Abstract:
|
This paper concerns the estimation of causal effects over networks, e.g., if you start smoking, will it influence your friends to smoke as well? Such contagion effects, i.e. the causal influence units on the network have on other units, are generically confounded with homophily, the tendency of connected units to share common (latent) traits. Did your friend start smoking because of you, or did you both start because you're in the same Humphry Bogart fan club? Estimating the causal effect requires adjusting for possible shared traits, which often are not observed. Nonetheless, we might expect the network itself to carry the required information: homophily occurs because the link structure reflects latent traits, so we may be able to infer these traits from the link pattern. We describe a method for using node embedding methods to extract this information and perform causal adjustment. By combining embedding techniques with general non-parametric estimators for network causal effects we arrive at an efficient adjustment procedure that requires only minimal parametric assumptions on how the network is generated, and that sidesteps the need to observe the traits of each unit.
|