"Tell me your friends and I will tell you who you are" my primary school teacher Mr Moriarty used to advise us. This could have simply been a statement about correlation but I think implicitly he was thinking of a treatment effects model: he wanted us to to stay away from the gurriers in the class (of whom there were some spectacular examples) as he thought they would lead us into bad ways. I actually didn't need any advice to avoid the thugs.
Anyway, this basic notion of the importance of peer effects is pervasive and in the academic literature goes back to the famous Coleman report in 1966. It spawned a big empirical literature which seems to have shown that, in education at least, peer effects are small and estimates quite fragile. Recent theoretical developments, by people such as Durlauf & Benabou, have highlighted the potential importance of such interactions.
A big problem in such models is identification: it is not a simple matter to identify parameters of interest. This is especially true if you want to distinguish between endogenous peer effects (when my outcome depends on your outcome) as well as contextual effects (when my outcome depends on your characteristics as well as my own). Identifying these separate effects is important if one wants the estimates to be useful for policy purposes since it is essential to know what is causing the peer effect. Endogenous membership (sorting into peer groups) as well as standard correlated ubobservables makes these quite tricky to estimate.
A really good outline of the issues involved is the paper by Robert Moffitt . A recent application in this area is a paper by Devereux et al.
1 comment:
Not literally peer effects, but an interesting paper about identifiability where homophily is competing with social network effects is discussed here. It targets the Framingham-style obesity/divorce/smoking-cessation social network contagion research.
Post a Comment