Causality in Social Network Analysis

Abstract

The role that causality can play in social network analysis is unclear. The author provides a broad characterization of social network analysis before considering the nature of causality. He distinguishes four types of causality: system causality, statistical causal-ity, mechanism causality, and algorithmic causality. Their potential places in network analysis are discussed. Understanding generative mechanisms—be they system, mechanism , or algorithmic—seems the most promising way to proceed. The role of statistical causality is a source of potential data analytic tools that can be mobilized within analyses conducted in the spirit of the other three types of causality. T he title Causality in Social Network Analysis is both precise and ambiguous. The ambiguity stems from the terms causality and social network analysis. I start with causality as a primitive term and then outline four avenues by which causality can be approached: system causality, statistical causality, (social) mechanism causality, and algorithmic causality. The potential roles each can play in social network analysis are then discussed.

Full text

SOCIOLOGICAL METHODS & RESEARCHDoreian / SOCIAL NETWORK ANALYSIS
The role that causality can play in social network analysis is unclear. The author pro-
vides a broad characterization of social network analysis before considering the nature
of causality. He distinguishes four types of causality: system causality, statistical causal-
ity, mechanism causality, and algorithmic causality. Their potential places in network
analysis are discussed. Understanding generative mechanisms—be they system, mecha-
nism, or algorithmic—seems the most promising way to proceed. The role of statistical
causality is a source of potential data analytic tools that can be mobilized within analy-
ses conducted in the spirit of the other three types of causality.
Causality in Social Network Analysis
PATRICK DOREIAN
University of Pittsburgh
T
he title Causality in Social Network Analysis is both precise
and ambiguous. The ambiguity stems from the terms causality
and social network analysis. I start with causality as a primitive term
and then outline four avenues by which causality can be approached:
system causality, statistical causality, (social) mechanism causality,
and algorithmic causality. The potential roles each can play in social
network analysis are then discussed.
1. SOCIAL NETWORK ANALYSIS
1.1. SOCIAL NETWORKS
The most straightforward definition of a social network is G=(V,
R), where Vis a set of social actors and Ris a social relation defined
over the elements of V. Each element of Ris a pair of elements from V
with R⊆V×V, the Cartesian product of Vwith itself. Put differently,
AUTHOR’S NOTE:I appreciate greatly the comments of Thomas Fararo, Tom Snijders,
Kayo Fujimoto, Lauren Springman, Esther Sales, and two anonymous reviewers. They
helped me clarify some ideas and improve the overall presentation. The errors and infe-
licities that remain are, alas, my own.
SOCIOLOGICAL METHODS & RESEARCH, Vol. 30 No. 1, August 2001 81-114
© 2001 Sage Publications
81

for i,jεV,iRj means (i,j)εR. The relational ties can be binary (0, 1) or
have magnitude.
Instead of labeling a relation as R, we can list the elements of Rand
call them lines. The network can be represented as N=(V,L), where L
is the set of lines in the relation R. For a symmetric relation, iRj ⇔jRi.
Such lines are undirected and are termed edges. For directed ties, the
lines are termed arcs. Networks having both arcs and edges can be
written N=(V,A,E), with Athe set of arcs and Ethe set of edges. A less
simple definition of a network is N=(V,R1,R2,...RK), where the social
actors have multiple (K) social relations defined over them.
Another type of network structure involving two kinds of social
actors is a membership or affiliation network. Letting Gand Hbe the
two sets of actors and λthe membership/affiliation relation, if g∈G
and h∈H, then gλhmeans that (g,h)∈λ. The data structure for this
type of relation (defined in terms of two sets of social actors) has been
termed “two mode,” in contrast to “one mode,” having only one kind
of social actor. In principle, there can be a specific data set. We could
consider three kinds of social actors (see Fararo and Doreian 1984).
Batchelder, Kumbasar, and Boyd (1997) provided an alternate treat-
ment of three-mode data. In a specific analysis of a network, we can
include multiple types of actors and multiple-mode relations. All that
is required for this discussion is that there can be multiple relations for
multiple types of actors in the definition of a social network.
Networks may be disjoint. They may overlap to differing extents,
and they may be nested within hierarchies of networks. In principle,
any structure that can be expressed in terms of social actors and social
relations can be represented in social network terms. This means that
empirical social networks can range from the very simple (e.g., an iso-
lated dyad) to the extremely complex (e.g., a nation-state). It may well
be the case that representing these social structures faithfully is
extremely difficult.
1.2. THE SOCIAL NETWORK ANALYTIC PERSPECTIVE
Social network analysts are committed, in a fundamental way, to
incorporating social network characteristics in their work. This leads
them to define their central concerns as the creation, maintenance,
transformation, and dissolution of social structures. (For Fararo
82 SOCIOLOGICAL METHODS & RESEARCH

1989:62, these are the main problems of general theoretical sociol-
ogy.) As Doreian and Stokman (1997) claimed, “network processes
are series of events that create, sustain and dissolve social structures”
(p. 3, italics added).
This commitment is paradigmatic. For network analysts, there is a
sharp distinction between information about the social actors and
information concerning the social structures within which these actors
are located. Without the latter, there are no network analyses.
Wellman (1988) distilled the central features of this paradigm, and the
following narrative focuses exclusively on his characterization.
First, “behavior is interpreted in terms of structural constraints on
activity rather than in terms of inner forces within [actors]” (Wellman
1988:20, italics added). For some analysts (me included), the “rather
than” can be replaced by “in addition to.” Second, “analyses focus on
the relations between [social actors]” (p. 20). The third feature is “A
central consideration is how the patterned relationships among multi-
ple actors jointly affect [italics added] network members’ behavior”
(p. 20). Here, the language has shaded into terms incorporating cau-
sality in some sense. The fourth feature is expressed as follows:
“Structure is treated as a network of networks that may or may not be
partitioned into discrete groups” (p. 20). Finally, “analytic methods
deal directly with the patterned, relational nature of social structure”
(p. 20).
Items 2, 4, and 5 are simply descriptions of what network analysts
do. In contrast, Items 1 and 3 are different: A primacy is placed on the
benefits of working within the network analytic paradigm. The lan-
guage Wellman (1988) used became more forceful when he outlined
the structural alternative.
Again, Wellman (1988) used five claims. In restating these claims,
emphasis has been placed on those words that seem relevant for dis-
cussing causality. “Structured social relationships are a more powerful
source of sociological explanation than person attributes of system
members” (p. 31, italics added). Whatever sociological explanation is,
it fares better, in Wellman’s view, when the (relevant) structural infor-
mation is incorporated into the explanation.
“Norms emerge [italics added] from location in structured systems
of social relationships” (Wellman 1988:33). This, from a nonnetwork
perspective, is radical. There is much more here than an emphasis on
Doreian / SOCIAL NETWORK ANALYSIS 83

structural characteristics. Often, in conventional actor-attribute–ori-
ented accounts, people’s behaviors are explained by their adherence
to norms (or values or other cultural prescriptions and proscriptions).
The structuralist argument claims that even if such accounts have
merit, the norms and so forth used in these explanations are created
structurally. The ultimate explanation, it seems, is found in social
structures.
“Social structures determine [italics added] the operation of dyadic
relationships” (Wellman 1988:35). The simplest network structure is
a dyad. At face value, a dyad involves two actors, their attributes, and
the relation(s) between them. It is possible that actor attributes have
some impact on the formation, maintenance, and dissolution of their
social ties. In a completely isolated dyad, this may be the case. How-
ever, Wellman (1988) argued that dyads are embedded in larger struc-
tures and these structures determine the operation and structure of
dyadic ties.
“The world is composed of networks, not groups” (Wellman
1988:37). We can make this (trivially) true by use of the affiliation
type of relational tie. However, as I read it, this was not Wellman’s
(1988) intent. If social actors are located in multiple networks, it is this
set of structural locations that comes first in trying to understand
behavior. Viewing the world as networks of networks—and not social
aggregates as social classes and age groups—is more fruitful than
struggling with such ill-defined aggregates in seeking to account for
behavior.
“Structural methods supplement and supplant [italics added] indi-
vidualistic methods” (Wellman 1988:38). There is no ambiguity here.
When network analysts express preferences about the units of analy-
sis, about using social relational information and constructing struc-
tural accounts, it is possible to think of structuralist methods supple-
menting whatever is used in sociological explanation. But the claim to
“supplant” other approaches is unequivocal and seems to suggest that
social scientists should use network analysis.
Consistent with this is Berkowitz’s (1982) claim that “structural
analysis does not represent a simple extension of existing social sci-
ence paradigms, but, signals instead the beginning of a scientific revo-
lution” (p. 158). The “ought” claim is now explicit. Rogers (1987)
drew attention to “how radically different network analysis is from
84 SOCIOLOGICAL METHODS & RESEARCH

