The Uses of Simulation in Social Network Analysis: Retrospect and Prospect

militaryzoologistΤεχνίτη Νοημοσύνη και Ρομποτική

1 Δεκ 2013 (πριν από 3 χρόνια και 8 μήνες)

63 εμφανίσεις

[simuseweb.doc]

1

The Uses of Simulation in Social Network Analysis: Retrospect and
Prospect


Edmund Chattoe
-
Brown

Department of Sociology

University of Leicester

University Road

Leicester

LE1 7RH

ecb18@le.ac.uk

http://www.le.ac.uk/sociology/staff/ecb18.html


Abstract


This

paper considers a variety of methods referred to as “simulation” and the ways in which
they have been applied to social network analysis (SNA). While all of these methods are
potentially useful (though not necessarily
equally

useful), it is argued that th
eir effectiveness
depends on clear distinctions between their goals, the uses to which they can be put and a full
understanding of their distinctive methodologies. In particular, it is argued that one of the
newer techniques (agent based social simulation
or ABSS) has far from fulfilled its potential
with regard to the most pressing challenges raised by SNA. This situation may result, in part,
from confusions that have arisen between it and older techniques also referred to as
“simulation”.


Introduction



Metaphysicians armed with vague generalities, had often tried their hands at the subject, and left it no more
advanced than they found it. … vagueness was not to be met by vagueness, but by definiteness and precision:
details were not to be encountered wit
h generalities, but with details. Nor could any progress be made, on such a
subject, by merely showing that existing things were bad; it was necessary also to show how they might be made
better.” (John Stuart Mill,
London and Westminster Review
, August 183
8.)


It might be argued that reception of simulation as a technique has become a victim of its own
longevity outside the mainstream. As I shall show in this paper, diverse techniques, goals and
methodologies co
-
exist under the same broad heading and can ea
sily become confused.
Anecdotally, in presenting my research in the area to general audiences, I am surprised at how
many people still associate the term (and its potential limitations) with Forrester’s system
dynamic modelling for the Club of Rome in the
early seventies (Forrester 1971).
1

Because of
the huge scope of social science, it is easy to see how an area may be held to account for its
shortcomings of thirty or so years ago, the last time the critic acquired any detailed knowledge
of it. (At that po
int, Forrester’s work was almost front page news, at least in the quality press.)
This is bad enough when the area has retained the same broad shape but the techniques and
methodology of the newest forms of simulation are so far from those of thirty (or ev
en
fifteen) years ago as to be virtually unrecognisable. This has led me to attempt a systematic
survey of uses of simulation in social network analysis (SNA). The intention is to distinguish
the methods, their goals and associated methodologies so the rea
der can see what each
technique can be used for, how it works and what its limitations are both on its own and in



1

This phenomena is an inte
resting one for innovation diffusion (Rogers 1995). Most models in that area assume
that progress occurs from ignorance to knowledge. In certain circumstances, innovation must begin with
dispelling incorrect beliefs arising from related ones.

[simuseweb.doc]

2

relation to other methods. In particular, my intention is to suggest that the potential of a
particular technique, Agent Based Social Simulati
on (ABSS) is very far from being realised
in meeting the most pressing challenges raised by SNA. In part, this may result from the kind
of confusion I have outlined. If am wrong then, following Mill’s quotation above, at least I
may have reduced the potent
ial vagueness in the area so that others may see more clearly why
this potential remains unfulfilled to date and whether or not this is regrettable or appropriate.


Methods


The argument of this paper is supported in part by a systematic literature review.

Using
SCOPUS, I searched for all journal articles classified as social science and containing the
search terms “social network*” and “simulat*”. Using wildcards in this way gave “social
networks”, “social network” and “social networking” for example. As o
f January 2010, the
search yielded 191 articles.
2

While systematic, the review was not exhaustive because that
would not actually help the argument. Unless there is reason to suppose that these search
terms will throw up a biased sample of the target liter
ature, obtaining a larger N won’t
actually make any substantive difference. For example, later in the paper, I claim that the
predominant use of simulation in SNA is instrumental rather than descriptive. Unless my
search strategy is actually excluding desc
riptive applications particularly, the fact that most
applications
in my sample

are instrumental still makes my point in a principled way. The only
qualification of this strategy involves the section discussing the relatively few applications of
ABSS with
good grounding in SNA research. Here is might be argued that I should do an
exhaustive search including, for example, web material since the function of this section is to
describe rather than classify. However, there is a counter argument. By comparing on
ly
journal articles, I am ensuring a reasonable standard of academic quality (compared to
conference proceedings, let alone online “self publication”). If the trends I identify apply even
to the “best” research which has been successfully reviewed and publ
ished then their
consequences are relevant to to the actual practice of SNA and ABSS.
3


Gaming


Dating back to the sixties, the technique called simulation actually encompassed two very
different approaches to understanding human behaviour and their differ
ence does not seem to
have been recognised at that time (judging, for example, by the way they are mixed together
in edited volumes). The first (which for clarity can more usefully be referred to as “gaming”,



2

A few paper
s were not examined because they didn’t fit the remit of the paper. Johnson
et al
. (2009) doesn’t
deal with human networks despite being catalogued as social science. McCloskey (1991) deals only with neural
networks. Squazzoni (2008) deals with the “networ
k society” of Castells. Mizruchi
et al
. (1986) and Shih (2008)
do not appear to contain the search term “simulat*”. Morel and Ramanujam (1999) is a “think piece” referring to
simulation and social networks but not reporting on either. Doreian (2006) refers

to simulations by Hummon
(2000), which are included in the sample, but does not conduct simulations himself. In the same way, Doreian
(2002) reprises a description of a simulation actually presented in Hummon and Doreian (2003) which is
sampled. Bothner
e
t al
. (2004) refer once to their analytical model “simulating” real world outcomes but appear
not to be using the term in any of the usual senses. It would seem that, as they understand the word, any theory
purportedly describing the social world is a simu
lation. Rytina and Morgan (1982) also refer to their
diagrammatic thought experiment about group flows as a simulation, presumably on the same grounds. Germann
et al
. (2006) only use the term network once and, on closer inspection, actually present a sophi
sticated random
matching model of infection with people co
-
located in different ways at different times of day which therefore,
arguably, does
not

represent social networks.

3

In fact, doing it this way gives a ready made opportunity for someone to try and

refute my claims scientifically
using “out of sample” testing on web or other classes of publications. My hypothesis is that the concerns I raise
will be even more evident in research that is less well peer reviewed.

