1. Top of page
  2. Abstract

A formal framework is introduced for a general class of assignment systems that can be used to characterize a range of social phenomena. An exponential family of distributions is developed for modeling such systems, allowing for the incorporation of both attributional and relational covariates. Methods are shown for simulation and inference using the location system model. Two illustrative applications (occupational stratification and residential settlement patterns) are presented, and simulation is employed to show the behavior of the location system model in each case; a third application, involving occupancy of positions within an organization, is used to demonstrate inference for the location system. By leveraging established results in the fields of social network analysis, spatial statistics, and statistical mechanics, it is argued that sociologists can model complex social systems without sacrificing inferential tractability.


  1. Top of page
  2. Abstract

Social systems take many forms and may be studied at multiple levels. Despite more than a century of effort by sociologists to subsume this diversity under a single unifying framework—the grand theoretical narratives of Spencer (1896), Sorokin (1957), Parsons (1951), and Coleman (1990) being classic examples—the goal of theoretical unification continues to prove elusive. While the possibility—or even the desirability—of a unified framework for studying social phenomena remains an open question, theorists such as Fararo and Skvoretz (1987) and Fararo (1989) persuasively argue that more modest progress can be made by identifying and exploiting connections among particular classes of social systems. Where common features are present, it may be possible to identify model families that are broad enough to permit cross-fertilization of findings while remaining narrow enough to be deployable in practical settings. This “middle range” approach (Merton 1957) is as much a methodological as it is a theoretical endeavor: to be of scientific use, even a limited unified framework requires a common set of formal representations, deductive methods, and measurement techniques. An important challenge, then, is to identify classes of social phenomena that admit such methodological and conceptual unity and to construct formal systems to facilitate their measurement and analysis.

One promising candidate for a more unified treatment is the broad class of social systems that involve the arrangement of social entities (here referred to generically as “objects”) with respect to “locations” of one sort or another. The notions of “object” and “location” employed here are intended to be quite general— for example, encompassing individuals, households, and organizations in the former case and social, economic, or physical positions in the latter. For instance, individuals may hold particular jobs in a market economy (Kelso and Crawford 1982; Sattinger 1993), households may occupy particular housing units (Massey and Denton 1993; Benenson 2004), and firms may elect to site their facilities in particular locales (Sweeney and Feser 2003; Feser, Sweeney, and Renski 2005). At a more abstract level, we may even think of relational structures as involving the allocation of individual entities to structural positions (Lorrain and White 1971; Doreian, Batagelj, and Ferlioj 2005), and discrete choice behavior as individuals' “allocation” of preference to options from a choice set (Luce 1959). What these otherwise disparate systems have in common is that they can be represented (to a first approximation, at least) in terms of a discrete set of objects, a discrete set of locations, and a mapping that assigns the elements of the first set to elements of the second set. Although the location of one object may be independent of the locations of others in the system, this is not generally the case: in many systems, the objects' location assignments depend on each other in complex ways. In the case of residential settlement, for instance, local tendencies toward homophily or xenophobia affect households' selection of housing units, leading in many cases to racial and/or ethnic segregation at large scales (Schelling 1969; Sakoda 1971; Fossett 2006). Likewise, compositional variations in the populations of jobs and job seekers interact to affect the entry of younger workers into high-ranking positions and can combine with discrimination by race or gender to maintain robust patterns of stratification within and across organizations (see Stewman [1988] for a review). Capturing the behavior of these systems thus requires modeling not only baseline tendencies for certain objects to occupy certain locations but also the impact of interdependencies and occupancy constraints.

The focus of this paper is the development of a general framework for modeling and analysis of systems that can be specified in terms of the arrangement of a finite set of objects with respect to a finite set of locations. As motivated by the examples mentioned above, the behavior of these “generalized location systems” (as we shall call them) may involve dependencies among objects and/or locations as well as constraints on which (and how many) objects can occupy particular locations at any given point in time. To deal with this challenge, we employ a core formalism (the discrete exponential family) that allows us to leverage the large literature on the stochastic modeling of systems with nontrivial dependence structures. This formalism also allows us to construct models that are applicable across a wide range of substantive contexts; that scale well to large social systems; that are readily simulated; that are specifiable in terms of directly measurable properties; and that support likelihood-based inference using (fairly) standard methods. Although framed holistically in terms of system-level behavior (an approach advocated by Mayhew [1980, 1981], among others), models generated under this framework can also be interpreted as arising from certain types of microlevel processes; where the appropriate assumptions are met, therefore, model parameters may be understood in terms of object-level behavior. While procedures for parametrization and efficient simulation of models for generalized location systems occupy the bulk of what is treated here, basic methods for likelihood-based inference on location system parameters from cross-sectional data will also be presented. These methods provide a direct mechanism for empirical evaluation of competing theoretical claims and are hence an important benefit of this approach. Finally, three illustrative examples will be shown, which demonstrate how processes of occupational stratification and residential settlement can be modeled within the location system framework.

1.1. Modeling Location Systems: Some Prior Approaches

As we might expect from the substantive diversity of generalized location systems, a number of distinct modeling approaches have been suggested for location systems in particular empirical contexts. Most commonly, researchers have modeled specific properties of occupied positions (or sets thereof) without attempting to capture the behavior of the system as a whole. Much of the literature on income and educational attainment is in this vein (e.g., Beck, Horan, and Tolbert 1978; Budig and England 2001; Joy 2003; Huffman and Cohen 2004), as is much of the work on residential segregation (Tauber and Tauber 1965; Massey and Denton 1993) and the compositional properties of neighborhoods or other regions (Galster 1982; Frey and Farley 1994). A rather different approach may be found in the family of stochastic choice models (Luce 1959; McFadden 1973), which in our terms model the allocation of a decision (object) to a set of possible options (locations). In addition to its applications in economic contexts (Loehman and De 1982; Corstjens and Gautschi 1983; Eckstein and Wolpin 1989), this approach is widely employed in geography and transportation engineering to model route selection (Bovy and Stern 1990; Oppenheim 1995). Common to all of these approaches are severe limits on the types of dependence among system elements that can be modeled. Regression-based approaches (e.g., income attainment models) do not generally treat the assignment process directly and they treat observations as independent conditional on object (and sometimes location) covariates. Spatial autocorrelation models (Anselin 1988) may be employed to capture certain types of dependence but still neglect factors such as occupancy constraints. Standard stochastic choice models, on the other hand, treat decisions as independent while accounting for the composition of the location set (though not limits on total occupancy). A much richer range of dependence can be captured by permutation or assignment models (Butts [this volume, page xxx building on the work of Hubert [1987]), although these are limited to cases where the object/location mapping is 1:1. We shall have more to say about this family later, since it can be viewed as a special case of the models studied here.

In addition to these statistically oriented families, various dynamic and agent-based approaches have also been used to study particular location systems. Some of the best known of these include Schelling's (1969) and Sakoda's (1971) models of residential settlement, which have their modern incarnation in models such as those of Benenson (2004) and Fossett (2006). White's (1970) classic treatise on “vacancy chains” can also be seen as proposing a number of dynamic models for location systems arising in organizational contexts, and it has spawned a large literature of its own (see Stewman [1988] for a review). A closely related literature in economics deals with matching models, which capture strategic interactions in contexts (e.g., labor markets) in which transactions involve discrete matches between traders (Shapley and Shubik 1971; Kelso and Crawford 1982; Roth and Sotomayor 1990). Although most research on matching models has focused on either game-theoretic solutions or large-population behavior in the deterministic limit, recent work by Zhang (2004) employs a boundedly rational stochastic choice model that recasts the Schelling model in terms of a family of potential games (Monderer and Shapley 1996; Young 1998). Such games have the property that the differences in any actor's utilities for unilateral strategy changes can be expressed as differences in a real-valued function that depends only on the change being considered and the strategies being employed by other actors (and not, for example, on which actor is involved). Potential games also have the appealing property of having well-defined equilibrium dynamics under stochastic choice (Young, 1998), which Zhang employs to deduce long-run system behavior given a particular family of utility functions. Although Zhang's (2004) discussion is limited to a specific case, the model in question belongs to a much larger family of stochastic processes that will be presented here; thus, it is possible to carry out simulation and inference for the Zhang model (and extensions of it) using the methods discussed in this paper.

Overall, then, prior efforts to understand location systems have generally been domain-specific and have broadly represented a tradeoff between models with well-understood inferential properties (e.g., regression models) and models that capture more complex interactions among elements (e.g., agent-based models). Although specialized models have merit in many situations, there is also value in pursuing a more unified approach; likewise, it is scientifically desirable to have standard methods for both deduction and inference within the same framework. A successful example of this unifying strategy can be found in the field of social network analysis, where a common formalism (the discrete exponential family) has been used to model a wide (and growing) range of structural phenomena. Because this approach is closely related to that described here, we briefly consider it before continuing to an in-depth discussion of the generalized location system framework.

1.2. Statistical Models for Dependent Systems: The Case of Network Analysis

Social networks pose substantial modeling challenges, due to the interdependence of their component parts. Formally, we may think of a network as consisting of a set of vertices (representing individual actors) that are connected by edges (representing ties between actors); such a structure is generically referred to as a graph. While the simplest models of network structure assume all ties to be independent (e.g., the famous random graph models of Erdös and Rényi [1960]), this is at best a loose approximation. Typical social networks display properties such as reciprocity or asymmetry (in which an edge from i to j depends on the state of the corresponding (j, i) edge), transitivity bias (in which the existence of an (i, j) edge is affected by the existence of a path from i to j through some intermediary, k), and even complex biases such as the avoidance of odd-length cycles (e.g., as in predominantly heterosexual networks ([Bearman et al. 2004]) or balanced negative-valence relations [Harary 1953]). Such properties may arise as an artifact of unobserved dyadic effects (such as homophily [McPherson, Smith-Lovin, and Cook 2001]; see Hoff, Raftery, and Handcock [2002] for a statistical discussion), or from intrinsic dynamics (e.g., see Carley [1991];Hummon and Doreian [2003]). Regardless of how they occur, however, these properties reflect potentially complex patterns of dependence among edges and advances in network modeling have thus required researchers to cope with this problem.

Over the past quarter-century, a great deal of progress has been made within the social network field toward developing practical models for social systems with complex dependence structures. This work began in earnest with the log-linear models of Holland and Leinhardt (1981) and Fienberg and Wasserman (1981), which were extended by various researchers (e.g., Fienberg, Meyer, and Wasserman 1985; Holland, Laskey, and Leinhardt 1983) to cover more complex cases. In a series of important developments (starting with the foundational work of Frank and Strauss [1986]), this approach was generalized to incorporate processes with at first local and then general dependence among edges (Strauss and Ikeda 1990; Wasserman and Pattison 1996; Pattison and Wasserman 1999; Robins, Pattison, and Wasserman, 1999). What united these various efforts was the use of discrete exponential families as a formalism for representing general classes of distributions on graphs. Given a random graph G drawn from a finite set of possible graphs inline image, we may write the distribution of G as Pr (G=g) ∝ exp (θTt(g)), where θ is a vector of real-valued parameters and t is a vector of real-valued functions on inline image may be chosen with very few constraints, although most common models involve counts of structural features such as mutual dyads, star formations, or triangles (some of the technical reasons for this are sketched in Wasserman and Robins [2005]). Since the log-probability of a graph is proportional to the weighted sum of its statistics, t may be intuitively understood as describing structural features that are either enhanced (where θ is positive) or suppressed (where θ is negative) by the model. Models represented in this way are known as exponential random graph (ERG) or p* models, although this is more properly understood as a description of their parametrization rather than their content; indeed, any fixed distribution on a finite inline image can be written in this manner (albeit not always parsimoniously). By providing a general framework for the parametrization of graph distributions with nontrivial dependence (via choice of t) and via its associated inferential theory (Barndorff-Nielsen 1978; Brown 1986), the discrete exponential family formalism has been central to progress in the network modeling area.

Recent work has expanded on these innovations with improved inferential strategies (Crouch, Wasserman, and Trachtenburg 1998; Snijders 2002; Hunter and Handcock 2006), new parameterizations for structural effects (Snijders et al. 2006), and an expanded understanding of the models themselves (Handcock 2003b; Robins, Pattison, and Woodcock, 2005). While these models have often been couched in purely methodological terms, it has become increasingly apparent that they can be employed to capture theoretically relevant local influences on structure formation (Robins et al. 2005; Robins and Pattison 2005), as well as (in some cases) mechanisms of structural evolution (Robins and Pattison 2001). Although much of this work is still in a fairly early stage of development, the foundations have arguably been laid for a minor revolution in structural analysis.

While time will tell if this promise is realized, the successes that have so far been obtained underscore the aforementioned value of synthesis in scientific research. Rather than arising in isolation, they have resulted from cross-application of work in fields as diverse as spatial statistics (e.g., Besag 1974, 1975) and statistical physics (Strauss 1986; Swendsen and Wang 1987), as well as innovations in computing technology and simulation methods (Geyer and Thompson 1992; Gamerman 1997). Although frequently motivated by substantive concerns (e.g., the desire to model balance-theoretic influences [Heider 1958]), modelers have also attempted to work with general formalisms that can be deployed on networks arising within many different substantive contexts. By drawing on results obtained by researchers studying structurally similar problems in other substantive areas, then, network researchers have been able to greatly accelerate development in their own field.

