Alexander Volfovsky - Duke Statistics
Causal inference in the presence of networks: randomization and observation.
Date: 02/23/2018 (Fri)
Time: 10:00am-11:00am
Location: Seminar will be held on-site: Soc Sci 111
Organizer: Matt Masten
Meeting Schedule: Login or email the organizer to schedule a meeting.
All meetings will be held in the same location as the seminar unless otherwise noted.
10:00am - Seminar Presentation (10:00am to 11:00am)
10:30am - Seminar
11:00am - Federico Bugni
11:30am - Matt Masten
12:00pm - Lunch w/Matt Masten, Federico Bugni, Jackson Bunting
12:30pm - Lunch
Additional Comments: ABSTRACT: Much of classical causal analysis relies on notions of independence. However, modern datasets on disease prevalence, social development, online advertising and business transactions come equipped with observed or estimable network information that links the units together, rendering these notions implausible. When designing randomized experiments, scientists must control for network interference and homophily in order to guarantee the theoretical properties of their estimators. Studying the direct treatment effect in networks, we describe a new class of randomizations that can guarantee unbiasedness and control the variance of the estimator. In situations where an experiment cannot be performed, causal analysis requires the use of matching techniques in order to protect against bias due to a lack of balance between treated and control units. We provide examples of the complications that arise when information about the network is disregarded and develop a matching technique that extends classical propensity scores to the realm of networks.