[simuseweb.doc]

3

another name by which is it frequently known) i
nvolves creating a setting for social actors
and then observing how they behave within it. For example, school children may be
instructed to take the parts of members of the UN Security Council and negotiate a response
to a particular situation presented t
o them as a “vignette”.
4

This activity can serve a number of
goals. It may have a pure entertainment value (“let’s pretend” played by very small children).
It can be used as a teaching tool to give insight into processes from direct experience.
(Playing su
ch games may easily be more engaging than reading about the UN though it may
not be more informative for reasons discussed below.) It can also, potentially, be used by
social scientists to learn about social behaviour. In the last two cases, it may be comp
ared
with laboratory experiments (Webster and Sell 2009). A recognisable weakness of these is the
problem of ecological validity (Brewer 2000). Can we be sure that discoveries about
behaviour made in the relatively “artificial” setting of a laboratory carr
y over to the real
world?
5

It is possible that gaming increases ecological validity by making the setting both
more engaging and more naturalistic but this is not guaranteed and the extent to which it does
so may be hard to establish. Even using relevant e
xperts (real members of the UN security
council) does not guarantee that they will take gaming as seriously as reality (or put on public
display behaviours they are normally allowed to keep private) and clearly asking children to
act as UN negotiators may
tell us something about how children negotiate (and indirectly
about what beliefs they have about the UN, as when they decide to settle a disagreement with
an arm wrestle) but probably nothing about the UN itself.
6


A relevant recent example of gaming is p
rovided by Money and Allred (2009).
7

They get
MBA students to take part in a complex negotiation game involving the construction of an
International Space Station. The ten person negotiating groups are broken up into five pairs
representing relevant actors

(NASA, Boeing, ESA and so on). These groups are allowed to
negotiate informally between formal meetings of the whole group where agreement is
supposed to be reached. The study then explores the relationships between four key variables:
satisfaction with t
he negotiation (measured by an index of survey questions for each group),
use of integrative negotiating strategies (rated subjectively by a standard scale both for ones
own group and the other groups), coalition building (measured by the number of times e
ach
group reported informal negotiations with another group) and “centrality” (which, contrary to
the way it is used in SNA, meant how many mentions each group received as being a “power
player” in the negotiations). The study was designed to test a set of

hypotheses about these
relationships formulated at the outset.





4

For example, the task is to draft a
resolution in response to the discovery that a small country has concealed
Weapons of Mass Destruction. For more on vignettes as a research technique, see Finch and Mason (1993).

5

A relevant example is provided by Mason
et al
. (2008) who create “well know
n” network structures in a lab
and then ask participants to make guesses about a hidden numerical value. Feedback on the “quality” of these
guesses is then offered not only to the guesser but to his or her ties and the propagation of guess accuracy in
diff
erent network structures is explored. While this is clearly a creative and fascinating experiment, it isn’t at all
clear which real social domains findings from “guess the number” can safely be applied to. For example,
scientific hypotheses have frameworks

of other beliefs around them which influence their acceptance rather than
simply working on a “warmer, cooler” basis.

6

There is also a design problem. One can present the “issue” to the children relatively easily but it is much
harder to give the child r
epresenting America “realistic” goals reflecting the country they are supposed to “be”
(even assuming the attributed goals aren’t themselves contested in the real world). Some gaming involves giving
different briefings to different players but it is hard t
o establish whether this actually puts them in the “mind set”
that the organiser wants them in. Even in quite austere laboratory settings with copious explanation, my
experience through “eavesdropping” is that experimental subjects see the situation very d
ifferently from the way
the experimenters intend.

7

Other examples involving gaming are Krackhardt and Stern (1988), xx.

[simuseweb.doc]

4

Clearly, such an approach suffers from all the problems previously outlined (as the authors
freely admit). What do we deduce about the real world from it? (At least, in this case, a set of
me
asurements and hypotheses are clearly stated so that these could be taken outside an
experimental setting.) How confident are we that the negotiators are both competent and
suitably “primed” to take the roles they have been assigned? Do we think an MBA stu
dent
asked to be NASA can be rendered “like NASA” by an instruction sheet when they could
equally well have been asked to be the ESA instead?


Nonetheless, with suitable regard to the objectives of the field, such a method clearly
could

contribute to SNA.
Experimental subjects have the advantage that they can simultaneously be
instructed, observed and surveyed (qualitatively or quantitatively). Provided attention is given
to the match between gaming roles and the identities of participants and what exactly
is to be
measured and why, the method may be very helpful in exploring the complex interplay
between individual attributes, structural factors and network properties where access in
naturalistic settings is problematic. Even if hypotheses generated in this

way are not regarded
as trustworthy (and in a sense, that is the whole point of hypotheses), the measures and
relations clearly established in the laboratory can subsequently be reliably tested in other
contexts. The main weakness of this study from an SN
A perspective (and from their own
perspective the researchers do what they set out to do) is that the network variables involved
really aren’t very relational as SNA understands social settings. However, this issue could
easily be addressed by collaboratio
n between SNA and experimental psychologists or
economists with similar interests.


Instrumental versus Descriptive Simulations


Probably the most fundamental distinction in the use of simulation in SNA regards the
goals

of the method. It is relatively eas
y to distinguish, at least informally, between the instrumental
and descriptive use of simulation. Instrumental simulation occurs when the analysis involved
in a piece of research (its ends) are determined first and simulation is only used subsequently
as
one possible tool (means) to achieve a technical goal which has already been established.
(For example, a differential equation may be solved analytically, using pencil and paper
approximation or using a computer to do the calculations.) Here, the value of

simulation is
measured by its ability to do the task set for it easily, quickly (or “cheaply”) and so on.
8

It is
not

measured either by the extent to which it “problematises” (potentially in a constructive
way) the ends of the research or by any intended
descriptive similarity between the way the
computer does something and the way social actors do it. The most intuitive example of the
second point comes not from SNA but from Artificial Intelligence. Adult humans can
recognise written characters. Computers

can be programmed to do this too. However, this can
either be done for instrumental reasons (because machine character recognition is quicker
and/or cheaper in mail sorting offices) or descriptive ones: Because we want to learn how
humans do the task. The

method best suited to the instrumental goal may be nothing like the
way humans do the task (for example, using brute force computation rather than heuristics).
Conversely, the descriptive goal may be best served by a programme that actually does the
task
poorly

in some respects, particularly if the errors mirror those observed in humans. (This
point will be important in discussing the methodology of ABSS below.) For example,
children start by inferring a general rule for plurals from examples (“add s”) tha
t means that,
for a while, they will misform some plurals (“sheeps”) until they learn the quite small set of



8

In some cases, the computer may be the only tool that can do the task but unless it has some part in defining
that task via an attemp
t at description of social behaviour, it is still being used instrumentally according to the
distinction proposed here.

[simuseweb.doc]

5

exceptions. This is actually an effective way to learn since the general rule will make them
“mostly right” (and generally comprehensible) initiall
y, with refinement added later.
Application of the general rule will also quickly elicit feedback on the exceptions: “No, not
sheeps, sheep.” One actually learns far more about whether a simulation is doing a task “like”
a human from its errors and omissio
ns than one would from perfectly reproducing the correct
answers (which effectively makes the functioning of the simulation a “black box”.)


It is fair to say that the great majority of simulation in SNA is instrumental.
9

Its task is to
explore the consequ
ences of theories (and particularly measures, statistical tests and models)
that have already been devised in other ways.


As an example of simulation used to examine the properties of a model, consider the work of
Mizruchi and Neuman (2008). They use inst
rumental simulation to explore the hypothesis
that estimating the network autocorrelation model (NAM, Doreian 1981) tends to produce
negatively biased parameter estimates. In a nutshell, the advantage of simulation here is that it
is possible to generate l
arge quantities of “simulated” data with any desired properties.
Normally, when one applies a test (or measure) or fits a model, one doesn’t know the
underlying properties of the Data Generating Process (DGP, Hendry 1995) Indeed, that is
precisely what one

is trying to find out on the presumption that the test, measure or model
“works”. Using simulated data, where the DGP is actually known, allows one to examine
whether, in fact, the measure, model or test is doing what it is supposed to (which in Mizruchi
and Mason’s case is produce unbiased parameter estimates).