conventional social science” (p. 288). Doreian (1995) suggested that it
was far too early to declare the arrival of a (structural) scientific revo-
lution. Thinking about causality reinforces that argument.
1.3. ANALYSIS OF NETWORK AND OTHER TYPES OF DATA
Before discussing issues of causality, some attention to data and
data analyses is useful. Conventional data sets are viewed in terms of
units of analysis and variables. They reflect a variable-centered
approach as described by Abell (1987). Such attribute data can be ana-
lyzed with a wide variety of statistical procedures including regres-
sion, structural equation modeling, and log-linear models. This can be
characterized as a statistics-only approach.
Based on the prospect of revolutionizing social science via net-
works—or just from the need to analyze network structures—many
special social network tools have been created to analyze network
data. Wasserman and Faust (1994) provided a compendium of such
procedures together with an informed interpretive narrative. A strict
network orthodoxy would insist that only data and analyses of this sort
are pertinent. This is a networks-only approach. Of course, statistical
tools have been used to analyze network data for network con-
cerns—see, for example, Wasserman and Pattison (1996)—and some
of the network tools have been used on conventional attribute data.
Even if this distinction between statistical and network analyses is too
simplistic, it is useful here.
It seems most flexible and fruitful to work with both network and
attribute data. Such uses can range from simply interpreting the out-
comes of one in terms of the other to the joint integrated use of tools to
analyze both types of data. Wasserman and Pattison’s (1996) adapta-
tion of logistic regression for use with network data, models of net-
work autocorrelation (Doreian 1989), and Snijders’s (1997) innova-
tive development of a class of statistical models for examining
network change provide examples. Together, they are parts of a joint
statistics and networks approach.
But useful for what? This signals Causality that it is time to come to
center stage. However, data analysis will not slink off the stage
because all discussions of causality—as an empirical matter—turn on
what we do with regard to data, our assumptions about the tools we
Doreian / SOCIAL NETWORK ANALYSIS 85

use, the nature of the world, and our assumptions about what we think
we know.
2. CAUSALITY
As libraries can be filled with the articles and books devoted to this
topic, the consideration here is woefully incomplete. The approach
taken here starts within an engineering perspective, one that provides
a conceptually straightforward approach to characterizing causality
by means of the idea of a system. I call this system causality. The Sys-
tem Causality section is followed by a section on statistical causality
that looks at the use of statistics to discern causality. Both of these sec-
tions are formulated within the variable-centered approach. Our dis-
cussion will then shift to consider social mechanism causality and
event sequences before a concluding subsection of algorithmic cau-
sality. The roles these types of causality play in network analysis is
considered in Section 3.
2.1. SYSTEM CAUSALITY
My point of departure is alien from the structuralist perspective as
described above. A (social) system can be described by a set of vari-
ables and one or more (social) processes. In a crude fashion, the opera-
tion of such a system transforms inputs at one point in time into out-
puts at a subsequent point in time. This is rather glib, as instantaneous
change is possible and to make such statements requires knowledge of
the time scales of the processes modeled.
The state (of the system) is some compact representation of the past ac-
tivity of the system complete enough to allow us to predict, on the basis
of the inputs, exactly what the outputs will be, and also update the state
itself. (Padulo and Arbib 1974:21)
Let x represent a set of, say, mvariables that characterize a system, and
let x(t) represent the state of the system at time t.
Suppose the system exists at time t0and receives inputs, represented
by z. If the state of the system at t0is x(t0), then the new state of the sys-
86 SOCIOLOGICAL METHODS & RESEARCH

tem is given by x(t1)=φ(t0,t1,x(t0), z) for some well-defined function
φ.
This function φis called the state transition map and it tells us that if we
specify two times t1and t0, a state ~
x, and an admissible input function z,
then, if we start the system in state ~
xat time t0and apply the input func-
tion z, the system will end up at state φ(t0,t1,~
x,z) at time t1. (Padulo and
Arbib 1974:26)
The output of the operation of the system is given by some function
y(t1)=η(t1,t0,x(t1), z). This characterization is for a deterministic sys-
tem. If, no matter how carefully the state of the system and the inputs
are specified, the subsequent state of the system and the output are not
predicted exactly, then the system is said to be stochastic.
If we focus on x(t), putting y(t) to one side, we can examine how
these variables might change. The x(t) values are not free to take on
any values. The state of the system constrains the possible future
states. Doreian and Hummon (1976) wrote of structural control mod-
els where the core statement of change is:
∆x(t) = γ(x*(t) – x(t))∆t, (1)
where ∆x(t) is an increment of change of the x values in an increment
of time ∆t. In this formulation, x*(t) represents the control variables
for the process and γis a parameter that represents the responsiveness
of the system to departures of x(t) from the values of the control vari-
ables, x*(t). These departures are (x*(t)–x(t)), and the specifications
are those of an equilibration model. The variables in x*(t) are taken to
be functions of some of the other variables in x(t) that represent the
state of the system. Dividing by ∆tand taking limits as ∆ttends to 0,
&
x(t) = γ(x*(t) – x(t)), (2)
where &
x(t) represents the derivatives of the system variables with
respect to time. If, for example, x*(t) = Ax + b, where A is a matrix of
parameters (with zeros on the main diagonal) and b is a vector of
parameters, the equation system representing change is
&
x(t) = γ(Ax(t) + b – x(t)). (3)
Doreian / SOCIAL NETWORK ANALYSIS 87

Equation (3) is particularly important as it expresses the coupling
of the xk(t) processes generating each of the xk. More generally, the A
and b may be time-varying parameters; even more generally, x*(t) can
be a set of nonlinear functions of the x(t).
Another source of model-building complexity is the incorporation
of exogenous impacts on the system (or inputs to the system).
Denoting these by z(t), they can be incorporated in the control variable
specification x*(t)=Ax+B+Γz(t) where Γcontains further
parameters.
Each of these differential equation systems can be represented sim-
ply as
&
x(t) = f(x(t), z(t), c), (4)
where c is a collection of parameters (or products of parameters)
drawn from γ,A,b,andΓ, and fis a well-defined function. In general,
such a differential equation system can be solved to yield
x(t) = g(x(t), z(t), c), (5)
where g is another well-defined function. If there are no exogenous
inputs, then equations (4) and (5) are written as &
x(t)=f(x(t), c) and
x(t) = g(x(t), c), respectively. Notationally and conceptually, these two
equations are those used by Fararo (1989:74-5) to represent a general
dynamical system.
Equations (4) and (5) (or their variants without exogenous inputs)
are particularly important as they describe the generators of the pro-
cesses. If the parameters, the initial conditions (at t0) and the exoge-
nous inputs, z(t), are all known, these equations can be used to gener-
ate the states of the system and its outputs at each point in time. Fararo
(1989) used the term “recursive generation” to label such processes.
By way of a partial summary, system causality is given by the oper-
ation of the generating equations, examples of which are equations (4)
and (5). With the parameters, the initial conditions, and the equations
all known, the trajectories of the system are generated. In this sense,
the idea of variables causing variables has a clear and simple meaning.
Approaching causality in this fashion permits important questions.
First, our attention is directed to the idea of initial conditions, the x(t0).
Knowledge of these conditions is needed to understand a trajectory
88 SOCIOLOGICAL METHODS & RESEARCH

(and potential trajectories). The basic (ordinary or partial) differential
equations can be viewed as providing a set of transition rules that gov-
ern the operation of the system, and knowledge of these rules also is
needed to understand the behavior generated by the system. With the
processes represented in this fashion, equilibria (if they exist) are
characterized by &
x(t) = 0. This implies f(x(t), c) = 0 which can be used
to locate the equilibrium state(s). If there are equilibria, the formal
machinery permits the determination of whether the equilibria are sta-
ble or unstable.
For these systems, there is a set of possible parameter values and a
set of possible system states. Together, these generate all conceptually
possible outcomes as results of the operation of system causality.
Actually using this idea of system causality implies two types of tasks.
One is theoretical, where the task is to explore the implications of dif-
ferent sets of parameter values. If the dynamic behavior of such sys-
tems is different for different sets of parameter values, understanding
how differences in parameters produce different behaviors and system
states is important theoretical knowledge. If the trajectories of the sys-
tem depend on the initial conditions, knowing the nature of this
dependency is also important theoretical knowledge.
The second type of task is empirical. Given an empirical system,
the task is to locate appropriate process models and estimate the val-
ues of the parameters of such models. (See Tuma and Hannan 1984 for
a discussion of such methods.) The operation of a specific empirical
system can be understood. To do this, the theoretical tasks become
particularly important. In general, a specific set of parameter values
can generate many trajectories depending on where the process is acti-
vated. Empirically, we observe a very small number of instantiations
(or realizations) of a generative process, and it is useful to locate these
trajectories in the space of possible trajectories.
2.2. STATISTICAL CAUSALITY
A causal law is a statement or proposition in a theory which says there
exist environments in which change in the value of one variable is asso-
ciated with a change in the value of another variable and can produce
this change without change of other variables in the environment.
(Stinchcombe 1968:31)
Doreian / SOCIAL NETWORK ANALYSIS 89

