This past week I came across a research paper by Nicholas Christakis and James Fowler in which they discuss their testing of a means for detecting disease outbreaks early in their development. They start from the idea that individuals who are more central to the network are more like likely to be close (as measured in terms of the number of edges) to the individual who introduces a disease into the network, therefore will catch the disease earlier.
Here’s the super-clever part. Looking at previous research on and attempts at heading off outbreaks, they concluded that actually mapping a social network is too slow and onerous a task to be useful. Epidemics can move quickly, and large social networks can be very difficult to map. Instead, they pursued a much easier process for determining who in a network is more central. Taking advantage of the “friendship paradox” (as they describe it, the strange fact that “your friends have more friends than you do”), they took a random sample of Harvard undergraduates and asked them to name a friend. This second sample, not the original, randomly chosen students but the friends of the randomly chosen students, should theoretically exhibit greater centrality than the random selection, meaning that this group will generally be closer to the source of the disease and therefore catch it sooner. They tested this using students at Harvard during a flu outbreak and found that they were correct; the epidemic peaked earlier in the second group than in the first.
What I find intriguing and important about this work is that they manage to reveal information about the network structure without actually having to map the network. As we’ve learned during this past semester, there are many common, stable properties of networks that could be used to make predictions about network behavior. However, this is often useful only so long as one has a map of the network. By teasing out an easily discoverable group that exhibits higher than average centrality, the researchers were able to use this group as a “sensor” for coming outbreaks of the flu. This reminds of me of what Dr. D when we were studying the adoption of new technologies that have network effects; our abstract quantitative understanding of the phenomenon, which is dependent on having information we would never have in the real world, points us in the direction of a qualitative understanding that is easily applied in the real world. Similarly, Christakis and Fowler’s approach allows for one to make potentially powerful predictions about the spreading of a contagious disease within a network without having to do the difficult-to-impossible work of precisely mapping the network.
(Christakis happens to also have given this related TEDTalk)