More specifically, what Mizruchi and Mason did was to generate “random” networks
probabilistically (thus allowing for different network densities) and “construct” the “real”
regression coefficient
s (actually only real in a world of simulated data, in other words a
simulated DGP). Then, the simulated data could be sampled repeatedly and the NAM applied
to see if the regression coefficients in the DGP were reproduced accurately by fitting the
model.
They found that over almost the whole range of parameters they explored, the
estimated coefficients under
-
reported the real values “known by construction”. In particular,
they found that the negative bias in estimates was dependent on the density of the ne
twork.





9

Examples are Anderson
et al
. (1999), Bala and Sorger (2001), Bolland (1988), Burk
et al
. (2007), Butts (2001,
2007), Carayol
et al
. (2
008), Carayol and Roux (2009), Costenbader and Valente (2003), Donninger (1986), van
Duijn
et al
. (2003, 2009), Field
et al
. (2006), Frank (1995), Girvan and Newman (2002), Gnutzmann (2008),
Goodreau (2007), Handcock
et al
. (2007), Iacobucci
et al
. (1999),

Johnson
et al
. (1989), Koskinen and Snijders
(2007), Krivitsky
et al
. (2009), Mayer and Puller (2008), Moody (2001), Páez
et al
. (2008), Pemantle and
Skyrms (2004), Poulin
et al
. (2000), Silenzio
et al
. (2009), Skvoretz (1985), Skvoretz and Fararo (1986),

Tallberg (2004, 2005), Vega
-
Redondo (2006), xx. An unusual example is Adamic and Adar (2005) which uses
real network and attribute data and then simulates the effectiveness of different search strategies on those.
However, the goal of the paper is to meas
ure the effectiveness of search algorithms and not to understand how
real people search networks. A borderline example is Ashworth and Carley (2006) which intends to explore the
explanatory power of competing resource based and network accounts of team per
formance. To do this, an
existing descriptive simulation (CONSTRUCT) is used to assess the effect on team performance of removing
individuals. This measure is then used in subsequent analysis. While the quality of the results stands or falls on
the quality

of the descriptive simulation, the research goal is not to develop that simulation but to say something
about competing theories of team performance. Here, unusually, a descriptive simulation is being used
instrumentally. Smith and Stevens (1999) was clas
sified as instrumental on “best guess”. Although the term
simulation is used frequently in the article, no clear algorithm is specified and therefore it appears that the task of
the computer is to solve sets of equations presented. Compare the length of th
is list with the one for descriptive
simulations (and particularly those that are calibrated and/or validated) below.

[simuseweb.doc]

6

Clearly, this kind of instrumental use for simulation is extremely important generally and
particularly in SNA where analytical solutions and standard statistical results are in short
supply.
10

At most, there are two rather minor concerns which coul
d be raised about the
approach as exemplified by this paper. Firstly, though it seems really rather unlikely, the use
of very “unrealistic” networks (in this case random ones) may mean that the results proved
don’t apply so well to real networks. It is
jus
t

possible that the parameter estimates in real
networks are less biased or even unbiased. Secondly, and this is a slightly more convincing
expression of the same concern, this use of simulated data (and its value) proceeds on the
presumption that the mode
l being fitted is actually “true” and “complete” (i. e. it
is

the DGP).
In other words, the simulated data is generated by exactly the same set of regression variables
that are then tested against it to see if they are unbiased. In some sense, to have any
value, this
is all this particular use of simulated data can do, being a purely formal analysis of the ability
that the test has to “reconstruct” underlying patterns known to exist. On the other hand, it
does leave one wondering what might happen in the mu
ch more likely case when, faced with
real data where the DGP is inaccessible, the test was being asked to reconstruct a DGP with a
different functional form or, if that seems too far fetched, at least one of the same functional
form that was mis
-
specified
in one or more variables. Most intriguing at all (for reasons that
will be explained and discussed in the sections on ABSS), what if the DGP is not a functional
form in variables at all? Thus, although it remains a useful instrumental application of
simula
tion, there is an interesting issue about why this approach is used to answer
such

an
instrumental question, namely what are the statistical properties of a model on the
presumption it is correct. From the perspective of a non statistician, a far more inte
resting
instrumental question would seem to be, assuming the model is
unlikely

to be correct,
11

how
badly harmed is it by mis
-
specification, incorrect assumptions about functional form and so
on? While it would be ridiculous to look for the Philosopher’s St
one of a test that defies errors
and mis
-
specifications of all kinds, it is not at all unreasonable that some kinds of tests might
be much more vulnerable than others on these grounds and thus used on real data with caution
and only as a last resort.


Simu
lation Techniques and Tacit Assumptions


One thing that proved to be absent from the simulation, actually rather to my surprise were
simulations using older approaches like System Dynamics (Forrester 1971). Nonetheless, it is
worth considering briefly why
some of these techniques might be argued to have been
superseded by ABSS in any event. The answer, in a nutshell, lies in their tacit assumptions.
The best example is provided by an approach (cellular automata, CA) which, if not exactly
superseded by ABSS,

at least has its domain of application severely circumscribed.


While acknowledging that they have rather serious limitations as representations of networks

Bonacich (2001), for some reason, chooses to represent exchange structures as CA (Goles and
Martin
ez 1999).
12

In these simulated systems, each square on a regular grid can be in one of N
states and the states of each square are determined in some way by the states of its more
-
or
-



10

Because networks are recognised to violate assumptions of independence and so on on, most existing
statistical results do not apply.

11

I
n a sense, why
would

it be? As Galileo put it: “Philosophy is written in this grand book


I mean the
universe


which stands continually open to our gaze, but it cannot be understood unless one first learns to
comprehend the language in which it is writte
n.” (Il Saggiatore [The Assayer], 1623) Why do we suppose that
the part of the book about the social world happens to be written in mathematics and, more particularly, linear
equation systems? This point has also been discussed by Abbott (2001). To argue t
his from the fit of the models
when the meaningfulness of the fit depends on the assumption itself is clearly circular.

12

Another example from outside SNA is Nettle (1999).

[simuseweb.doc]

7

less immediate neighbours. There is no movement, no network dynamics and ea
ch agent has
exactly the same number of “ties” for which some common properties of real networks (like
transitivity) cannot apply for geometrical reasons. The broad findings could probably have
been anticipated without simulation, namely that if people can

move it will subvert the power
arising from limited exchange opportunities. Unfortunately, his more specific findings about
how movement can lead to dynamically stable configurations are probably very sensitive to
the particular simulation specified. Ther
e would seem to be an opportunity here for a “full”
ABSS model of exchange networks without the restrictive assumptions built into CA. On the
principle that “the assumptions you don’t realise you are making are the ones that will do you
in”, it isn’t clear

why CA (with its unavoidable and rather strange assumptions about network
architecture) is a suitable simulation technique for SNA. Similar arguments might be put
forward (but are probably too technical and lengthly for this article where they will not fu
rther
the argument particularly) for discrete event simulation (as used, for example, by Hummon
and Fararo 1995).


What is ABSS?


Although several SNA papers (particularly Hummon 2000, Hummon and Doreian 2003,
Zeggelink 1994, 1995) provide explanations of
ABSS, it is appropriate to revisit the details of
the method here. This is a new approach with fast moving associated technologies and even
accounts from 5 years ago will be considerably dated in their details. I shall begin with a brief
definition and the
n offer a specific example.