Cross-application of concepts and methods has led to great strides in network analysis, but there is more that can be done. As promising as the developments cited above have been, few if any attempts have been made to extend them to problems other than network formation and diffusion. It has already been noted that processes such as stratification, settlement patterns, migration, firm siting, and occupational segregation pose similar challenges of complex dependence, but they are currently studied through a variety of (generally incompatible) modeling frameworks. A more generic approach would facilitate the cross-application of findings and techniques, thereby laying the groundwork for cumulative theoretical development and, ultimately, unification (Fararo and Skvoretz 1987; Fararo 1989). As we shall see in Section 3.1, the tools for creating such a unifying framework can be found in the same modeling techniques now being employed for social networks; this paper is intended as a first step in this direction.

1.3. A Brief Comment on Notation

We here outline some general notation, which will be used in the material that follows. A graph, G, is defined as G= (V, E), where V is a set of vertices and E is a set of edges on V. When applied to sets, | · | represents cardinality; thus |V| is the number of vertices (or order) of G. In some cases (particularly when dealing with valued graphs), it will be useful to represent graphs in adjacency matrix form, where the adjacency matrix X for graph G is defined as a |V| × |V| matrix such that Xij is the value of the (i, j) edge in G. By convention, Xij= 0 if G contains no (i, j) edge. A tuple of graphs (G1, …, Gn) on common vertex set V may be similarly represented by a n× |V| × |V| adjacency array, X, such that Xi·· is the adjacency matrix for Gi.

When referring to a random variable, X, we denote the probability of a particular event x by Pr (X=x). More generically, Pr (X) refers to the probability mass function of X (where X is discrete). Expectation is denoted by the operator E, with subscripts used to designate conditioning where necessary. Thus, the parametric pmf Pr (X|θ) leads to the corresponding expectation Eθ(X). (Likewise for variance, written Var θ(X).) When discussing sequences of realizations of a random variable X, parenthetical superscript notation is used to designate particular draws—for example, (x(1), … , x(n)).

On occasion, some specialized vector notation will also be employed. For vector x, xi refers to all elements of xother than the ith. Thus, Pr (X=x | Xi=xi) refers to the probability that random vector X is equal to x, conditional on the non-ith values of X being equal to xi. In addition to the above, we will also at times need to refer to a vector for which specific values have been replaced (all others remaining unchanged). To this end, the expression i,jx is used to denote a vector whose ith element has been fixed to j and whose other elements are equal to x.


  1. Top of page
  2. Abstract

Our focus here is on what we shall call generalized location systems, which represent the allocation of arbitrary entities (e.g., persons, objects, organizations) to “locations” (e.g., physical regions, jobs, social roles). While our intent is to maintain a high level of generality, we will limit ourselves to systems for which both entities and locations are countable and discrete and for which it is meaningful to treat the properties of entities and locations as relatively stable (at least for purposes of analysis). Relaxation of these constraints is possible but will not be pursued here; as will be shown, the present framework still allows for a great deal of flexibility.

We begin our development by assuming a system that consists of n identifiable objects, O= (o1, … , on), each of which may reside in exactly one of m identifiable locations, L= (l1, … , lm). The current state of this system is given by a configuration vector, ∈{1, … , m}n, which is defined such that i=j iff oi resides at location lj. The set of all such configuration vectors that are realizable is said to be the set of accessible configurations and is denoted inline image. One very important parametrization of inline image with which we will deal is in terms of occupancy constraints. We define the occupancy function of a location system as

  • image(1)

where I is the standard indicator function. The vectors of maximum and minimum occupancies for a given location system are composed of the minimum/maximum values of the occupancy function for each state under inline image (respectively). That is, we require that PiP(i, ) ≤P+i for all i∈ 1, …, m, ℓ∈inline image, where P, P+ are the minimum and maximum occupancy vectors. If Pi=P+i= 1 ∀i∈ 1, … , m, then it follows that is a permutation vector on 1, … , n, in which case we must have m=n for non-empty inline image. This is an important special case, particularly in organizational contexts (White 1970). By contrast, it is frequently the case in geographical contexts (e.g., settlement) that Pi= 0 and P+i > ni∈ 1, … , m, in which case occupancy is effectively unconstrained.

In addition to configurations and labels, objects and locations typically possess other properties of scientific interest. We refer to these as features, with FO being the set of object features and FL being the set of location features. While we do not (initially) place constraints on the feature sets, it is worth highlighting two feature types that are of special interest. Feature vectors provide ways of assigning numerical values to individual objects or locations—for example, age, average rent level, or wage rate. Adjacency matrices can also serve as important features, encoding dyadic relationships among objects or locations. Examples of such relationships can include travel distance, marital ties, or demographic similarity. Because relational features allow for coupling of objects or locations, they play a central role in the modeling of complex social processes (as we shall see).

To draw the above together, we define a generalized location system by the tuple inline image. The state of the system is given by inline image, which will be of primary modeling interest. Various specifications of inline image are possible, but particular emphasis is placed on occupancy constraints, which specify the range of populations that each location can support. With these elements, it is possible to model a wide range of social systems and it is to this problem that we now turn.


  1. Top of page
  2. Abstract

Although many approaches to location system modeling are possible, we will here focus on models for the observation of configuration vectors at arbitrary times. In this sense, our focus is on the stochastic equilibrium behavior of the system: if we take a snapshot of a system at any given instant, what is the probability of observing one configuration rather than another? While this perspective can be expanded upon, it nevertheless allows us to say a fair amount regarding system behavior. In modeling state probabilities, it is also essential that the models be constructed in such a way as to allow for inference from extant data; while this may seem to be a self-evident constraint, even a brief perusal of the sociological literature reveals that this condition is often unsatisfied. Finally, it must be the case that the location model be capable of capturing the sorts of complex dependencies that are known to operate within large-scale social systems. These include homogeneity effects, density dependence, homophily/propinquity, and capacity constraints, in addition to more prosaic attraction/repulsion mechanisms.

In this section, we provide a modeling framework that satisfies these constraints. The core of this framework is a discrete exponential family of distributions that is closely linked to related models employed in spatial statistics, statistical mechanics, and social network analysis. Although our focus will be on the modeling of location systems per se, we frequently draw upon results from these fields. Our treatment of the topic begins with the development of the general location system model and proceeds to a specific family of submodels that incorporates a range of substantively important effects in a reasonably simple fashion. Some simple inferential properties of the location system are also discussed as well as methods for simulating draws from the location system model.

3.1. The General Model

We here define a stochastic model for the equilibrium state of a generalized location system. In particular, we assume that the system is ergodic and—given a set of accessible configurations, inline image—the system will be found to occupy any particular configuration, , with some specified probability.1 Our primary interest is in the modeling of these equilibrium probabilities, although some dynamic extensions are possible.

Given the above, we first define the set indicator function

  • image(2)

The equilibrium probability of observing a given configuration can then be written as

  • image(3)

,where S is the random state and inline image is a quantity called the social potential (defined below). The sum

  • image(4)

is the normalizing factor for the location model and corresponds to the partition function of statistical mechanics (Kittel and Kroemer 1980). Clearly, equation (3) defines a discrete exponential family with support on inline image and it is generic in the sense that any distribution on inline image can be written in the form of equation (3). There are several benefits to working within such a framework. First, as noted, the framework is complete with respect to the underlying location system. Second, much is known about models with the form of equation (3). We have already seen that a parallel formalism (the ERG) is widely used to construct models of networks with complex dependence and the formalism is similarly common in both physics (Kittel and Kroemer 1980) and mathematical statistics (e.g., see Besag 1974, 1975; Barndorff-Nielsen 1978; Stoyan, Kendall, and Mecke 1987). However, the most important property of the exponential family framework is perhaps the third: given an appropriate parametrization of inline image, there are existing results that permit principled inference from empirical data (Johansen 1979; Brown 1986). While its deductive value is also important, the availability of viable inferential tools is a major motivation for our approach; models that can capture complex social processes are of little use if they cannot be evaluated on readily available data.

While equation (3) can represent any distribution on inline image, its scientific utility clearly lies in the specification of inline image. Intuitively, the social potential for any given configuration is equal to its log-probability, up to an additive constant. Thus, the location system is more likely to be found in areas of high potential, and/or (in a dynamic context) to spend more time in such states. While any number of forms for inline image could be proposed, we begin with a constrained family that incorporates a number of features of known substantive importance for a variety of social systems. This form is introduced in the section that follows.

3.2. A Family of Social Potentials

As noted above, we seek a family of functions inline image such that inline image. This family should incorporate as wide a range of substantively meaningful effects as possible; since it is not reasonable to expect effects to be identical in every situation, the family should be parametrized so as to allow differential weighting of effects. Ideally, the social potential family should also be easily computed and its structure easily interpreted. For our present purposes, these latter qualities will guide our construction of inline image. It should be noted, however, that alternative approaches are possible (such as the direct use of dependence graphs [Besag 1974; Pattison and Robins 2002; Wasserman and Robins 2005]) and the potentials so generated may be deployed in the same manner as those discussed here.

An obvious initial solution to this problem is to construct inline image from a linear combination of sufficient statistics (i.e., deterministic functions of , FL and FO). Employing such a potential function within equation (3) leads to a regular exponential family on inline image (Johansen 1979), which has a number of useful statistical implications. Of course, such an approach also has the usual virtues of linear families (additivity of effects, nesting, etc.) familiar to most social scientists. The canonical ERG parametrization shown in Section 1.2 is of this form as well, although some recent parameterizations employ nonlinear constraints on the manner in which sufficient statistics are weighted (Hunter and Handcock 2006; Snijders et al. 2006).

Even if a linear form is supposed, however, we are left with a more important question: what statistics should be included in the social potential? Obviously, these statistics must be parametrized as functions of the location and object features. Since the impact of the social potential is invariant up to an arbitrary additive constant, per equation (3), it follows that properties that are invariant over inline image—i.e., those that depend only on aspects of the populations of objects or locations without regard to the assignment of objects to locations—can be safely ignored. All relevant statistics for the construction of inline image must thus involve some interaction between object features and the features of locations to which they are assigned; “main effects” of object or location features, in the usual sense of the term, are not meaningful in this context. With respect to the features themselves, these may include both attributes (features of the individual location or object per se) and relations (features of object or location sets). Here, we will limit ourselves to relations that are dyadic (i.e., defined on pairs) and single-mode (i.e., that do not mix objects and locations). Thus, our effects should be functions of feature vectors, and/or (possibly valued) graphs.

While this may seem to leave innumerable possibilities, we can further focus our attention by noting that the purpose of inline image is ultimately to control the assignment of objects to locations. This suggests immediately that the effects of greatest substantive importance will be those that draw objects toward or away from particular locations. Table 1 provides one categorization of such effects by feature type. In the first (upper-left) cell, we find effects that express direct attraction or repulsion between particular objects and locations, based on their attributes. In the second (upper-right) cell are effects that express a tendency for objects linked through connected locations to be particularly similar or distinct. (Spatial autocorrelation is a classic example of such an effect.) The converse family of effects is found in the third (lower-left) cell; these effects represent a tendency for objects to be connected to other objects with similar (or different) locations. A tendency for husbands and wives to make similar career choices—where careers are interpreted as “locations”—serves as an example of a location homogeneity effect. Note that the essential difference between cells two and three lies in whether the clustering/dispersion of similar objects is being assessed (cell 2) versus the assignment of connected (but not necessarily similar) objects to similar/dissimilar locations (cell 3). This will be discussed in greater detail below. Finally, in the fourth (lower-right) cell we have effects based on the tendency of location relations to align (or disalign) with object relations. Propinquity, for example, is a tendency for adjacent objects to reside in nearby locations.

Table 1.  Elements of the Social Potential
 Location AttributesLocation Relations
Object AttributesAttraction/repulsion effectsObject homogeneity/heterogeneity effects (through locations)
Object RelationsLocation homogeneity/heterogeneity effects (through objects)Alignment effects

Taken together, these four categories of effects combine to form the social potential. Under the assumption of linear decomposability, we thus posit four subpotentials (one for each category) such that

  • image(5)

,where inline image is the potential associated with attraction/repulsion effects, inline image is the potential associated with object heterogeneity effects, inline image is the potential associated with location heterogeneity effects, and inline image is the potential associated with alignment effects. We now consider each of these functions in turn.

3.2.1. Attraction/Repulsion Potential

The first class of effects that must be represented in any practical location system are global attraction/repulsion—also called “push/pull”—effects. Residential locations, potential firm sites, occupations, and the like have features that make them generally likely to attract or repel certain objects (be they persons, organizations, or other entities). Such effects are naturally modeled via product-moments of attributes. Let inline image be exogenous features reflecting location and object attributes (respectively) and let inline image be a parameter vector. Then we may define inline image as

  • image(6)
  • image(7)

where tα is a vector of sufficient statistics.

The behavior of equation (7) is quite intuitive. For instance, let Qi be a location feature and let Xi= (1, …, 1) be a constant object feature. Then αi > 0 and αi < 0 produce attraction and repulsion effects (respectively) based on Qi. If the effect in question is stronger or weaker for particular objects, this may in turn be produced by allowing Xi to vary.

One substantively important case of such an effect is discrimination. Discrimination may be understood as a conditional tendency for individuals with certain features to be placed in (or denied access to) certain positions. In terms of social potential, this is simply a push/pull effect where Qi describes the location feature with respect to which discrimination is occurring and Xi encodes the individual feature or group membership that is the basis of discrimination. Such an approach is operationally similar to the treatment used in conventional regression analyses of wage discrimination (e.g., Huffman and Cohen 2004), although there is an important difference in interpretation. While a wage discrimination effect represents a marginal increase/decrease in wages for persons with certain features,2 a discrimination effect within the location model represents a conditional tendency for persons with certain features to be differentially assigned to particular positions (or positions with particular features). The difference between the two may be appreciated by contemplating a hypothetical change in which Xi becomes identical for all actors. This would lead a conventional wage discrimination effect model to predict a mean shift in population wages, while such a shift need not occur under the location model. This last is because a location system model of wage attainment effectively models the process of competition among workers for a fixed set of potential jobs (not all of which have to be filled), rather than taking wage to be a property that arises from the intrinsic properties of the actors themselves (irrespective of the jobs available). The ability to capture such institutional constraints is an attractive property of the location approach. Of course, discrimination effects need not be confined to wages—any tendency for differential assignment may be included in the same manner.

