Spring 2021 complete list with abstracts
Tuesday, June 8, 2021: Leon Bottou (Facebook)
Title: Learning Representations Using Causal Invariance
Discussant: Dominik Rothenhäusler (Stanford University)
Abstract: Learning algorithms often capture spurious correlations present in the training data distribution instead of addressing the task of interest. Such spurious correlations occur because the data collection process is subject to uncontrolled confounding biases. Suppose however that we have access to a few distinct datasets exemplifying the same concept but whose distributions exhibit different biases. Can we learn something that is common across all these distributions, while ignoring the spurious ways in which they differ? Can we go beyond setting this problem as a multi-criterion optimization problem and learn something that not only works well for our datasets, but also works all distributions that exercises the same biases in possibly more extreme form? One way to achieve this goal consists of projecting the data into a representation space in which training on any of our datasets leads to exactly the same solution. This idea differs in important ways from previous work on statistical robustness or adversarial objectives. Similar to recent work on invariant feature selection, this is about discovering the actual mechanism underlying the data instead of modeling its superficial statistics. The presentation provides some evidence that this can work and also discusses some of the current problems with this approach.
[Video]Tuesday, June 1, 2021: Niklas Pfister (University of Copenhagen)
Title: Statistical Testing under Distributional Shifts
Discussant: Thomas Berrett (University of Warwick)
Abstract: Statistical hypothesis testing is a central problem in empirical inference. Observing data from a distribution P, one is interested in testing whether P lies in a given null hypothesis while controlling the probability of false rejections. In this talk, we will introduce a framework for statistical testing under distributional shifts. Our goal will be to test a target hypothesis P in H0 using observed data from a distribution Q, where we assume that P is related to Q through a known distributional shift. We propose a general testing procedure that first resamples from the observed data to construct an auxiliary data set (mimicking properties of P) and then applies an existing test in the target domain. We prove that this procedure holds pointwise asymptotic level if the target test holds pointwise asymptotic level, the size of the resample is at most of order square root n, and the resampling weights are well-behaved. We will see that testing under distributional shifts allows us to tackle a diverse set of problems, such as problems in reinforcement learning, causal inference and conditional independence testing.
The talk is based on joint work with Nikolaj Thams, Sorawit Saengkyongam and Jonas Peters.
[Video] [Paper] [Slides]Tuesday, May 25, 2021: Razieh Nabi (Johns Hopkins University)
Title: Semiparametric inference for causal effects in graphical models with hidden variables
Discussant: Eric Tchetgen Tchetgen (University of Pennsylvania)
Abstract: There exist sound and complete algorithms for identifying causal effects in causal graphical models with unmeasured confounders. However, these algorithms remain underused due to the complexity of estimating the identifying functionals that they output. In this work, we bridge the gap between identification and estimation of population-level causal effects involving a single treatment and a single outcome. The majority of the talk focuses on semiparametric estimation of effects that are identified in a large class of causal graphical models with unmeasured confounders, where the treatment satisfies a simple graphical criterion that we call primal fixability. This class includes the adjustment and front-door functionals as special cases. We derive several types of estimators including the influence function-based doubly robust estimators. Such estimators allow us to incorporate flexible ML methods into the causal inference pipeline in a variety of settings where the standard, but often unreasonable, assumption of conditional ignorability does not hold. We also provide necessary and sufficient conditions under which the underlying statistical models are non-parametrically saturated and discuss the efficiency bounds. Finally, we provide a sound and complete identification algorithm that directly yields a weighting-based estimation strategy for any identifiable effect, even if the treatment is not primal fixable.
[Video] [Slides] [Discussant slides]Tuesday, May 18, 2021: Speaker: Ramesh Johari (Stanford University)
Title: Experimental design in two-sided platforms: an analysis of bias
Discussant: Panos Toulis (University of Chicago)
Abstract: We develop an analytical framework to study experimental design in two-sided marketplaces. Many of these experiments exhibit interference, where an intervention applied to one market participant influences the behavior of another participant. This interference leads to biased estimates of the treatment effect of the intervention. We develop a stochastic market model and associated mean field limit to capture dynamics in such experiments, and use our model to investigate how the performance of different designs and estimators is affected by marketplace interference effects. Platforms typically use two common experimental designs: demand-side ("customer") randomization (CR) and supply-side ("listing") randomization (LR), along with their associated estimators. We show that good experimental design depends on market balance: in highly demand-constrained markets, CR is unbiased, while LR is biased; conversely, in highly supply-constrained markets, LR is unbiased, while CR is biased. We also introduce and study a novel experimental design based on two-sided randomization (TSR) where both customers and listings are randomized to treatment and control. We show that appropriate choices of TSR designs can be unbiased in both extremes of market balance, while yielding relatively low bias in intermediate regimes of market balance.