In a nutshell, one key thing that distinguishes ABSS from other approaches is the distinctive
way in which it represents social processes (theories). Most social scientists are used to
theories presented as narratives (for exam
ple ethnographies) or as equation systems (for
example econometric models). ABSS presents social theories as computer programmes. This
allows some of the richness of qualitative data to be combined with much of the rigour of
quantitative methods.
13

The reas
on the method is called “agent based” is because the
particular form of the programme specification involves descriptions of the properties of each
agent and how these change through interaction with other agents and the environment
without the kind of imp
licit “organising assumptions” built into the architectures of older
methods. (For example, an agent may acquire food by foraging, information by gossiping or
gold coins by brute force.) The interaction of agents produces emergent patterns. For
example, im
agine agents who acquire skills in work and then apply for different jobs with
associated “class positions” which they may hear about either through social networks or
“public” advertising. What is the emergent match between skill levels and class arising
from
repeated processes of skill acquisition, selection of (and application for) jobs with acceptance
or rejection, network formation through employment and so on?
14

These associations are of
the kind that is often measured quantitatively but it can be seen

in this example why it may be
hard to do this robustly. Any empirical regularities between individual attributes and
structural features (in this case the ecology of the labour market) are shorthand for a complex
(and possibly changing) set of underlying
social processes. It should not be thought, however,



13

Narratives are very rich but this very richness allows for lacunae and inconsiste
ncies to creep into the logical
and causal structure of theories. Equation systems avoid this but at the cost of making strong simplifying
assumptions (for solubility) which may not correspond to the real social domain. ABSS do not need to be
“solved”, onl
y “processed” and, as such, they retain logical and causal rigour without requiring the strong
simplifying assumptions.

14

For example, the old distinction between “spiralists” (those who ascend the class hierarchy by geographical
mobility) and “burgesses”
(who don’t, Watson 1964) might involve different associated networks across the set
of potential employers.

[simuseweb.doc]

8

that ABSS is limited to representing the kind of individual attributes dealt with by
quantitative research. As well as these (gender,
15

ethnicity, age, wealth), the same process
description can be applied

to changes in relational data (how do people make, change and
break ties) and structural phenomena (who fills which job in an organisation or goes self
-
employed and creates their own company).
16

An additional element of the ABSS approach is
the general avo
idance of aggregate variables or concepts unless these are legitimate parts of
the social world the simulation is supposed to represent. It is reasonable, for example, to
suggest that an agent decides to save a fixed fraction of its income. It is also like
ly that if this
is how people save, the fractions will differ across agents (and quite likely that not all agents
will use the same rule, a possibility that ABSS can easily represent). Thus an ABSS would
not attempt to make any use of the aggregate savings

fraction which often forms part of
simple macroeconomic models. This aggregate is a theoretical or practical “construct” (like
market price) with no correspondence in the real world. As such, using it to attempt to
describe the real world is problematic.
To the extent that savings fractions exist at all (and this
may be nothing like how people save), they reside in the heads of individuals as part of their
decision process. However, not all aggregate variables are like this. The interest rate is also an
ag
gregate but it is one that is “fed back” to individual savers by banks (institutions can also be
agents in ABSS) and can thus influence behaviour. Here, individual saving and borrowing
decisions affect the rates set by banks and these in turn affect indivi
dual decisions. Thus not
all aggregate variables are merely theory constructs, some do have real world referents and
can accordingly be represented legitimately in ABSS.


An specific example by Abdou and Gilbert (2009) with a definite network component wil
l
probably make these general points easier to grasp. The intention of this model is to
understand the distribution of different groups in labour markets. The case study concerns
Copts and Muslims in Egypt. To develop this understanding, the simulation spe
cifies that
each agent has a maximum number of ties and that it forms ties probabilistically depending on
the local context of Copts and Muslims. This means that the work ties formed by a Copt in a
Copt dominated firm will be very different from those of a

Copt in a Muslim dominated firm.
Each agent is assumed to have a preference for homophily which influences tie formation.
This preference itself evolves (so a Copt with high homophily may become more tolerant if
he spends time in a Muslim work place). Man
y links form through work contexts but a
smaller proportion form through existing networks (which may include non work ones) and a
very small proportion form randomly (“meeting on a train”.) It is assumed, though I shall
return to the point, that homophily

is the main motivation for changing jobs. Workers will
leave employers where the composition of the work force makes them uncomfortable. There
is considerably more to the behavioural description of the ABSS than this (and the reader can
consult the publis
hed article for a full specification) but a description of the components
relevant to the social network should make the point about ABSS adequately.
17




15

The agent and process based approach also requires treating such variables differently. A quantitative study
may look at the effect of “being a wo
men” on educational success. An ABSS will consider the various processes
by which the gender attribute may impact on that (selection processes, formation of different peer groups,
distinct socialisation and so on). Clearly, being a women does not “cause” e
ducational success directly. Instead,
it feeds into a dynamic social process at various points, the
outcome

of which is educational success.

16

Debate about whether this is structural (rather than relational in a different manner) is not fruitful. The self
-
employment example is clearly structural even if the job filling example is not.

17

This description illustrates the earlier question about what happens when the DGP is not an equation system. If
a descriptive behavioural specification of a social domain ca
n be built and encoded as an ABSS (which is clearly
not

an equation system), it would be surprising if the standard statistical tests (whose properties are investigated
on the presumption that they accurately represent the DGP) could characterise it robust
ly. This point is explored
in Chattoe
-
Brown and Edmonds (2009).

[simuseweb.doc]

9


Several objections can be made to the use of ABSS and I shall try to deal with all of them.
The major on
e, which is the “scientific” status of such models will be addressed separately in
the next section which discusses ABSS methodology. Another important objection might go
like this: “You have been rude about variables in statistical models, but here you ar
e
introducing exactly the same kinds of things into your simulation. What is homophily or the
network constraint if not a variable?” There are three answers to this. The first has already
been touched on. Although homophily is a variable, it is clearly “ap
propriately located” in the
individual agent using ABSS rather than at some aggregate level with no real correspondence
to the social domain as it might be, for example, in an econometric model. Variables are not
inappropriate
per se
, only operating at the

wrong level of description or aggregation. The
second is that, although it is presented here as a variable, it seems reasonable that something
like it could be measured relatively straightforwardly using standard social science methods
(in this case, perh
aps a standard psychological test battery) and thus become “a constant”.
Variables should not be equated with “fiddle factors” which can take any value to get the
desired result although these do exist in poor quality ABSS research. A properly designed
ABS
S should only contain measurable variables for reasons discussed in the next section.
Thirdly, to avoid “theories of everything” (like the notorious geographers of Borges who
made a map as big as the world which was therefore useless), we have to treat som
e aspects
of any model as “exogenous”.
18

Here, when measuring homophily, we might reasonably
expect to find that, as a psychological “disposition”, it did not change very fast and actors
were unable to give accounts of the process by which it changed. In th
ese circumstances, we
might reasonably treat it as a variable but not because, as with most quantitative analysis, this
is what the method requires but, given the goals of the model and the measurements available,
this could be shown to be an appropriate w
ay to treat it. (It is an empirical matter whether
something changes fast or slowly and whether it can be explained using a process articulated
by the social actor or not.) A final set of objections might involve the specific assumptions
made
19

but it is im
portant not to mistake the implications arising from this kind of objection.
20

A simulation containing assumptions that turn out to be wrong (or even are obviously
implausible) no more counts against simulation as a method that applying the wrong
significan
ce test to a set of data counts against statistics as a method.
21


The Methodology of ABSS and Why It Matters to SNA




18

There are general reasons for supposing that bounded social subsystems must exist. If we had to understand
the whole social world to understand any of it, we would not be able to function as

everyday social actors let
alone social scientists.