3.2.2. Object Homogeneity/Heterogeneity Potential

A second class of effects concerns object homogeneity/heterogeneity—that is, the conditional tendency for associated locations to be occupied by objects with similar (or different) features. Let inline image be a matrix of object attributes, inline image be an adjacency array on the location set, and inline image be a parameter vector. We then define the object homogeneity/heterogeneity potential by

  • image(8)
  • image(9)

where, as before, tβ is a vector of sufficient statistics. It should be noted that the form of tβ is closely related to Geary's C, a widely used index of spatial autocorrelation (Cliff and Ord 1973). tβ is based on absolute rather than squared differences and it is not normalized in the same manner as C, but its behavior is qualitatively similar in many respects.

As a simple illustration of inline image, let L be a set of disjoint spatial regions with contiguity matrix Bi··· Let O represent a population of households and let Y·i be a vector representing an object feature (e.g., a categorical code for racial self-identification of the primary household informant). Then βi < 0 corresponds to a tendency for households with similar features (here, race) to be contiguously located, while βi > 0 favors a heterogeneous assignment. Put another way, negative β values induce homogeneity or segregation, while positive β values induce heterogeneity or supra-random mixing. This situation can be complicated further by allowing B to take on arbitrary values: the magnitude of Bijk controls the strength of connection between the j, k locations on the ith feature, while the sign of Bijk determines whether βi > 0 induces heterogeneity (Bijk > 0) or homogeneity (Bijk < 0). Thus, it is possible to model both effects within the same relation. Similarly, a diagonal Bi·· matrix can be used to model homogeneity/heterogeneity within locations in the absence of cross-location ties. Such a structure may be employed, for instance, when attempting to model occupational segregation; in this case, L represents the set of occupations and setting Bi·· equal to the identity matrix allows βi to directly parametrize the extent of “segregation pressure” within the system.

3.2.3. Location Homogeneity/Heterogeneity Potential

The parallel case to inline image is inline image, which models the effect of location homogeneity or heterogeneity through objects. Let inline image be a matrix of location features, inline image be an adjacency array on the object set and inline image be a parameter vector. We then define inline image as follows:

  • image(10)
  • image(11)

As implied by the above, tγ is the vector of sufficient statistics for location homogeneity. tγ is at core similar to tβ, save in that the role of object and location are reversed: absolute differences are now taken with respect to location features and are evaluated with respect to the connections between the objects occupying said locations.

While inline image may seem less intuitive than inline image, its utility is easily demonstrated via a simple example. Consider, for instance, the case of wage rates within married couples. To set up the problem, we begin by letting Ai·· be a matrix representing all marital ties among members of the sample; this will consist of a set of isolated symmetric dyads, accompanied by isolates if the sample includes unmarried persons. L is taken in this case to be a collection of jobs, each of which is associated with a wage rate (contained in R·i). For inline image then places more weight on job allocations that increase the within-couple wage rate differences (ceteris paribus), and γi < 0 produces the opposite effect (i.e., within-couple wage homogeneity). Processes leading to within-couple wage heterogeneity have been postulated by Becker (1991), among others; by turns, several processes identified by social capital theorists (Granovetter 1973; Calvo-Armengol and Jackson 2004) would be expected to lead to within-couple homogeneity in wage rates. Such effects can be modeled directly through inline image, above and beyond other allocative mechanisms.

3.2.4. Alignment Potential

The final element of the social potential is the alignment potential, inline image, which expresses tendencies toward alignment or disalignment of object and location relations. Given object and location adjacency arrays inline image and inline image (respectively) and parameter vector inline image, the alignment potential is given by

  • image(12)
  • image(13)

where, as in the prior cases, tδ represents the vector of sufficient statistics. The form chosen for tδ is Hubert's gamma, which is the standard matrix cross-product moment (see Hubert [1987] for a range of applications).

Although the alignment potential has been utilized in prior work on graph comparison (see Butts, this volume, page xxx), our application is more concerned with modeling the direct impact of relations on location assignment. As the name implies, the alignment potential captures the extent to which relations among objects are mirrored by relations among their associated locations. Consider, for instance, a collection of disjoint spatial regions with travel distance matrix Di··, and a population of actors whose kinship network is represented by the adjacency matrix Wi··· Where δi < 0, the kinship network is propinquitous; that is, actors tend to reside (ceteris paribus) in locations that are physically proximate. By contrast, δi > 0 would indicate a dispersal effect, in which actors who are tied to one another tend to occupy more distant locations. (Such an effect might be expected, for instance, among firms that are tied to one another via production of similar products.)

Another important alignment effect is density dependence—the tendency for objects to cluster (positive dependence) or disperse (negative dependence) with respect to locations. To model density dependence, we create an object relation Wi·· representing a complete graph and employ the identity matrix for Di··· Under this construction, tδi indexes the extent to which objects are clustered in a small number of locations; δi > 0 increases this tendency, while δi < 0 inhibits it. Replacing the identity matrix with an inverse distance matrix allows for a more general form of spatial dependence, but the general intuition is similar.

As a final point, it may be noted that the alignment effect is “generic” for the potential family employed here in the sense that the attraction/repulsion, object heterogeneity, and location heterogeneity statistics can be written as alignment statistics on suitably transformed input matrices. We may thus usefully characterize the present social potential family as that composed of all potential linear combinations of matrix cross product-moments between object and location features. While this also means that (given appropriate data transformations) inline image can be written entirely in terms of inline image, we continue to separate the subpotentials throughout the paper. One reason for this is substantive: as Table 1 shows, each subpotential arises from a conceptually distinct combination of object and location features and is most easily understood in this fashion. Another reason for the separation is computational; specifically, there are computational shortcuts that are available for other effects, which cannot be realized in the generic inline image case. While we continue to draw such distinctions, then, it should be borne in mind that they are not essential in character.

3.2.5. Combined Linear Potential

We are now ready to form the combined linear social potential. Substituting the quantities of equations (7) through (13) into equation (5) gives us

  • image(14)

in terms of sufficient statistics, or

  • image(15)

in terms of the underlying covariates. Together with equation (3), equation (15) specifies a regular exponential family of models for the generalized location system. As we have seen, this family allows for the independent specification of attraction/repulsion, heterogeneity/homogeneity, and alignment effects (including differential attractiveness, segregation, homophily/propinquity, and density dependence as special cases). We now proceed to a consideration of some of the properties of this model family, before turning to the problem of simulation.

3.2.6. Interpreting the Social Potential

Per equation (3), the social potential is a real-valued function inline image on inline image such that the probability of observing any given configuration inline image is proportional to inline image. As we have seen, inline image can be constructed so as to incorporate a variety of effects, ranging from simple attraction/repulsion to the enhancement or suppression of heterogeneity. These effects are parametrized via one or more sufficient statistics, which are weighted by coefficients. Beyond the narrower discussion of Section 3.2, it is useful to consider some of the ways in which the social potential per se (and any associated parameters) may be interpreted. In the discussion that follows, we economize notation slightly by employing the general parameter vector θ= (α, β, γ, δ) in place of the four separate parameter vectors: attraction, object heterogeneity, location heterogeneity, and alignment parameter vectors. We will likewise concatenate the four sufficient statistic vectors as t= (tα, tβ, tγ, tδ), leading to the more compact social potential expression inline image. Given this, we now consider the interpretation of inline image in terms of total system behavior and of hypothetical microlevel dynamics.

Holistic Interpretation Since inline image can be properly regarded as providing an expression for the joint behavior of the location system as a whole. In particular, the system is more likely to be observed in configurations of high potential than in configurations of low potential, which leads to an immediate interpretation for a given parameter, θi. Where θi > 0, configurations for which the associated statistic ti is large are given higher potential (ceteris paribus); thus positive values of θi indicate a general tendency of the location system to exhibit configurations with higher ti values than would be otherwise observed. Similarly, negative values of θi imply that configurations with large values of ti are suppressed. This is easily seen by considering the probability ratio for two hypothetical configurations inline image under potential inline image:

  • image(16)
  • image(17)
  • image(18)

Thus, every unit change in ti multiplies the odds of observing ℓ′ versus ℓ by a factor of emath image. This observation provides a direct quantitative interpretation of potential function parameters, although it should be borne in mind that many statistics may not be fully separable in practice. This phenomenon is well-known in the context of ERG models, for which the intrinsic relationships between statistics can be particularly strong (e.g., see Handcock 2003b); it would be a mistake to view this as a unique property of models for dependent systems, however, since it will arise in any probability model whose sufficient statistics are potentially related to one another (e.g., OLS regression with correlated predictors). When dealing with heavily correlated statistics, it may be more useful to interpret the effects of parameters in batches (much as we might jointly interpret the parameters of a polynomial regression). Alternately, it may in some cases be helpful to reparameterize t so as to produce statistics that are closer to being orthogonal over inline image. Regardless of any intra-t relationships, however, the relationship between relative state probability and changes in t described in equation (18) remains a valid means of interpreting the effect of θ as a whole.

In keeping with the above, it should be emphasized that equation (3) may be interpreted as a model for the distribution of a single cross-sectional observation of a location system, even where no dynamic equilibrium interpretation is appropriate. In this case, inline image simply parametrizes the dependence among system elements within the observed cross-section, irrespective of the putative generating mechanism. Similar “fall-back” interpretations exist for other models in this general class (for example, ERGs) and are not unique to the location system framework.

Microdynamic Interpretation In building the location system model, it will be noticed that little has been said about the underlying microprocesses that give rise to the configuration vector, or about the detailed evolution of ℓ under equilibrium conditions (other than distributional properties). This omission is deliberate: the location system model may be viewed as process agnostic, in that there are many conceivable microprocesses that would give rise to the same equilibrium distribution. Nevertheless, there are some useful statements that can be made about dynamic aspects of location models and we review several of these here.

First, we posit a family of microprocesses whose long-run dynamics give rise to the equilibrium distribution of equation (3). Consider a process in which, at finite (but otherwise arbitrary) time intervals, a random object X is drawn with some fixed distribution such that all objects are selected with positive probability. Let ℓ be the pre-draw system state and let X,i for i∈ 1, …, m be equal to ℓ save for the assignment of object X to state i. The system then transitions from ℓ to X,i with probability inline image (with ℓ to ℓ being an acceptable “transition”), the realization of which becomes the base state for the next transition event. So long as it is possible to transition from any given state ℓ to state ℓ′ in a finite series of moves, the states of the above process form an irreducible Markov chain on inline image. Furthermore, the transition probabilities at each step can be recognized as the conditional distribution of X, given θ and the other elements of ℓ. (This follows directly from equation [3], where the set of accessible states is restricted to those that involve changing only X.) A Markov chain of this type is commonly known as a Gibbs sampler (Gilks, Richardson, and Spiegelhalter 1996a), and its equilibrium distribution is the joint distribution of ℓ; here, this is simply the distribution of equation (3), demonstrating that a process in which each object moves randomly in proportion to the relative exponentiated potential of its possible locations will generate a global distribution of states that is compatible with the location system model. If the time intervals between transition events are independent of the system state, it also immediately follows that the expected fraction of time spent in each accessible state is similarly proportional to exp inline image (where such expectations exist).

One type of microlevel process with this behavior stems from a class of potential games (as alluded to in Section 1.1). Let us consider an n player “assignment game” (with the object set O= (o1, …, on) corresponding to the players), in which each player has m potential strategies (corresponding to the choice among the elements of L). Since each player's strategy corresponds to a choice of location, we can represent an assignment of strategies to all players by ℓ. We assume that the preferences of a given player oi are represented by the utility ui (ℓi, ℓi), where the first argument is trivially the strategy of oi and the second denotes the strategies of all other players. Now define inline image to be the set of all accessible i (irrespective of i). Given this, we say that the assignment game belongs to the class of potential games if there exists a function inline image such that

  • image(19)

for all inline image. (This follows immediately from the definition of Young [1998: 36].) Intuitively, the above implies that an assignment game is a potential game if and only if we can posit some function (the potential, ρ) such that the change in utility for a unilateral strategy shift is equal to the change in potential. While assuming identical utilities for all actors will clearly result in such a game, this is not a necessary condition: for instance, utilities that are identical up to an additive constant will also satisfy the above definition.

For our purposes, the above is most relevant when we take inline image. In this case, the potential for the assignment game is equal to our social potential and changes in utility for unilateral moves are equal to the corresponding changes in inline image. Let us further posit that the assignment game is played by boundedly rational actors who select strategies according to a stochastic choice model. Specifically, given an opportunity to move, we posit that actor oi selects his or her location as a random variable Ki with probability mass

  • image(20)

(where i,j denotes the assignment formed by adding i=j to i). This can be rewritten as

  • image(21)

and substitution of utility differences for potential differences then yields

  • image(22)

This is immediately recognizable as the transition mechanism for the Gibbs sampler (as shown above); hence, the state distribution arising from strategy choice in the assignment game will be proportional to inline image, so long as the irreducibility condition of the Gibbs sampler is met. This may be satisfied in a number of ways, including (1) sequential moves by each actor in turn and (2) sequential moves by randomly chosen actors, where every actor is chosen from a fixed distribution with positive probability (Gilks 1996). Heuristically speaking, the long-run dynamics of such a system are generally insensitive to the details of the movement opportunity process, so long as actions are selected in accordance with equation (20).

