞ ¼ ðy i ðtÞ þ 1Þ=2g.

With somePalgebra,  from (A.4) and noting that the mean interaction strength in set g(k) is
ðkÞ
J0 =N k  j2gðkÞ Jij =N k ; we get
" #!
X X
pi ðsi ðtÞjJij ; #i ðtÞ; b; gðkÞÞ ¼ S bð2si ðtÞ  1Þ 4 Jij sj ðt  1Þ  2 Jij þ 2#i ðtÞ
j2gðkÞ j2gðkÞ
" #!
X ðkÞ
¼ S bð2si ðtÞ  1Þ 4 Jij sj ðt  1Þ  2J0 þ 2#i ðtÞ . ðA:7Þ
j2gðkÞ
C.J. Emmanouilides, R.B. Davies / European Journal of Operational Research 177 (2007) 1253–1274 1273

ðkÞ
Setting J ij ¼ 4Jij and hi ðtÞ ¼ 2ð#i ðtÞ  J0 Þ we get model (1) of Section 2. Note that in the new formulation,
ðkÞ ðkÞ ðkÞ
the interaction strengths Jij have mean J 0 =N k ¼ 4J0 =N k . The unscaled mean interaction strength J0 of the
interaction set g(k) in which agent i belongs is absorbed into the constant term of hi(t), thus not altering its
interpretation as a private utility component.

References

[1] G.A. Akerlof, Social distance and social decisions, Econometrica 65 (5) (1997) 1005–1027.
[2] S.P. Anderson, A. DePalma, J.F. Thisse, Discrete Choice Theory of Product Differentiation, MIT Press, Cambridge, MA, 1992.
[3] G. Baltas, P. Doyle, Random utility models in marketing research: A survey, Journal of Business Research 51 (2001) 115–125.
[4] F.M. Bass, A new product growth model for consumer durables, Management Science 5 (January) (1969) 215–227.
[5] A.C. Bemmaor, Modelling the diffusion of new durable goods: Word-of-mouth effect versus consumer heterogeneity, in: G. Laurent,
G.L. Lilien, B. Pras (Eds.), Research Traditions in Marketing, Kluwer, Dordrecht, 1993, pp. 201–229.
[6] J.J. Binney, N.J. Dowrick, A.J. Fisher, M.E.J. Newman, The Theory of Critical Phenomena: An Introduction to the Renormalization
Group, Oxford University Press, Oxford, 1995.
[7] W.A. Brock, S.N. Durlauf, Interactions-based models, in: J. Heckman, E. Leamer (Eds.), Handbook of Econometrics, vol. 5, North-
Holland, Amsterdam, 2001.
[8] L.A. Brown, Innovation Diffusion: A New Perspective, Methuen, New York, 1981.
[9] M. Chandrashekaran, R.K. Sinha, Isolating the determinants of innovativeness: A split-population tobit (SPOT) duration model of
timing and volume of first and repeat purchase, Journal of Marketing Research XXXII (August) (1995) 444–456.
[10] R. Chatterjee, J. Eliashberg, The innovation diffusion process in a heterogenous population: A micromodelling approach,
Management Science 36 (9) (1990) 1057–1079.
[11] N.A.C. Cressie, Statistics for Spatial Data. Wiley Series in Probability and Mathematical Statistics, John Wiley & Sons, New York,
1991.
[12] A. Crisanti, G. Paladin, A. Vulpiani, Products of Random Matrices, in: Statistical Physics. Springer Series in Solid-State Sciences, vol.
104, Springer-Verlag, New York, 1993.
[13] P.A. David, A contribution to the theory of diffusion, Research Center in Economic Growth, Memorandum No. 71, Stanford
University, 1969.
[14] S. Davies, The Diffusion of Process Innovations, Cambridge University Press, Cambridge, 1979.
[15] P. Diggle, K. Liang, S.L. Zeger, Analysis of Longitudinal Data, Oxford University Press, Oxford, 1994.
[16] C.J. Emmanouilides, A mean field model with social interactions and network effects in technology diffusion, in: Proceedings of the
26th EMAC Conference, Warwick Business School, 20–23 May 1997, pp. 416–436.
[17] C.J. Emmanouilides, Micro-level modelling of new product adoption in the presence of social interactions, Ph.D. Thesis, University of
Lancaster, UK, 1999.
[18] C.J. Emmanouilides, S. Kasderidis, J.G. Taylor, A random asymmetric temporal model of multi-agent interactions: Dynamical
analysis, Physica D (2003) 102–120.
[19] K.H. Fischer, J.A. Hertz, Spin Glasses, Cambridge University Press, Cambridge, 1991.
[20] H. Gatignon, J. Eliashberg, T.S. Robertson, Modelling multinational diffusion patterns—An efficient methodology, Marketing
Science 8 (3) (1989) 231–247.
[21] H.A. Gatignon, T.S. Robertson, Integration of consumer diffusion theory and diffusion models: New research directions, in: V.
Mahajan, Y. Wind (Eds.), Innovation Diffusion Models of New Product Acceptance, Series in Econometrics and Management
Sciences, vol. 5, Ballinger, Cambridge, MA, 1985, pp. 37–60.
[22] D. Geman, Random Fields and Inverse Problems in Imaging. Lecture Notes in Mathematics, vol. 1427, Springer-Verlag, 1990.
[23] General Household Survey, Office of Population Censuses and Surveys, HMSO, London, 1971–1994.
[24] P.E. Gill, M.A. Murray, M. Wright, Practical Optimization, Academic Press, New York, 1981.
[25] T. Hagerstrand, Innovation Diffusion as a Spatial Process, University of Chicago Press, Chicago, 1969.
[26] J.J. Heckman, Statistical models for discrete panel data, in: C.F. Manski, D. McFadden (Eds.), Structural Analysis of Discrete Data
with Econometric Applications, MIT Press, Cambridge, MA, 1981.
[27] J.J. Heckman, B. Singer, Social science duration analysis, in: J.J. Heckman, B. Singer (Eds.), Longitudinal Analysis of Labour
Market Data, Economic Society Monographs, vol. 10, Cambridge University Press, 1985.
[28] D. Hedeker, R.D. Gibbons, A random-effects ordinal regression model for multilevel analysis, Biometrics 50 (1994) 933–944.
[29] K. Helsen, K. Jedidi, W.S. DeSarbo, A new approach to country segmentation utilizing multinational diffusion patterns, Journal of
Marketing 57 (4) (1993) 60–71.
[30] M. Karshenas, P.L. Stoneman, Rank, stock, order, and epidemic effects in the diffusion of new process technologies: An empirical
model, RAND Journal of Economics 24 (4) (1993) 503–528.
[31] M.L. Katz, C. Shapiro, Technology adoption in the presence of network externalities, Journal of Political Economy 94 (4) (1985) 822–
841.
[32] R. Kindermann, J.S. Snell, Markov Random Fields and their Applications. Series in Contemporary Mathematics, vol. 1, American
Mathematical Society, Providence, RI, 1980.
[33] P. Krugman, The Self-Organizing Economy, Blackwell, Cambridge, MA, 1996.
[34] N.T. Longford, Random Coefficient Models. Oxford Statistical Science Series, vol. 11, Oxford University Press, Oxford, 1993.
1274 C.J. Emmanouilides, R.B. Davies / European Journal of Operational Research 177 (2007) 1253–1274

