Hi all,
Our next virtual meeting will be at 12pm (EST) Wednesday, February 24,
where James Robins presents research on "Estimation of Optimal Testing and
Treatment Regimes under No Direct Effect (NDE) of Testing."
*Abstract*:
In this talk I describe new, highly efficient estimators of optimal joint
testing and treatment regimes under the no direct effect assumption that a
given laboratory, diagnostic, or screening test has no effect on a
patient's clinical outcomes, except through the effect of the test results
on the choice of treatment. The proposed estimators attain high efficiency
because they leverage this "no direct effect of testing" (abbreviated as
NDE) assumption.
What is surprising and, indeed, unprecedented in my experience, is that, in
a substantive study of HIV infected subjects, our new estimators delivered
a 50-fold increase in efficiency (and, thus, a 50 fold reduction in
required sample size) compared to estimators that fail to leverage the NDE
assumption! In this talk I review the results of this HIV study, describe
the new estimators, and provide guidance as to when such large gains in
efficiency are to be expected.
Areas in which our new, more efficient estimators should be particularly
important is that of cost-benefit analyses wherein the costs of diagnostic
tests (such as MRIs to screen for lung cancer, mammograms to screen for
breast cancer, and urinary cytology to screen for bladder cancer) are
weighed against the clinical value of the information supplied by the test
results. In a political science context, candidates often conduct private
polls and focus groups to help update campaign outreach decisions such as
the number and content of social media and television ad buys to be
allocated to various demographic groups. One can view private polls and
focus groups as tests and outreach decisions as treatments that together
satisfy no direct effect of testing on the outcome of the election except
through the test results on the choice of treatment.
*Message to the audience*:
The speaker Jamie Robins suggests you read the following less technical
sections of the paper
<https://projects.iq.harvard.edu/files/applied.stats.workshop-gov3009/files/…>
that
cover many of the main ideas.
Section 1: Introduction
Section 4: The NDE of Testing Assumption
Section 5.1
Section 7
Section 8 and 8.1 stopping after remark 7
*Link to the paper*: URL
<https://projects.iq.harvard.edu/files/applied.stats.workshop-gov3009/files/…>
*Zoom link*:
https://harvard.zoom.us/j/97787602526?pwd=Uzh3bVVVS0F4TEVYQTJlV3BQNjcydz09
*Schedule of the workshop*:
https://projects.iq.harvard.edu/applied.stats.workshop-gov3009
Looking forward to seeing you all on Wednesday!
Best,
Soichiro
--
Soichiro Yamauchi
PhD candidate
Harvard University
URL: https://soichiroy.github.io/
Hi all,
Our next virtual meeting will be at 12pm (EST) Wednesday, February 17
(tomorrow), where I (Soichiro Yamauchi) will present "Adjusting for
Unmeasured Confounding in Marginal Structural Models with Propensity-Score
Fixed Effects." This is joint work with Matthew Blackwell (Harvard
University).
*Abstract*:
Marginal structural models are a popular tool for investigating the effects
of time-varying treatments, but they require the assumption that there are
no unobserved confounders between the treatment and outcome. With
observational data, this assumption may be difficult to maintain, and in
studies with panel data, many researchers use fixed effects models to purge
the data of time-constant unmeasured confounding. Unfortunately,
traditional linear fixed effects models are not suitable for marginal
structural models, since they can only estimate lagged effects under
implausible assumptions. To resolve this tension, we propose a novel
inverse probability of treatment weighting estimator with propensity-score
fixed effects to adjust for time-constant unmeasured confounding in
marginal structural models. We show that, in spite of the incidental
parameters problem, these estimators are consistent and asymptotically
normal when the number of units and time periods grow at a similar rate.
Unlike traditional fixed effect models, this approach works even when the
outcome is only measured at a single point in time as is common in marginal
structural models. We apply these methods to estimate the effect of
negative advertising on the electoral success of candidates for statewide
offices in the United States.
*Zoom link: *
https://harvard.zoom.us/j/97787602526?pwd=Uzh3bVVVS0F4TEVYQTJlV3BQNjcydz09
*Schedule of the workshop:*
https://projects.iq.harvard.edu/applied.stats.workshop-gov3009
Looking forward to seeing you all tomorrow!
Best,
Soichiro
--
Soichiro Yamauchi
PhD candidate
Harvard University
URL: https://soichiroy.github.io/
Hi all,
Our next virtual meeting will be at 12pm (EST) Wednesday, February 10,
where we will hear Dean Knox <http://www.dcknox.com/> (University of
Pennsylvania) presents research on "ε-sharp Bounds for Partially Observed
Causal Processes: Testing for Racial Bias in Policing by Fusing Incomplete
Records."
*Authors:*
Guilherme Duarte, Dean Knox, Jonathan Mummolo
*Abstract*:
Social scientists often possess fragmented information about subsets and
aspects of the complex causal processes they study. Research on
police-civilian interactions, for example, is complicated not only by
undocumented interactions, but inconsistent recording of events within
documented interactions. These data constraints can lead to a proliferation
of incompatible analytic approaches relying on contradictory unstated
assumptions, impeding scientific progress on important questions like the
severity of racial bias in policing. Nonparametric sharp bounds, or the
tightest possible range of answers consistent with available data, offer a
path forward: claims outside the bounds can be immediately rejected, and
claims inside the bounds must explicitly justify the additional assumptions
that enable tightening. However, we show proving sharpness is NP-hard for
broad classes of data constraints and causal quantities, rendering this
approach computationally infeasible for even moderately sized causal
processes. We present an efficient spatial branch-and-bound procedure with
a theoretical guarantee that we term "ε-sharpness," indicating the
worst-case looseness factor of the relaxed bounds relative to the (unknown)
completely sharp bounds. The procedure is guaranteed to attain complete
sharpness with sufficient computation time. We present results on
asymptotic validity of and conservative statistical inference for ε-sharp
bounds. The technique is illustrated using simulations using common
research designs in the study of policing.
*Zoom link: *
https://harvard.zoom.us/j/97787602526?pwd=Uzh3bVVVS0F4TEVYQTJlV3BQNjcydz09
*Schedule of the workshop:*
https://projects.iq.harvard.edu/applied.stats.workshop-gov3009
Looking forward to seeing you all on Wednesday!
Best,
Soichiro
Hi all,
Our next virtual meeting will be at 12pm (EST) Wednesday, February 3
(tomorrow), where we will hear Alex Tarr (Princeton University) presents
research on "Estimating Average Treatment Effects with Support Vector
Machines."
*Abstract*:
Support vector machine (SVM) is one of the most popular classification
algorithms in the machine learning literature. We demonstrate that SVM can
be used to balance covariates and estimate average causal effects under the
unconfoundedness assumption. We show that the SVM cost parameter controls
the trade-off between covariate balance and subset size, and as a result,
existing SVM regularization path algorithms can be used to compute the
balance-sample size frontier. We then characterize the bias of causal
effect estimation arising from this tradeoff, connecting the proposed SVM
procedure to the existing kernel balancing methods. Finally, we conduct
simulation and empirical studies to evaluate the performance of the
proposed methodology and find that SVM is competitive with the
state-of-the-art covariate balancing methods.
*Zoom link*:
https://harvard.zoom.us/j/97787602526?pwd=Uzh3bVVVS0F4TEVYQTJlV3BQNjcydz09
*Schedule of the workshop*:
https://projects.iq.harvard.edu/applied.stats.workshop-gov3009
Looking forward to seeing you all tomorrow!
Best,
Soichiro
--
Soichiro Yamauchi
PhD candidate
Harvard University
URL: https://soichiroy.github.io/