As these results imply, the generalized location system model is compatible with certain microprocess interpretations, including potential games like those considered by Zhang (2004) and Young (1998). In the latter circumstance, specific model parameters can be interpreted as reflecting the partial utilities associated with changes in their corresponding statistics. Although it should be emphasized that these are not the only processes that can lead to distributions of the type explored here, they nevertheless serve as useful examples of how such micro/macro connections (strongly endorsed by Coleman [1990]) may be made in practice.

3.3. Simulation

For purposes of both prediction and inference, it is necessary to simulate the behavior of the location system model for arbitrary covariates and parameter values. While it is not generally possible to take draws from the location system model directly, approximate samples may be readily obtained by means of a Metropolis algorithm.3 Given that numerous accessible references on the Metropolis algorithm are currently available (e.g., see Gamerman 1997; Gilks et al. 1996b; Gelman et al. 1995), we will focus here on issues that are specific to the model at hand. Fortunately, the location system model is not especially difficult to simulate, although certain measures are necessary to ensure scalability for large systems.

To review, a Metropolis algorithm proceeds in the following general manner (see Gilks et al. [1996b] for further details). Let S be the (random) system state. We begin with some initial state inline image and propose moving to a candidate state (1), which is generally chosen so as to be in a neighborhood of (0). (Some additional constraints—for example detailed balance—apply to the candidate distribution, but these do not affect the results given here.) The candidate state is then “accepted” with probability inline image. If accepted, the candidate becomes our new base state and we repeat the process for (2). If rejected, (1) is replaced by a copy of (0) and again the process is repeated. This process constitutes a Markov chain whose equilibrium distribution (under certain fairly broad conditions) converges to the target distribution (here, inline image). It is noteworthy that this process requires only that the target distribution be computable up to a constant factor; this feature makes Metropolis algorithms (and related MCMC techniques) very attractive to those working with exponential family models (e.g., Strauss 1986; Snijders 2002; Butts, this volume, page x).

To implement the Metropolis algorithm, then, our core concern is computation of the probability ratio between states (a problem encountered earlier in Section 3.2.6). Given a current state, (i), the probability of accepting a candidate state, (i+1), is then

  • image(23)
  • image(24)
  • image(25)

Thus, the log-probability of a state change is simply the difference in social potentials between the two assignments. In the case of the linear potential, substituting the potential function from equation (14) further gives us

  • image(26)

or, substituting from equation (15),

  • image(27)

Although equation (27) can be used to compute the potential difference directly, equation (26) demonstrates that the same quantity can be expressed in terms of a fixed linear combination of differences in sufficient statistics. For purposes of simulation, then, we need only track such differences. (This process is an exact analog to the “changescore” methods used in ERG simulation tools such as Handcock et al. [2003].) As this implies, we can speed computation by choosing our proposals so as to facilitate difference calculations; an obvious choice in this regard is a proposal mechanism that reassigns a randomly chosen object to a randomly selected location. In addition to simplicity of implementation, this proposal density admits considerable improvement in computational efficiency over the iterated calculation of equation (27). In particular, let ℓ be the current state and ℓ′ the proposal formed by assigning object j to location k. Then the respective differences in sufficient statistics are as follows:

  • image(28)
  • image(29)
  • image(30)
  • image(31)

Calculation of equation (26) for a single reassignment using equations (28) through (31) is an inline image operation. This is a substantial improvement over the inline image complexity for direct application of equation (27) in the arbitrary case, particularly for large-n systems. As a side note, it should be mentioned that occupancy constraints may not allow single assignments to take place. (The permutation case is a trivial example, since the smallest change possible is the dyadic exchange.) In this case, the proposal mechanism may need to include multiple reassignments in a single step; however, it is still the case that the above computations can be performed for each such reassignment and the resulting complexity is still linear in n so long as the number of reassignments per step is bounded by a constant. Even fairly complex schemes can thus be reduced to an iterated application of the reassignment calculation.

3.3.1. Estimating the Partition Function

Though the above provides the essential elements needed to simulate draws from the location system model, the approach used bypasses calculation of the partition function. This is deliberate: Z is not directly computable in polynomial time and the unevenness of the Boltzmann factor (inline image) renders simple Monte Carlo strategies hopelessly inefficient. What is to be done, however, when the partition function (or its derivatives) is needed for a specific application (such as the deviance calculations discussed in Section 3.4)? In this case, we employ the fact that we are able to simulate draws from the location system model to produce an importance sample, thereby allowing efficient Monte Carlo quadrature of Z.

To begin, we assume that a sample of M draws (denoted (1), …, ℓ(M)) have been taken from the location system model with combined parameter vector θ= (α, β, γ, δ) and vector of sufficient statistic functions t= (tα, tβ, tγ, tδ). Our interest is in estimating Z(θ′, C), where θ′ is a combined parameter vector that is close to θ (in the sense that |θ′−θ| is small). Our estimator of the partition function is based on the result that

  • image(32)

This result may be shown as follows. First, we note that, from the standard Monte Carlo theorem in the discrete case (Kalos and Whitlock 1986),

  • image(33)

where convergence is almost sure and in mean square, so long as the function inline image has a finite second moment. Setting f(ℓ) = exp (θ′Tt(ℓ)) then gives us

  • image(34)
  • image(35)

While this gives us an expression for inline image on the right hand side, it requires us to know the value of inline image and thus is of little immediate use. Since the partition function does not depend on ℓ, however, it may be pulled out of the initial summand:

  • image(36)


  • image(37)

Dividing through by inline image then gives us

  • image(38)

To see the value of this, let us return to the left-hand side of equation (32). Immediately, we note that

  • image(39)

As M[RIGHTWARDS ARROW], we have already seen that inline image. Therefore, the above becomes (for large M)

  • image(40)
  • image(41)

Now, what of the factor on the right? Returning to equation (33), we simply take f(ℓ′) = 1, which yields

  • image(42)
  • image(43)

and thus, by substitution,

  • image(44)
  • image(45)

It therefore follows that we can estimate the partition function directly, given a sample from the location model. Since the approximation works well for any θ′ that is close to θ, only a single sample is usually needed to compute numerical derivatives.4 In the special case for which we are solely interested in inline image (given an importance sample based on θ), equation (32) further simplifies to

  • image(46)

This last follows immediately by substitution.

3.3.2. Expected Value Estimates

An important use for the partition function approximation of equation (32) is the calculation of approximate expected values for the sufficient statistics of the location system model. While these can obviously be estimated directly via simulation (per Section 3.3), it is in some cases more efficient to calculate expected statistics by means of the partition function. In doing so, we exploit the standard result (e.g., Brown [1986]) that, for a regular exponential family with parameter vector θ and sufficient statistics t, the expected value of t under θ is equal to the first derivative of the logarithm of the partition function with respect to θ. Allowing t= (tα, tβ, tγ, tδ) and θ= (α, β, γ, δ), this translates to the relation

  • image(47)

for the location system model. Since direct computation of Z is infeasible, this would seem to be of little use; however, we have already seen that Z may be approximated from importance sample draws using equation (32). Here, too, it may at first blush seem that using a sample to approximate a samplable quantity provides no particular advantage. However, consider the case in which one must evaluate Eθt(ℓ) at several different points, all of which are reasonably close to one another. In such situations, equation (32) may be employed to quickly calculate multiple approximations from a single sample. This is particularly useful when seeking to calculate derivatives of the expectations themselves, as for the moment matching method described in Section 3.4.2 below.

Let us begin by assuming that we have drawn an importance sample (1), …, ℓ(M) from a location system model with support inline image and concatenated parameter vector θ (e.g., using a Metropolis algorithm). To calculate the expectation of some statistic ti under parameter vector θ′, it is natural to replace the derivative of equation (47) by a standard finite difference approximation (Press et al. 1992). For some ε<< 1 (ε= 0.00001 being a value that gives good precision in most cases),

  • image(48)
  • image(49)

Clearly, this is not directly computable. However, we may substitute the importance sampling estimate of equation (32) to obtain

  • image(50)

which simplifies to

  • image(51)

Repeating this calculation for each i gives the entire vector of expectations (if desired). Each such calculation requires inline image time, whereas drawing the initial sample requires inline image time (or worse, for example, if subsampling—“thinning”—of a larger set of draws is performed prior to use); this can be a substantial savings, in practice. Such gains do not occur without cost, of course. The tradeoff here lies in the accuracy of the approximation, which is in turn governed primarily by the accuracy with which inline image is estimated. As noted earlier, accuracy of the partition function estimator is generally high when θ is close to θ′, degrading as the distance between θ and θ′ grows large. For this reason, it is usually wise to employ equation (51) only when evaluating expectations in the immediate vicinity of the sampled parameters. Comparison of directly estimated expectations with those estimated by this method can be used to evaluate the degree of closeness required for a particular model, where this is a significant concern.

3.4. Inference

As noted elsewhere in this paper, an important benefit of the use of discrete exponential families in the modeling of location systems is the ability to leverage existing inferential theory (for example, Barndorff-Nielsen 1978; Brown 1986). Given a location system specified by the tuple inline image with parametric social potential inline image, the joint likelihood of observed state ℓ is given by equation (3). Since this quantity is well-defined, principled inference using Bayesian or maximum likelihood methods would seem to be a straightforward affair; the computational expense of directly calculating the required normalizing factor, inline image makes this task less trivial than it might be, however. Fortunately, a number of methods exist for circumventing this problem. Here, we will consider two: estimation based on pseudo-likelihoods and maximum likelihood estimation via first moment matching. These two approaches build on each other (with the former providing initial estimates to be refined by the latter) and will hence be considered in the above order. Other alternatives are also available (the Monte Carlo approach of Geyer and Thompson [1992] being an obvious possibility; see Butts [this volume, page xxx] for a closely related application) and in general it should be possible to apply any method for inference on regular discrete exponential families to the location system model.

3.4.1. Maximum Pseudo-Likelihood Estimation

While we have seen that it is possible to estimate the normalizing factor for the likelihood of equation (3) using importance sampling, this process is too computationally expensive to permit direct maximization of the likelihood surface (though see Geyer and Thompson [1992] for an effective ratio-based approach). One alternative approach (originating with Besag [1975] but better known to sociologists from Strauss and Ikeda [1990]) involves approximating the joint likelihood of the data by a product of conditional likelihoods; the parameter vector that maximizes this pseudo-likelihood is then used as an estimator of the unknown true parameters. Such a vector is known as a maximum pseudo-likelihood estimator, or MPLE. Put more formally, let Y be a vector of k random variables, with parameter vector φ. Then the pseudo-likelihood of realization y of Y is given by

  • image(52)
  • image(53)

Note that the pseudo-likelihood is equal to the true likelihood when the elements of Y are independent given φ. (This follows immediately from the fact that p(Yi=yi|yi, φ) =p(Yi=yi|φ) and p(Y=y |φ) =∏ki=1p(Yiyi| φ) under conditional independence.) Where this assumption does not hold, inline image will depart from the true likelihood to some extent. Nevertheless, the maximum of inline image with respect to φ is often close to the corresponding maximum on the likelihood surface, making it a potentially viable estimator when inline image is easily calculable. Alternatively, the MPLE can be used as an initial approximation to the true maximum likelihood estimator (MLE), to be subsequently refined using other methods. This last strategy has proved practical in the estimation of ERG models (as implemented, for example, by Handcock et al. [2003]) and is that suggested here.

To define a pseudo-likelihood function for the location system model, it is first necessary to select a partition of the joint distribution into individual elements. The conditional likelihoods of these elements are then multiplied to produce the corresponding pseudo-likelihood. In the present case, the most natural decomposition of the random system state S is in terms of the individual object assignments. The appropriate conditional probability in this case is then

  • image(54)

leading to the corresponding pseudo-likelihood function

  • image(55)

Note that the evaluation of inline image requires inline image potential difference computations. Since these can be performed in inline image operations (per Section 3.3), the total computational complexity of inline image is inline image. This is moderately expensive, but obviously far better than inline image! To form the MPLE, we simply find the parameter vectors that maximize inline image— that is,

  • image(56)

This can be done using standard heuristic optimization methods, such as Newton's method, simulated annealing, or the like (Press et al. 1992; Acton 1990).

Although the above partition of S is perhaps most natural, it cannot be applied for certain choices of inline image. In particular, if there exists at least one location for which P+i=Pi > 0, then not all objects can be unilaterally assigned. In this case, a reasonable choice for the decomposition of S is in terms of the set of dyadic exchanges on ℓ (i.e., location swaps). In this case, we may treat the ordering of each dyad as our variable of interest, leading to the conditional distribution

  • image(57)

To form the associated pseudo-likelihood, we then take the product of the conditional distributions over all dyads in O:

  • image(58)

(Note that this is directly analogous to the permutation model pseudo-likelihood of Butts [this volume, page x].) The complexity of this calculation is inline image, which will be better than the complexity of the single-state pseudo-likelihood when m > n. On the other hand, inline image conditions on the occupancy structure of ℓ and as such may represent less information than inline image. To find the MPLE under dyadic exchange, we simply substitute inline image for inline image in equation (56) and solve in the same manner as the single move case. This allows us to compute MPLEs for most choices of inline image, including the important special case of 1:1 matchings (i.e., permutations).

Although the MPLE is not guaranteed to have good frequentist properties in the general case, it is sometimes used directly (e.g., see Wasserman and Pattison [1996]; Pattison and Wasserman [1999]; Robins et al. [1999]; Contractor et al. [2006]). Where this is to be done, it is strongly recommended that draws from the estimated model be simulated (using the method of Section 3.3) and that the first moments of the simulated sufficient statistics be compared to the same statistics on the observed data. To the extent that substantial disparities are observed (e.g., with respect to a t or similar statistic), the MPLE should be employed cautiously (if at all). On the other hand, a close match between the mean simulated statistics and the observed statistics indicates that the MPLE is in fact functionally close to the MLE; we shall see the basis for this conclusion in the next section.