I think Stinchcombe (1968), in a very influential document, is writing
about “propagated effects” in the sense of Doreian and Hummon
(1976) and can be viewed, in a limited way, as consistent with the op-
eration of system causality. Yet, the book was influential also in the
causal modeling literature.
2.2.1. Structural Equation Modeling
Two of the most used tools in sociological research are (linear)
regression and structural equation modeling. In this context, there is
some sense that an effect is obtained from a cause with the expression
of some model with parametric values. For regression, the most gen-
eral statement of a model is
y = Xβ+⑀, (6)
where ⑀is independently distributed as N(0, σ2I). The variables in X
are used to predict y through the estimation of the parameters (βand σ)
in the model. Estimating the parameters of this type of model is more
straightforward than estimating parameters in systemic causality
models—which may account for their popularity. Some asymmetry is
imposed by specifying the predictor variables and the response vari-
able(s). For structural equation modeling (SEM), the observed vari-
ables are partitioned into those that are exogenous (whose variation is
taken as a given) and the endogenous variables (whose variation is to
be explained). The structural model can be stated as
η=βη +Γξ +ζ, (7)
where ηis a vector of unobserved endogenous variable, ξis a vector
of unobserved exogenous variables, ζis a vector of disturbance terms,
and βand Γare matrices of parameters to be estimated. There is also a
measurement model given by
y = Λyη+⑀and x = Λxξ+υ, (8)
where εand υare vectors of white noise terms. The two Λmatrices
specify which unobserved variables have which observed indicators.
90 SOCIOLOGICAL METHODS & RESEARCH

The immense appeal here is the prospect of estimating equations with
linked latent variables. Together, the structural model and measure-
ment model, with specific numerical values, can be used to construct a
covariance matrix for the observed variables,
$
Σ. This may or may not
be close to the observed covariance matrix, S. If
$
Σis sufficiently close
to S, the specific model (with its parameter values) is said to fit the
data. And, if the correspondence between
$
Σand S is poor, then the
model with its parameter values does not fit. The empirical task is to
specify a substantively based model, estimate it, and assess its fit. (See
Bollen and Long 1993 for a collection of methods for assessing the fit
of an estimated SEM.) Models that do not fit tend to be respecified,
reestimated, and tested again, an iterative process that ends with a fit-
ted model (or no model at all). In the following, it is assumed that for a
given covariance structure, there is at least one model that fits. It is in
this sense that I use the term statistical causality.
2.2.2. Causal Modeling
For my purposes here, causal modeling is the use of statistical
machinery to determine or locate causal relationships between vari-
ables. This includes sifting partial correlations, uses of regressions,
path analysis, and the use of SEM. Although there are many ways of
approaching this topic, I will focus on three.
From correlations to causality. Morgan (1997) provided a con-
structive history of the sifting of economic statistics in the form of
time series. She described changes in the collective mindsets of those
who sifted this kind of evidence. Initially, there was a concern to un-
derstand whole series of one (or more) variables with an emphasis on
examining variation (as do causal modelers today.) Morgan argued
that these (early) practitioners understood that “causes were generally
treated as being historically contingent” (p. 61). There was a recogni-
tion also of a need to disentangle multiple causes and a recognition
that different causes can operate in different periods.
Over time, it seemed that the mind-set shifted so that if a causal
relation was assumed, the “correlation (could be) interpreted in terms
of the strength of an atemporal cause-effect or functional relationship”
(Morgan 1997:72). Morgan (1997) used the (delightful) term “causal
Doreian / SOCIAL NETWORK ANALYSIS 91

stories” for substantive tales woven around particular correlational
analyses and concluded that “different causal stories could still be
advanced and supported by appeal to the same set of data” (p. 73).
Yet another change in the mind-set came in the 1930s and 1940s
with the recognition that correlations were rather limited and that mul-
tiple equations were needed. “The general causal mechanism was now
understood to be embedded in the relationships between [italics
added] the individual equations of the model” (Morgan 1997:75-6).
This line of thought is reflected in the formulation and estimation of
SEM models as described above. Although it may not be clear that the
formulation of a set of equations, even when estimated simulta-
neously, captures the nature of the causality, the idea is reasonable.
Morgan (1997) went on to observe that “such structural models pro-
vided the basis for a causal story in terms of causal processes which
defined the time order and interaction of the multiple causes at work”
(p. 76, italics added). There are two key ideas here: (1) the time order is
defined (or assumed) and (2) a model is specified and estimated and
then a causal story is told. I have no quarrel with this procedure when
our collective ambition is modest. The claim to be able to detect
causes through sifting covariance structures speaks to a much greater
ambition.
Regression Equations. Clogg and Haritou (1997) focus attention on
the use of regression equations in the enterprise of detecting casual
connections and the problems that occur in such efforts. Problems
arise when we look at data that are generated non-experimentally with
the ambition of substituting statistical control—via incorporating
variables—for randomization. As they noted, arguments that apply
for one equation can be extended to the use of many equations in a
multiple equation system.
Consider the equation y=βX+εas a simple bivariate regression
equation (taken as an unconditional form with centered variables).
The unconditional expectation is E[y]=βE[X]+E[ε], which becomes
E[y]=βE[X] with the assumption E[ε] = 0. If the bivariate regression
equation is multiplied by Xand expectations taken, we have σyx =βσx
2+
σxε, where σyx is the covariance between Xand y,σx
2is the variance of x,
and σxεis the covariance of Xand ε. Although both σyx and σ2are
92 SOCIOLOGICAL METHODS & RESEARCH

available from the data, this equation cannot be solved unless σxεis
known. If we assume (by convention) that σxε=E[Xε] = 0, then β=
σyx/σ2. Clogg and Haritou (1997) called E[Xε]=0acausal assump-
tion. If we start with equation (6) as the regression equation in matrix
form for a set of regressors, multiply by X′and take expectations, we
reach E(X′y)=βE(X′X) + E(X′ε). Again, if we assume E(X′ε)=0,
obtaining βis possible. The specification E(X′ε) = 0 becomes a set of
causal assumptions. Following the statement of the regression model
(equation (6) in matrix form), the OLS regression estimator of βis
given by
$
β=(X′X)–1(X′y) where, along the way, the assumptions
E(X′ε) = 0 have been made. These assumptions can never be checked
because the OLS procedure produces residuals, $
ε, uncorrelated with
all of the variables in X. The drawback with using more control vari-
ables, Z, is that they too must be specified as being uncorrelated with
ε:E[Z′ε] = 0. Put differently, even if the concern is with the causal
impact of X on y, the number of causal assumptions goes up with the
inclusion of Z.
In an analysis for the parameter estimates, Clogg and Haritou
(1997) showed that the numerical values of the
$
βalso turn on the corre-
lations between X and ε(which are usually assumed to be zero.) They
“conclude that nothing can be learned about the causal effect of X on y
from the correlation(s) between X and y unless something is really
known about the correlation between X and ε” (pp. 95-6). They
observed that it would be more useful to think of plausible nonzero
values for these correlations than to cling blindly to the notion that
they are all zero.
Clogg and Haritou’s (1997) conclusions are bleak in their implica-
tions for causal modeling: (1) “Causal inferences are driven more by
assumptions than by data” (p. 101); (2) “It is impossible from the data
to know which causal assumptions are true” (p. 103); and (3) “Finding
the ‘best’ statistical (conditional) regression has no bearing whatso-
ever on the problem of making a causal inference about the effect of X
on y” (p. 104). Their analysis extends to multiple equation systems: “It
is difficult to distinguish between what is assumed and what is inferred
from complex ‘supermodels’” (p. 106). This makes it hard to keep the
faith in making sound causal inferences from statistical procedures
applied to conventional data.
Doreian / SOCIAL NETWORK ANALYSIS 93

Automated causal detection systems. Spirtes, Glymour, and
Scheines (1993) have proposed imaginative methods for detecting
causal models through the sifting of data in a linear world. They com-
bine three fundamental ideas: (1) the use of directed acyclic graphs
(DAGs) (which describe a causal structure), (2) probability theory
(with a focus on conditional independence), and (3) the idea of causal-
ity. Although some SEM devotees may chafe at the apparent restric-
tion of attention to recursive systems where the βin equation (7) is
lower triangular, this does not seem a pressing problem. Given a
causal model (as a DAG), a set of probability distributions for the pre-
dictor variables and the disturbance terms, the joint probability distri-
bution of all of the included variables can be generated (as is the case
for SEM). In particular, conditional independence configurations im-
plied in the model can be assessed. At this level, different causal dia-
grams can be drawn and examined. For each, implied (conditional) in-
dependencies can be checked: Some of the causal diagrams will have
patterns that are consistent with the empirical evidence whereas others
will be inconsistent with the evidence.
There seems to be serious disagreement about the intent of
so-called automated causal detection systems—whether in the form
of old style causal modeling (sifting partial correlations and playing
with path coefficients) or the newer, far more sophisticated methods.
A simple purpose is to specify a specific causal model and check it out.
This can be extended to having some causal models that are checked.
At some point, there will be a practical and conceptual limit as to how
many such models can be checked. There is an obvious question:
Would it not be grand if we did not have to specify these models and
yet still check them all automatically? An automated causal model
detector would do just that and the use of TETRAD (Spirtes et al.
1993) can be seen as an artful way of accomplishing this task. But
somewhere, a boundary between assessing a small number of (sub-
stantively based) models and checking them en masse has been
crossed.
Freedman (1997) is among the critics who have expressed grave
doubts about the methods developed by Spirtes et al. (1993). The spir-
ited exchange between Freedman and Spirtes and Scheines (1997) is
delightful and instructive for both what is written and what is not.
94 SOCIOLOGICAL METHODS & RESEARCH

