Contagious Treatments: The Policy Relevance of Social Networks

The United States government has increasingly used insights from behavioral science to

inform government policy. However, policy interventions often ignore one of the most

valuable lessons from behavioral science: people change their attitudes and behavior in

response to those around them. We know that even the most targeted policy interventions

have impacts not just on treated individuals, but can spillover onto their friends, families,

neighbors, and their communities. These changes can spread over social networks, with

those closest to a policy intervention most directly affected. Ignoring these spillover effects,

we underestimate the effects of programs, and fail to accurately measure the ways policies

can benefit–and harm–communities.

 

When evaluating policy interventions, in general we would like to identify the causal effect

of the intervention–that is, the difference between people’s outcomes under the intervention

and in the absence of the intervention. If the program changes the outcomes of those directly

treated under the program and their communities, we need to consider spillover to get

an unbiased estimate of the total effects of the program. Additionally, we may think that

community-level effects are most significant when programs are targeted to certain types of

people, or that spillover effects become most important once a certain share of a person’s

peers have been treated. My research will advance methods to measure and compare

different kinds of spillover effects, to help answer some of these questions. Developing and

applying better tools to accurately measure how program effects spread across networks will

give us a more comprehensive picture of how programs affect communities, and will help

policy-makers target treatments most effectively.

 

We can think about two general kinds of spillover effects: spillover that happens at the group

level, and spillover that happens on the individual level. To measure spillover that happens

at the group level, some programs vary the proportion of treated individuals in a community.

This is most relevant when the share of the group treated affects an individual’s outcomes,

but interactions between treated and untreated individuals don’t matter. For example, in

their study of intestinal worms, Edward Miguel and Michael Kremer note that the share

of children receiving deworming medication affects the number of worms in the local

environment. Consequently, this treatment indirectly affects the probability for both treated

and untreated children of contracting worms in the future. Children can’t “catch” worms

from each other, so spillover is not person-to-person, but via an environmental factor. We

generally think about level of vaccination in a similar way, and often consider a group-level

immunity threshold, above which even unvaccinated individuals are extremely unlikely to

contract a disease.

 

For behavioral outcomes, however, we know that individuals are most affected by the

attitudes and behaviors of those closest to them. Using information about social ties,

network analysis can give us a more nuanced understanding of spillover effects, as we can

consider exposure not just at the community level in terms of the share of treated people,

but also at the individual level in terms of the number or characteristics of treated people

that a given person has contact with.

 

For example, suppose a school is implementing a smoking cessation program that targets

individual students, and the administrators of the program care about the effect the program

has on whether students smoke. The program may directly affect students who participate,

but these effects may also spillover from to the friends of students who went through the

program. We can imagine that a student who didn’t go through the program might be more

or less likely to change their smoking behavior based on how many of their close friends

went through the program, or based on characteristics of the treated students, such as how

popular or social they are.

 

We can also learn more about the conditions for spillover. Spillover effects flow through

two general channels: “outcome interference” and “treatment interference.” Outcome

interference occurs when a person’s outcomes are affected by the outcome of their peers. If

a student who went through the smoking cessation program stops smoking because of it,

their friends may also be more likely to stop smoking due to the social nature of the activity.

Treatment interference occurs when a person’s outcomes are affected by the treatment of

their peers, regardless of their peers’ outcomes. Even if a student who went through the

smoking cessation program doesn’t stop smoking, they may pass on information about the

program or discuss the topic with their friends, potentially resulting in behavior changes in

their friends.

 

These channels are of course not mutually exclusive. But refining methodological tools

for network analysis will help us understand the extent to which each of these channels

contributes to behavior change, and how this varies with different types of outcomes and

different types of programs.