BEGIN:VCALENDAR
X-WR-CALDESC:Yale Department of Computer Science
X-WR-CALNAME:Yale CS Events
BEGIN:VEVENT
SUMMARY:Dissertation Defense - Christopher Harshaw
DTSTART;VALUE=DATE-TIME:20210923T141500
DTEND;VALUE=DATE-TIME:20210923T151500
DESCRIPTION:Event description:\nDissertation Defense\nChristopher Harshaw\
n\nTitle: Algorithmic Advances for the Design and Analysis of Randomized
Experiments\n\nAdvisors: Daniel Spielman and Amin Karbasi\n\nOther commit
tee members:\nSekhar Tatikonda\nFredrik Sävje\n\nAbstract:\n\nRandomized
experiments are the gold standard for establishing the causal effect of a
treatment on a population. In this dissertation\, I present algorithmic ad
vances for three different problems arising in the design and analysis of
randomized experiments: covariate balancing\, variance estimation\, and bi
partite experiments.\n\nIn the first part of the defense\, we describe an
inherent trade-off between covariate balancing and robustness\, which is t
hen formulated as a distributional discrepancy problem. In order to navig
ate this trade-off\, we present the Gram–Schmidt Walk Design which is ba
sed on the recent discrepancy algorithm of Bansal\, Dadush\, Garg\, and Lo
vett (2019). By tightening the algorithmic analysis\, we derive bounds on
the mean squared error of the Horvitz–Thompson estimator under this desi
gn in terms of a ridge regression of the outcomes on the covariates\, whic
h we interpret as regression by design. We carry out further analysis incl
uding tail bounds on the effect estimator\, methods for constructing confi
dence intervals\, and a kernel method extension of the design which accomm
odates non-linear responses.\n\nIn the second part of the dense\, we study
the problem of estimating the variance of treatment effect estimators und
er interference. It is well-known that due to the fundamental problem of c
ausal inference\, unbiased variance estimation is impossible without stron
g assumption on the outcomes. Thus\, we study a class of conservative esti
mators which are based on variance bounds. We identify conditions under wh
ich the variance bounds themselves are admissible and provide a general al
gorithmic framework to construct admissible variance bounds\, according to
the experimenter’s preferences and prior substantive knowledge.\n\nTime
permitting\, we will present methodology for the newly proposed bipartite
experimental framework\, where units which receive treatment are distinct
from units on which outcomes are measured\, and the two are connected via
a bipartite graph. We propose the Exposure Re-weighted Linear (ERL) estim
ator which we show is unbiased in finite samples and consistent and asympt
otically normal in large samples provided the bipartite graph is sufficien
tly sparse. We provide an variance estimator which facilitates confidence
intervals based on the normal approximation. Finally\, we present Exposure
-Design\, a correlation clustering based design for improving precision of
the ERL estimator.\n\n\nhttps://cpsc.yale.edu/event/dissertation-defense-
christopher-harshaw
LOCATION:Zoom Presentation
STATUS:CONFIRMED
URL:https://cpsc.yale.edu/event/dissertation-defense-christopher-harshaw
END:VEVENT
END:VCALENDAR