Dear Workshop Community,
Our next meeting will be *Wednesday October 30*, where Xun Pang will
present research on* "**A Bayesian Generalized Synthetic Control Approach
for Causal Inference With TSCS Data: A Dynamic Latent Factor Model With
Hierarchical Shrinkage**"*.
*Abstract:* This paper proposes a Bayesian synthetic control method
(a.k.a., the latent multifactor model approach) for causal inference with
observational time-series cross-sectional (TSCS) data. We develop a
state-space latent factor model and make dynamic and multilevel extensions
to the widely-applied difference-in-differences estimator, the synthetic
control approach, and latent factor models. We adopt a fully Bayesian
prior-to-posterior approach to parameter estimation and counterfactual
prediction. Compared with existing frequentist approaches, our method has
several advantages. First, it assigns unit- and time-specific weights to
outcomes and features of the donor pool to flexibly model the response
surface and exploits high-order relationships between treated and control
time series. Secondly, by combining dense modeling with latent factor
analysis and sparse modeling with Bayesian shrinkage, the method achieves a
good balance between correcting bias and avoiding overfitting. Thirdly,
based on Bayesian posterior distributions of counterfactuals, the proposed
method generates easily interpretable finite-sample inference for causal
quantities at an individual, group, or global level, which has long been a
challenge for the synthetic control method and its extensions. As a
model-based semi-parametric approach, the proposed method is highly
flexible and relax restrictive requirements on the data structure. We test
the method with Monte Carlo simulations and apply it to several empirical
studies to illustrate how to implement the method and to compare it with
some widely-applied alternative approaches. Those applications demonstrate
that the proposed method can help the researcher test causal effect
generated by complicated causal mechanisms and with substantively important
and methodologically thorny timing issues.
*Where:* CGIS Knafel Building, Room K354 (see this link
<https://map.harvard.edu/?bld=04471&level=9> for directions).
*When: *Wednesday, October 30 at 12 noon - 1:30 pm.
All are welcome. Lunch will be provided.
Best,
Georgie
Dear Workshop Community,
Our next meeting will be *Wednesday October 23*, where Xiang Zhou will
present research on* "**Effect Decomposition in the Presence of
Treatment-induced Confounding: A Regression-with-Residuals Approach"*.
*Abstract:* Analyses of causal mediation are often complicated by
treatment-induced confounders of the mediator-outcome relationship. In the
presence of such confounders, the natural direct and indirect effects of
treatment on the outcome, into which the total effect can be additively
decomposed, are not identified. An alternative but similar set of effects,
known as randomized intervention analogues to the natural direct effect
(R-NDE) and the natural indirect effect (R-NIE), can still be identified in
this situation, but existing estimators for these effects require a
complicated weighting procedure that is difficult to use in practice. We
introduce a new method for estimating the R-NDE and R-NIE that involves
only a minor adaptation of the comparatively simple regression methods used
to perform effect decomposition in the absence of treatment-induced
confounding. It involves fitting (a) a generalized linear model for the
conditional mean of the mediator given treatment and a set of baseline
confounders and (b) a linear model for the conditional mean of the outcome
given the treatment, mediator, baseline confounders, and a set of
treatment-induced confounders that have been residualized with respect to
the observed past. The R-NDE and R-NIE are simple functions of the
parameters in these models when they are correctly specified and when there
are no unobserved variables that confound the treatment-outcome,
treatment-mediator, or mediator-outcome relationships. We illustrate the
method by decomposing the effect of education on depression at midlife into
components operating through income versus alternative factors. R and Stata
packages are available for implementing the proposed method.
*Where:* CGIS Knafel Building, Room K354 (see this link
<https://map.harvard.edu/?bld=04471&level=9> for directions).
*When: *Wednesday, October 23 at 12 noon - 1:30 pm.
All are welcome. Lunch will be provided.
Best,
Georgie
Dear Workshop Community,
Our next meeting will be Wednesday October 16, where Pedro Rodriguez will present co-authored work with Arthur Spirling on "Word Embeddings: What works, what doesn’t, and how to tell the difference for applied research".
Abstract: We consider the properties and performance of word embeddings techniques in the context of political science research. In particular, we explore key parameter choices—including context window length, embedding vector dimensions and the use of pre-trained vs locally fit variants—in terms of effects on the efficiency and quality of inferences possible with these models. Reassuringly, with caveats, we show that results are robust to such choices for political corpora of various sizes and in various languages. Beyond reporting extensive technical findings, we provide a novel crowd-sourced “Turing test”-style method for examining the relative performance of any two models that produce substantive, text-based outputs. Encouragingly, we show that popular, easily available pre-trained embeddings perform at a level close to---or surpassing---both human coders and more complicated locally-fit models. For completeness, we provide best practice advice for cases where local fitting is required.
Where: CGIS Knafel Building, Room K354 (see this link <https://map.harvard.edu/?bld=04471&level=9> for directions).
When: Wednesday, October 16 at 12 noon - 1:30 pm.
All are welcome. Lunch will be provided.
Best,
Georgie
Dear Workshop Community,
Our next meeting will be Wednesday October 9, where Max Goplerud will present his job market paper "Modelling Heterogeneity Using Bayesian Structured Sparsity”.
Abstract: How to estimate heterogeneity, e.g. the effect of some variable differing across observations, is a key question in political science. Methods for doing so make simplifying assumptions about the underlying nature of the heterogeneity to draw reliable inferences. This paper allows a common way of simplifying complex phenomenon (placing observations with similar effects into discrete groups) to be integrated into regression analysis. The framework allows researchers to (i) use their prior knowledge to guide which groups are permissible and (ii) appropriately quantify uncertainty. The paper does this by translating work on "structured sparsity" from a penalized likelihood approach into a Bayesian prior and deriving theoretical results on posterior propriety and inference. It shows that this method outperforms state-of-the-art methods for estimating heterogeneous effects when the underlying heterogeneity is grouped and more effectively identifies groups of observations with different effects in observational data. A link to the paper can be found at j.mp/goplerud_sparsity <http://j.mp/goplerud_sparsity>.
Where: CGIS Knafel Building, Room K354 (see this link <https://map.harvard.edu/?bld=04471&level=9> for directions).
When: Wednesday, October 9 at 12 noon - 1:30 pm.
All are welcome! Lunch will be provided.
Best,
Georgie