Some interesting examples of field experiments with social networks.

Sharique Hasan // sharique.org

Fuqua School of Business, Duke University

Department of Sociology, Duke University (Courtesy)

Field: Real people; real meaningful outcomes.

Experiment: Randomization into conditions, done by experimenter or naturally in field setting.

There are hundreds of questions that we can try to answer about networks using field experiments.

I will focus on four.

Peer effects: Does j influence the behavior/outcomes of i?

Network Formation: What affects whether i forms a network tie with j?

Designing networks: Which network structures maximize network-level outcomes?

​

Using information in networks: Can we use networks to capture unobservable information for better decisions?

​

There are hundreds of questions that we can try to answer about networks using field experiments.

Peer effects: Does j influence the behavior/outcomes of i?

Alice

Bob

Alice has good grades.

By interacting with Alice, will Bob also get good grades?

Influence

Y(i) = b0 + b1*(Y(j)) + e

Selection: Similar people are more likely to interact. � We are picking up homophily, not influence.

Reflection: The regression specifies the influence of A on B.

We are picking up the influence of B on A.

There are at least three problems with this approach:

Common shocks: People who are interacting have common contexts and experiences.

We are picking up the effect of common experiences (e.g., noisy hallways or extra coffee on the dorm floor).

In essence, there is imbalance between the treatment and control groups.

peer influence

peer influence

peer influence

Something�Else

Entirely

Something�Else

Something�Else

The identification problem: � How much of b1 is actually peer influence?

Selection: Similar people are more likely to interact.

Include controls that account for dimensions on which people may decide to form a connection.

Reflection: The regression specifies the influence of A on B.

Lag the independent variable.

The standard approach to dealing with this identification problem:

Common shocks: People who are interacting have common contexts and experiences.

Include features of the context that may influence the performance of both A and B.

Y(i,t) = b0 + b1*(Y(j,t-1)) + Controls + e

This seems like a reasonable strategy, but:

How many other factors can you actually account for with “controls?”

Y(i,t) = b0 + b1*(Y(j,t-1)) + e

Why randomization?:

​

1. Creates balance between treatment and control groups.�
2. Reduces the likelihood that unobserved factors of selection or context are causing the observed effects.�
3. This is because, treatment condition and the characteristics of pairs are now uncorrelated.

Y(i,t) = b0 + b1*(Y(j,t-1)) + Controls + e

No evidence for a causal peer effect. The coefficient is basically “0.”

Alice

Bob

Alice has good grades.

By interacting with Alice, will Bob also get good grades?

Influence

Much of this was driven by two thing: (1) subject specific learning; (2) peer effects were driven primarily by students who actually studied together.

“the messages not only influenced the users who received them but also the users’ friends, and friends of friends. The effect of social transmission on real-world voting was greater than the direct effect of the messages themselves, and nearly all the transmission occurred between ‘close friends’ who were more likely to have a face-to-face relationship. These results suggest that strong ties are instrumental for spreading both online and real-world behaviour in human social networks”

Alice

Bob

I encourage Alice to vote by showing an online message.

Does Bob, who is connected to Alice online, also vote?

Influence

Our estimates show that peer influence causes more than a 60% increase in odds of buying the service due to the influence coming from an adopting friend. In addition, we find that users with a smaller number of friends experience stronger relative increase in the adoption likelihood due to influence from their peers as compared to the users with a larger number of friends.

Alice

Bob

I encourage Alice to buy an online product.

Does Bob, who is connected to Alice online, also buy it?

Influence

Founders who received advice from peers who actively managed their employees---with regular meetings, goal setting, and feedback---grew their firms to be 28% larger and were 10% less likely to fail as compared to those who got advice from peers with a passive people-management approach. However, entrepreneurs with MBAs or accelerator experience did not respond to the advice of either active or passive peers, suggesting that formal training can limit the spread of informal management advice from peers.

Network Formation: What affects whether i form a network tie with j?

Alice

Bob

What encourages Alice and Bob to form a tie?

We run a randomized field experiment on a major North American online dating website, where 50,000 of 100,000 randomly selected new users are gifted the ability to anonymously view profiles of other users. Compared with the control group, the users treated with anonymity become disinhibited, in that they view more profiles and are more likely to view same-sex and interracial mates.

Anonymous users, who lose the ability to leave a weak signal, end up having fewer matches compared with their nonanonymous counterparts. This effect of anonymity is particularly strong for women, who tend not to make the first move and instead rely on the counterparty to initiate the communication.

“I was looking back to see if you were looking back at me to see me looking back at you”

Vary search costs for pairs of potential collaborators by randomly assigning individuals to 90-minute structured information-sharing sessions as part of a grant funding opportunity.

​

We estimate that the treatment increases the probability of grant co-application of a given pair of researchers by 75%. The findings suggest that matching between scientists is subject to considerable friction, even in the case of geographically proximate scientists working in the same institutional context.

Alice

Bob

Alice

Cal

Alice

Bob

Cam

Dev

Does Alice’s network grow in �proportion to her connections?

We find strong evidence that interacting with random, but well-connected, roommates causes significant growth of a focal student’s network. Further, we find that this growth also implies an increase in how close an actor moves to a network’s center and whether that actor is likely to serve as a network bridge.

Overall, network ties formed after a randomized interaction account for about one-third the individuals a participant knows, of their friendships, and their advice relations. Nevertheless, roughly 90% of randomized interactions never become social ties of friendship or advice. A key result from our research is that while joint tasks may serve to structure the social consideration set of possible connections, individual preferences strongly shape the structure of networks.

Designing networks: Which network structures maximize network-level outcomes?

“behavior spread farther and faster across clustered-lattice networks than across corresponding random networks.”

In order to induce farmers to adopt a productive new agricultural technology, we apply simple and complex contagion diffusion models on rich social network data from 200 villages in Malawi to identify seed farmers to target and train on the new technology

Alice

Bob

Alice has good grades.

By interacting with Alice, will Bob also get good grades?

Influence

Y(i,t) = b0 + b1*(Y(j,t-1)) + b2*(Y(i,t-1)) +b3*(Y(j,t-1)*Y(i,t-1)) + e

We provide evidence that within our “optimally” designed peer groups, students avoided the peers with whom we intended them to interact and instead formed more homogeneous subgroups.

​

These results illustrate how policies that manipulate peer groups for a desired social outcome can be confounded by changes in the endogenous patterns of social interactions within the group.

Using information in networks: Can we use networks to capture unobservable information for better decisions?

​

Alice

Bob

X1…xn (tacit knowledge)

ROI = b0 + b1*(X-bar) + b2*(Random $) +b3*(X-bar * Random $) + e

Sam

Chris

X1…xn (tacit knowledge)

X1…xn (tacit knowledge)

There are hundreds of questions that we can try to answer about networks using field experiments.

Four types of network experiments

Peer effects: Does j influence the behavior/outcomes of i?

Network Formation: What affects whether i forms a network tie with j?

Designing networks: Which network structures maximize network-level outcomes?

​

Using information in networks: Can we use networks to capture unobservable information for better decisions?

​