Abstract
We know that social influence or social contagion exists, and one of the most important problems in network science is to actually measure it. Doing so faces a fundamental obstacle: contagion effects are unidentified in the presence of latent homophily, which is to say almost always. Understanding where this identification problem comes from makes it clear why it invalidates almost all published observational studies on contagion, and why none of the obvious fixes work. There seem to be two ideas which aren't yet known to fail: partial identification or bounding, and control via community discovery. Neither of them will get rid of the fundamental problem, but they might help sometimes.