19

For example, it strikes me that many people move jobs as a result of changes in family circumstance
(marriage, children) and/or better opportunities rather than the ethnic composition of the place they

work.
However, I may turn out to be wrong about this and, even if incorporated into the simulation, this process
might

not change the outcome. The burden of proof remains on me to show, based on evidence, that the assumption
presented is wrong, and, ideal
ly, that it does have an impact on the outcome of the simulation. The least fruitful
kind of criticism (not just in simulation but also in standard qualitative and quantitative research) is one where
the critic expects their victim to do huge amounts of ex
tra work on the basis of no more than a vague intuition.
Agreeing the “burden of proof” in specific contexts of academic debate is an important way to improve the
quality of social science.

20

However, I
shall

deal with objections arising from the way that
a lot of ABSS “stereotypes” findings from
SNA as these form a barrier to fruitful collaboration.

21

“The process of tracing regularity in any complicated, and at first sight confused, set of appearances, is
necessarily tentative; we begin by making any supp
osition, even a false one, to see what consequences will
follow from it; and by observing how these differ from the real phenomena, we learn what corrections to make in
our assumption.” (John Stuart Mill,
A System of Logic, Ratiocinative and Inductive
, 185
8, p. 295).


[simuseweb.doc]

10


One of the most important things that prevents ABSS being just like quantitative research
with its variables differently aggregated (as som
e people appear to believe, thus also
potentially confusing it with the different technique of microsimulation) it is its methodology.
We can see how, in the Abdou and Gilbert paper, different kinds of data would be collected in
different ways. Insight int
o the effects of homophily, social tie composition and the
“atmosphere” of particular work places would be collected using individual level surveys,
ethnographic observation, psychological tests, vignettes and so on. By contrast, aggregate
level phenomena
like the distribution of segregation levels across employers would be
understood by reference to quantitative analysis and aggregate data. In a nutshell (more detail
is available in Gilbert and Troitzsch 2005, pp. 15
-
18), rather than fitting a model to dat
a on
the presumption that it is correct, an ABSS is a large joint hypothesis about the properties of
individual agents, the environment and their interactions not dependent on any presumption of
correctness. The test of the joint hypothesis is the ability
of simulated data emerging from
agent and environmental interaction to mirror as many different kinds of real aggregate data
as closely as possible. For example, given assumptions about how individuals make and break
ties and move jobs, can we reproduce in

the simulation the observed distribution of employer
segregation fractions or the unemployment rate of the minority group? Of course, like any
methodology, this is only a general procedure. We can no more ask “how much” similarity in
“how many” dimensions

is adequate than we can ask “how much” statistical fit is adequate
but clearly the better the fit in the greater number of dimensions, the more convincing the
simulation is.


However, the question of fit raises an important methodological issue. There are

two common
ways in which this methodology can lose its apparent rigour. The first occurs if the behaviour
to be explained is “too simple”. If all we know about a particular social domain is that there is
a linear relationship between two variables, then w
e cannot possibly justify an ABSS with
500 parameters. This (but unfortunately without the equivalent formal results) is simply the
notion of power in statistical tests. A model of any given number of variables requires a
corresponding sample size to ensur
e its fit is meaningful. With too many free variables in a
model relative to the available data, you can prove anything, thus proving nothing. In exactly
the same way, there are probably an infinite number of ABSS of reasonable size (furthermore
with signi
ficantly different process descriptions) that can output a particular linear
relationship between two variables. The second observed failure of rigour occurs when a
simulation is either designed so its parameters cannot be measured or, in practice, too man
y of
them simply aren’t measured. In this case, even with a research domain that is sufficiently
rich to discriminate competing theories, a particular simulation can be made to do almost
anything by suitable arbitrary choice of parameter values.


Thus two
key design principles are essential to making the ABSS methodology work. Firstly,
parameter values should be measurable in principle and as many as possible should be
assigned some value, however approximate, before the quality of the model can be assessed
.
22

Secondly, the domain to be explained must have sufficient complexity to justify an ABSS



22

In scientific terms what makes a model falsifiable is that an experiment to test it can be done, not that it has
already has been done. Thus, providing it is clear how a parameter would be measured (or measured more
precisely if only an appr
oximate value has been used), a model can be considered a scientifically respectable
provisional hypotheses pending further research. This contrasts with simulations in which it isn’t clear how
parameters would be measured or where almost no parameters hav
e even approximately grounded real values.
These can rightly be called “toy” or “fudge factor” models.

[simuseweb.doc]

11

approach and, at the same time, to discriminate simulations effectively.
23

In both regards
(with one exception which I will discuss), SNA is admirably suited as a dom
ain for ABSS. A
given real network can be characterised in an extremely large number of (probably somewhat
orthogonal) ways in terms of ego networks, dyadic, triadic, n
-
ary and whole network
properties. Simulating such a network in a way that matches a con
siderable number of those
measures is passing an extremely tough test thus making it rather likely that the ABSS has
really captured some underlying truth about social processes rather than just “got lucky”. At
the same time, SNA has devoted considerable e
ffort to a diverse range of measurement
techniques appropriate to networks (though with at least one significant gap to be discussed
below) thus leaving little excuse for toy models.


On the basis of these methodological points, it is possible to make an e
ffective division of the
types of ABSS that have been published with a significant network component (either
originating in SNA or ABSS itself). There are seven of these but only four will be discussed
in detail.
24

The largest, but also least convincing cla
ss are toy models which are neither
significantly calibrated nor validated, exploring the properties of purely abstract systems
which may have no applicability to the real world.
25


Quite a large proportion of models in this class originating in SNA seem to

share a Rational
Choice/game theoretic framework.
26

Agents play games on networks (for which they must
pay costs to maintain each tie) and alter their strategies by imitation of successful ones. In
addition, they may make or break ties (or move geographica
lly) in response to cooperative or
non
-
cooperative play (Buskens
et al
. 2008, Eguíliz
et al
. 2005, Hanaki
et al
. 2007, Helbing
and Yu 2009, Huisman and Snijders 2003, Marwell
et al
. 1988, Hummon 2000).
27


Models in this class have some appealing properties
. Agents can make myopic decisions that
are nonetheless recognisably rational. Imitating the best strategy amongst your neighbours is
clearly sensible as is deciding whether to change strategy or break a tie (in models where both



23

In the context of SNA, this concern may specifically apply to Zeggelink’s (1994, 1995) simulations of very
small friendship groups and the assumptions
behind Cook
et al
. (1983). It may be that these small networks with
relatively few ties are simply not complicated enough to falsify the corresponding theories.