[Video] [Paper] [Slides] [Discussant slides]Tuesday, May 11, 2021: Corwin Zigler (University of Texas at Austin)
Title: Bipartite inference and air pollution transport: estimating health effects of power plant interventions
Discussant: Forrest Crawford (Yale)
Abstract: Evaluating air quality interventions is confronted with the challenge of interference since interventions at a particular pollution source likely impact air quality and health at distant locations and air quality and health at any given location are likely impacted by interventions at many sources. The structure of interference in this context is dictated by complex atmospheric processes governing how pollution emitted from a particular source is transformed and transported across space, and can be cast with a bipartite structure reflecting the two distinct types of units: 1) interventional units on which treatments are applied or withheld to change pollution emissions; and 2) outcome units on which outcomes of primary interest are measured. We propose new estimands for bipartite causal inference with interference that construe two components of treatment: a "key-associated" (or "individual") treatment and an "upwind" (or "neighborhood") treatment. Estimation is carried out using a semi-parametric adjustment approach based on joint propensity scores. A reduced-complexity atmospheric model is deployed to characterize the structure of the interference network by modeling the movement of air parcels through time and space. The new methods are deployed to evaluate the effectiveness of installing flue-gas desulfurization scrubbers on 472 coal-burning power plants (the interventional units) in reducing Medicare hospitalizations among 22,603,597 Medicare beneficiaries residing across 23,675 ZIP codes in the United States (the outcome units).
[Video] [Paper] [Slides]Tuesday, May 4, 2021: Sara Magliacane (University of Amsterdam, MIT-IBM Watson AI Lab)
Title: Domain adaptation by using causal inference to predict invariant conditional distributions
Discussant: Dominik Rothenhäusler (Stanford University)
Abstract: An important goal common to domain adaptation and causal inference is to make accurate predictions when the distributions for the source (or training) domain(s) and target (or test) domain(s) differ. In many cases, these different distributions can be modeled as different contexts of a single underlying system, in which each distribution corresponds to a different perturbation of the system, or in causal terms, an intervention. We focus on a class of such causal domain adaptation problems, where data for one or more source domains are given, and the task is to predict the distribution of a certain target variable from measurements of other variables in one or more target domains. We propose an approach for solving these problems that exploits causal inference and does not rely on prior knowledge of the causal graph, the type of interventions or the intervention targets. We demonstrate our approach by evaluating a possible implementation on simulated and real world data.
[Video] [Paper] [Slides] [Discussant slides]Tuesday, April 27, 2021: Issa Dahabreh (Harvard University)
Title: Causally interpretable meta-analysis: transporting inferences from multiple randomized trials to a target population
Discussant: Eloise Kaizar (The Ohio State University)
Abstract: We present methods for causally interpretable meta-analyses that combine information from multiple randomized trials to estimate potential (counterfactual) outcome means and average treatment effects in a target population. We consider identifiability conditions, derive implications of the conditions for the law of the observed data, and obtain identification results for transporting causal inferences from a collection of independent randomized trials to a new target population in which experimental data may not be available. We propose an estimator for the potential (counterfactual) outcome mean in the target population under each treatment studied in the trials. The estimator uses covariate, treatment, and outcome data from the collection of trials, but only covariate data from the target population sample. We show that the estimator is doubly robust, in the sense that it is consistent and asymptotically normal when at least one of the models it relies on is correctly specified. We study its finite sample properties in simulation studies and demonstrate its implementation using data from a multi-center randomized trial.
Joint work with Sarah Robertson, Lucia Petito, Miguel A. Hernán, and Jon Steingrimsson.
[Slides] [Discussant slides]Tuesday, April 20, 2021: Alberto Abadie (MIT)
Title: A Penalized Synthetic Control Estimator for Disaggregated Data
Discussant: Stefan Wager (Stanford)
Abstract: Synthetic control methods are commonly applied in empirical research to estimate the effects of treatments or interventions on aggregate outcomes. A synthetic control estimator compares the outcome of a treated unit to the outcome of a weighted average of untreated units that best resembles the characteristics of the treated unit before the intervention. When disaggregated data are available, constructing separate synthetic controls for each treated unit may help avoid interpolation biases. However, the problem of finding a synthetic control that best reproduces the characteristics of a treated unit may not have a unique solution. Multiplicity of solutions is a particularly daunting challenge when the data includes many treated and untreated units. To address this challenge, we propose a synthetic control estimator that penalizes the pairwise discrepancies between the characteristics of the treated units and the characteristics of the units that contribute to their synthetic controls. The penalization parameter trades off pairwise matching discrepancies with respect to the characteristics of each unit in the synthetic control against matching discrepancies with respect to the characteristics of the synthetic control unit as a whole. We study the properties of this estimator and propose data-driven choices of the penalization parameter.