Freedman’s point of departure, like that of Clogg and Haritou (1997),
is regression. “Many treatments of regression seem to take for granted
that the investigator knows the relevant variables, their causal order,
and the functional form of the relationships among them; measure-
ments of the independent variables are measured without error”
(Freedman 1997:113). Clearly, Freedman sees many problems here.
However, if all of these claims (or assumptions) are appropriate, then
there is little problem in using regression and related techniques. Nor
is the assumption of measuring independent variables without error a
major problem, and this objection, in principle, is met through the use
of measurement models, in the form of equation (8), specified with
SEM. Of course, the arguments of Clogg and Haritou will give one
pause regarding the claims of knowing the causal order of all of the
relevant variables.
Freedman (1997) went further when he wrote, “I see no cases in
which regression equations, let alone the more complex methods,
have succeeded as engines for discovering causal relations” (p. 114).
In part, this rests on the nature of the tools used and our assumptions
concerning the empirical world. In the model-building and model-
fitting realm, the tasks are mathematical and statistical. However, as
Freedman noted, statements of cause are statements about the empiri-
cal world. The use of a specific DAG as a representation of some phe-
nomenon appears to bring cause into the model by fiat. By itself, this is
not a problem if the model is being tested. But, claims about the detec-
tion of causal models go beyond this. Freedman provided examples of
nonsense models “detected” by using TETRAD. At face value, this is
damaging to the case for the causal detection program. But, from a
substantive view, this is not a serious problem as nonsense models can
be dismissed. More problematic are models that are plausible but also
incorrect. Although simulation studies can provide comfort that
causal models can be detected when the linear causal model “generat-
ing” the data is known, the deeper problem is that empirically, the true
model is unknown. Suddenly, the problems of detecting causal models
begin to rival those of estimating parameters in a system causality
model.
However, there is a sharp difference between these kinds of models.
One tries to capture the generative mechanisms of social phenomena
Doreian / SOCIAL NETWORK ANALYSIS 95

while the other seeks a numerical summary in the form of a set of
linked equations and their estimated parameters. If there is a concern
with prediction, it seems that the system causality models have a better
chance of success relative to the statistical causality models. Of
course, this holds only if they can be estimated. There is one way to
couple system causality ideas and the tools used in statistical causality
modeling. When equations like equation (4) are solved to yield equa-
tions like equation (5), these estimation equations could be estimated
via SEM procedures. I write “could” because it will be necessary to
assume equilibrium conditions. Otherwise, the parameters in the
regression or SEM equations are functions of time and other parame-
ters. Including such transients greatly complicates estimation.
2.3. MECHANISM CAUSALITY AND SEQUENCES OF EVENTS
Looking at social mechanisms and sequences of events provides
another approach to the issues raised by using causality in providing
explanations of social organization. The link between statistical cau-
sality and this discussion is provided by N. Cartwright (1997). Her
point of departure is the Spirtes et al. (1993:45) definition of a causal
structure as
an ordered pair, < V,E>, where Vis a set of variables, and Eis a set of
ordered pairs of Vwhere < V,E>isinEif and only if Xis a direct cause
of Yrelative to V.
Cartwright wrote, “alternatively, Vcan be a set of events. But we
should not be misled into thinking we are talking about specific events
occurring at particular times and places [italics added]” (p. 343).
A focus on specific events draws us into the arena of social mecha-
nisms. As is the case with causality, social mechanism is not defined
uniquely. But, the general idea as well as the theory goals are clear. For
Hedström and Swedberg (1998), the theory goal is “to explicate the
social mechanisms that generate and explain observed associations
with events” (p. 1). They added that an approach “should not be con-
fused with a purely descriptive approach that seeks to account for the
unique chain of events that lead from one situation or event to another”
(p. 1). This is a very subtle distinction: We observe sequences in the
96 SOCIOLOGICAL METHODS & RESEARCH

empirical world and seek to account for them in terms of generalized
mechanisms that are also sequences of some kinds of events.
Hedström and Swedberg insisted that the distinction be made as they
went on to observe that their vision for explanatory sociology includes
the creation of “an ensemble of such fundamental mechanisms that
can be used for explanatory purposes” (p. 2). For them, a simple
description of a sequence of events is not an explanatory account.
This is consistent with Stinchcombe’s (1998) characterization of a
mechanism as
bits of “sometimes true theory” or “model” that represent a causal pro-
cess, that have some actual or possible empirical support separate from
the larger theory in which it is a mechanism, and that generate increased
precision, power or elegance in the large-scale theories. (p. 267)
He offered the example of monopoly power and competitive environ-
ments as mechanisms (that can also be coupled and help generate eco-
nomic structures). Another example (Chase 1982; Fararo, Skvoretz,
and Kosaka 1994) has fighting and bystander effects as two mecha-
nisms generating dominance hierarchies among primates. Schelling’s
(1998) definition of an institution is close to that of Stinchcombe
(1998): “A social mechanism is a plausible hypothesis, or set of hy-
potheses, that could be the explanation of some social phenomenon,
the explanation being in terms of interactions between individuals and
other individuals, or between individuals and some social aggregate”
(pp. 32-3). (Schelling’s formulation has obvious implications for net-
work analysts.)
Another definition of a mechanism comes from Elster (1998):
“Mechanisms are frequently occurring and easily recognizable casual
patterns that are triggered under generally unknown conditions or
with indeterminate consequences” (p. 44). A major substantive task
for Elster is to establish mechanisms (compare Hedström and
Swedberg 1998) and the conditions under which they are triggered.
In an earlier analysis, Elster (1983) asserted that the antonym of a
mechanism is a black box. This is an image that social mechanism the-
orists appear to like: i1→[M1]→o1, where i1and o1are inputs and
outputs, respectively, and M1is a black box. In causal modeling, iand
oare variables, and the black box is present as a magical object
Doreian / SOCIAL NETWORK ANALYSIS 97

accepted by causal modelers. An emphasis on mechanisms takes us
inside the black box and helps explain phenomena (and not variables—
or, rather, covariances of variables).
The attack on the black box aspect of variable-oriented causal mod-
eling is rather obvious. According to Elster (1998), commenting on
the use of statistics, it is difficult to discern causality from correlations
(something the causal modelers would not deny), and he claimed this
is another reason why statistical explanations are weak and vulnera-
ble. He is joined by Sørensen (1998) in an extended critique of the
(blind) use of regression analysis where linear equations are confused
for theory. However, Sørensen had a different definition of social
mechanism, one that I think Elster would disavow: “It is an account of
how a change in some variables [italics added] is brought about”
(p. 240). Even so, because Sørensen’s specific examples are much
closer in spirit to those of the social mechanism theorists than to those
of the causal modelers, I suspect that the primary feature of a social
mechanism has little to do with the presence or absence of variables.
Mechanisms as bits of theory or understandings are pieces of substan-
tive knowledge. A commitment to social mechanisms does not ex-
clude the idea of variables nor analyses of data constructed in terms of
variables. If this is the case, then much of the attack on causal model-
ing, regression, and SEM for involving variables is misplaced. Also, I
suspect that once inside the black box [M1], we will encounter i2→
[M2]→o2. This implies that the sting in the argument of how bad it is
to not go inside the black box when using regression and SEM is rather
mild.
2.4. ALGORITHMIC CAUSALITY
Discussions of social mechanisms suggest the relevance of rules as
elements that govern social action and interaction. Hummon and
Fararo (1995) described the emergence of computation as a crucial
part of a scientific triad whose elements are “theory,” “empirical,” and
“computation.” Theory and computation are coupled to form simula-
tions: Theoretical ideas inform the construction of simulated social
actors that embody rules for action. When these rules are invoked, the
simulation generates outcomes implied by the theory built into the
rules of the simulation. These rules—deterministic or stochastic—are
98 SOCIOLOGICAL METHODS & RESEARCH