24

The other three classes are policy simulations, heuristic simulations and simulations that ar
e calibrated but not
validated. Carley (2006) uses CONSTRUCT to look at network destabilisation strategies for terrorist networks.
The value of her policy suggestions stands or falls on the descriptive adequacy of CONSTRUCT which is not
established in the
paper. Bonacich (1990) uses a simple descriptive simulation with quite arbitrary looking
assumptions to generate hypotheses for experimental testing. However, it isn’t clear why he proceeds in this
way. There don’t seem to be any instances in the sample of

simulations that are calibrated but not validated
except, arguably, Cook
et al
. (1983) which I will discuss further below.

25

This claim can be made more rigorous by looking at the pattern of citations in the relevant papers. Simulations
stated to be “abou
t” something like social mobility far too frequently contain no citations to empirical work in
that field. Other clues are citations to extremely general works, other toy simulations (so that assumptions come
to be acceptable by “convention”) or “empirical
” insights which can be traced back to theoretical works which,
in their turn, acquired them at second hand. (This is often the way that “factoids” enter social science.) Overall,
this kind of casual analysis (which could nonetheless be made more systemati
c) is a proxy for “closeness to
data”.

26

Examples simulations not following this trend in subject matter but still neither calibrated nor validated are
Abell and Ludwig (2009), Billari
et al
. (2007), Hummon and Fararo (1995), Hummon and Doreian (2003),
Mor
one and Taylor (2004). Zeggelink (1994, 1995) is a borderline case. She deals with friendship networks but
the model she presents has some elements of “rational optimisation” based on strong information requirements.
However, there is no strategic dimensio
n.

27

Lozano
et al
. (2008) is unusual in using real network data as a substrate for simulation but the behavioural
assumptions of the model are still arbitrary. Glanville and Bienenstock (2009) also use arbitrary assumptions but
focus much more on access to

reputation information through friendship networks as a way of choosing
appropriate strategies against new co
-
players.

[simuseweb.doc]

12

options exist) on Cost
-
Ben
efit grounds. However, taken as a whole, the weaknesses of the toy
approach become rather clear. By comparison, certain assumptions in particular models seem
rather implausible. For example, Buskens allows change in the network to occur under three
differe
nt regimes: one strategy change or network change per period (based on comparison of
costs and benefits), a single strategy change and network change per period in alternating
periods and multiple network changes in a period with a single strategy change.
Even taking
this space of possibilities as given, these alternative looks arbitrary: Why not a single network
change and multiple strategy changes? In a similar vein, Eguíliz
et al
. (2005) assume that an
agent will sever a tie to a defecting strategy when
imitating that strategy but no rationale for
this rather specific mechanism is provided. Following this logic, it would be consistent to
make a tie when imitating a cooperative strategy.


The general problem with these toy models is that they are simply no
t scientific in failing to
follow a recognised evaluation strategy. However, this style of research also leads to other
problems. Because each combination of assumptions is seen as no more or less plausible than
any other (because none of them are confront
ed with data at any point), there is no pressure
for progressive research. Each paper stands alone with results conditional on an arbitrary set
of assumptions that cannot be evaluated against any other set.
28

As worrying evidence of this,
although papers by

Buskens
et al
., Hanaki
et al
. and Helbing and Yu display striking
similarities, none cites any of the others.
29


At this point it should be made clear that the great majority of ABSS simulations also
currently fall into the toy class at least with regard t
o their network components.

While ABSS
like Acosta
-
Michlik and Espaldon (2008) incorporate social networks, these are not models of
social networks in the sense that SNA would understand them. This article contains no
references to published SNA and makes
assumptions that would appear to be mixtures of
“common sense” (always a risky ingredient), ABSS “tradition” and (one suspects)
expedience. In fact, there is a kind of ABSS publication that is rapidly turning into a cliché
(fortunately it increasingly rare
ly gets published), one in which three or four different network
architectures (often chosen by reading popular science books authored by physicists!) which
are regarded as somehow “generic” are slotted into a simulation of some social phenomenon
to explor
e the effects. The conclusion is along the lines of “If the world is like this, then this
happens, but if it is like that, then that happens.” The exasperated response of the reviewer is
“But isn’t one of the jobs of the social scientist [and SNA definitel
y qualifies in this regard] to
find out how the world
is

in fact!”
30


However, it is necessary to draw the right conclusions from this apparent failure to move
beyond toy models. SNA might reasonably take the view that, if faced with a choice between



28

In fact, ABSS can help with this by deliberately building integrative models to replicate results. Unfortunately,
this is not yet done

as often as it should be.

29

In a similar vein, Billari
et al
. (2007) use a threshold model of marriage choice but don’t cite Granovetter
(1978).

30

Other examples of papers including social networks but making only “folk” assumptions not based on cited
SNA

research are Barnaud
et al
. (2008), Barreteau and Bousquet (2000), Bloomquist (2006), Hu and Wang
(2009), Kenrick
et al
. (2003),
Radosavljevic
et al
. (2009)
, Spiekermann (2009), Zimbres
et al
. (2009), Recent
examples of papers slotting in different networ
k architectures but citing no SNA research include Bakker
et al
.
(2010), xx. A further group of papers “barely” cite SNA, referring to standard texts, overviews, other modelling
papers or non social science research rather than empirical studies, for examp
le Ackland and Shorish (2009),
Burr and Chowell (2008), Furtado
et al
. (2009), Grebel
et al
. (2003), Huang and Tsai (2010), Janssen and Jager
(2001, 2002), Kaufmann
et al
. (2009), xx. SNA would now seem to be in the unfortunate position with regard to
ABS
S that people know enough to realise that networks are important but not enough to find out what SNA has
already discovered about them!

[simuseweb.doc]

13

its ow
n arbitrary assumptions and those in ABSS, it might as well stick with its own. The
importance of ABSS is that it now has (albeit only a few) exemplars that follow the
methodology effectively and thus display full scientific status. There is a real danger
that
SNA by looking unsystematically at “typical” ABSS will come away with an unduly
pessimistic view of its capabilities.


Before looking at why two papers (by Abdou and Gilbert and, to a lesser extent, Ormerod and
Wiltshire) exemplify the ABSS methodolog
y and its advantages, there is a small, interesting,
but slightly tangential group of simulations of networks which provide a useful lesson. These
(Cook
et al
. 1983, Bonacich 2001) have arisen from experimental studies of exchange
networks. Cook
et al
. dev
eloped and tested a set of hypotheses experimentally and then
developed a simulation representing them. They were then able to show the corresponding
results for networks that were too large to study experimentally while, at the same time,
calibrating the
simulated and experimental results for smaller simulated networks. Despite its
neatness, the difficulties of this approach have already been discussed. In a sense, this paper is
validating the simulation against the laboratory setting rather than against a

real social domain
so we still have no idea how applicable either the experiments or the simulations are. The
Bonacich paper has already been discussed but it is interesting to consider whether its rather
particular assumptions could be replicated in a la
boratory for validation.


In a sense, there is little to say about the Abdou and Gilbert paper except that it shows that
both calibration and validation of ABSS are perfectly feasible.
31

Microdata was gathered from
existing surveys and also directly by Abdo
u to fill gaps identified by the process of building
the simulation (2009, pp. 188
-
189), a common situation and another heuristic advantage of
applying a new method. (This is another area of opportunity for novel methodological
development, collaborative r
esearch by data collectors and modellers where modellers attend
to the constraints of data collection and data collectors work with the construction of a
simulation very much in mind.) Simulated aggregates (unemployment rates and measures of
segregation) w
ere compared with real data, displaying impressive similarity (
ibid
., p. 189). In
the same vein, though Ormerod and Wiltshire (xxxx) do not calibrate
and

validate their
model, they are a step up from toy models in that they compare a set of potential netwo
rk
structures involved in binge drinking and show that the outputs of a simulation with a
particular network structure are most consistent with the real data.
32

Nonetheless,
unfortunately, models attaining this standard of rigour, particularly with a signif
icant social
network component are still a small minority. It may be that explicitly drawing attention to
this fact may improve matters!


