Why Communication Studies Should Embrace Complexity and Agent-Based Modeling:

An Interactive Essay

Joe A. Wasserman

@joewasserman | joewasserman.com

A brief but skippable introduction

This essay started with an observation and a personal interest.

The observation. Often, theories are more nuanced than the research into those theories. Uses and gratifications (U&G) theory is a case in point: the theory includes feedback loops among media consumers and producers, processes that unfold over time, interactions among people, and both media and non-media behaviors. Most U&G research, however, primarily focuses narrowly on the gratifications people seek from consuming media.

The personal interest. The tools and methods of complexity science present new ways to model and understand processes in which behaviors and outcomes arise from difficult-to-predict interactions among many individual agents, including humans. These tools include agent-based modeling, which are the focus of this interactive essay.

Why Communication Studies Should Embrace Complexity and Agent-Based Modeling: An Interactive Essay

It has been argued that the tools of complexity science could help communication studies scholars study and understand communication as a dynamic process (Sherry, 2015). Complexity science studies systems with interacting components that generate behaviors that are difficult to predict by observing single components in isolation. Models and theories that involve this kind of interdependence among individuals contravene the typical individual-level focus of communication studies (see Walter, Cody, & Ball-Rokeach, 2018). Complexity science provides methods for studying dynamic processes that account for the interdependence of individuals, including agent-based modeling. Agent-based modeling is a promising entree to complexity science for communication scholars because agent-based models are specified at the level of individual agents—and sociopsychological communication scholars are used to thinking about individuals. After describing complexity science and agent-based modeling, this interactive essay explores the suitability of uses and gratifications theory (Katz, Blumler, & Gurevitch, 1973) for agent-based modeling as a demonstration of the promise and challenge of moving from traditional approaches in communication studies to complexity science and agent-based modeling.

Current Communication Studies Paradigm

The dominant empirical paradigm in communication studies has been sociopsychological (Craig, 1999). In other words, it has been primarily concerned with predicting and explaining the cognitive, affective, and behavioral dimensions of the influences on and consequences of communication at an individual level. Macro-scale studies of communities or societies have been rare, and research that crosses micro (individual) and macro (society) scales even rarer (Walter et al., 2018). In the sociopsychological paradigm, typical analyses statistically analyze small sets of quantitative variables measured from individual people.

While the sociopsychological paradigm has undoubtedly yielded great insights into media consumption and its effects, Sherry (2015) has argued that this paradigm has treated human communication in terms of static snapshots. In contrast to these static snapshots, communication, like other complex systems:

  • is a dynamic process that unfolds over time,
  • involves the interaction of interdependent heterogeneous individuals with different characteristics and behaviors, and
  • involves recursive feedback processes (Sherry, 2015).

Communication research that focuses strictly on independent individuals makes it hard to say anything about more macro-level phenomena that involve groups of individuals. Complexity science is a cross-disciplinary scientific paradigm that holds great promise for advancing understandings of communication processes as complex systems (Sherry, 2015).

Complexity Science and Complex Systems

Although the precise definition of "complex system" is still debated, there are several useful ways to understand what makes complex systems distinct from simple systems. For example, complex systems have been defined as “system[s] in which large networks of components with no central control and simple rules of operation give rise to complex collective behavior, sophisticated information processing, and adaptation via learning or evolution…[or]…that exhibit nontrivial emergent and self-organizing behaviors” (Mitchell, 2009, p. 13). This definition is dense, so let's break it down with an interactive example.

Consider the simulation in Figure 1, a toy model of flocking dynamics that is prototypical of simulations in complexity science. In this model, the components (or agents) are represented by triangular shapes that move about a two-dimensional plane that can be thought of as abstractions of birds. No central control guides the behavior of these bird agents. Instead, each bird individually observes its local environment and adjusts its behavior based on a set of simple rules. Specifically, each bird (1) turns away from birds that are too close, (2) turns to face in the same direction as other nearby birds, and (3) turns toward nearby birds that are not too close. After running for a brief period of time, these three simple rules produce complex collective behavior (i.e., nontrivial emergent or self-organizing behaviors): flocks of birds that travel together in the same direction. Emergence, which occurs when “simple rules produce complex [macroscopic] behavior in hard-to-predict ways” (Mitchell, 2009, p. 13), is a hallmark of complex systems. Crucially, this emergent behavior is self-organized, in that the “behavior arises without an internal or external controller or leader” (Mitchell, 2009, p. 13). Without any kind of leader or controller to arrange individual birds into flocks, the birds nevertheless flock on their own. Adding a small number of additional rules can produce even more complex behaviors, such as flocking in V-shaped patterns (see Wilkerson-Jerde, Stonedahl, & Wilensky, 2009).

Figure 1. Flocking model in NetLogo (Wilensky, 1998a). To run the simulation, click “setup” and then “go.”