built into the code of the simulation, hence the term algorithmic cau-
sality. Notions of object-oriented programming and parallel process-
ing are used to generate rich processes. The rules can represent activi-
ties whose precise order is not determined in the simulation, yet the
global outcomes depend on the order in which activities occur. If theo-
ries are expressed in parametric form, the behavior generated by dif-
fering combinations of the parameters can be explored. In short, the
parametric space is explored to learn the implications of theory. See,
for example, Flache and Macy (1997) or Fararo (1978).
Hummon and Fararo (1995) described the link between computa-
tion and empirical as “data analysis.” This can be split into two parts.
One comprises the conventional data analyses used to estimate param-
eters, make inferences, and the like. The second part is the computa-
tion done to generate simulation outcomes into which specific
instantiations can be located. In the empirical world, we study a lim-
ited number of instantiations and, in general, cannot really locate
those from the one empirical world in the set of all of the potential
empirical worlds governed by the same process rules. The key notions
here are twofold: (1) We are better able to look at our empirical world
if we know more about how a single process operates under different
(parametric or algorithmic) conditions and (2) our simulations (and
theory expressed within them) improve when we examine their out-
comes in relation to what empirical evidence we have.
3. CAUSALITY IN SOCIAL NETWORK ANALYSIS
The distinctions between the four types of causality are not truly
tight. At some level, all are concerned with prediction and a broad
term of predictive causality covers them. Although the algorithmic
rules could be consistent with system causality in a broad view, the
rules themselves might not be embodied in generating equations.
Mechanism causality shades into algorithmic causality if the algorith-
mic rules capture the operation of the social mechanisms. Both system
causality and statistical causality provide explicit foundations for pre-
cise predictions. To the extent that a dynamic system is specified cor-
rectly, the foundations for prediction seem more secure. But, in talk-
ing about the future and trying to predict what would happen if
Doreian / SOCIAL NETWORK ANALYSIS 99

variables and a small number of parameters (usually one) change, they
are both vulnerable. Using counterfactuals—see, for example, Cox
and Wermuth (1996)—carries great risk if other parameters change in
a hypothetical future.
In general, experiments where variables are controlled and/or ran-
domization is used are not very relevant in this discussion. True, the
communication experiments of Leavitt (1951) and Bavalas (1948)
have near-legendary status among network analysts for providing evi-
dence that network structures of groups have relevance for collective
outcomes and the location of actors in a network has relevance for
actor outcomes. And, exchange theory driven experiments—for
example, those discussed in the special issue of Social Networks
(1992, Vol. 14, Nos. 3 and 4)—have provided great insight into studies
of power and bargaining. However, the brutal truth is that most empiri-
cal studies of social networks are done in uncontrolled situations. For
this realm of empirical work, experimental evidence provides, at
most, suggestive insights.
3.1. STATISTICAL CAUSALITY AND NETWORK ANALYSIS
Attempts to study network behavior require the use of both network
and attribute data. Even when a great emphasis is placed on the net-
work within which the actor is embedded, it seems overly narrow to
exclude actor attributes. If large organizations behave differently than
small organizations (in some contexts), excluding size from the items
used to understand organizational conduct seems risky. Of course, if
we do have a very long sequence of data, we may be able to model the
processes by which organizations change in size and learn that the dis-
tribution of size is fundamentally a network process. After all, as net-
work analysts believe that location in a network can facilitate or
restrict access to resources, this is not unreasonable. Similar argu-
ments can be put forward for the operation of children’s networks. At
very young ages, there are many between-gender ties. These diminish
as children get older—almost to the point of friendship choice’s being
completely segregated by gender in networks. As they get a little
older, between-gender ties increase. This description is a kind of
causal story where gender plays a role.
100 SOCIOLOGICAL METHODS & RESEARCH

For networks of actors with attributes, there are two kinds of
explanatory accounts. The ultimate network explanation is that every-
thing—for example, gender (of people) and size (of organizations)—
is generated through network processes. To empirically test this, we
will need very, very long time series of observations. The second kind
of explanation concerns phenomena as we study them without having
to go back to primordial times. Large organizations do act differently
than small organizations with regard to both intraorganizational struc-
tural processes and interorganizational structural processes. To
explain the behavior of General Motors now with regard to say, indus-
trialization and the global movement of capital, going back to the ori-
gins of the automobile industry at the end of the 19th century and the
beginning of the 20th century is, at best, a distraction. It seems that
men and women do behave differently—for whatever reason—in
many social contexts. Gender is thought to have an impact on these
processes and conditions them. The inescapable point is that social
network analysts will have to deal with both attribute and network
data.
Network analytic life might be more perverse. (Some) network
analysts construct measures of centrality, prominence, status, or
standing from network data for use as...variables to help characterize
actors! These variables have been known to go into statements such as
“The greater the centrality of an actor, the greater (or lesser). . . .” Does
statistical causality help us determine causal relations? Probably not:
All of the reasons that are cited for not being able to determine statisti-
cal causality unambiguously extend to the use of variables constructed
from network data.
The conventional regression model y = Xβ+⑀can be modified to
incorporate structural (relational) information. This is exactly what
network autocorrelation models (Doreian 1980, 1989) are designed to
do. Given a network, some matrix W can be constructed to capture the
relevant interdependency of actors. An effects model uses y = ρW+
Xβ+⑀where ρis a parameter to be estimated. The disturbance term, ⑀,
is specified in the same way as for a regression model. A disturbances
model is one where the regression model is unchanged but the distur-
bance is specified as ⑀=ρW⑀+υ,with υrepresenting the white noise
disturbance terms. A combined effects and disturbances model can be
constructed as well as models with multiple autocorrelation regimes.
Doreian / SOCIAL NETWORK ANALYSIS 101

Regardless, network information can be brought into statistical equa-
tions in the narrow sense used here. But, is this useful for providing
causal explanations?
Again, the answer is “probably not.” Here the argument is stronger.
Suppose Wexpresses in some fashion structural or regular equiva-
lence and the estimate of ρis significant. It would suggest that struc-
turally or regularly equivalent actors exhibit similarity. If the data
points are nations and Wis defined in terms of trade, the causal story is
obscure. If the colonies of England (or of France) are similar, under-
standing comes with incorporating mechanisms. If the tie between
colonies and the colonial powers is one where the latter exploit the for-
mer, the mechanism provides the explanation of why colonies do not
fare well. And, if the colonies of France and England exhibit the same
properties, invoking arguments of regular equivalence seems an
incomplete account. Again, colonial mechanisms seem more fruitful
as an explanatory account than some appeal to equivalence with
regard to structural location.
Statistical methods are used for network analysis in a variety of
other ways. A rich tradition starting with the Holland and Leinhardt
(1981) family of p1models, leading to Feinberg and Wasserman’s
(1981) application of log linear modeling tools, to Wasserman and
Pattison’s (1996) use of logit modeling methods to estimate p* models
provide clear evidence of the fruitful use of statistical procedures to
analyze social networks.
Given the uses of statistical methods, it seems that statistical cau-
sality has obvious relevance for network analysis. The downside of
this observation is that all of the limitations of statistical causality with
regard to establishing cause apply. It will be extremely difficult to
establish causality through the use of statistical tools when network
data are analyzed. One interesting approach is that of Leenders
(1997), who disentangles processes of selection (actors choosing
actors like themselves) and contagion (where actors become similar to
the actors to whom they are linked). An even stronger argument can be
made. Within the “covariance approach” there is a systematic exami-
nation of causal diagrams (Pearl 1995) as ways of integrating statisti-
cal and substantive ideas in an attempt to discern causal relations. The
focus is on reasoning with conditional probabilities (see, e.g., Pearl,
Geiger, and Verma 1990) in a philosophical context where relations
102 SOCIOLOGICAL METHODS & RESEARCH

between variables—represented in a DAG—are paramount. At a min-
imum, network analysts will need to construct—or adapt— such a
framework to examine statistical causality for networks. Cox and
Wermuth (1996) provided a helpful discussion of issues involved in
the context of conventional variables.
3.2. SYSTEM CAUSALITY AND NETWORK ANALYSIS
System causality was cast in terms of generative rules or prediction
equations for variables. One fragment of studying this concerns the
relational ties between social actors. For two actors, iand j, let one of
the ties between them be aij. If we abandon a static approach to
describing networks, the ties between iand jare better written as aij(t)
to allow that these ties can change through time. Thus, if the relational
tie concerns liking, aij(t) represents the strength of i’s liking of jat time
t. Or, for the nations and the relation exports, aij(t) is the volume of
exports from ito jat time t. Expressed in these terms, the ties them-
selves are variables, and the language of both system causality and sta-
tistical causality apply to the relational items. Certainly, there will be
greater difficulty in trying to establish formal models in explicit sys-
tem terms and estimating them. Although Doreian (1979) provided a
discussion of structural control models applied to the ties between
actors in a small group, there was no discussion of estimating such
models.
The general image is one of a system of nactors that are character-
ized by {xr(t), ys(t)} as variables that change through time. If there is a
set of relations {Rm(t)} whose elements are also time varying, we have
a system where both the relational ties and actor attributes can vary
through time. The state space is characterized by variables measuring
attributes and structures, all of which are to be modeled.
Stepping back from the differential equation imagery, Fararo
(1989:80) described four choice sets: The state space variables can be
continuous or discrete, the parameter space can be continuous or dis-
crete, the time domain can be continuous or discrete, and the generator
can be deterministic or stochastic. All combinations of these items can
be used to describe dynamic systems. In principle, while these can be
modeled, there are very few attempts to do so empirically.
Doreian / SOCIAL NETWORK ANALYSIS 103

