I went to a guest lecture by Michael Macy on network autocorrelation. its not fair for me to tell you what he was saying (so look him up if you want), but I will summarize.

there is a lot of research done with surveys, and most of these surveys have an inherent assumption: that the people live on an island and are not autocorrelated.

so he created a false network of people with random fake political views, etc and added network autocorrelation. they found that random views would cluster with demographics and were very statistically significant - but the views were random.

in a similar experiment, a music sharing website to collect data on its users separated users into 'worlds'. users could see what songs were downloaded by other users within the world. and each world had its own popular songs despite being made up of random people.

problem with networks (people) is that even a few missing nodes makes analyzing the network almost useless.