3.4.2. Maximum Likelihood Estimation

Maximum likelihood estimation for statistical exponential families has been extensively studied (see Johansen [1979] or Brown [1986] for reviews) and is attractive on both frequentist and approximate Bayesian grounds (the MLE appearing as a noninformative limit of Bayes estimators in many settings [Robert 1994]). Under the linear social potential proposed in Section 3.2, the maximum likelihood estimator for the parameters of inline image given observed state obs is

  • image(59)

where this maximum exists. Under the proposed social potential, equation (3) defines a regular exponential family and it is thus a standard result that the MLE exists (and is unique) if and only if the elements of t are finite and affinely independent and if t(ℓ) belongs to the relative interior of the convex hull of t over inline image (Barndorff-Nielsen 1978). In practice, nonexistence of the MLE arises where one or more statistics (or a linear combination thereof) are maximally extreme (e.g., all women occupying the highest paid position and all men occupying the lowest paid position in an occupational model with a gender/wage effect). In this case, the MLE effectively diverges (in our example, the apparent gender effect is unbounded) and no finite estimate exists. Fortunately, such extreme arrangements are unlikely to occur in large systems and even then practical approximations (e.g., truncating the diverging parameters at values of very large, but finite, magnitude) will usually permit reasonable estimates of the remaining parameters (see Handcock [2003a] for a discussion).

Due to the expense of approximating Z, direct maximization of the likelihood is generally infeasible. Since (3) is a regular exponential family, however, it is a standard result that

  • image(60)

where (tαobs, tβobs, tγobs, tδobs) is the vector of observed sufficient statistics, provided that the MLE exists. This result motivates a method of moments technique, in which heuristic search is used to equate the (simulated) expected sufficient statistics to their observed values; the parameter vector that gives rise to these values is the MLE. Although sometimes slow, this approach can be quite efficacious and has been successfully employed by Snijders (2002) in the context of exponential random graph families. In practice, convergence can be accelerated by initiating the search procedure with an initial approximation to inline image—the MPLE, inline image, is suggested for this purpose. Simulation costs in estimating inline image can also be reduced by making use of the fact that inline image as was discussed in Section 3.3.2. Specifically, an initial sample from the current point estimate (using the method of Section 3.3) can be used as an importance sample for the numerical estimation of the derivatives of inline image (e.g., by applying the method of finite differences to the logged partition function estimator of Section 3.3.1). This same sample can be reused multiple times to obtain the derivatives of inline image in the vicinity of the original estimate, which can then be used to project a new estimate that leads to a better approximation of (tαobs, tβobs, tγobs, tδobs) (e.g., via Newton's method). A new sample is then drawn from the refined estimate and the process is repeated until the desired degree of convergence is obtained.

Given the resulting MLE inline image, the associated deviance can also be estimated by means of the importance sampling method of Section 3.3.1. Specifically, let (1), …, ℓ(M) be draws from inline image (e.g., taken via the Metropolis algorithm of Section 3.3). Then substitution from the partition function estimator of equation (46) into the likelihood of equation (3) yields

  • image(61)
  • image(62)

. The estimated deviance may then be used to compare models using standard selection criteria, such as the AIC or BIC (see Bozdogan [2000] and Wasserman [2000] for comparative reviews).

With respect to estimates of uncertainty, it should also be noted that standard asymptotics hold for location model MLEs in the case of independent observations from the same social system. (This is a consequence of the fact that the location model forms a regular exponential family; e.g., see Johansen [1979].) Whether similar asymptotic results can be obtained in the limit of increasing system size is not known. This problem is essentially equivalent to the problem of asymptotics for exponential random graph models, which is also unsolved at this time. Where asymptotic results cannot be relied upon, however, Monte Carlo procedures can be employed to obtain standard errors and p-values for classical tests. (See Hunter and Handcock ]2006[ for a parallel case involving ERG models.) Thus, the standard tools of likelihood-based inference avail themselves here. Bayesian treatment of the location model is another possibility, although posterior simulation is greatly complicated by the difficulty of computing the likelihood function. Approximation methods based on curvature of the posterior near the mode (Gelman et al. 1995) would seem to provide an obvious starting point.


  1. Top of page
  2. Abstract

One of the positive features of the location system model is the great number of substantive problems for which it may be employed. Here, we illustrate some of the behaviors of the model by means of two simple applications, one involving economic inequality and the other involving residential segregation. While both are simplified for purposes of exposition, it should be emphasized that slightly elaborated versions can be fit to data from survey or archival sources using the tools of Section 3.4. Simulation studies such as these can thus form the basis for subsequent empirical investigation, without the necessity of adding cumbersome operationalization assumptions. As a third example, we also demonstrate the inferential use of the location system model to study position occupancy in organizations, using a data set of Lazega (2001).

4.1. Job Segregation, Discrimination, and Inequality

Our first example demonstrates the use of the location system in modeling occupational stratification. In the interest of clarity, we restrict our analysis to a simplified “microeconomy” of 100 workers (objects) matched with 100 distinct jobs (locations) on a 1:1 basis. Workers are evenly divided by gender and are randomly allocated to (heterosexual) couples such that all members of the set have exactly one partner. Finally, workers are ranked on a unidimensional “human capital” score, with ranks assigned randomly in alternating fashion by gender. (Thus, rank distributions are effectively identical by gender and random with respect to partners.) Jobs are ranked by “wage” and are organized into ten contiguous occupational categories. Thus, the ten highest-paying jobs are in one category, followed by jobs ranked 11 through 20, etc. While this setting is heavily stylized, it nevertheless allows us to capture basic interactions between occupational segregation, household effects, and factors such as discrimination. Such a model could be elaborated to include hierarchical job categories, distinct unemployed states, additional job/worker attributes, and relaxations of assumptions such as 1:1 matching, as appropriate to the data in hand.

In order to represent the above within the location system framework, we begin by translating selected features into the elements of the social potential of Section 3.1. We here seek to model two attraction/repulsion effects: the tendency of workers of particular genders to be assigned to higher/lower wage jobs (“discrimination”); and the tendency of workers with higher levels of human capital to be assigned to higher/lower wage jobs (“merit”). For this purpose, we set X such that Xi1 and Xi2 are respectively the gender (coded dichotomously with male = 1) and the human capital rank (coded from 1 to 100, with 100 representing the highest value) of worker i. The corresponding position feature matrix, Q, is defined such that Qi1 and Qi2 are both equal to the wage rank of job i (from 1 to 100 in ascending order); thus, tα1(ℓ) is the sufficient statistic for the gender/wage interaction, while tα2(ℓ) captures the corresponding human capital effect. For occupational segregation, we posit a single statistic tβ arising from a single-column object feature matrix Y consisting of dichotomous gender codes (as with X·1) interacting with a dichotomous location-location array B such that B1ij= 1 if job i belongs to the same occupational category as job j. The corresponding location heterogeneity statistic that will be of interest here is couple-level wage heterogeneity (described further below), a property that is parametrized via a statistic tγ formed from a dichotomous object relation array A such that A1ij= 1 if worker i and worker j are members of the same couple and single-column location attribute matrix R containing the wage rank of each job (as with the columns of Q). We do not posit any alignment effects for this particular, model and hence this completes our specification of t.

To get a sense of the behavior of the location model, we begin by demonstrating some of the job assignments that can arise under various parameter values. Figure 1 depicts simulated draws from the location model under a variety of conditions. Initially, we shall limit our consideration to configurations arising from manipulation of the discrimination (α1) and (anti-)segregation (β) parameters, with all other parameters held to 0. For each choice of parameter values, the corresponding panel of Figure 1 shows 250 Metropolis draws from the associated model; these were uniformly thinned (i.e., subsampled) from a total of 25 million draws in each case, following a (discarded) burn-in sample of size 200,000. Jobs (ordered by wage rank) are shown on the vertical axis, with occupant gender indicated by color (light corresponds to male). For ease of reference, job category boundaries are shown by horizontal dashed lines; thus the gender composition of each category may be determined by examining the fraction of light versus dark cells between the appropriate lines for a given vertical slice.


Figure 1. Location model draws, job assignment model.

Download figure to PowerPoint

Parameter values for each panel of Figure 1 are interpreted as follows. α1 effects for all panels parametrize the strength of association between gender and wage, with positive values reflecting a stronger tendency to sort males into high-ranking wage positions. Thus α1 acts as a discrimination parameter. (Negative α1 values reverse the sorting direction but produce otherwise identical results; only positive values are considered here.) The object homogeneity/heterogeneity parameter, β, reflects the tendency of jobs within the same occupational category to be occupied by persons of the same gender (i.e., occupational segregation). As per equation (9), β < 0 indicates a tendency toward segregation (homogeneity), while β > 0 indicates a tendency toward desegregation (heterogeneity). Zero values for any parameter imply an absence of the corresponding sorting effect. Thus 1, β) = (0, 0) results in a null model of uniform assignment.

Examination of the panels of Figure 1 hints at the diversity of configurations that can result from even a small number of interacting regimes. Panel 1 displays the baseline condition of random assignment: men and women appear across the spectrum of wage ranks, in essentially even numbers. The remaining panels show various types of sorting by gender, reflecting an interaction of segregation and discrimination effects. Panel 2, for instance, depicts a “block random” pattern, in which men and women are concentrated into uniform blocks which are otherwise randomly allocated across the wage ordering. Panels 3 and 4, on the other hand, show a clear pattern of stratification, in which men tend to be sorted into higher wage positions. What accounts for these patterns? Clearly, sorting by gender is driven by α1 and is only observed for samples in which α > 0. Heterogeneity within occupational categories is controlled by β, however and hence can act in ways which are distinct of discrimination per se. In Panel 2, for instance, we have pressure toward homogeneity/segregation (β < 0), which tends to force each occupational category to collapse into a single preferred gender. Since the model is here indifferent to which gender occupies any given occupational category, however, there is no net tendency for men or women to be sorted into higher-wage positions. By contrast, Panel 4 illustrates the interaction of a strong wage discrimination effect with a powerful tendency toward heterogeneity/antisegregation (β > 0). While the former effect seeks to sort men into high-wage positions and women into low-wage positions (as in Panel 3), the latter resists the accompanying necessity of producing gender-homogeneous occupational categories. The result is a structural “compromise,” in which an overall gender/wage gradient is somewhat attenuated by the inclusion of both genders within each occupational stratum. Interestingly, the impact of the discrimination parameter on variation within occupational categories creates a tendency to reproduce a miniature version of the male/female wage gradient inside each occupational category; thus the well-known tendency for macrolevel stratification patterns to reproduce themselves at multiple levels is shown here to be an emergent property of the interaction of a global sorting process (here, discrimination) with mechanisms favoring local heterogeneity (here, an antisegregation effect).

In addition to the general types of configurations found under different assignment regimes, it is useful to consider the quantitative impact of model parameters on outcomes of interest. In the present case, consider the difference in mean wage rank by gender. By construction, discrimination effects must lead to an exaggeration of such differences, but these effects must interact with other processes as well. For instance, in a competitive labor market, we generally expect workers with greater human capital to obtain positions with higher wage rates. Depending on the relationship between human capital and gender, this interaction may strengthen or weaken inequality in wage attainment. An example of this well-known phenomenon is shown in Figure 2, which presents differences in mean wage rank (by gender) for the location model with α effects for discrimination (gender by wage) and merit (human capital by wage).5 As the figure clearly shows, the impact of discrimination is attenuated by merit effects where human capital is uncorrelated with gender. In addition to weakening the local impact of mild discrimination (i.e., 1| small), this attenuation softens the transition from a mixed-wage environment to a “frozen” environment in which wages are strictly stratified by gender. In the absence of competing factors, even a fairly small amount of discrimination is adequate to lock the system into a stratified state; an intervention with the intent of inhibiting stratification by reducing discrimination is thus unlikely to prove effective unless discrimination can be tightly controlled. To draw on a physical analogy, it may be noted that the effect of 1/α is directly analogous to a temperature parameter (as pointed out in other exponential family contexts by Strauss [1986] and Robins et al. [2005]). The stratification system “solidifies” at fairly low temperatures (|α| large) and is thus difficult to force into other, less stratified, states by modest changes in discrimination alone. By contrast, an intervention that attempts to inhibit stratification by introducing selective factors uncorrelated with gender could prove effective even with relatively high levels of residual discrimination (by lowering the relevant “melting point”). While it is perhaps intuitive that the introduction of competing selective factors would attenuate discrimination, the quantitative impact of these effects is particularly clear under the location system model.


Figure 2. Gender difference in mean wage rank, discrimination, and merit effects.

Download figure to PowerPoint

If human capital effects are inhibitory of stratification in our scenario, what of segregation? Removing the human capital effect and adding in a β effect for gender segregation by occupational category yields the wage rank difference relationship of Figure 3.6 The impact of segregation on stratification is both clear and striking: segregation strongly exacerbates discrimination, while desegregation mildly inhibits it. The mechanism involved is a combination of those observed in panels 2 and 3 of Figure 1—namely, the tendency of gender-typed job categories to be “sorted” by wage in the presence of a background discrimination effect. The net effect is a substantially higher degree of stratification than would be obtained by discrimination alone. Desegregation pressure, by contrast, reduces the extent to which high or low wage categories can become male or female dominated, thereby “flattening” the wage distribution (as was seen in panel 4 of Figure 1). Interventions such as affirmative action programs can be understood as acting through mechanisms of this type; interestingly, Figure 2 would seem to suggest that most of the impact of such interventions is likely to come through the elimination of active segregation pressure rather than through pressure for desegregation per se.