3.3. EVENT SEQUENCES AND SOCIAL NETWORK MECHANISMS
It is possible to describe event sequences of both network and
attribute data. A simple example is found in institutionalist theories of
interorganizational relations and organizational forms. Consider the
dyad of a mental health office (MHO), a large major funding organiza-
tion for mental health care and a residential care provider (RHOME).
Suppose further that the RHOME is critically dependent on the MHO
for funds and clients. In general, there will be an interorganizational
network containing multiple large funding organizations and many
residential care facilities. Suppose we have data that show the residen-
tial facilities are structured in very similar ways. Consider the follow-
ing hypothesis, one that fits well with statistical causality: “The
greater the dependence of an organization on another organization, the
more similar it will become to that organization in structure, climate,
and behavioral focus” (DiMaggio and Powell 1991:74). Substan-
tively, this is a nonstarter as it would be absurd for a small residential
facility to structure itself as if it were a funding organization. A more
compelling hypothesis would claim, “The greater the dependence of
an organization on another organization, the more similar it will
become in structure, climate and behavioral focus to yet other organi-
zations that are dependent—in the same fashion—on the same (focal)
organization.”
We could construct an argument cast in terms of social mecha-
nisms. Representatives of the focal organization (the MHO) could
impose criteria—minimum bed sizes, small numbers of patients in a
room, cleanliness standards, accounting procedures, and so
forth—that the RHOME must satisfy if it is to receive clients and
funding. There could be a sequence of negotiations (interactions) as a
set of working arrangements is hammered out. Institutional theorists
would recognize this as an example of a coercion mechanism. Stan-
dards are set and an organization must conform to them if it wants to
receive further funding. If the RHOME falls out of compliance, fur-
ther interactions are triggered. The actual sequences of interactions
are empirical descriptions. Although they are of narrative interest, the
explanation comes from seeing them as an integral part of a general-
ized (coercive) mechanism. Is this a causal explanation? It takes the
form of identifying an empirical event sequence as an instantiation of
104 SOCIOLOGICAL METHODS & RESEARCH

the coercion mechanism together with a specification of how this
mechanism works. If there is causality involved, it comes from the
general claim as to how the mechanism operates.
Is it possible to specify events that trigger the mechanism? Usually,
it is tied to a (predictable, routinized) budgetary cycle, but it can be
triggered by an unpredictable event (e.g., a patient dies or leaves the
residential home without permission). Once invoked, it is assumed
that the mechanism operates in a certain (causal) way, and a causal
story can be told. But, is it truly causal? If the mechanism operates in
the same fashion every time and always produces the same organiza-
tional structure, there is some basis for accepting a causal argument.
Causality has been brought into the presumed operation of the coer-
cive mechanism. This seems appropriate, but if the operation of the
coercive mechanism does not lead to the same outcome every time,
belief about the operation of a causal mechanism becomes fragile. The
director of the RHOME could seek alternative sources of funding and
not comply with the demands of the MHO. No doubt, a second-order
mechanism could be constructed, but the analysis verges on being
another just-so story.
Using social mechanisms as part of a causal explanation becomes
much more difficult if there are multiple social mechanisms that could
be operating. If the substantive issue is to understand how organiza-
tions (in a specific domain) are so alike, the operation of an institution-
alized coercive mechanism provides one account. Consider a school
of social work (SSW) that is coming up for accreditation from some
licensing authority. The timetable for these events is fixed, and the
accrediting rituals occur at predictable times. Part of the process is
coercive: If the SSW does not have certain courses, does not have
enough faculty with certain types of degrees, does not have a coherent
field placement system, and so forth, it will not be accredited fully.
But, is this a sufficient account for isomorphic structures among
places such as the SSW? No, because there are rival mechanisms.
There is a mimetic mechanism whereby the practices of successful
organizations are copied by other organizations wanting to be suc-
cessful (or wanting just to survive), and there are normative mecha-
nisms whereby norms are followed because they specify the appropri-
ate behaviors. This could be a normative mechanism—both with
regard to the creation of norms (compare Wellman 1988, above) and
Doreian / SOCIAL NETWORK ANALYSIS 105

their role in constraining behavior. Put differently, a set of mecha-
nisms is an element in a profession mechanism that governs behavior
of professionals. In the SSW example, all three mechanisms could be
operating. The explanatory task becomes one on which the mecha-
nisms are disentangled in their operation and combined in their (joint)
impact.
Abell (1987, 1993) advocated the use of narrative methods as a
route toward the construction of social theory. He assumed “a social
world to be comprised of a web of interconnected human actions and
forbearances” (Abell 1993:94). With a focus on action, he noted that
action “is characteristically situated in at least four senses”: (1) Action
is a/the consequence—usually in part—of prior actions, (2) action
may be contingent on a set of circumstances, (3) beliefs and values of
an actor inform the actions of an actor, and (4) actors can involve stra-
tegic calculation. He argued that laying out a sequence (or sequences)
of actions and interactions constitutes a narrative. At this level, such a
narrative is just an empirical sequence and, in Hedström and
Swedberg’s (1998) terms, cannot be viewed as a social mechanism.
However, Abell (1987) sought to compare narratives by homomor-
phically mapping them to a set of well-defined and known structures.
He talked of successive levels of abstraction that may bring his analy-
ses into the realm of social network mechanisms. Where this is located
in a sequence of homomorphic mappings from a very concrete empiri-
cal narrative to successive higher levels of abstracted structures is not
clear. However, Abell (1987) constructed tools for establishing a set of
homomorphic reductions of complexity into abstracted structures that
can be formally compared. There is a remarkable parallel between
Abell’s diagrams connecting events and the diagrams of Cox and
Wermuth (1996) connecting variables. This seems a parallel worthy of
exploration.
3.4. ALGORITHMIC CAUSALITY AND NETWORK ANALYSIS
We can step further back and consider an algorithm as any well-
defined computational procedure that takes some value (or a set of val-
ues) as input and produces a value (or a set of values) as output (Cormen,
Lieserson, and Rivest 1990). This output can be viewed as including
the new state of the system. Put differently, an algorithm is a sequence
106 SOCIOLOGICAL METHODS & RESEARCH

of computational steps that transforms inputs to outputs and generates
new system states. The transition rules considered in Section 2.1. are
equational. However, a well-defined sequence of computational steps
can capture process rules for many processes without having explicit
equational generators. Indeed, the whole domain of theory-driven
simulations (see, e.g., Heise 1995, Fararo and Hummon 1994) can be
viewed as examples where dynamic generating processes transform
systems. Whether equations are used is secondary. What truly counts
is the set of process rules, the study of which helps us understand the
operation of social processes. The research agenda of Fararo and
Skvoretz (1986) on E-state structuralism is an attempt to use expecta-
tion states theory to unify a set of structuralist theories via a set of
coherent rules governing structural action.
One successful area of structuralist theory is structural balance.
When Heider (1946) laid out the foundations for balance theory, it was
clear that empirical domain for the structural balance dynamics were
in actors’ heads. The generalization proposed by D. Cartwright and
Harary (1956) was a brilliant example of formalizing theory. But, it
came with a cost: The location of the dynamics was taken out of the
minds of individuals and located at the level of a social group. Their
basic structural theorem stated that if a group structure was balanced,
the group members belong to two subgroups (later called plus-sets)
where all of the positive ties were within subgroups and all of the nega-
tive ties were between subgroups. Davis (1967) generalized balance
by defining the all-negative triad as balanced and established a second
structure theorem: If a group is balanced, there are two or more
plus-sets such that all positive ties are within plus-sets and all of the
negative ties are between plus-sets. Doreian and Mrvar (1996) pro-
posed a method for locating partitions of signed networks that are as
close as possible to an ideal partition based on structural balance. They
used the line index of imbalance as a general measure to be minimized
in locating the set of best fitting partitions.
Hummon and Doreian (2001) developed a simulation based on
structural balance theory and the set of partitions that are closest to a
balanced configuration. All actors in the network have images of the
overall signed network in which they are located. The simulation runs
on two levels. One is inside the minds of the actors and a balance
mechanism is specified as a set of rules whereby actors move toward
Doreian / SOCIAL NETWORK ANALYSIS 107