[35] V. Mahajan, E. Muller, F.M. Bass, New product diffusion models: A review and directions for research, Journal of Marketing 54
(January) (1990) 1–26.
[36] V. Mahajan, E. Muller, F.M. Bass, New product diffusion models, in: J. Eliashberg, G.L. Lilien (Eds.), Marketing, Handbooks in
Operations Research and Manufacturing Science, vol. 5, Elsevier, Amsterdam, 1993, pp. 349–407.
[37] V. Mahajan, E. Muller, Y. Wind, New-Product Diffusion Models. International Series in Quantitative Marketing, Kluwer Academic
Publishers, Massachusetts, 2000.
[38] E. Mansfield, Industrial Research and Technological Innovation: An Economic Analysis, Norton, New York, 1968.
[39] E. Mansfield, The diffusion of industrial robots in Japan and the United States, Research Policy 18 (1989) 183–192.
[40] C.F. Manski, Identification of endogenous social effects: The reflection problem, Review of Economic Studies 60 (1993) 531–542.
[41] J. Micklewright, The analysis of pooled cross-sectional data: Early school leaving, in: A. Dale, R.B. Davies (Eds.), Analysing Social
and Political Change: A Casebook of Methods, Sage, London, 1994.
[42] R. Morrill, G.L. Gaile, G.I. Thrall, Spatial Diffusion, Sage, Newbury Park, CA, 1988.
[43] M.J.D. Powell, A tolerant algorithm for linearly constrained optimization calculations, DAMPT Report NA17, University of
Cambridge, England, 1988.
[44] C.J. Preston, Gibbs States on Countable Sets. Cambridge Tracts in Mathematics, vol. 68, Cambridge University Press, Cambridge,
1974.
[45] J.H. Roberts, J.M. Lattin, Disaggregate-level diffusion models, in: V. Mahajan, E. Muller, Y. Wind (Eds.), New-Product Diffusion
Models, International Series in Quantitative Marketing, Kluwer Academic Publishers, Massachusetts, 2000.
[46] E.M. Rogers, Diffusion of Innovations, third ed., Free Press, New York, 1983.
[47] R.K. Sinha, M. Chandrashekaran, A split hazard model for analysing the diffusion of innovations, Journal of Marketing Research
XXIX (February) (1992) 116–127.
[48] M. Sherman, Variance estimation for statistics computed from spatial lattice data, Journal of the Royal Statistical Society, Series B 58
(3) (1996) 509–523.
[49] P.L. Stoneman, The Economic Analysis of Technology Policy, Oxford University Press, Oxford, 1987.
[50] S. Weerahandi, S.R. Dalal, A choice-based approach to the diffusion of a service: Forecasting fax penetration by market segments,
Marketing Science 11 (1) (1992) 39–53.
[51] J.M. Yeomans, Statistical Mechanics of Phase Transitions, Oxford University Press, Oxford, 1992.

Potrebbero piacerti anche