Figure 3. Gender difference in mean wage rank, discrimination, and segregation effects.

Download figure to PowerPoint

While occupational segregation is a factor of obvious importance for stratification outcomes, a less well-studied issue is the impact of within-couple inequality on macroscopic wage differences. Effects as diverse as social influence (Freidkin 1998), homophily on unobserved characteristics (McPherson et al. 2001), and diffusion of opportunity through social ties (Calvo-Armengol and Jackson 2004) can potentially lead to a net tendency for similarity of within-couple wage rates; on the other hand, mechanisms such as market/home production specialization (Becker 1991), normative pressures for intensive parenting (Jones and Brayfield 1997), and the like can generate pressure for heterogeneous wage rates. To explore this within the location system model, we repeat the simulations of Figure 3, replacing the β effect with a γ effect for couple-level wage homogeneity/heterogeneity. The results are shown in Figure 4. As might be expected, heterogeneity pressure exacerbates discrimination. The effect, however, is slight and the marginal impact declines rapidly in γ. By contrast, the impact of couple-level homogeneity is profound: even a small amount of within-couple homogeneity pressure virtually eliminates the impact of discrimination, even when the latter is exceedingly strong. By tying together the fortunes of individual men and women, couples can act to counteract large-scale selection pressures toward wage inequality.


Figure 4. Gender difference in mean wage rank, discrimination, and couple-level homogeneity effects.

Download figure to PowerPoint

While these simulation results merely scratch the surface of what is possible when using the location system to model occupational inequality, they nevertheless suggest some interesting and nonobvious effects. Of particular import is the relative power of couple-level homogeneity effects in suppressing labor market discrimination, a finding that suggests a stronger connection between processes such as mate selection and marital bargaining with macrolevel stratification than might be supposed. The exacerbation of discrimination effects by segregation is also noteworthy, along with the somewhat less powerful inhibiting effect of active desegregation. These phenomena highlight the importance of considering dependencies—both among individuals and among jobs—when modeling inequality in labor market settings. Such effects can be readily captured by the location system, facilitating a more complete theoretical and empirical treatment of stratification within the occupational system.

4.2. Settlement Patterns and Residential Segregation

Another domain of long-standing interest to sociologists, geographers, and economists has been the role of segregation within residential settlement patterns (Schelling 1969; Bourne 1981; Massey and Denton 1993; Zhang 2004). Here, we illustrate the use of the location model on a simplified settlement system involving 1000 households (objects) allocated to regions on a uniform 20 by 20 spatial grid (locations). Unlike the job allocation system described above, this system places no occupancy constraints on each cell; however, “soft” constraints may be implemented via density dependence effects. For purposes of demonstration, each household is assigned a random “income” (drawn independently from a log-normal distribution with parameters 10 and 1.5) and an “ethnicity” (drawn from two types, with 500 households belonging to each type). Households are tied to one another via social ties, here modeled simply as a Bernoulli graph with mean degree of 1.5. The Bernoulli graph is a random structure in which ties between actors are independent and it is commonly used as a simple null model of network structure (Anderson, Butts, and Carley 1999). Regions, for their part, relate to one another via their spatial location. Here, we will make use of both Euclidean distances between regions and Queen's contiguity (for purposes of segregation). To obtain interregional distances, we treat each location as a 1 unit by 1 unit square planar region and take Euclidean distances between centroids; household position is modeled only up to the cell level, in analogy with data observed at the level of areal units such as census tracts. Similarly, two regions are considered to be contiguous under the Queen's rule if they border one another at either a point or an edge. Finally, each region is also assigned a location on a “rent” gradient, which is proportional to the inverse square of centroid distance the region in question to the center of the grid.

With these building blocks, a number of mechanisms can be explored. An obvious attraction/repulsion effect that plays an important role in household settlement patterns is the tendency for low-income households to “avoid” (or be excluded from) high-rent areas. In our case, this is equivalent to an attraction effect between household income and location rent level and hence we model it via a statistic tα based on a single vector object feature matrix X containing household incomes and a single vector location feature matrix Q containing location rent levels. Ethnic segregation—a tendency for households to avoid settling in areas dominated by households of different ethnicity—is a form of object heterogeneity and it is captured here by a single statistic tβ formed from the interaction of a single-column matrix of dichotomous ethnicity codes (Y) and a dichotomous B array such that B1ij= 1 if region i is contiguous with region j (using the Queen's contiguity rule, as noted above). Although we do not posit any location heterogeneity effects in the present case, we consider two alignment statistics. The first is a crowding, or density dependence effect. To define the statistic associated with this effect, we let the object-object relational feature matrix W1·· be a matrix containing only 1s and the location-location relational feature matrix D1·· be the identity matrix (i.e., a matrix with 0s on off-diagonal cells and 1s on the diagonal). The sufficient statistic tδ1 formed by these matrices is then equal to the sum of squared population sizes for each location and serves to parametrize pressures toward (or, for δ1 < 0, against) crowding. The second alignment statistic, negative propinquity, expresses a tendency for households that are socially tied to one another to reside in spatially distant locations. (We use the term “negative” propinquity here to emphasize that the natural alignment statistic technically measures the inverse of propinquity in its usual sense; δ2 < 0 thus generates pressure toward propinquity per se.) To form the negative propinquity statistic, we simply let W2ij= 1 if household i is tied to household j (and 0 otherwise), with D2ij being equal to the Euclidean distance between the ith and jth regions. This gives us our second tδ statistic and completes our specification of t for this model.

Several examples of configurations resulting from the above mechanisms are shown in Figure 5. Each panel shows the 400 regions comprising the location set, with household positions indicated by circles. (Within-cell positions are jittered for clarity.) Household ethnicity is indicated by color and network ties are shown via edges. While each configuration corresponds to a single draw from the location model, a burn-in sample of 100,000 draws was taken (and discarded) prior to sampling. Configurations shown here are typical of model behavior for these covariates and parameter values.


Figure 5. Location model draws, spatial model.

Download figure to PowerPoint

The panels of Figure 5 nicely illustrate a number of model features. In Panel 1, a model has been fit with an attraction parameter based on an interaction between rent level and household income (α= 0.0001), balanced by a negative density dependence parameter δ1=−0.01. Although the former effect tends to pull all households toward the center of the grid, the density avoidance effect tends to prevent “clumping.” As a result, high-income households are preferentially clustered in high-rent areas, with lower-income households displaced to outlying areas. Note that without segregation or propinquity effects, neither ethnic nor social clustering are present; this would not be the case if ties were formed homophilously, and/or if ethnicity was correlated with income. Clustering can also be induced directly, of course, as is shown in panel 2. Here, we have added an object homogeneity effect for ethnicity through Queen's contiguity of regions (β=−0.5), which tends to allocate households to regions so as to reduce local heterogeneity. As can be seen, this induces strong ethnic clustering within the location system; while high-income households are still preferentially attracted to high-rent areas, this sorting is not strong enough to overcome segregation effects. Another interesting feature of the resulting configurations is the nearly empty “buffer” territory that lies between ethnic clusters. These buffer regions arise as a side effect of the contiguity rule, which tends to discourage direct contact between clusters. As this suggests, the neighborhood over which segregation effects operate can have a substantial impact on the nature of the clustering that results. This would seem to indicate an important direction for empirical research.

A rather different sort of clustering is generated by adding a propinquity effect to the original attraction and density model. As noted earlier, we implement propinquity as an alignment effect between the interhousehold network and the Euclidean distance between household locations 2=−1). As one might anticipate, the primary effect of propinquity (shown in Panel 3 of Figure 5) is to pull members of the giant component together. Since many of these members also happen to be strongly attracted to high-rent regions, the net effect is greater population density in the area immediately surrounding the urban core. Another interesting effect, however, involves households on the periphery: since propinquity draws socially connected households into the core, peripheral households are disproportionately those with few ties and/or which belong to smaller components. The model thus predicts an association between social isolation and geographical isolation. Ironically, this situation is somewhat attenuated by the reintroduction of a residential segregation effect (Panel 4). While there is still a tendency for social isolates to be forced into the geographical periphery, the consolidation of ethnic clusters limits this somewhat. Because ties are uncorrelated with ethnicity, propinquity also acts to break the settlement pattern into somewhat smaller, “band-like” clusters with interethnic ties spanning the intercluster buffer zones. (One would not expect to observe this effect in most empirical settings, however, due to the strong ethnic homophily of most social ties (McPherson et al. 2001).)

Quantitative information on the interaction between propinquity and segregation can be obtained by simulating draws from the location system with systematically varied parameters. Figure 6, for instance, shows the average value of the heterogeneity statistic for ethnicity difference by Queen's contiguity (tβ) as a function of (anti-)segregation (β) and (negative) propinquity (δ2) effects. (Each data point represents a mean of 500 Metropolis draws uniformly thinned from a total sample of size 100,000, with a burn-in of 20,000 draws. All other parameters have been set to 0.) In the absence of a propinquity effect, the location system undergoes a sharp behavioral transition at β= 0; extreme homogeneity is observed below this threshold, with high levels of heterogeneity immediately above it. While propinquity seems to have little effect on heterogeneity in the segregated regime, its effect in the desegregated regime is uniformly inhibitory. Whether positive or negative, propinquity effects tend to weaken heterogeneity (with stronger effects being observed in the dispropinquitous case). This phenomenon stems from the fact that propinquity affects which actors can be clustered together, with dispropinquitous systems tending to force large numbers of actors to remain in different locations. Such constraints make it more difficult to maximize local heterogeneity, which is most easily accomplished by the formation of dense, ethnically diverse clusters. Although the positive β, δ2 regime may seem unlikely to arise in most residential contexts, it may still appear in related contexts such as firm siting, in which firms benefit from a heterogeneous market environment but simultaneously seek to avoid being placed too close to competitors. Interactions such as those of Figure 6 may thus have implications for the appearance and survival of locally competitive markets in a spatial context.


Figure 6. Mean segregation statistic, by segregation and propinquity effects.

Download figure to PowerPoint

In addition to clustering, segregation has implications for population density. This is clearly illustrated by Figure 7, which shows the mean concentration statistic (tδ1) formed by the alignment of the identity matrix on locations with the complete graph on objects. (Simulations for this figure are as for Figure 6, with concentration replacing propinquity.) As the figure shows, population distribution for the location system tends toward one of two regimes: a highly concentrated regime (in which most households are packed into a very small number of regions) and a diffuse regime (in which households tend to be widely dispersed). Transitions between these regimes are moderately abrupt, with some additional consolidation occurring within the high-concentration regime for increasing δ1. While we might at first imagine that segregation would enhance population concentration, this is not the case. Rather, segregation inhibits population concentration, with desegregation actively promoting it. Intuitively, this is due to the fact that the corresponding heterogeneity statistic can be most effectively increased by placing a diverse population within a small area. By contrast, concentrated, segregated population distributions are relatively difficult to produce (since any heterogeneous “incursions” are amplified by the local population level). As this implies, inhibition/promotion does not manifest here through an alteration of the extremity of the two regimes; instead, segregation effects shift the concentration “temperature” at which the transition occurs. Such a result is suggestive of the behavior of binary mixtures (particularly eutectic mixtures), which can solidify at temperatures that differ greatly from those of their constituents (Kittel and Kroemer 1980).


Figure 7. Mean concentration statistic, by density and segregation effects.

Download figure to PowerPoint

As Schelling (1969) long ago noted, even mild tendencies toward local segregation can result in residential segregation at larger scales. While the location system model certainly bears this out, the model also suggests that factors such as population density and interhousehold ties can interact with segregation in nontrivial ways. Using the location system framework, such interactions are easy to examine and the strength of the relevant parameters can be readily estimated from census or other data sources. It is also a simple matter to introduce objects of other types (e.g., firms) that relate to households and to each other in distinct ways (as represented through additional covariates). In an era in which geographical data is increasingly available, such capabilities create the opportunity for numerous lines of research.

4.3. Empirical Example: Lazega's Lawyers

As a final sample application, we here demonstrate the use of location system models in an empirical context. The data for this example comes from Lazega's (2001) study of a midsized U.S. corporate law firm. The relevant population of the firm consists of 71 lawyers, for whom gender, tenure within the firm, age, specialty (litigation versus corporate), and law school background (here Ivy-league versus non-Ivy) are measured. Lazega also reports on relationships among the lawyers in question, of which we will here use attributions of friendship. Positions within the firm vary on two salient dimensions: seniority level (associate versus partner) and office (three work sites being present). While Lazega considers this case in great depth, the present analysis will be limited to a simple analysis of the factors associated with position occupancy within the firm. Like the other examples provided here, its purpose is more illustrative than substantive.

In studying actors' positions within the firm we will condition on the composition of both the actor and location populations; ℓ is thus taken to be a 1:1 mapping and inline image to be the set of permutations on o1, …, o71. We initially hypothesize that the attraction/repulsion of actors to positions is governed by the interaction of the five individual attributes with position seniority. X is thus composed of these five attribute vectors (in columns), with dichotomous coding for gender (male = 1), specialty (corporate = 1), and law school (Ivy = 1). Since Q involves only one variable, it is composed of five identical columns (with Qij being the seniority of li, partner coded as 1). In considering potential object heterogeneity effects, one obvious possibility is a tendency toward gender segregation by work site; to test for this, we posit an object heterogeneity term with object attribute matrix Y consisting of a column matrix of gender codes and location adjacency matrix B coded dichotomously such that B1ij= 1 if positions li and lj are at the same work site. In addition to object heterogeneity effects, we also note the possibility of location heterogeneity within the firm. One obvious effect of this type is the potential for friendships to be stratified by seniority—that is, for friends to occupy the same strata within the firm. Such an effect can be captured by a location heterogeneity term whose object relation matrix, A, is the adjacency matrix of the friendship network and whose location attribute matrix R is a single vector of position seniority codes. Finally, we conjecture (per Lazega) that there will be a net tendency for friendship to align with work site co-membership. This is modeled via an alignment statistic whose W matrix is the adjacency matrix for the friendship network and whose D matrix is coded as B above.