balance by changing one line (either its sign or deleting it) and report-
ing the change to the group (as a second level). The actor making the
change can report to the group in a variety of ways: to the person
involved in the specific changed tie, to actors connected via a positive
tie (friends), to actors connected by any tie (positive or negative), or to
the whole group (broadcasting). The group “signed structure” is con-
structed from the reports of tie changes. The relevant point for this dis-
cussion is that the group network can be different to the actor images
of the network, and the actor images can differ from one another.
The partitions of the group network, obtained from a group-level
process, are returned to the actors as information to be incorporated by
them as they reflect on their image of the signed network. In general,
an actor has more than one option in trying to move toward balance.
This ambiguity is resolved by having the actors select one option ran-
domly. For this discussion, there are two important implications.
First, the balance mechanism is specified and operates: All individ-
ual images as well as the group structure do move toward balance.
This is not surprising nor is there much learned from the existence of
change in a given direction. What is a surprise is the sheer number of
terminal states: One process with fixed algorithmic rules can generate
many outcomes even when the same process operates. Furthermore,
the process can generate different outcomes from the same starting
point (as initial conditions). This has serious implications in the
theory-empirical-computation triad of Hummon and Fararo (1995)
and especially for the empirical-theory link. (It has implications also
for system causality and statistical causality.) In the empirical world,
we observe a relatively small number of instantiations of an assumed
process. Certainly, the number of observed outcomes will be tiny
compared to the number of outcomes that could be generated by the
same process. In the simulation study, one outcome was a single posi-
tive plus-set whereas another had many plus-sets and a large measure
of imbalance. Yet another comprised a small number of plus-sets that
were mutually hostile. Some of these outcomes were balanced, but
others were not. The same process rules generated all of them. Is this
analysis causal? The same arguments hold as in Section 2. Causality
comes from mapping a presumed mechanism generating outcomes to
an empirical network. Causality is presumed in the operation of the
network mechanism and not detected. In short, a causal story can be
108 SOCIOLOGICAL METHODS & RESEARCH

constructed to go with the presumed operation of a mechanism with a
very large number of potential outcomes.
The second implication concerns the vexing problem in balance
theory with regard to its major empirical hypothesis that signed struc-
tures move toward balance. Often, the groups did not reach exact bal-
ance. Some of the simulation outcomes are networks where each indi-
vidual’s image of the structure was balanced but the group network
was not balanced. This is substantively appealing as it is an account of
how the balance mechanism can operate without leading to a balanced
group network. The simulation is driven by an operationalized social
network mechanism in the form of algorithmic rules. The large num-
ber of potential outcomes suggests a stochastic mechanism rather than
a deterministic one. The specific change rule once a tie is selected for a
change is deterministic—for the single change of a single tie by one
actor. But, the changes can be executed in different sequences in dif-
ferent runs. The sequencing is important: Different individual and
group trajectories are generated. Staying within the realm of a simula-
tion, explorations of theories are possible and appealing. But, does it
establish cause empirically? Not really: If an empirical outcome
matches one simulation outcome, it does not follow that the empirical
outcome was generated by the kind of process built into the simula-
tion. Nor does an empirical outcome that differs from the known simu-
lation outcomes mean it was not generated by that process.
4. CONCLUSIONS
To get at causes requires a temporal framework. Some sort of pro-
cess has to be defined and seen to be operating to generate a sequence
of outcomes. This suggests that some combinations of system causal-
ity, mechanism causality, or algorithmic causality have the most
promise. Yet, this promise seems fragile—and our grasp of it tenuous.
System causality requires that we get all of the process rules right and
can estimate all (or enough) of the relevant parameters. Mechanism
causality is the most seductive as it seems to permit a simultaneous
description of the form of networks together with events that generate
both actor and network outcomes. This line of analyses builds the
causes into the functioning of the mechanisms. To the extent that
Doreian / SOCIAL NETWORK ANALYSIS 109

mechanisms are bits of sometime true theory and represent our current
state of knowledge, this has considerable appeal. If the empirical
realm is mapped cleanly—in the sense that the mechanism is relevant
and veridical for the process—“causes” for the outcomes can be dis-
cerned. The obvious drawback is that there is much circularity in this
kind of argument. Algorithmic causality is very flexible as it uses a set
of rules and the operation of these rules generates specific outcomes.
But, there is a huge amount of indeterminacy as the operation of the
algorithmic rules can generate many different trajectories—and many
of these will never be observed.
Our basic goal is to look at the empirical world and construct some
understanding as to how it works. Within the social network frame-
work, social network analysts can continue to examine network struc-
tures to discern the consequences for actors located in networks and
outcomes for the whole network. We will continue to describe net-
works and attempt to link these descriptions to network outcomes.
Asking if these analyses are therefore causal in the sense of determin-
ing causes of either actor or network outcomes may be an impossibly
broad question. The basic problem is that causal explanations are con-
structed on the basis of empirical evidence expressed in descriptions.
They can take the form of equations, mechanisms, or rules—or com-
binations of these. But, they remain specific descriptions. For all of the
types of causality discussed here, moving from a formulated causal
machine to generated data that reflect the generative mechanisms is
easy. Traveling in the reverse direction is much harder with the diffi-
culty verging on the impossible. This may help explain why simula-
tion is a popular approach for exploring the implications of a theory.
The intrinsic empirical difficulties can be ignored safely, and many of
the insights from simulations are appealing and oh so plausible.
A narrative as a straightforward description of a sequence of events
has considerable appeal. Most network analysts who study empirical
phenomena use narrative. In part, it is window dressing, but it has
more than surface interest. The risk is that the narrative becomes yet
another just-so story with events following each other in time under
convenient stage management. Once it is recognized that the only real
connection between the described events is merely temporal, the
causal enterprise is shaken. If a different event could follow a given
110 SOCIOLOGICAL METHODS & RESEARCH

event—which happens—the coupling of the events in a narrative is
loosened. And, if there could be other outcomes between two hitherto
sequential events that appear in a set of narratives, the tight coupling
between events is lost again. Here, Abell’s (1987) tools have potential at
some homomorphic level to provide rigor for constructing narratives.
The most hard-nosed assessment is that truly establishing causality
in network analysis is impossible—just as it is in the realm of statisti-
cal causality. Perhaps this answer is premature, and the question is ill
posed. There is a far more constructive conclusion: There needs to be a
very tight coupling of theory, mechanisms, and credible empirical
information before we can delineate the actual operation of causes in
the empirical world before we can tell causal stories. Understanding
generative mechanisms—be they system, mechanism, or algorithmic—
seems the most promising way to proceed. The role of statistical cau-
sality is a source of potential data analytic tools that can be mobilized
within analyses conducted in the spirit of the other three types of cau-
sality. The rigor of thought in the form of causal diagrams will be
another very useful import from the domain of statistical causality.
Two further points follow: (1) Scorning such a fruitful source is fool-
ish and (2) until we do make significant progress with the incorpora-
tion of causality into network analysis, it seems best not to crow about
the alleged supremacy of network analytic accounts of how the world
works.
REFERENCES
Abell, P. 1987. The Syntax of Social Life. Oxford, UK: Clarendon.
Abell, P. 1993. “Some Aspects of Narrative Method.” Journal of Mathematical Sociology
18:93-134.
Batchelder, W. H., E. Kumbasar, and J. Boyd. 1997. “Consensus Analysis of Three-Way Social
Network Data.” Journal of Mathematical Sociology 22:29-57.
Bavalas, A. 1948. “A Mathematical Model of Group Structure.” Human Organization7:16-30.
Berkowitz, S. D. 1982. An Introduction to Structural Analysis: The Network Approach to Social
Research. Toronto, Canada: Butterworths.
Bollen, K. A., and J. S. Long. 1993. Testing Structural Equation Models. Newbury Park, CA:
Sage.
Cartwright, D., and F.Harary. 1956. “Structural Balance: A Generalization of Heider’s Theory.”
Psychological Review 63:277-92.
Doreian / SOCIAL NETWORK ANALYSIS 111

