Funded Grants

Experimental Investigations into the Effects of Network Structure on Social Contagions

Most collective behaviors spread through social contact. From the emergence of social norms, to the adoption of technological innovations, to the growth of social movements, social networks are the pathways along which these "social contagions" propagate. Studies of diffusion dynamics have demonstrated that the structure (or topology) of a social network can have significant consequences for the patterns of collective behavior that emerge.

Over the last thirty-five years, questions about how the structure of social networks affects the dynamics of diffusion have been of increasing interest to social scientists. Granovetter's (1973) "Strength of Weak Ties" study, ushered in an era of unprecedented interest in how network dynamics, and in particular diffusion on networks, affect every aspect of social life, from the organization of social movements, to school segregation, to immigration. Granovetter's study showed that "weak ties" between casual acquaintances can be much more effective at promoting diffusion and social integration than "strong ties" between close friends. This is because although casual friendships are relationally weak, they are more likely to be formed between socially distant actors with few network "neighbors" in common. These "long ties" between otherwise distant nodes provide access to new information and greatly increase the rate at which information propagates, despite the relational weakness of the tie as a conduit.

More recently, the explosion of network science across disciplines such as physics, biology, and computer science has produced many important advances for understanding how the structure of social networks affect the dynamics of diffusion. The full impact of Granovetter's original insight was not realized until Watts and Strogatz's (1998) "small world" model demonstrated that bridge ties connecting otherwise distant nodes can dramatically increase the rate of propagation across a network by creating "shortcuts" between remote clusters. Further, these results showed that surprisingly few "long ties" are needed to give even highly clustered networks the "degrees of separation" characteristic of a "small world." This model of network dynamics has had a tremendous impact on fields as diverse as computer science, physics, epidemiology, sociology, and political science.

Yet, despite the remarkable progress of networks research across the disciplines, there may be serious risks in generalizing results from the physical, biological, and even information sciences to the diffusion of collective behaviors. This is because the most important theoretical advances in the study of cascades on large networks (including Granovetter's original study) have been made for simple contagions, such the spread of disease or information, in which a single activated node is sufficient to trigger the activation of its neighbors. However propagation dynamics of social behaviors are typically complex, which means that node activation requires simultaneous exposure to multiple activated neighbors. Social behaviors have the important property that a bystander's probability of joining an activity increases with the level of local participation by her neighbors. Unlike disease or information, one neighbor acting alone is rarely sufficient to pass the contagion to another. This distinctive property of behavioral cascades led us to question whether the effects of small worlds and weak ties can be generalized from simple contagions like disease and information to the complex contagions that characterize the spread of collective behaviors.

In a series of computational studies, our research led to a startling discovery about the spread of collective behaviors: If the credibility of information or the willingness to adopt an innovation depends on receiving independent confirmation from multiple sources, increasing the fraction of "weak ties" or "bridges" in a social network may not only fail to increase the rate of diffusion, but can even preclude diffusion entirely. This surprising reversal of the small world principle suggests that across a wide variety of conditions there may be significant differences in the diffusion dynamics of simple contagions (such as information and disease) versus complex contagions (such as norms and collective behavior).

This has striking implications for any kind of social, commercial, or health intervention designed to promote the spread of a collective behavior. For example, casual contact sexual networks have been shown to have small world and "scale-free" properties, which allow for the rapid diffusion of simple contagions such as HIV/AIDS. Public health officials interested in promoting safe sex behavior have a natural interest in exploiting these causal contact networks since the long ties that accelerate the spread of disease are also the channels along which preventative information can quickly propagate. However, where public health innovations contravene existing social norms, health reform is likely to require social reinforcement, not simply access to information. While word of mouth transmission of new ideas may travel as quickly as the spread of a disease, the information may have little effect in changing entrenched yet risky behaviors without the social reinforcement provided by additional contacts. Thus, efforts to change behavioral norms through peer influence may reach greater numbers with greater speed by targeting networks with seemingly less desirable characteristics for diffusion (e.g., tightly-knit residential networks) rather than the complex networks through which disease is more rapidly transmitted.

These implications highlight the need for more careful research on the effects of network topology on the spread of social contagions. They also highlight a growing disparity between the increasingly powerful tools for developing theoretical insights at the intersection of social science and complex systems, and the empirical resources for testing and evaluating these insights. The empirical problem of collecting social network data is a difficult one. But, for the study of diffusion dynamics, this problem is compounded by the need for dynamic, longitudinal data with extremely fine-grained resolution, without which we cannot track the flow of a behavior through a social network. Further, we neither have systematic data on individual thresholds for adoption based on social network contacts, nor on how social connectedness affects the pathways of diffusion. All of which makes the empirical relevance of theoretical research on social networks profoundly uncertain: without a means for testing the implications of theoretical models, we have no method for arbitrating between them, or for developing cumulative research on the diffusion dynamics of social behavior. Traditional methods of data collection in the social sciences are simply not up to the task.

This proposal focuses on the development of new experimental methods for studying how the structure of a population affects the diffusion dynamics of collective behaviors across it. This research will pioneer the use of an Internet-based experimental design for studying the diffusion of behavior through on-line communities. A pilot study for this proposal, which was developed to test the effectiveness of this approach, has proved to be remarkably successful at generating social cascade dynamics. This experimental design will also generate data on 1) how thresholds for adoption interact with network topology, and 2) how the timing effects of signaling (i.e., how long on average between signals people wait before responding) interact with actors' positions in the social network. While these quantities can significantly affect the dynamics of diffusion, current network models make simplifying assumptions about them because there previously has not been a method for reliably collecting these data.

The pilot study intentionally has a very simple design, however its success opens the door to a host of interesting research questions that previously could not be studied in an experimental setting. For example, the effects of status, power, affective strength of ties, assortative interaction (i.e., homophily), and tie valence on the dynamics of social diffusion have all been very difficult to approach through a purely theoretical lens because of the lack of empirical data on how these variables behave and interact in the context of network diffusion. Similarly, these factors have been difficult to disentangle with available empirical instruments because of their frequent co-occurrence, leading to speculations about their respective roles in the process of diffusion without any systematic means for determining their independent effects. The method of experimental research that we propose pursuing will allow us to disambiguate these factors so that we can isolate and identify the effects of different structural properties of social networks on the dynamics of emergent social behavior.

Our goal with this project is not just to test existing theory, but to develop new theoretical models based on the dynamic data we collect. Thus, we will use the findings from our network experiments to develop new computational and analytical models of the spread of collective behavior. While there is a long tradition of social dynamics research, the study of complex social dynamics is still in its infancy - we know very little about complex contagions, or about how well our general theoretical models map onto various domains of application. The proposed research leverages the new medium of on-line communities to develop a novel approach for empirically testing theoretical models of social diffusion, thereby increasing the power and applicability of research on the dynamics of collective behavior.