When you recommend a new product to your friend and she tries it out, is it because you've persuaded her, or because she shares your preferences? If you're like me, you don't generally bother recommending things to people unless you know they happen to share your tastes along some dimension. Yet in principle there are some people out there whose opinions are "influential" -- coaxing otherwise disinterested people to try new things despite having different tastes. In principle, that's part of why professional athletes and celebrities are paid to endorse products. But teasing out exactly how much of the adoption of a new product is due to influence and how much is due to shared characteristics (i.e., "homophily") in the network can be difficult.
One cut at this question is forthcoming in PNAS in an article titled, "Distinguishing Influence Based Contagion from Homophily Driven Diffusion in Dynamic Networks" by Sinan Aral, Lev Muchnik,and Arun Sundararajan (as noted on the OrgTheory blog) (try emailing Aral for a copy):
In the context of product adoption, peer influence is associated with the presence of adopters in one‟s local network... However, identification of causal peer influence effects is complicated by the unobservability problem. Each user either has adopter friends or not, making it impossible to observe whether those with adopter friends ... would have adopted had they not had adopter friends. Homophily in this case creates a selection bias because treatments are not randomly assigned: adopters are more likely to be treated due to similarity with their neighbors. Thus, frequently used methods such as regression analysis, which can only establish correlation, are insufficient. Causal treatment effects can on the other hand be estimated via matched sampling, which controls for confounding factors and overcomes selection bias by comparing observations that have the same likelihood of treatment.
Toward this end we adapt matched sample estimation for use in dynamic networked settings. Conditioning matches on a vector of observable characteristics, behaviors and attributes yields influence estimates that account for the homophily that may make product adoption decisions cluster in the network even if no influence exists. This procedure establishes upper bounds on the degree to which influence (rather than homophily) explains assortative mixing and temporal clustering in networks. Since influence can vary over time, our framework provides estimates of its evolution. We can also assess the marginal influence of having an additional adopter friend for any arbitrary number of friends.
The upshot of this work is that shared characteristics among individuals accounts for almost all of the "viral" spread of new product adoption -- i.e., that preferences are largely fixed within the time-scale of weeks or months after a new product is introduced. There exists, nonetheless, a small group of people who are "influenced" into adopting a new product -- i.e., whom adopt the product through social interactions with other adopters with whom few characteristics are shared.
At least on its face, the work seems to affirm the use of fixed preferences as an approximation in this context. This validates the use of this approximation by economists. It also demonstrates, however, that the approximation is not without exception. Also of interest to economists.
One question I had on considering the work is whether it would be possible to single out the "influenced" and "influencers" in this survey, and determine whether or not such people occupy a position as bridges or "weak-tie" links within the social networks. Along the lines of Granovetter, the idea is that influence here is largely a function of whether or how often individuals come into contact with others with characteristics far different from their own. It would be interesting to know whether these individuals could be identified using this approach, which isolates "influence" from homophily.


Recent Comments