The Prospects for a Fruitful Interaction Between SNA and ABSS


Based on the foregoing discussion, these can be divided

into three broad areas:


THE “MISSING DIMENSION” OF BEHAVIOURAL SNA: I suggested earlier that there
was one area in which SNA had not really focused its attention to measurement and data
collection. This is the behavioural underpinnings of networks. It is

easy to see why,



31

Interestingly, there is one very old example which seems to have been forgotten even within the ABSS
community (Hägers
trand 1965). The age of this example supports the view that rigour in simulation is a matter
of attitude to the importance of data rather than advanced simulation techniques.

32

This is less conclusive because it is a joint hypothesis with all the other non
-
network assumptions of the
model. It is possible (but who knows how unlikely) that the success occurs because faulty assumptions about the
network and about other aspects of the social process cancel out!

[simuseweb.doc]

14

institutionally, this might have occurred. As a predominantly quantitative and statistical
method, it may have followed the prevailing assumption in those areas that the model
was

the
DGP and viewed the way that ties form and break as if
they
were

simply probabilistic
transitions (because there was a fit with models
representing them

thus). ABSS shows how
we can say more than this and create an interplay between behavioural specifications and new
research. For example, what is it that allo
ws strangers to start conversations and thus become
acquaintances or friends? We share public spaces with hundreds of people every day and yet
meetings with strangers (particularly those that lead to permanent ties) are exceptionally rare.
What ramificatio
ns do the social conventions of conversation (regular practices that can be
understood and programmed into an ABSS) have on the nature of these ties? (Are they more
or less homophilous than ties formed in other ways for example?) I have argued elsewhere,
t
hat, despite believing otherwise, qualitative research in its traditional forms often does not
deliver the kind of processual information needed to explore these issues (Chattoe
-
Brown
2009) and there are therefore creative opportunities for novel applicati
ons of research
methods (biographical interviews on friendships, ethnographic observation with this issue
specifically in mind, vignettes) combined with ABSS and SNA to gain a deeper
understanding of the generative processes resulting in network structure.


PROPERLY INTEGRATIVE THEORIES OF SOCIAL PROCESSES: One difficulty that
SNA (along with all other methods) faces is judging the limitations of its own applicability.
Understandably methods tend to focus on showing where they work well and not on the
warni
ng signs suggesting they shouldn’t even be applied. By understanding relational
systems, SNA has a unique contribution to make to social science but this needs to be
balanced against the equally unique contributions of models exploring individual differenc
e,
cognitive process, structure, geography and so on.
33

Disciplinary boundaries often fragment
these processes in arbitrary ways. For example, how much does social network structure owe
to geographical space? (This issue is considered, in an extremely preli
minary way, in xx.)
ABSS can serve as a relatively “neutral” setting where contributions from all these distinct
approaches can be integrated and, when the ABSS methodology is followed, assessed. There
is nothing about the technology of ABSS that requires
it to favour Rational Choice, networks,
bureaucratic hierarchies or other “partial theories” as the dominant mode of social
explanation. In fact, by focusing on detailed process, ABSS suggests exactly why such partial
theories are unlikely to be correct.


MUTUAL LEARNING: SNA and ABSS both things to contribute to their mutual advantage.
ABSS has its methodology which ensures the rigorous falsification of competing models.
SNA has its extensive knowledge of network structures and library of techniques for th
eir
measurement. However, as the survey above shows, this exchange has barely begun. Further,
each discipline may have to think about itself in slightly new ways to facilitate their mutual
development. For example, in comparing real and simulated data, we
are doing something
slightly different from model fitting. Can SNA propose the most effective index of network
measures that will establish the similarity between two networks whether these are both real
or one real and one simulated? What “generic” proper
ties of networks has SNA discovered
which can be used to populate ABSS? My experience has been that SNA (perhaps oddly for a
quantitative field) is still disposed to think of each network as an ethnographically unique
event. If we believe in the existence
of regular social practices which generate networks,
would we not expect a meta
-
analysis of network studies to yield at least some “generic
properties?” (This appears to have happened in the context of small world networks but I am



33

SNA is already developing techniques to deal wit
h some of these (like the interplay between attributes and
networks) but others (like organisational structure and geography) remain, to my knowledge, totally neglected.

[simuseweb.doc]

15

worried by the evidence
base for their generic nature and wonder if the idea that many
networks are small world is actually becoming a weakly grounded conventional belief.)
ABSS, for its part, with SNA as with many other substantive domains will have to learn to
take the trouble
to access what is already known!


References


Abbott, A. (2001) ‘Transcending General Linear Reality’, in
Time Matters: On Theory and
Method

(Chicago, IL: The University of Chicago Press), pp. 37
-
63.


Acosta
-
Michlik, L. and Espaldon, V. (2008) ‘Assessing V
ulnerability of Selected Farming
Communities in the Philippines Based on a Behavioural Model of Agent’s Adaptation to
Global Environmental Change’,
Global Environmental Change
,
18
(4), pp. 554
-
563.


Adamic, L. and Adar, E. (2005) ‘How to Search a Social Net
work’,
Social Networks
,
27
(3),
pp. 187
-
203.


Anderson B. S., Butts, C. and Carley, K. (1999) ‘The interaction of size and density with
graph
-
level indices’,
Social Networks
,
21
(3), pp. 239
-
267.


Ashworth, M. J. and Carley, K. M. (2006) ‘Who You Know versus

What You Know: The
Impact of Social Position and Knowledge on Team Performance’,
Journal of Mathematical
Sociology
,
30
(1), pp. 43
-
75.


Bala, V. and Sorger, G. (2001) ‘A Spatial
-
Temporal Model of Human Capital Accumulation’,
Journal of Economic Theory
,
96
(
1
-
2), pp. 153
-
179.


Billari, F. C., Prskawetz, A., Diaz, B.A. and Fent, T. (2007) ‘The “Wedding
-
Ring”: An
Agent
-
Based Marriage Model Based on Social Interaction’,
Demographic Research
,
17
, pp.
59
-
81.


Bolland, J. M. (1988) ‘Sorting Out Centrality: An Analy
sis of the Performance of Four
Centrality Models In Real And Simulated Networks’,
Social Networks
,
10
(3), pp. 233
-
253.


Bonacich P. (2001) ‘The Evolution of Exchange Networks: A Simulation Study’,
Journal of
Social Structure
,
2
(5), <xx>.


Bothner, M. S., S
tuart, T. E. and White, H. C. (2004) ‘Status Differentiation and the Cohesion
of Social Networks’,
Journal of Mathematical Sociology
,
28
(4), pp. 261
-
295.


Brewer, M. (2000) ‘Research Design and Issues of Validity’, in Reis, H. T. and Judd, C. M.
(eds.)
Han
dbook of Research Methods in Social and Personality Psychology

(Cambridge:
Cambridge University Press), pp. 3
-
15. [Also has Computer Simulation Models for Social
Psychology by Hastie and Stasser.]