To fit the location system model to the Lazega data, we employ the moment-matching method of Section 3.4. Initial parameter estimates were obtained using the permutation MPLE and were iteratively refined until the Euclidean distance between the mean simulated statistics and the vector of observed statistics was less than 0.1. Means for convergence testing were taken from a sample of 3000 Metropolis draws, uniformly thinned from a total sample of size 6,000,000, and 20,000 burn-in draws were taken (and discarded) before taking sample draws from each chain. A sample from the (converged) MLE was also used to estimate the deviance for each tested model (per Section 3.3.1). The deviance estimate was then used to compute the corresponding AIC scores, with model degrees of freedom equal to the number of included statistics (and hence parameters).

To assess the general properties of the assignment system, we begin by fitting each of our four effect types to the data as separate blocks. Parameter estimates and goodness-of-fit information for each such model are shown in Table 2, together with equivalent statistics for the null model of random assignment and a model containing all effects simultaneously. Under random assignment, each assignment vector ℓ has an equal chance of being observed and hence the data likelihood under this model is fixed at (n!)−1. Adding parameters improves fit, at the cost of parsimony—thus, models should be compared via the AIC (which penalizes the deviance by the number of parameters) for purposes of model selection. To provide some additional sense of the differences in fit, p-values are also provided for a standard χ2 test of improvement in deviance versus the null model. Given that the usual asymptotic argument for the χ2 as a reference distribution when comparing nested regular exponential family models (e.g., Johansen 1979) is not immediate in this case, these p-values should be interpreted heuristically; they nevertheless provide some baseline against which to assess quantitative differences in fit.

Table 2.  Estimated Models for the Lazega Data
EffectNull Modelα Onlyβ Onlyγ Onlyδ OnlyAll Effects
Attraction/Repulsion Effects
 Gender × Seniority 2.39162.35491.36860.9780
 Tenure × Seniority 0.79130.75300.34930.5277
 Age × Seniority 0.13920.1264−0.08540.0033
 Specialty × Seniority 0.18350.2763−0.3647−1.2983
 School × Seniority 0.38520.3564−0.0453−0.0117
Object Heterogeneity Effects
 Gender × Same Office 0.04020.0284 −0.04090.0227
Location Heterogeneity Effects
 Friendship × Seniority −0.6575−0.3158 −0.6633−0.4207
Alignment Effects
 Friendship × Same Office 1.06740.74261.01200.8702
Deviance 469.403 454.245 468.272 438.151 437.271 393.823
Model degrees of freedom 0 5 1 1 1 8
AIC 469.403 464.245 470.272 440.151 439.271 409.823
Null deviance–model 0 15.158 1.131 31.252 32.132 75.580
χ2p-value (vs. Null)   0.0097 0.2876 <0.0001 <0.0001 <0.0001

An examination of Table 2 reveals that all of the effect blocks other than object heterogeneity appear to have nonnegligible marginal effects, in the sense that each model is preferred to the null model under the AIC. (These differences in fit are also large compared to a χ2 baseline, as indicated by the associated p-values.) The strongest effect per parameter clearly occurs for the location heterogeneity effect, followed by the alignment and attraction/repulsion effects (respectively). By contrast, object heterogeneity appears to have little impact on the assignment process, resulting in no significant improvement in fit (as assessed either by AIC or the χ2 reference distribution). This immediately suggests that the dominant relationships within this position system are the tendency for lawyers to be tied to others of similar seniority and for friendships to coincide with shared work sites. Weaker (but still prominent) influences involve general tendencies for lawyers with particular attributes to be assigned to more (or less) senior positions; on the other hand, it does not seem to be the case that assignment within work sites is gender-homophilous.

For a closer look at the properties of the location system, we turn to the parameter estimates of Table 2. Point estimates for each fitted model are shown in two columns, containing the maximum pseudo-likelihood and maximum likelihood estimates for each included parameter (respectively). As expected, the MPLEs are generally in the neighborhood of the MLEs for these models (with sign changes observed only for small-magnitude coefficients). On the other hand, it is immediately evident from inspection of the estimates that the MPLEs are more extreme than the MLEs, being of greater absolute magnitude in 13 of 16 cases. This extremity bias results from the fact that the MPLE considers only the local information involved in pairwise exchanges and hence underweights the impact of parameters on assignment arising from higher order dependence. The phenomenon is a fairly general one, although the effect is exacerbated for systems with stronger dependence among positions.

With regard to the estimates themselves, many of the effects are immediately intuitive. Men and those with long tenure within the firm are more likely to be found in senior positions (even controlling for relational effects). Friendship is homophilous by seniority and propinquitous by work site, as demonstrated by the consistently negative location heterogeneity and positive alignment parameters (respectively). A more subtle set of influences are clearly at work with the effects of age, specialty, and school, whose parameter estimates reverse once the effects of friendship are controlled for. Marginally, workers who are older, who specialize in corporate law, and who attended an Ivy-league school are more likely to be found in senior positions. When controlling for interpersonal relationships, however, we find that the impact of age and school diminish by more than an order of magnitude, contributing little to the relative probability of assignment. Specialty, on the other hand, both reverses sign and increases dramatically: controlling for relational effects, specialists in corporate law are actually less likely to be found in senior positions than would be expected under a “neutral” assignment process. This phenomenon appears to result from the strong tendency of friendship ties to be homophilous by both legal specialty and seniority (QAP p≤ 0.0001 for both relationships). We here condition on the former relationship (both specialty and friendship being object features) and controlling for the second (via tδ) leads to a model in which seniority status, ceteris paribus, should tend to be heavily clustered by specialty. Given this relationship (and the coincidence of specialization with other attributes), the observed rate at which corporate law specialists occupy senior positions is below that which would be otherwise expected. Whether this is causal is not trivial to determine from cross-sectional data, but the location system model allows us to identify such relationships for possible future examination. As Mayhew and Levinger (1976) famously note, marginal relationships can be very misleading in the presence of strong baseline effects resulting from structural biases; the location system allows us to detect and control for many such interactions.

As a final comment, it should be noted that the magnitudes of many of the effects estimated here are fairly substantial. Given a pair of lawyers of which one is male and the other female, of whom one occupies a senior position, the odds of finding the male (rather than the female) in the senior position are approximately four times what they would be if both individuals were of the same gender. That this effect is retained even when controlling for firm composition, basic covariates, and relational structure clearly suggests the presence of gender discrimination in hiring or promotion, although a more extensive analysis would be desirable to rule out alternative explanations. Likewise, each year of tenure difference within a pair increases the relative odds of the longer-tenured individual “beating out” a more recent hire for a senior position by approximately 70 percent. Since the average tenure difference within this firm is 10.48 years, this effect can be profound — at the mean tenure difference, the odds of an “upset” given otherwise identical contenders for a senior position are reduced by a factor of over 250! Similarly, our strong estimates for γ and δ suggest that the odds of observing a potential move that increases the net number of mismatches of friendships by work site or seniority versus some baseline configuration are reduced by a factor of 40 percent to 50 percent (respectively) per net discordant edge. Of course, all things are rarely equal and the exact probability of observing any given configuration will depend on the interaction of multiple factors. Nevertheless, it is immediately apparent that the occupancy of positions within this firm is highly structured by both demographic and relational forces. Parameter estimates such as those shown here allow us to determine the strength and direction of those forces by exploiting structural signatures in the observed data. These estimates, in turn, can serve as the basis for subsequent simulation (e.g., to perform “what-if” analyses of firm structure under hypothetical perturbations), theory building (e.g., by mapping the estimated social potential onto a process such a potential game), or theory testing (e.g., by comparing the estimated parameters with those predicted by competing theories).


  1. Top of page
  2. Abstract

We have shown a general framework (the generalized location system) that can be used to characterize a range of social systems. An exponential family of distributions was developed for modeling such systems, allowing for the incorporation of both attributional and relational covariates. This family belongs to a class of distributions that are well-known in the statistical literature (the regular exponential families) and it also has strong parallels with models developed for physical systems. Drawing on these established results, methods were shown for simulation and inference using the location system model. Three illustrative applications (occupational stratification, residential settlement patterns, and position occupancy within a law firm) were presented and simulation or inference was employed to show the potential utility of the location system model in each case. While there are a number of issues that have been treated here briefly or not at all—including dynamics, compatibility with low-level mechanisms, and endogenization of covariates—the material presented is sufficient to permit deployment of location system models in a wide range of empirical contexts. It is hoped that by cross-applying tools from other domains, the location system will allow for a more thorough and general treatment of complex problems than could be obtained using domain-specific methods.

  • 1

    If this assumption does not hold, the model may still be interpreted in terms of the probability of a single observation arising from a static process with the distribution of equation (3). For purposes of exposition, however, ergodicity (and the equilibrium interpretation) will be assumed.

  • 2

    Note that this “increase” may be interpreted as a difference in means rather than a causal difference. The latter is commonly employed, however.

  • 3

    Obviously, the Gibbs samplers discussed in Section 3.2.6 can also be used in this capacity; we focus here on the Metropolis algorithm due to its superior scalability.

  • 4

    More properly, the “closeness” of θ′ to θ can be assessed in terms of the variance of exp ((θ′−θ)Tt(ℓ)). To the extent that this function is flat, the associated integral will be well-approximated (Kalos and Whitlock 1986).

  • 5

    For these simulations, β=γ= 0, means are based on 750 Metropolis draws uniformly thinned from samples of 750,000, after a burn-in period of 100,000 draws.

  • 6

    All simulation runs for Figures 3 and 4 were conducted in the same manner as those for Figure 2.


