Please join Information Science with colloquium guest, Dean Eckles. Dean Eckles is a social scientist, statistician, and member of the Data Science team at Facebook. He studies how interactive technologies affect human behavior by mediating, amplifying, and directing social influence — and the statistical methods to study these processes. His current work uses large field experiments and observational studies. His research appears in peer reviewed proceedings and journals in computer science, marketing, and statistics. Dean holds degrees from Stanford University in philosophy (BA), cognitive science (BS, MS), and statistics (MS), and communication (PhD).
Title: Peer effects and interventions in networks: Causal inference with and without experiments
Abstract: Peer effects (i.e., social interactions, interference, social influence, spillovers) are common in many settings of interest to social scientists, epidemiologists, system designers, and policy-makers. Researchers often aim to estimate these peer effects themselves and/or estimate what would happen under a global (i.e., network-wide) treatment that functions partially through peer effects. In this talk, I consider multiple strategies for learning about peer effects in online social networks using a variety of experimental and non-experimental designs.
For estimating peer effects, we use experimental designs that either modulate mechanisms by which peer effects occur or encourage peers to engage in the behaviors of interest. This is illustrated with examples from online advertising and information sharing. (http://arxiv.org/abs/1206.4327)
When experimentation is not possible, observational analyses require often implausible assumptions to identify peer effects. Nonetheless, adjustment and matching estimators may reduce bias enough to be informative, if not unbiased. We use a large experiment that identifies peer effects in information and media sharing as a "gold standard" for assessing the bias of observational studies of peer effects. High-dimensional models adjusting for thousands of past behaviors provide the greatest bias reduction, such that the full model reduces bias by over 70%.
For estimating effects of global treatments, estimates from simple random assignment and standard analyses can suffer from substantial bias. We use experimental designs that reduce bias by producing treatment assignments that are correlated in the network through the use of methods for graph partitioning. We provide theoretical and simulation results showing this substantially reduces bias and total error. (http://arxiv.org/abs/1404.7530)
Together, this work illustrates the utility of large datasets, experimentation platforms, and modern statistical learning for improving causal inference in social science and decision making. This covers joint work, especially with Eytan Bakshy, Brian Karrer, and Johan Ugander.