Experiments on Social Networks
27 Feb 2017 16:30
(A placeholder, while I think things through. Much of this would apply to many other networks where the homophily vs. influence problem has parallels.)
Issues: randomization of the network vs. randomization of treatments applied to nodes; protocols for randomizing senders vs. receivers of influence.
— I was fortunate enough to hear lots of these papers presented at the 2013 LARC workshop on statistical and machine learning approaches to network experiments.
See also: Causal Inference
- Recommended:
- Peter Aronow and Cyrus Samii, "Estimating average causal effects under general interference", arxiv:1305.6156
- Eytan Bakshy, Dean Eckles, Rong Yan, Itamar Rosenn, "Social Influence in Social Advertising: Evidence from Field Experiments", arxiv:1206.4327
- Eytan Bakshy, Itamar Rosenn, Cameron Marlow and Lada Adamic, "The Role of Social Networks in Information Diffusion", WWW 2012, arxiv:1201.4145
- Jake Bowers and Mark M. Fredrickson, and Costas Panagopoulos, "Reasoning about Interference Between Units: A General Framework", Political Analysis 21 (2013): 97--124 [Preprint via Prof. Bowers]
- David S. Choi, "Estimation of Monotone Treatment Effects in Network Experiments", arxiv:1408.4102
- Charles F. Manski, "Identification of Treatment Response with Social Interactions", Econometrics Journal 16 (2013): S1--S23 [Thanks to Arun Chandrasekhar for the pointer]
- Andrew C. Thomas and Michael Finegold, "Protocols for Randomized Experiments to Identify Network Contagion" [PDF]
- Johan Ugander, Brian Karrer, Lars Backstrom, and Jon Kleinberg, "Graph cluster randomization: network exposure to multiple universes", arxiv:1305.6979
- Not altogether recommended, but not dis-recommended either:
- Panos Toulis and Edward Kao, "Estimation of Causal Peer Influence Effects", ICML 2013 [The definition of causal effects under interference seems fine to me. But it is not clear to me that their randomization estimator is consistent, and I am quite sure that their Bayesian estimator is only consistent if there is no unmeasured homophily, and the treatment is received independently of node attributes. (Also, pettily, section 1.1 contains an error about the content of my paper with Andrew Thomas.)]
- Robert M. Bond, Christopher J. Fariss, Jason J. Jones, Adam D. I. Kramer, Cameron Marlow, Jaime E. Settle and James H. Fowler, "A 61-million-person experiment in social influence and political mobilization", Nature 489 (2012): 295--298 [A truly massive experiment and a heroic undertaking in both social engineering and data analysis. However, I believe there is a subtle but substantial technical flaw in the way the experiment was designed, which I hope to explain in a paper. PDF reprint via Prof. Fowler.]
- To read:
- Dean Eckles, Brian Karrer, Johan Ugander, "Design and analysis of experiments in networks: Reducing bias from interference", arxiv:1404.7530
- Paul R. Rosenbaum, "Interference Between Units in Randomized Experiments", Journal of the American Statistical Association 102 (2007): 191--200
- To write:
- CRS, "Do Not Adjust Your Receiver: Egocentric and Altercentric Experimental Designs"