Butts C. T. (2007) ‘Models For Generalized Location System
s’,
Sociological Methodology
,
37
(1), pp. 283
-
348.


[simuseweb.doc]

16

Chattoe
-
Brown, E. (2009) ‘
Gin and Tonic or Oil and Water? Can We Integrate Research
Methods Rather Than Just “Mixing” Them? A Case Study of Social Mobility Research
’,
Draft Paper, Department of Sociology,
University of Leicester.


Chattoe
-
Brown, E. and Edmonds, B. (2009) ‘Using Agent
-
Based Simulation to Understand
Social Networks’, Draft Paper,

Department of Sociology, University of Leicester and

Centre
for Policy Modelling, Manchester Metropolitan Universi
ty.


Donninger, Ch. (1986) ‘The Distribution of Centrality in Social Networks’,
Social Networks
,
8
(2), pp. 191
-
203.


Doreian, P. (1981) ‘Estimating Linear Models with Spatially Distributed Data’,
Sociological
Methodology
,
11
, pp. 359

388.


Doreian, P. (200
2) ‘Event Sequences as Generators of Social Network Evolution’,
Social
Networks
,
24
(xx), pp. 93

119.


Doreian, P. (2006) ‘Actor Network Utilities and Network Evolution’,
Social Networks
,
28
(2),
pp. 137
-
164.


Finch, J. and Mason, J. (1993)
Negotiating Famil
y Responsibilities

(London:
Tavistock/Routledge).


Forrester, J. W. (1971)
World Dynamics

(Cambridge, MA: Wright
-
Allen Press).


Germann, T. C., Kadau, K., Longini Junior, I. M. and Macken, C. A. (2006) Mitigation
Strategies for Pandemic Influenza in the Un
ited States’,
Proceedings of the National Academy
of Sciences of the United States of America
,
103
(15), pp. 5935
-
5940.


Girvan, M. and Newman, M. E. J. (2002) ‘Community Structure in Social and Biological
Networks’,
Proceedings of the National Academy of S
ciences of the United States of America
,
99
(12), pp. 7821
-
7826.


Granovetter, M. (1978) ‘Threshold Models of Collective Behavior’,
American Journal of
Sociology
,
83
(xx), pp. 1420
-
1443.


Hägerstrand, T. (1965) ‘A Monte Carlo Approach to Diffusion’,
Archives

Européennes de
Sociologie
,
6
(1), pp. 43
-
67.


Hendry, David (1995)
Dynamic Econometrics
, Advanced Texts in Econometrics (Oxford:
Oxford University Press).


Hummon, N. P. (2000) ‘Utility and Dynamic Social Networks’,
Social Networks
,
22
(3), pp.
221
-
249.


Hu
mmon, N. P. and Doreian, P. (2003) ‘Some Dynamics of Social Balance Processes:
Bringing Heider Back into Balance
Theory’,
Social Networks
,
25
(xx), pp. 17
-
49.


Johnson, J. C., Luczkovich J. J., Borgatti, S. P. and Snijders, T. A. B. (2009) ‘Using Social
Net
work Analysis Tools in Ecology: Markov Process Transition Models Applied to the
[simuseweb.doc]

17

Seasonal Trophic Network Dynamics of the Chesapeake Bay’,
Ecological Modelling
,
220
(22), pp. 3133
-
3140.


Koskinen, J. H. and Snijders, T. A. B. (2007) ‘Bayesian Inference for D
ynamic Social
Network Data’,
Journal of Statistical Planning and Inference
,
137
(12), pp. 3930
-
3938.


Krackhardt, D. and Stern, R. (1988) ‘Informal Networks and Organizational Crises: An
Experimental Simulation’,
Social Psychology Quarterly
,
51
(2), pp. 123
-
140.


Goles, Eric and Martinez, Servet (1999)
Cellular Automata and Complex Systems

(Boston,
MA: Kluwer Academic Publishers).


Mason, W. A., Jones, A. and Goldstone, R. L. (2008) ‘Propagation of Innovations in
Networked Groups’,
Journal of Experimental Psy
chology: General
,
137
(3), pp. 422
-
433.


McCloskey, M. (1991) ‘Networks and Theories: The Place of Connectionism in Cognitive
Science’,
Psychological Science
,
2
(6), pp. 387
-
395.


Mizruchi, M. S. and Neuman, E. J. (2008) ‘The Effect of Density on the Level o
f Bias in the
Network Autocorrelation Model’,
Social Networks
,
30
(3), pp. 190
-
200.


Mizruchi, M. S., Mariolis, P., Schwartz, M. and Mintz, B. (1986) ‘Techniques for
Disaggregating Centrality Scores in Social Networks’,
Sociological Methodology
,
16
, pp. 26
-
48.


Money, R. B. and Allred, C. R. (2009) ‘An Exploration of a Model of Social Networks and
Multilateral Negotiations’,
Negotiation Journal
,
25
(3), pp. 337
-
356.


Morel, B. and Ramanujam, R. (1999) ‘Through the Looking Glass of Complexity: The
Dynamics of
Organizations as Adaptive and Evolving Systems’,
Organization Science
,
10
(3),
pp. 278
-
293.


Ormerod and Wiltshire (xxxx)


Rogers, E. M. (1995)
Diffusion of Innovations
, fourth edition (New York, NY: Free Press).


Rytina, S. and Morgan, D. L. (1982) ‘The Ar
ithmetic of Social Relations: The Interplay of
Category and Network’,
American Journal of Sociology
,
88
(1), pp. 88
-
113.


Shih, H.
-
Y. (2008) ‘Contagion Effects of Electronic Commerce Diffusion: Perspective from
Network Analysis of Industrial Structure’,
Tec
hnological Forecasting and Social Change
,
75
(1), pp. 78
-
90.


Squazzoni, F. (2008) ‘A (Computational) Social Science Perspective on Societal Transitions’,
Computational and Mathematical Organization Theory
,
14
(4), pp. 266
-
282.


Tallberg, C. (2004) ‘Testing
Centralization in Random Graphs’,
Social Networks
,
26
(3), pp.
205
-
219.


[simuseweb.doc]

18

Tallberg, C. (2005) ‘A Bayesian Approach to Modeling Stochastic Blockstructures with
Covariates’,
Journal of Mathematical Sociology
,
29
(1), pp. 1
-
23.


Vega
-
Redondo F. (2006) ‘Building
Up Social Capital in a Changing World’,
Journal of
Economic Dynamics and Control
,
30
(11), pp. 2305
-
2338.


Watson, W. (1964) ‘Social Mobility and Social Class in Industrial Societies’, in Gluckman,
M. and Devons, E. (eds.)
Closed Systems and Open Minds: The

Limits of Naivety in Social
Anthopology

(Edinburgh: Oliver and Boyd), pp. xx
-
xx.


Webster, M. Junior and Sell, J. (eds.) (2007)
Laboratory Experiments in the Social Sciences

(Amsterdam: Academic Press/Elsevier).


Zeggelink, E. P. H. (1994) ‘Dynamics of St
ructure: An Individual Oriented Approach’,
Social
Networks
,
16
(xx), pp. 295
-
333.


Zeggelink, E. (1995)

Evolving Friendship Networks: An Individual
-
Oriented Approach
Implementing Similarity’,
Social Networks
,
17
(xx), pp. 83
-
110.