Separating homophily and peer influence with latent space

We study the impact of peer behavior on the adoption of mobile apps in a social network. To identify social influence properly, we introduce latent space as an approach to control for latent homophily, the idea that "birds of a feather flock together." In a series of simulations, we show that latent space coordinates significantly reduce bias in the estimate of social influence. The intuition is that latent coordinates act as proxy variables for hidden traits that give rise to latent homophily. The approach outperforms existing methods such as including observed covariates, random effects, or fixed effects. We then apply the latent space approach to identify social influence on installation of mobile apps in a social network. We find that peer influence accounts for 27% of mobile app adoptions, and that latent homophily inflates this estimate by 40% (to 38%). In some samples, ignoring latent homophily can result in overestimation of social effects by over 100%.