More recently, complex systems have been defined as those that “display behavior across a wide range of scales” (Allen, Stacey, & Bar-Yam, 2017, sec. 1 para. 5) or in which “the elements act neither fully independently nor fully coherently, and the separation of scales breaks down” (Bar-Yam, 2017, para. 11). Scale is defined as “the number of entities or units acting in concert” (Allen et al., 2017, sec. 3 para. 1). Again using the flocking model in Figure 1 as an example, the elements or entities are birds. The scale of this model ranges from each bird individually (micro scale) to all of the birds in combination (macro scale). If they were fully independent, the movement of each bird would be unrelated to each other. If they were fully coherent, the movement of all birds would be identical. In the flocking model, separation of scales breaks down because it's not informative to individually describe the behavior of every bird (micro scale) or to take an average of the movement of all of the birds combined (macro scale). Both micro (individual birds) and macro (averaged) descriptions of Figure 1 miss an important characteristic of the system: the emergence of flocks. Even if one were only interested in individual birds, because their behavior partially depends on the behavior of other birds, even the behavior of a single bird cannot be fully explained by examining that bird in isolation from its neighbors.

Many aspects of human communication appear to be complex systems (Sherry, 2015): they involve agents (i.e., humans) who act partly interdependently and partly coherently, such that separation of scales breaks down. Communication processes are central to the construction and maintenance of aggregates of humans including groups, firms, societies, and nation-states. None of these aggregates can be fully described by each of the individuals comprising them or by some global average of those individuals. Instead, the differences among individuals, their relationships, and their communication produce emergent characteristics of the aggregate. These emergent phenomena influence the behaviors of individuals in turn.

Examples of human communication phenomena in which separation of scales breaks down abound. The diffusion of innovations throughout some group involve heterogeneity throughout the population as different individuals adopt an innovation (or not) at different times, are variously influenced depending on their social and communication networks, and different individuals hold greater or lesser influence (Rogers, 1983). The competitiveness of individuals and firms is related to their position in a structure of relationships between individuals and firms (Burt, 1995). Media tastes and consumption practices are structured at a group-level (Gans, 1974). More recently, the proliferation of Internet networks have called to scholars’ attention to communication and information networks connecting individuals (Castells, 2010). Only networks in which each individual is equally connected with every other individual would be fully coherent, and only non-networks of entirely unconnected individuals would be fully interdependent. More realistic network structures lie somewhere in between (e.g., Watts & Strogatz, 1998), including both connected and unconnected individuals. These network structures are typical of complex systems that have network structures. Before turning to a detailed consideration of applying a complexity perspective to a single communication studies theory, a promising complexity science method for communication scholars is described: agent-based modeling.

Studying Complex Communication Systems with Agent-Based Models

Agent-based models (ABMs; see Wilensky & Rand, 2015) are a promising type of dynamic simulation for merging complexity science with communication studies. In ABMs, agents are simulated entities whose behaviors are determined by algorithms that are typically executed by a computer (Wilensky & Rand, 2015). To translate a verbal theory to an ABM, it is necessary to encode the theory as a set of algorithms for agents’ behavior. ABMs have several advantages for communication scholars over equation-based approaches to complexity:

  • ABMs are defined at the level of individuals’ behaviors, which aligns with the often individual-level focus of existing communication scholarship;
  • ABMs are able to identify individual-level behaviors that plausibly produce emergent phenomena at meso- and macro-levels; and
  • the process of developing an ABM can identify gaps in knowledge and theory that can drive future research and theory development.

The aforementioned flocking model in Figure 1 is an example of an ABM. Building ABMs requires several kinds of definitions:

  • defining the types of agents in a model and their properties;
  • defining the possible behaviors of the agents in a model;
  • defining the environment of the agents, which are typically spatial or network-based;
  • defining the interactions among agents, between agents and their environments, and among environmental elements;
  • defining agents’ cognition, which combines agent properties, their behaviors, and their environments: how agents respond to their properties and environments, maximize utility functions, pursue goals, adapt, learn and/or evolve (see Wilensky & Rand, 2015).

The contexts in which creating an ABM is most beneficial map well to many phenomena of interest to communication researchers. ABM is most beneficial when:

  • a situation contains a moderate number of agents, roughly on the order of tens to millions;
  • agents are heterogeneous, i.e., they do not all have identical properties or behaviors;
  • agents interact with each other;
  • agents interact with their environments; and
  • processes of interest have temporal dynamics, i.e., they occur over time (see Wilensky & Rand, 2015).

Many communication phenomena seem to correspond with these criteria. Outside of modeling dyads, small groups, and entire countries (or larger), most communication practices involve somewhere between 10 and several millions of individuals, such as social media usage (Ahern, 2017). Individuals have different characteristics that influence how they behave, such as the influence of individuals’ partisanship on hostile media perceptions (Vallone & Ross, 1985). Humans interact with each other, often via communication. Humans also interact with their environments, for example by consuming media in the pursuit of goals (Katz et al., 1973). Finally, these processes unfold over time.