  1. Top of page
  2. Abstract
  • Acton, Forman S. 1990. Numerical Methods That Work. 3d ed. Washington , DC : Mathematical Association of America.
  • Anderson, Brigham S., Carter T. Butts, and Kathleen M. Carley. 1999. “The Interaction of Size and Density with Graph-Level Indices. Social Networks 21:23967.
  • Anselin, Luc. 1988. Spatial Econometrics: Methods and Models. Norwell , MA : Kluwer.
  • Barndorff-Nielsen, Ole. 1978. Information and Exponential Families in Statistical Theory. New York : Wiley.
  • Bearman, Peter S., James Moody, and Katherine Stovel. 2004. “Chains of Affection: The Structure of Adolescent Romantic and Sexual Networks. American Journal of Sociology 110:4491.
  • Beck, Elwood M., Patrick M. Horan, and Charles M. Tolbert. 1978. “Stratification in a Dual Economy: A Sectoral Model of Earnings Determination. American Sociological Review 43:70420.
  • Becker, Gary S. 1991. A Treatise on the Family. Cambridge , MA : Harvard University Press.
  • Benenson, Itzhak. 2004. “Agent-Based Modeling: From Individual Residential Choice to Urban Residential Dynamics.” Pp. 6795 in Spatially Integrated Social Science, edited by Michael F.Goodchild and Donald G.Janelle. Oxford : Oxford University Press.
  • Besag, Julian. 1974. “Spatial Interaction and the Statistical Analysis of Lattice Systems. Journal of the Royal Statistical Society, Series B 36:192236.
  • Besag, Julian. 1975. “Statistical Analysis of Non-Lattice Data. The Statistician 24:17995.
  • Bourne, Larry S. 1981. The Geography of Housing. New York : V. H. Winston.
  • Bovy, Piet H. L., and Eliahu Stern. 1990. Route Choice: Wayfinding in Transport Networks. Dordrecht , Netherlands : Kluwer Academic Publishers.
  • Bozdogan, Hamparsum. 2000. “Akaike's Information Criterion and Recent Developments in Information Complexity. Journal of Mathematical Psychology 44:6191.
  • Brown, Lawrence D. 1986. Fundamentals of Statistical Exponential Families, with Applications in Statistical Decision Theory. Hayward , CA : Institute of Mathematical Statistics .
  • Budig, Michele J., and Paula England. 2001. “The Wage Penalty for Motherhood. American Sociological Review 66:20425.
  • Calvo-Armengol, Antoni, and Matthew O. Jackson. 2004. “The Effects of Social Networks on Employment and Inequality. American Economic Review 94:42654.
  • Carley, Kathleen M. 1991. “A Theory of Group Stability. American Sociological Review 56:33154.
  • Cliff, Andrew D., and J. Keith Ord. 1973. Spatial Autocorrelation. London : Pion.
  • Coleman, James S. 1990. Foundations of Social Theory. Cambridge , MA : Harvard University Press.
  • Contractor, Nosh, Stanley Wasserman, and Katherine Faust. 2006. “Testing Multi-theoretical Multilevel Hypotheses About Organizational Networks: An Analytic Framework and Empirical Example. Academy of Management Review 31:681703.
  • Corstjens, Marcel L., and David A. Gautschi. 1983. “Formal Choice Models in Marketing. Marketing Science 2:1956.
  • Crouch, Bradley, Stanley Wasserman, and F. Trachtenburg. 1998. “Markov Chain Monte Carlo Maximum Likelihood Estimation for p* Social Network Models. Presented at the Eighteenth International Sunbelt Social Network Conference, Sitges, Spain .
  • Doreian, Patrick, Vladimir Batagelj, and Anuska Ferlioj. 2005. Generalized Blockmodeling. Cambridge , England : Cambridge University Press.
  • Eckstein, Zvi, and Kenneth Wolpin. 1989. “The Specification and Estimation of Dynamic Stochastic Discrete Choice Models. Journal of Human Resources 24:56298.
  • Erdös, P., and A. Rényi. 1960. “On the Evolution of Random Graphs. Public Mathematical Institute of Hungary Academy of Sciences 5:1761.
  • Fararo, Thomas J. 1989. “The Spirit of Unification in Sociological Theory. Sociological Theory 7:17590.
  • Fararo, Thomas J., and John Skvoretz. 1987. “Unification Research Programs: Integrating Two Structural Theories. American Journal of Sociology 92:1183209.
  • Feser, Edward J., Stuart H. Sweeney, and Henry C. Renski. 2005. “A Descriptive Analysis of Discrete U.S. Industrial Complexes. Journal of Regional Science 45:395419.
  • Fienberg, Steven E., M. M. Meyer, and Stanley Wasserman. 1985. “Statistical Analysis of Multiple Sociometric Relations. Journal of the American Statistical Association 80:5167.
  • Fienberg, Steven E., and Stanley Wasserman. 1981. “Categorical Data Analysis of Single Sociometric Relations.” Pp. 15692 in Sociological Methodology, vol. 12. San Francisco , CA : Jossey-Bass.
  • Fossett, Mark. 2006. “Ethnic Preferences, Social Distance Dynamics and Residential Segregation: Theoretical Explanations Using Simulation Analysis. Journal of Mathematical Sociology 30:185274.
  • Frank, Ove, and David Strauss. 1986. “Markov Graphs. Journal of the American Statistical Association 81:83242.
  • Freidkin, Noah. 1998. A Structural Theory of Social Influence. Cambridge , England : Cambridge University Press.
  • Frey, William H., and Reynolds Farley. 1994. “Latino, Asian and Black Segregation in U.S. Metropolatin Areas: Are Multi-Ethnic Metros Different Demography 33:3550.
  • Galster, George C. 1982. “Black and White Preferences for Neighborhood Racial Composition. Real Estate Economics 10:3966.
  • Gamerman, Dani. 1997. Markov Chain Monte Carlo: Stochastic Simulation for Bayesian Inference. London : Chapman and Hall.
  • Gelman, Andrew, John B. Carlin, Hal S. Stern, and Donald B. Rubin. 1995. Bayesian Data Analysis. London : Chapman and Hall.
  • Geyer, Charles J., and Elizabeth A. Thompson. 1992. “Constrained Monte Carlo Maximum Likelihood Calculations” (with discussion). Journal of the Royal Statistical Society, Series C, 54:65799.
  • Gilks, Walter R. 1996. “Full Conditional Distributions”. Pp. 7588 in Markov Chain Monte Carlo in Practice, edited by WalterR. Gilks, Sylvia Richardson, and David J.Spiegelhalter. London : Chapman and Hall.
  • Gilks, Walter R., Sylvia Richardson, and David J. Spiegelhalter. 1996a. “Intoducing Markov Chain Monte Carlo.” Pp. 120 in Markov Chain Monte Carlo in Practice, edited by Walter R.Gilks, SylviaRichardson, and David J.Spiegelhalter. London : Chapman and Hall.
  • Gilks, Walter R., Sylvia Richardson, and David J. Spiegelhalter, eds. 1996b. Markov Chain Monte Carlo in Practice. London : Chapman and Hall.
  • Granovetter, Mark. 1973. “The Strength of Weak Ties. American Journal of Sociology, 78:136980.
  • Handcock, Mark S. 2003a. “Assessing Degeneracy in Statistical Models of Social Networks.” CSSS Working Paper No. 39, University of Washington .
  • Handcock, Mark S. 2003b. “Statistical Models for Social Networks: Inference and Degeneracy.” Pp. 22940 in Dynamic Social Network Modeling and Analysis, edited by RonBreiger, Kathleen M.Carley, and PhilippaPattison. Washington , DC : National Academies Press.
  • Handcock, Mark S., David R. Hunter, Carter T. Butts, Steven M. Goodreau, and Martina Morris. 2003. “The Statnet Library for R.” Software library (
  • Harary, Frank. 1953. “On the Notion of Balance of a Signed Graph. Michigan Mathematical Journal 3:3741.
  • Heider, Fritz. 1958. The Psychology of Interpersonal Relations. New York : Wiley.
  • Hoff, Peter D., Adrian E. Raftery, and Mark S. Handcock. 2002. “Latent Space Approaches to Social Network Analysis. Journal of the American Statistical Association 97:109098.
  • Holland, Paul W., Kathryn Blackmond Laskey, and Samuel Leinhardt. 1983. “Stochastic Blockmodels: First Steps. Social Networks 5:10937.
  • Holland, Paul W., and Samuel Leinhardt. 1981. “An Exponential Family of Probability Distributions for Directed Graphs” (with discussion). Journal of the American Statistical Association 76:3350.
  • Hubert, Lawrence J. 1987. Assignment Methods in Combinatorial Data Analysis. New York : Marcel Dekker.
  • Huffman, Matt L., and Philip N. Cohen. 2004. “Racial Wage Inequality: Job Segregation and Devaluation Across U.S. Labor Markets. American Journal of Sociology 109:90236.
  • Hummon, Norman P., and Patrick Doreian. 2003. “Some Dynamics of Social Balance Processes: Bringing Heider Back into Balance Theory. Social Networks 25:1749.
  • Hunter, David R., and Mark S. Handcock. 2006. “Inference in Curved Exponential Family Models for Networks. Journal of Computational and Graphical Statistics 15:56583.
  • Johansen, Søren. 1979. Introduction to the Theory of Regular Exponential Families. Copenhagen : University of Copenhagen.
  • Jones, Rachael, and April Brayfield. 1997. “Life's Greatest Joy?: European Attitudes Toward the Centrality of Children. Social Forces 75:123969.
  • Joy, Lois. 2003. “Salaries of Recent Male and Female College Graduates: Educational and Labor Market Effects. Industrial and Labor Relations Review 56:60621.
  • Kalos, Malvin H., and Paula A. Whitlock. 1986. Monte Carlo Methods. Vol. 1, Basics. New York : Wiley.
  • Kelso, Alexander S., and Vincent P. Crawford. 1982. “Job Matching, Coalition Formation and Gross Substitutes. Econometrica 50:1483504.
  • Kittel, Chales, and Herbert Kroemer. 1980. Thermal Physics. 2d ed. New York : W. H. Freeman.
  • Lazega, Emmanuel. 2001. The Collegial Phenomenon: The Social Mechanisms of Cooperation Among Peers in a Corporate Law Partnership. Oxford , England : Oxford University Press.
  • Loehman, Edna, and Vo Hu De. 1982. “Application of Stochastic Choice Modeling to Policy Analysis of Public Goods: A Case Study of Air Quality Improvements. Review of Economics and Statistics 64:47480.
  • Lorrain, Francois, and Harrison C. White. 1971. “Structural Equivalence of Individuals in Social Networks. Journal of Mathematical Sociology 1:4980.
  • Luce, R. Duncan. 1959. Individual Choice Behavior. New York : Wiley.
  • Massey, Douglas S., and Nancy A. Denton. 1993. American Apartheid: Segregation and the Making of the Underclass. Cambridge , MA : Harvard University Press.
  • Mayhew, Bruce H. 1980. “Structuralism Versus Individualism: Part I, Shadowboxing in the Dark. Social Forces 59:33575.
  • Mayhew, Bruce H. 1981. “Structuralism Versus Individualism: Part II, Ideological and Other Obfuscations. Social Forces 59:62748.
  • Mayhew, Bruce H., and Roger L. Levinger. 1976. “Size and Density of Interaction in Human Aggregates. American Journal of Sociology 82:86110.
  • McFadden, Daniel. 1973. “Conditional Logit Analysis of Qualitative Choice Behavior.” Pp. 10542 in Frontiers in Econometrics, edited by P.Zarembka. New York : Academic Press.
  • McPherson, J. Miller, Lynn Smith-Lovin, and James M. Cook. 2001. “Birds of a Feather: Homophily in Social Networks. Annual Review of Sociology 27:41544.
  • Merton, Robert K. 1957. Social Theory and Social Structure. Glencoe , IL : Free Press.
  • Monderer, Dov, and Lloyd S. Shapley. 1996. “Potential Games. Games and Economic Behavior 14:12443.
  • Oppenheim, Norbert. 1995. Urban Travel Demand Modeling: From Individual Choices to General Equilibrium. New York : Wiley.
  • Parsons, Talcott. 1951. The Social System. London : Routledge.
  • Pattison, Philippa, and Garry Robins. 2002. “Neighborhood-Based Models for Social Networks.” Pp. 3013 in Sociological Methodology, vol. 32, edited by Ross M.Stolzenberg. Boston , MA : Blackwell Publishing.
  • Pattison, Philippa, and Stanley Wasserman. 1999. “Logit Models and Logistic Regressions for Social Networks: II. Multivariate Relations. British Journal of Mathematical and Statistical Psychology 52:16993.
  • Press, William H., Saul A. Teukolsky, William T. Vetterling, and Brian P. Flannery. 1992. Numerical Recipes: The Art of Scientific Computing. 2d ed. Cambridge , England : Cambridge University Press.
  • Robert, Christian P. 1994. The Bayesian Choice: A Decision-Theoretic Motivation. New York : Springer.
  • Robins, Garry, and Philippa Pattison. 2005. “Interdependencies and Social Processes: Dependence Graphs and Generalized Dependence Structures.” Pp. 192214 in Models and Methods in Social Network Analysis, edited by Peter J.Carrington, JohnScott, and StanleyWasserman. Cambridge , England : Cambridge University Press.
  • Robins, Garry, Philippa Pattison, and Stanley Wasserman. 1999. “Logit Models and Logistic Regressions for Social Networks, III. Valued Relations. Psychometrika 64:37194.
  • Robins, Garry, Philippa Pattison, and Jodie Woolcock. 2005. “Small and Other Worlds: Network Structures from Local Processes. American Journal of Sociology 110:894936.
  • Robins, Garry L., and Philippa Pattison. 2001. “Random Graph Models for Temporal Processes in Social Networks. Journal of Mathematical Sociology 25:541.
  • Roth, Alvin E., and Marilda Sotomayor. 1990. Two-Sided Matching: A Study in Game-Theoretic Modeling and Analysis. Cambridge , England : Cambridge University Press.
  • Sakoda, James M. 1971. “The Checkerboard Model of Social Interaction. Journal of Mathematical Sociology 1:11932.
  • Sattinger, Michael. 1993. “Assignment Models of the Distribution of Earnings. Journal of Economic Literature 31:83180.
  • Schelling, Thomas C. 1969. “Models of Segregation. American Economic Review 59:48393.
  • Shapley, Lloyd, and Martin Shubik. 1971. “The Assignment Game I: The Core. International Journal of Game Theory 1:11130.
  • Snijders, Tom A. B. 2002. “Markov Chain Monte Carlo Estimation of Exponential Random Graph Models. Journal of Social Structure 3 (
  • Snijders, Tom A. B., Philippa E. Pattison, Garry L. Robins, and Mark S. Handcock. 2006. “New Specifications for Exponential Random Graph Models.” Pp. 99153 in Sociological Methodology, vol. 36, edited by Ross M.Stolzenberg. Boston , MA : Blackwell Publishing.
  • Sorokin, Pitrim A. 1957. Social and Cultural Dynamics. Boston : Porter Sargent.
  • Spencer, Herbert. 1896. Social Statics. New York : D. Appleton.
  • Stewman, Shelby. 1988. “Organizational Demography. Annual Review of Sociology 14:173202.
  • Stoyan, Dietrich, Wilfrid S. Kendall, and Joseph Mecke. 1987. Stochastic Geometry and Its Applications. 2d ed. New York : Wiley.
  • Strauss, David. 1986. “On a General Class of Models for Interaction. SIAM Review 28:51327.
  • Strauss, David, and Michael Ikeda. 1990. “Pseudolikelihood Estimation for Social Networks. Journal of the American Statistical Association 85:20412.
  • Sweeney, Stuart H., and Edward J. Feser. 2003. “Business Location and Spatial Exernalities: Tying Concepts to Measures.” Pp. 23962 in Spatially Integrated Social Science: Examples in Best Practice, edited by M.Goodchild and D.Janelle. Oxford , England : Oxford University Press.
  • Swendsen, Robert H., and Jian-Sheng Wang. 1987. “Non-universal Critical Dynamics in Monte Carlo Simulation. Physical Review Letters 58:8688.
  • Tauber, Karl E., and Alma F. Tauber. 1965. Negroes in Cities: Residential Segregation and Neighborhood Change. Chicago , IL : Aldine.
  • Wasserman, Larry. 2000. “Bayesian Model Selection and Model Averaging. Journal of Mathematical Psychology 44:92107.
  • Wasserman, Stanley, and Philippa Pattison. 1996. “Logit Models and Logistic Regressions for Social Networks: I. An Introduction to Markov Graphs and p*. Psychometrika 60:40126.
  • Wasserman, Stanley, and Garry Robins. 2005. “An Introduction to Random Graphs, Dependence Graphs and p*.” Pp. 192214 in Models and Methods in Social Network Analysis, edited by Peter J.Carrington, JohnScott, and StanleyWasserman. Cambridge , England : Cambridge University Press.
  • White, Harrison C. 1970. Chains of Opportunity: System Models of Mobility in Organizations. Cambridge , MA : Harvard University Press.
  • Young, H. Peyton. 1998. Individual Strategy and Social Structure: An Evolutionary Theory of Institutions. Princeton , NJ : Princeton University Press.
  • Zhang, Junfu. 2004. “A Dynamic Model of Residential Segregation. Journal of Mathematical Sociology 28:14770.