Summer 2021 complete list with abstracts
Tuesday, August 10, 2021: Maggie Makar (University of Michigan); Xiaojie Mao (Tsinghua University)
Talk #1: Causally motivated shortcut removal using auxiliary labels (Maggie Makar)
Abstract: Robustness to certain forms of distribution shift is a key concern in many ML applications. Often, robustness can be formulated as enforcing invariances to particular interventions on the data generating process. Here, we study a flexible, causally-motivated approach to enforcing such invariances, paying special attention to shortcut learning, where a robust predictor can achieve optimal i.i.d generalization in principle, but instead it relies on spurious correlations or shortcuts in practice. Our approach uses auxiliary labels, typically available at training time, to enforce conditional independences between the latent factors that determine these labels. We show both theoretically and empirically that causally-motivated regularization schemes (a) lead to more robust estimators that generalize well under distribution shift, and (b) have better finite sample efficiency compared to usual regularization schemes, even in the absence of distribution shifts. Our analysis highlights important theoretical properties of training techniques commonly used in causal inference, fairness, and disentanglement literature.
This is based on joint work with Alex D'amour.
[Video] [Slides]
Talk #2: Controlling for Unmeasured Confounding in Panel Data Using Minimal Bridge Functions: From Two-Way Fixed Effects to Factor Models (Xiaojie Mao)
Abstract: We develop a new approach for identifying and estimating average causal effects in panel data under a linear factor model with unmeasured confounders. Compared to other methods tackling factor models such as synthetic controls and matrix completion, our method does not require the number of time periods to grow infinitely. Instead, we draw inspiration from the two-way fixed effect model as a special case of the linear factor model, where a simple difference-in-differences transformation identifies the effect. We show that analogous, albeit more complex, transformations exist in the more general linear factor model providing a new means to identify the effect in that model. In fact many such transformations exist, called bridge functions, all identifying the same causal effect estimand. This poses a unique challenge for estimation and inference, which we solve by targeting the minimal bridge function using a regularized estimation approach. We prove that our resulting average causal effect estimator is root-N consistent and asymptotically normal, and we provide asymptotically valid confidence intervals. Finally, we provide extensions for the case of a linear factor model with time-varying unobserved confounders.
This is based on a joint work with Guido Imbens and Nathan Kallus.
[Video] [Paper] [Slides]Tuesday, August 3, 2021: Anish Agarwal (MIT) and Dennis Shen (Berkeley)
Title: Synthetic Interventions
Discussant: Jason Poulos (Harvard)
Abstract: Consider a setting where there are N heterogeneous units (e.g., individuals, sub-populations) and D interventions (e.g., socio-economic policies). Our goal is to learn the potential outcome associated with every intervention on every unit (i.e., N x D causal parameters). Towards this, we present a causal framework, synthetic interventions (SI), to infer these N x D causal parameters while only observing each of the N units under at most two interventions, independent of D. This can be significant as the number of interventions, i.e, level of personalization, grows. Importantly, our estimator also allows for latent confounders that determine how interventions are assigned. Theoretically, under a novel tensor factor model across units, measurements, and interventions, we formally establish an identification result for each of these N x D causal parameters, and establish finite-sample consistency and asymptotic normality of our estimator. Empirically, we validate our framework through both experimental and observational case studies; namely, a large-scale A/B test performed on an e-commerce platform, a phase 3 clinical trial data from a pharmaceutical company, and an evaluation of mobility-restricting policies on COVID-19. We believe this has important implications for program evaluation and the design of data-efficient RCTs with heterogeneous units and multiple interventions.
[Video] [Paper] [Slides] [Discussant slides]Tuesday, July 27, 2021: Johannes Textor (Radboud University)
Title: Causal Inference using the R package DAGitty
Abstract: The R package "DAGitty" is a port of the online tool "dagitty.net" to the R platform for statistical computing. It provides access to graphical causal identification methods such as adjustment sets and instrumental variables, and has capabilities for simulating data from pre-specified graphical causal models. In this talk, I will explain the history and design principles behind the package, and show how it can be used in conjunction with other powerful packages such as ggdag, bnlearn, pcalg, and causaleffect. Special attention will be paid to the issue of testing graphical causal models via implied d-separation or tetrad implications.
[Video] [Slides]Tuesday, July 20, 2021: Fiammetta Menchetti (Universita degli Studi di Firenzi); Armeen Taeb (ETH Zürich)
Talk 1: Estimating the causal effect of an intervention in a time series setting: the C-ARIMA approach (Fiammetta Menchetti)
Abstract: The Rubin Causal Model (RCM) is a framework that allows to define the causal effect of an intervention as a contrast of potential outcomes. In recent years, several methods have been developed under the RCM to estimate causal effects in time series settings. None of these makes use of ARIMA models, which are instead very common in the econometrics literature. We propose a novel approach, C-ARIMA, to define and estimate the causal effect of an intervention in a time series setting under the RCM. We first formalize the assumptions enabling the definition, the estimation and the attribution of the effect to the intervention. In the empirical application, we use C-ARIMA to assess the causal effect of a permanent price reduction on supermarket sales.
[Video] [Slides]
Talk 2: Perturbations and causality in Gaussian latent variable models (Armeen Teb)
Abstract: With observational data alone, causal inference is a challenging problem. The task becomes easier when having access to data from perturbing the underlying system, even when the perturbations are happening in an unspecific and non-randomized way. We provide results that enable causal discovery in this setting, and also allow for the presence of latent variables. In particular, we examine a perturbation model for interventional data over a collection of Gaussian variables. Given access to data arising from perturbations, we will introduce a regularized maximum-likelihood framework that determines the class of equally representative DAGs, and uniquely identifies the underlying causal structure under sufficiently heterogeneous data. We illustrate the effectiveness of our framework on synthetic data as well as real data involving California reservoirs.