This is joint work with Jérémy L'Hour.
[Video] [Paper] [Slides]Tuesday, April 13, 2021: Andrea Rotnitzky (Di Tella University, Buenos Aires)
Title: Optimal adjustment sets in non-parametric graphical models
Discussant: Ema Perkovic (University of Washington)
Abstract: We consider the selection of potential confounding variables at the stage of the design of a planned observational study. Given a tentative non-parametric graphical causal model, possibly including unobservable variables, the aim is to select the set of observable covariates that both suffices to control for confounding under the model and yields a non-parametric estimator of the causal contrast of interest with smallest variance. For studies without unobservables aimed at assessing the effect of a static point exposure we show that graphical rules recently derived for identifying optimal covariate adjustment sets in linear causal graphical models and treatment effects estimated via ordinary least squares also apply in the non-parametric setting. Moreover, we show that, in graphs with unobservable variables, but with at least one adjustment set fully observable, there exist adjustment sets that are optimal minimal (minimum), yielding non-parametric estimators with the smallest variance among those that control for observable adjustment sets that are minimal (of minimum cardinality). In addition, although a globally optimal adjustment set among observable adjustment sets does not always exist, we provide a sufficient condition for its existence. We provide polynomial time algorithms to compute the observable globally optimal (when it exists), optimal minimal, and optimal minimum adjustment sets. For studies aimed at assessing the effects of interventions at multiple time points, static or dynamic, we derive graphical rules for comparing certain pairs of time dependent adjustment sets but we show that no global graphical rule is possible for determining optimal time dependent adjustment sets, even in graphs without unobservables. Finally, for graphs without unobservables and point interventions, we provide a sound and complete graphical criterion for determining when a non-parametric optimally adjusted estimator of the population average causal effect contrast is semiparametric efficient under the non-parametric causal graphical model. Joint work with Ezequiel Smucler and Facundo Sapienza.
[Video] [Slides] [Discussant slides]Tuesday, April 6, 2021: Richard Berk (University of Pennsylvania)
Title: Firearm Sales in California Through the Myopic Vision of an Interrupted Time Series Causal Analysis
Discussant: John Donohue (Stanford)
Abstract: There have been many claims in the media and a bit respectable research about the causes of variation in firearm sales. The challenges for causal inference can be quite daunting. In this talk, I report on an analysis of daily firearm sales in California from 1996 through most of 2018 using an interrupted time series design and analysis. The design was introduced to social scientists in 1963 by Campbell and Stanley, analysis methods were proposed by Box and Tiao in 1975, and more recent treatments are easily found (Box et al., 2016). But this approach to causal inference can be badly overmatched by the data on firearm sales, especially when the causal effects of gun control measures are estimated. For example, there can be dramatic responses to a wide variety of abrupt “shocks” to the sales data that can introduce serious and unanticipated confounding (e.g., the mass shooting in Las Vegas in 2017). Perhaps more important for this online gathering are fundamental oversights in the standard statistical methods employed. Test multiplicity problems are introduced by adaptive model selection built into recommended practice. The challenges are computational and conceptual. Some progress is made on both problems that arguably improves on past research, but the take-home message may be to reduce aspirations about what can be learned.
[Video] [Slides] [Discussant slides]Tuesday, March 30, 2021: Elizabeth Stuart (Johns Hopkins University)
Title: Using stacked comparative interrupted time series to estimate opioid policy effects
Discussant: Laura Hatfield (Harvard)
Abstract: Many opioid policies are being implemented at the state level; as one example, 37 states have passed laws limiting the dose and/or duration of opioid prescriptions. However, studying state policy effects can be challenging, especially when states that do and don’t implement the policies differ from one another, and when states implement laws across time (staggered implementation); recent work has shown that standard “two way fixed effects” analysis approaches can lead to substantial bias, and the methodological literature providing solutions to this problem is growing rapidly. In this work we take a design-based approach, called “stacked comparative interrupted time series” (CITS), which defines cohorts of states that implemented a policy at the same time, finds comparison states at that point in time, compares outcomes between the intervention and comparison states, and then “stacks” a number of these individual CITS designs on top of each other to account for the staggered implementation. Benefits of the approach include careful attention to design, the ability to examine balance during the pre-period, and a design that ensures conditioning only on pre-treatment measures. This talk will give an overview of some of the recent methodological innovations, and also discuss the practical challenges that arise when using these methods in practice. The work is particularly motivated by studies using large-scale medical insurance claims data to data to estimate the effects of opioid policies on prescribing patterns and overdoses, which raise questions around topics including variability in policy implementation and how to take advantage of the individual-level data available.
[Slides]