ABMs have been used to model human phenomena, ranging from early, simple toy models (for an accessible review, see Rauch, 2002) to more recent highly detailed ABMs that incorporate real-world data (see Waldrop, 2018). For example, consider the ABM in Figure 2. In this ABM, which was based on Schelling’s work on segregation (e.g., Schelling, 1971), each square is an agent representing a household in a neighborhood. The color of each square can be thought of as an abstraction of that household's race. The “%-similar-wanted” slider determines the minimum percent of same-colored neighbors each household desires. If that number is not reached, the household will be unhappy (represented by an “X”) and move to a random open space. At first, on average, households are randomly positioned and are bordered by about 50% same-colored households, which is displayed on the “Percent Similar” graph. This value can be thought of as a metric of the amount of segregation among neighborhoods.

In this example, a moderate number of agents (households) that are heterogeneous (different colors) and interact (by observing their neighbors). Over time, these agents' cognition (deciding whether to move or stay put) and behaviors (moving or staying) produce emergent outcomes (segregated neighborhoods).

Like many complex systems, this ABM is sensitive to initial starting conditions. Small differences can lead to dramatically different outcomes. For example, for values of “%-similar-wanted” less than or equal to 25%, percent similar barely increases over 50%. However, with “%-similar-wanted” at 26%, percent similar usually reaches 70% or more. In other words, a small increase in preferences for similar neighbors yields dramatically greater segregation—even when that preference is fairly low.

Figure 2. Segregation model in NetLogo (Wilensky, 1997).

A Complexity Approach toward Uses and Gratifications Theory

This section explores the suitability of a classic communication studies theory, uses and gratifications (U&G), for agent-based modeling as a demonstration of the promise and challenges of moving from traditional approaches in communication studies to complexity science and agent-based modeling. U&G is “concerned with: (1) the social and psychological origins of (2) needs, which generate (3) expectations of (4) the mass media or other sources, which lead to (5) differential patterns of media exposure (or engagement in other activities), resulting in (6) need gratifications and (7) other consequences, perhaps mostly unintended ones” (Katz et al., 1973, p. 510). The majority of U&G research has focused on (a) typologies of media gratifications, or the expected need satisfaction to be experienced via media usage, and (b) the relationship between those expectations and differential patterns of media consumption (Rubin, 2009). U&G research has also addressed social, psychological, and contextual influences on media consumption, as well as how media usage motivations influence media effects (Rubin, 2009). However, nearly all U&G research has neglected the reciprocal manner in which the consequences of media usage might reciprocally influence individuals’ attitudes, behaviors, and social relationships—and thereby the very social and environmental contexts that are thought to produce needs in the first place. U&G theory and research has not been immune to critique (for a review, see Ruggiero, 2000), including that it is overly individualistic. While this critique may be fair for the majority of U&G research, U&G theorizing includes supra-individual elements. Thus, individual-focused U&G research has not fully captured all of the dimensions of U&G theory.

Rosengren’s (1974) formulation of the U&G paradigm (Figure 3) provides one of the most nuanced treatments of U&G theory that includes supra-individual elements. For the most part, advances in U&G have involved more precisely specifying the nature of and relationships among theoretical constructs, rather than substantially altering the proposed theory (see Rubin, 2009). It specified a process in which individuals’ needs produce perceived problems and solutions to those problems, which in turn produce motives for consuming media and engaging other behaviors. These motives drive media consumption that yield gratifications (or non-gratifications) related to prior needs. Media consumption, other behavior, and gratification/non-gratification in turn influence individuals’ characteristics as well as social structures. Crucially, social structures and individual characteristics were specified as influencing all of these processes, thus incorporating feedback loops and scale-crossing phenomena into the model.

Figure 3. Graphical depiction of uses and gratifications, adapted from Rosengren (1974).

Rosengren’s (1974) formulation of U&G appears to describe a complex system. Individuals’ behaviors influence their environments, which in turn influence their behavior. Although these environments were under-theorized by Rosengren (1974), they can be reconceptualized as emergent properties of aggregates of individuals. Thus, U&G could involve partially independent actors whose behaviors also exhibit a degree of coherence. In other words, U&G involves phenomena in which separation of micro and macro scales breaks down.

U&G appears to meet all five criteria for benefitting from agent-based modeling. First, U&G involves a moderate number of agents (media consumers). Second, individuals are heterogeneous, having different personal characteristics. Third, individuals’ interactions with each other influence the process described by Rosengren (1974) and potentially gratify perceived needs. Fourth, individuals interact with their environments when they consume media and their environments influence the U&G process. Finally, these processes occur over time. Indeed, Slater’s (2007) reinforcing spirals model of the mutual, reciprocal influences of media usage and beliefs, attitudes, and behaviors can be interpreted as a cognate of U&G. Although Slater (2007) represented reinforcing spirals as a cross-lagged path model, he explicitly referred to general systems theory (Bertalanffy, 2015), an antecedent of complexity science, and feedback loops, a common element of complex systems. What, then, might it look like to construct an ABM based on U&G? The following description of converting a verbally-described theory into a fairly simplistic ABM (Figure 4) will highlight some of the choices necessary to make during this process, the high level of specificity required for modeling, and the challenges involved.