[Video] [Slides]Tuesday, July 13, 2021: Alexander Volfovsky (Duke University)
Title: Online experimentation for studying political polarization
Discussant: Edo Airoldi (Temple University)
Abstract: Social media sites are often blamed for exacerbating political polarization by creating “echo chambers” that prevent people from being exposed to information that contradicts their preexisting beliefs. We conducted a field experiment during which a large group of Democrats and Republicans followed bots that retweeted messages by elected officials and opinion leaders with opposing political views. Republican participants expressed substantially more conservative views after following a liberal Twitter bot, while Democrats’ attitudes became slightly more liberal after following a conservative Twitter bot.
To successfully study such pressing societal questions, it is imperative that experimental designs take into account new data paradigms. In modern settings where experiments are commonly run on online networks or when studying naturally networked phenomena standard randomization schemes do not exhibit the same theoretical properties. To address these issues we develop randomization schemes that are able to take into account violations of the no-interference and no-homophily assumptions. Under a scheme for studying direct treatment effects, we demonstrate the existence of unbiased estimators with bounded variance. We provide a simplified and computationally tractable randomized design which leads to asymptotically consistent estimators of direct treatment effects under both dense and sparse network regimes. We further extend this line of work by proposing experimental designs that efficiently target peer effects.
[Video] [Paper #1] [Paper #2] [Slides] [Discussant slides]Tuesday, July 6, 2021: Isaiah Andrews (Harvard University)
Title: Inference on Winners
Discussant: Will Fithian (UC Berkeley)
Abstract: Many empirical questions concern target parameters selected through optimization. For example, researchers may be interested in the effectiveness of the best policy found in a randomized trial, or the best-performing investment strategy based on historical data. Such settings give rise to a winner’s curse, where conventional estimates are biased and conventional confidence intervals are unreliable. This paper develops optimal confidence intervals and median-unbiased estimators that are valid conditional on the target selected and so overcome this winner’s curse. If one requires validity only on average over targets that might have been selected, we develop hybrid procedures that combine conditional and projection confidence intervals to offer further performance gains relative to existing alternatives.
[Video] [Paper] [Slides] [Discussant slides]Tuesday, June 29, 2021: Sam Pimentel (UC Berkeley)
Title: Optimal tradeoffs in matched designs comparing US-trained and internationally-trained surgeons.
Discussant: Magdalena Bennett (UT Austin)
Abstract: Does receiving a medical education outside the United States impact a surgeon's performance? We study this question by matching operations performed by internationally-trained surgeons to those performed by US-trained surgeons in reanalysis of a large health outcomes study. An effective matched design must achieve several goals, including balancing covariate distributions marginally, ensuring units within individual pairs have similar values on key covariates, and using a sufficiently large sample from the raw data. Yet in our study, optimizing some of these goals forces less desirable results on others. We address such tradeoffs from a multi-objective optimization perspective by creating matched designs that are Pareto optimal with respect to two goals. We provide general tools for generating representative subsets of Pareto optimal solution sets and articulate how they can be used to improve decision-making in observational study design. In the motivating surgical outcomes study, formulating a multi-objective version of the problem helps us balance an important variable without sacrificing two other design goals, average closeness of matched pairs on a multivariate distance and size of the final matched sample.
[Video] [Paper] [Slides]Tuesday, June 22, 2021: Stefan Wager (Stanford University)
Title: Treatment Effects in Market Equilibrium (joint work with Evan Munro and Kuang Xu)
Discussant: Fredrik Sävje (Yale University)
Abstract: In order to evaluate social and economic policy, it is important to measure policy effects within a market economy, where individuals interact by buying and selling various goods at the prevailing market price. In this setting, there is a direct policy effect on individual outcomes, and an indirect effect through resulting changes in equilibrium prices, which makes inference through standard randomized trials impossible. We define a stochastic general equilibrium model where interference occurs only through the equilibrium price and find that, in the mean field limit, non-parametrically defined direct and indirect treatment effects of small price changes converge to estimable elasticities with respect to prices and policy. We then show how to design a single-market experiment that enables us to estimate the total policy effect.Tuesday, June 15, 2021: Guido Imbens (Stanford University)
Title: Using Experiments to Correct for Selection in Observational Studies
Discussant: Nathan Kallus (Cornell University)
Abstract: In the social sciences there has been an increase in interest in randomized experiments to estimate causal effects, partly because their internal validity tends to be high, but they are often small and contain information on only a few variables. At the same time, as part of the big data revolution, large, detailed, and representative, administrative data sets have become more widely available. However, the credibility of estimates of causal effects based on such data sets alone can be low. In this paper, we develop statistical methods for systematically combining experimental and observational data to improve the credibility of estimates of the causal effects. We focus on a setting with a binary treatment where we are interested in the effect on a primary outcome that we only observe in the observational sample. Both the observational and experimental samples contain data about a treatment, observable individual characteristics, and a secondary (often short term) outcome. To estimate the effect of a treatment on the primary outcome, while accounting for the potential confounding in the observational sample, we propose a method that makes use of estimates of the relationship between the treatment and the secondary outcome from the experimental sample. We interpret differences in the estimated causal effects on the secondary outcome between the two samples as evidence of unobserved confounders in the observational sample, and develop control function methods for using those differences to adjust the estimates of the treatment effects on the primary outcome. We illustrate these ideas by combining data on class size and third grade test scores from the Project STAR experiment with observational data on class size and both third and eighth grade test scores from the New York school system.
[Video] [Slides] [Discussant slides]