Hi everyone!
This week at the Applied Statistics Workshop we will be welcoming *Michael
Bronstein*, Radcliffe Fellow at the Harvard Institute for Advanced Study.
He will be presenting work entitled *Deep Learning on Graphs: Going Beyond
Euclidean Data*. Please find the abstract below and on the Applied Stats
website here
<https://projects.iq.harvard.edu/applied.stats.workshop-gov3009>.
As usual, we will meet at noon in CGIS Knafel Room 354 and lunch will be
provided. See you all there!
-- Dana Higgins
*Title:* *Deep Learning on Graphs: Going Beyond Euclidean Data*
*Abstract:* In the past decade, deep learning methods have achieved
unprecedented performance on a broad range of problems in various fields
from computer vision to speech recognition. So far research has mainly
focused on developing deep learning methods for Euclidean-structured data.
However, many important applications have to deal with non-Euclidean
structured data, such as graphs and networks. Such data are becoming
increasingly important in computer graphics and 3D vision, sensor networks,
drug design, biomedicine, high energy physics, recommendation systems, and
web applications. The adoption of deep learning in these fields has been
lagging behind until recently, primarily since the non-Euclidean nature of
objects dealt with makes the very definition of basic operations used in
deep networks rather elusive. In this talk, I will introduce the emerging
field of geometric deep learning on graphs, overview existing solutions and
applications for this class of problems, and outline the key difficulties
and future research directions.
Hi everyone!
This week at the Applied Statistics Workshop we will be welcoming *Cassandra
Pattanayak*, Director of the Quantitative Analysis Institute and lecturer
at Wellesley College. She will be presenting work entitled *Effects of
Mentors and Advisors on Major Choice: A Naturally Re-Randomized Experiment
and a Subclassified Observational Study*. Please find the abstract below
and on the Applied Stats website here
<https://projects.iq.harvard.edu/applied.stats.workshop-gov3009>.
As usual, we will meet at noon in CGIS Knafel Room 354 and lunch will be
provided. See you all there!
-- Dana Higgins
*Title:* *Effects of Mentors and Advisors on Major Choice: A Naturally
Re-Randomized Experiment and a Subclassified Observational Study*
*Co-authored with Kelly Kung (Boston University)*
*Abstract:* We examine the effects of a peer mentor’s major and faculty
advisor’s department on a student’s choice of major at a liberal arts
women’s college. In particular, we study whether assignment to a peer
mentor or faculty advisor from STEM v. non-STEM fields during the first
year of college influences students’ choice to major in STEM v. non-STEM.
Our approach to the mentoring component illustrates a novel application of
rerandomization techniques to a natural experiment. The advising component
is an observational study designed for causal inference via
subclassification. Students’ STEM v. non-STEM choices did not appear to be
affected by the STEM status of first year peer mentors or faculty advisors
at this college.
Hi everyone!
This week at the Applied Statistics Workshop we will be welcoming *Christopher
Lucas*, graduate student in the Harvard Government Department. He will be
presenting his job talk entitled *A Model for Political Video: The Audio
Video Neural Network*. Please find the abstract below and on the Applied
Stats website here
<https://projects.iq.harvard.edu/applied.stats.workshop-gov3009>.
As usual, we will meet at noon in CGIS Knafel Room 354 and lunch will be
provided. See you all there!
-- Dana Higgins
*Title:* *A Model for Political Video: The Audio Video Neural Network*
*Abstract:* In the 2016 election cycle political campaigns spent over $4.4
billion on television advertisements. What type of content do campaigns use
in their efforts to influence voter behavior? Past efforts to answer
questions such as these, where media and television content are the topic
of interest, require researchers to engage in herculean efforts to manually
watch and classify video content. This is time consuming and prohibitively
expensive for most researchers. I develop a novel, general approach to
video classification, the Audio-Video Neural Network (AVNN), which is the
first contribution from political science to deep learning. The AVNN
recovers subtle categories of interest to political scientists like “fear”
and “negativity” while also successfully learning to parse more topical
classes like “political advertisement” and “cable news.” Importantly, the
model I propose can learn from both visual and audio features. I
demonstrate the AVNN by analyzing campaign advertisements in American
elections, first showing that my model can be used to separate political
campaign advertisements from cable news. Second, I show how subtle video
labels like “appeals to fear” can also be recovered. All methods described
here are implemented in easy-to-use Rand Python packages
Hi everyone!
This week at the Applied Statistics Workshop we will be welcoming *Anton
Strezhnev*, graduate student in the Harvard Government Department. He will
be presenting his job talk entitled *Robust principal scores for estimating
survivor causal effects with application to analyses of litigation outcomes*.
Please find the abstract below and on the Applied Stats website here
<https://projects.iq.harvard.edu/applied.stats.workshop-gov3009>.
As usual, we will meet at noon in CGIS Knafel Room 354 and lunch will be
provided. See you all there!
-- Dana Higgins
*Title:* *Robust principal scores for estimating survivor causal effects
with application to analyses of litigation outcomes*
*Abstract:* In many studies, outcomes are unobserved or undefined for some
units under analysis. This is common in empirical studies of the outcomes
of legal disputes where a large fraction of cases settle before a decision
is rendered. Analyses that condition on settlement failure are biased for
the causal effect even when the treatment of interest is randomly assigned
if that treatment also affects the probability of settlement. When outcomes
are truncated in this way, a valid causal quantity of interest is the
treatment effect among the ``principal stratum'' of units that would fail
to settle regardless of treatment. Principal score methods estimate these
effects by assuming ignorability of stratum membership given observed
covariates and weighting to eliminate covariate imbalances across strata.
These weights are estimated using a model for principal stratum membership
and can be highly sensitive to changes in model specification. I develop an
estimator for principal score weights that is more robust to
mis-specification of the principal score model by directly incorporating
known covariate balance conditions using a generalized method-of-moments
approach. I illustrate this new approach in a study of win-rates in
international investor-state arbitration.
Hi everyone!
This week at the Applied Statistics Workshop we will be welcoming *Molly
Roberts*, Assistant Professor at the University of California, San Diego.
She will be presenting work entitled *How to Make Causal Inferences Using
Text*. Please find the abstract below and on the Applied Stats website here
<https://projects.iq.harvard.edu/applied.stats.workshop-gov3009>.
As usual, we will meet at noon in CGIS Knafel Room 354 and lunch will be
provided. See you all there!
-- Dana Higgins
*Title:* *How to Make Causal Inferences Using Text*
(with Naoki Egami, Christian Fong, Justin Grimmer and Brandon Stewart)
*Abstract:* Texts are increasingly used to make causal inferences: either
with the document serving as the treatment or the outcome. We introduce a
new conceptual framework to understand all text-based causal inferences,
demonstrate fundamental problems that arise when using manual or
computational approaches applied to text for causal inference, and provide
solutions to the problems we raise. We demonstrate that all text-based
causal inferences depend upon a latent representation of the text and we
provide a framework to learn the latent representation. Estimating this
latent representation, however, creates new risks: we may unintentionally
create a dependency across observations or create opportunities to fish for
large effects. To address these risks, we introduce a train/test split
framework and apply it to estimate causal effects from an experiment on
immigration attitudes and a study on bureaucratic responsiveness. Our work
provides a rigorous foundation for text-based causal inferences, connecting
two previous disparate literatures.