Cartwright, N. 1997. “What Is Causal Structure?” Pp. 343-57 in Causality in Crisis: Statistical
Methods and the Search for Causal Knowledge in the Social Sciences, edited by V. R.
McKim and S. Turner. Notre Dame, IN: University of Notre Dame Press.
Chase, I. 1982. “Dynamics of Hierarchy Formation: The Sequential Developmentof Dominance
Relationships.” Behavior 80:218-40.
Clogg, C. C., and A. Haritou. 1997. “The Regression Method of Causal Inference and a Dilemma
Confronting This Method.” Pp. 83-112 in Causality in Crisis: Statistical Methods and the
Search for Causal Knowledge in the Social Sciences, edited by V. R. McKim and S. Turner.
Notre Dame, IN: University of Notre Dame Press.
Cormen, T. H., C. E. Lieserson, and R. L. Rivest. 1990. Introduction to Algorithms. Cambridge,
MA: MIT Press.
Cox, D. R., and N. Wermuth. 1996. Multivariate Dependencies: Models, Analysis and Interpre-
tation. London: Chapman and Hall.
Davis, J. A. 1967. “Clustering and Structural Balance in Graphs.” Human Relations 20:181-87.
DiMaggio, P. J., and W. W. Powell. 1991. “The Iron Cage Revisited: Institutional Isomorphism
and Collective Rationality in Organizational Fields.” Pp. 63-107 in The NewInstitutionalism
in Organizational Analysis, edited by W. W. Powelland P. J. DiMaggio. Chicago: University
of Chicago Press.
Doreian, P. 1979. “Structural Control Processes for Group Processes.” Pp. 201-21 in Perspec-
tives on Social Network Research, edited by P. W. Holland and S. Leinhardt. San Diego, CA:
Academic Press.
Doreian, P. 1980. “Linear Models with Spatially Distributed Data: Spatial Disturbances or Spa-
tial Effects?” Sociological Methods & Research 9:29-60.
Doreian, P. 1989. “Network Autocorrelation Models: Problems and Prospects.” Pp. 369-89 in
Spatial Statistics: Past, Present and Future, edited by D. A. Griffith. Institute of Mathemati-
cal Geography.
Doreian, P. 1995. “Social Network Analysis as a Scientific Revolution: Thinking in Circles or
Genuine Progress?” Pp. 1-21 in International Conference on Social Networks, edited by M.
G. Everett and K. Ronnolls. London: University of Greenwich.
Doreian, P., and N. P. Hummon. 1976. Modeling Social Processes. New York: Elsevier.
Doreian, P., and A. Mrvar. 1996. “A Partitioning Approach to Structural Balance.” Social Net-
works 18:149-68.
Doreian, P., and F. Stokman, eds. 1997. Evolution of Social Networks. New York: Gordon and
Breach.
Elster, J. 1983. Explaining Technical Change. Cambridge, UK: Cambridge University Press.
Elster, J. 1998. “A Plea for Mechanisms.” Pp. 45-73 in Social Mechanisms: An Analytic Ap-
proach to Social Theory, edited by P. Hedström and R. Swedberg. Cambridge, UK: Cam-
bridge University Press.
Fararo, T. J. 1978. “An Introduction to Catastrophies.” Behavioral Science 23:291-317.
Fararo, T.J. 1989. The Meaning of General Theoretical Sociology: Tradition and Formalization.
New York: Cambridge University Press.
Fararo, T. J., and P. Doreian. 1984. “Tripartite Structural Analysis: Generalizing the
Breiger-Wilson Formalism.” Social Networks 6:141-76.
Fararo, T. J., and N. P. Hummon. 1994. “Discrete Event Simulation and the Theoretical Models
in Sociology.” Advances in Group Processes 11:25-66.
Fararo, T.J., and J. Skvoretz. 1986. “Unification Research Programs: Integrating Two Structural
Theories.” American Journal of Sociology 92:1183-1209.
Fararo, T. J., J. Skvoretz, and K. Kosaka. 1994. “Advances in E-State Structuralism: Further
Studies in Dominance Structure Formation.” Social Networks 16:233-65.
112 SOCIOLOGICAL METHODS & RESEARCH

Feinberg, S. E., and S. S. Wasserman. 1981. “Categorical Data Analysis of Single Sociometric
Relations.” Pp. 156-92 in Sociological Methodology,1981, edited by S. Leinhardt. San Fran-
cisco: Jossey-Bass.
Flache, A., and M. W.Macy. 1997. “The Weaknessof Strong Ties: Collective Action Failure in a
Highly Cohesive Group. Pp. 19-44 in Evolution of Social Networks, P. Doreian and F.
Stokman. New York: Gordon and Breach.
Freedman, D. A. 1997. “From Association to Causation From Regression.” Pp. 113-61 in Cau-
sality in Crisis: Statistical Methods and the Search for Causal Knowledge in the Social Sci-
ences, edited by V. R. McKim and S. Turner. Notre Dame, IN: University of Notre Dame
Press.
Hedström, P., and R. Swedberg. 1998. “Social Mechanisms: An Introductory Essay.” Pp. 1-31 in
Social Mechanisms: An Analytic Approach to Social Theory, edited by P. Hedström and R.
Swedberg. Cambridge, UK: Cambridge University Press.
Heider, F. 1946. “Attitudes and Cognitive Organization.” Journal of Psychology 21:107-12.
Heise, D. R. 1995. Journal of Mathematical Sociology: Sociological Algorithms. New York:
Gordon and Breach.
Holland, P. W., and S. Leinhardt. 1981. “An Exponential Family of Probability Distributions for
Directed Graphs.” Journal of the American Statistical Association 76:33-65.
Hummon, N. P., and P. Doreian. 2001. Dynamics of Social Balance Processes. Pittsburgh, PA:
University of Pittsburgh.
Hummon, N. P., and T. J. Fararo. 1995. “The Emergence of Computational Sociology.” Journal
of Mathematical Sociology 20:79-87.
Leavitt, H. J. 1951. “Some Effects of Certain Communication Patterns on Group Performance.”
Journal of Abnormal Psychology 46:38-50.
Leenders, R.T.A.J. 1997. “Longitudinal Behavior of Network Structure and Actor Attributes:
Modeling Interdependence of Contagion and Selection.” Pp. 149-84 in Evolution of Social
Networks, edited by P. Doreian and F. Stokman. New York: Gordon and Breach.
Morgan, M. S. 1997. “Searching for Causal Relations in Economic Statistics: Reflections From
History.” Pp. 47-80 in Causality in Crisis: Statistical Methods and the Search for Causal
Knowledge in the Social Sciences, edited by V. R. McKim and S. Turner. Notre Dame, IN:
University of Notre Dame Press.
Padulo, L., and M. Arbib. 1974. System Theory: A Unified State-Space Approach to Continuous
and Discrete Systems. Philadelphia: W. B. Saunders.
Pearl, J. 1995. “Causal Diagrams for Empirical Research.” Biometrika 82:669-710.
Pearl, J., D. Geiger, and T. Verma. 1990. “The Logic of Influence Diagrams.” Pp. 67-87 in Influ-
ence Diagrams, Belief Nets and Decision Sciences, edited by R. M. Oliver and J. Q. Smith.
New York: John Wiley.
Rogers, E. M. 1987. “Progress, Problems and Prospects for Network Research: Investigating
Relationships in the Age of Electronic Communication Technologies.” Social Networks
9:285-310.
Schelling, T. C. 1998. “Social Mechanisms and Social Dynamics.” Pp. 32-44 in Social Mecha-
nisms: An Analytic Approach to Social Theory, edited by P. Hedström and R. Swedberg.
Cambridge, UK: Cambridge University Press.
Snijders, T. A. 1997. “Stochastic Actor-Oriented Models for Network Change.” Pp. 185-208 in
Evolution of Social Networks, edited by P. Doreian and F. Stokman. New York: Gordon and
Breach.
SPrensen, A. B. 1998. “Theoretical Mechanisms and the Empirical Study of Social Processes.”
Pp. 238-66 in Social Mechanisms: An Analytic Approach to Social Theory, edited by
P. Hedström and R. Swedberg. Cambridge, UK: Cambridge University Press.
Doreian / SOCIAL NETWORK ANALYSIS 113

Spirtes, P., C. Glymour, and R. Scheines. 1993. Causation, Prediction, and Search. Lecture
Notes in Statistics 81. New York: Springer-Verlag.
Spirtes, P., and R. Scheines. 1997. “Reply to Freedman.” Pp. 163-76 in Causality in Crisis: Sta-
tistical Methods and the Search for Causal Knowledge in the Social Sciences, edited by V. R.
McKim and S. Turner. Notre Dame, IN: University of Notre Dame Press.
Stinchcombe, A. L. 1968. Constructing Social Theories. New York: Harcourt, Brace and World.
Stinchcombe, A. L. 1998. “Monopolistic Competition as a Mechanism: Corporations, Univer-
sities, and Nation-States in Competitive Fields.” Pp. 267-305 in Social Mechanisms: An An-
alytic Approach to Social Theory, edited by P. Hedström and R. Swedberg. Cambridge, UK:
Cambridge University Press.
Tuma, N., and M. T. Hannan. 1984. Social Dynamics: Models and Methods. Orlando, FL: Aca-
demic Press.
Wasserman, S., and K. Faust. 1994. Social Network Analysis: Methods and Applications. Cam-
bridge: Cambridge University Press.
Wasserman, S., and P. Pattison. 1996. “Logit Models and Logistic Regressions for Social Net-
works: An Introduction to Markov Graphs and p*.” Psychometrika 61:401-25.
Wellman,B. 1988. “Structural Analysis: From Method and Metaphor to Theory and Substance.”
Pp. 19-61 in Social Structures: A Network Approach, edited by B. Wellman and S. D.
Berkowitz. University of Cambridge, Cambridge, England.
Patrick Doreian is a professor and the chair of the Department of Sociology at the Uni-
versity of Pittsburgh. He is also the editor of the Journal of Mathematical Sociology. His
research interests include social network analysis, mathematical sociology and macro
social change. Running throughthese interests is a broad concern for understanding and
modeling the evolution of socially structured systems that range from small groups to the
world system. Recent publications include “Symmetric-Acyclic Decompositions of Net-
works” (with V. Batagelj and A. Ferligoj) in the Journal of Classification (2000) and
“Pre-Transitive Balance Mechanisms for Signed Social Networks” (with D. Krackhardt)
in the Journal of Mathematical Sociology (2001).
114 SOCIOLOGICAL METHODS & RESEARCH