A General Agent-Based Model of Social Learning

by Sarah A. Nowak, Luke J. Matthews, Andrew M. Parker

This Article

RAND Health Quarterly, 2017; 7(1):10

Abstract

When engaging in behaviors that may entail risks or outcomes that are unknown or uncertain, individuals often look beyond their own experiences (including past behaviors and subsequent outcomes) to consider the experiences of others in their immediate social networks. This social influence at the micro-scale (i.e., the way in which individuals are influenced by their immediate social networks) can affect change in the greater social web in such a way that social networks may have profound effects on decisionmaking at the population level. Such micro-level social influence is central to many theories of individual decisionmaking and behavior. Observations of population-level dynamics at the macro-level demonstrate the end result of these processes—for example, over time, people's behavior tends to look more like that of their peers. This article describes a general agent-based model (ABM) for studying social influence, and uses that general ABM to explore the relationship between micro-influence and macro-dynamics for broad classes of problems. We also describe an approach to tailor the general ABM to model a specific behavior influenced by social learning, which we illustrate using surveys designed to inform the ABM. The framework we developed could be useful for studying any system in which social learning may occur. But while our general ABM can produce dynamics reminiscent of those that might result from many different types of behaviors, it will typically need to be tailored when used to model any particular behavior.

For more information, see RAND RR-1768-NIH at https://www.rand.org/pubs/research_reports/RR1768.html

Full Text

When engaging in behaviors that may entail risks or outcomes that are unknown or uncertain, individuals often look beyond their own experiences to consider the experiences of others in their immediate social networks. This is social influence at the micro-scale, and it can bring about change in the greater social web in such a way that local social networks may have profound effects on behaviors at the population level. This article describes a general agent-based model (ABM) for studying the ways in which micro-level social influence gives rise to population-level dynamics at the macro-level. ABMs consist of interacting "agents," which in our case are individuals. These individuals are connected to each other through a social network and can share information about behaviors and experiences with others to whom they are connected. In turn, individuals' behaviors can be influenced by information they receive from other agents in their social networks. The general ABM we developed can be used on its own to examine how social networks influence generic classes of behaviors—that is, groups of behaviors that share certain characteristics (e.g., behaviors for which outcomes are always self-reinforcing)—or it can be tailored in conjunction with surveys to examine specific behaviors. This study discusses results from the general ABM and describes the two ways in which we used a national survey to tailor the general ABM in the past, and the additional possibilities for how we might tailor the model in the future.

We examined the following different classes of behaviors the ABM could produce. We evaluated cases in which individuals were engaging in one of two mutually exclusive behaviors A and Ā (such as vaccinating or not vaccinating):

  1. The outcomes of both A and Ā suggest that A is the better behavior; for example, A could represent the adoption of a new, superior technology. We find that if memory, information transmission, and the effect of learning about network outcomes are relatively large, most individuals in the network will end up engaging in behavior A. If these parameters are small, most individuals will revert to a default behavioral state.
  2. The outcomes of behavior A suggest that behavior A is a good choice and outcomes from Ā suggest that Ā is a good choice, as would be the case with membership in one of two equally attractive groups (e.g., becoming a fan of different sports teams). This type of social learning mechanism can lead to distinct clusters of behavior in the population if there are few long-range connections between individuals and if initial conditions include large, relatively homogeneous clusters of individuals engaging in behavior A or Ā. Small clusters and large numbers of long-range social network connections can destabilize initial clusters.
  3. The outcome of behavior A suggests that behavior Ā is the better choice and the outcome of behavior Ā suggests that A is the better choice, as may be the case with two unsatisfying cable service providers. This results in dynamics in which individuals alternate between behaviors A and Ā. If individuals are highly connected, they may switch from one behavior to another in groups. If they are not, they may switch as individuals.
  4. Behavior A results in many positive outcomes that reinforce behavior A, but occasionally lead to very bad side effects that suggest behavior Ā is better—that is, behavior A leads to some outcomes with positive influence and other outcomes with negative influence. An example of this would be a medication that is generally helpful, but has a rare and severe side effect. In this case, behavior A will initially spread rapidly in a population, but may temporarily decrease in frequency, or may no longer occur entirely as individuals begin to experience adverse outcomes.

Additionally, we note that while our general ABM can produce dynamics reminiscent of those that might result from many different types of behaviors, the general ABM will need to be tailored when it is used to model any particular behavior.

The research described in this article was performed by RAND Health.

RAND Health Quarterly is produced by the RAND Corporation. ISSN 2162-8254.