Fall 2021 complete list with abstracts

• Tuesday, December 14, 2021: Ruoxuan Xiong  (Emory University)
  - Efficient Treatment Effect Estimation in Observational Studies under Heterogeneous Partial Interference
  - Discussant: Fabrizia Mealli (University of Florence)
  - Abstract: In many observational studies in social science and medical applications, subjects or individuals are connected, and one unit’s treatment and attributes may affect another unit’s treatment and outcome, violating the stable unit treatment value assumption (SUTVA) and resulting in interference. To enable feasible inference, many previous works assume the “exchangeability” of interfering units, under which the effect of interference is captured by the number or ratio of treated neighbors. However, in many applications with distinctive units, interference is heterogeneous. In this paper, we focus on the partial interference setting, and restrict units to be exchangeable conditional on observable characteristics. Under this framework, we propose generalized augmented inverse propensity weighted (AIPW) estimators for general causal estimands that include direct treatment effects and spillover effects. We show that they are consistent, asymptotically normal, semiparametric efficient, and robust to heterogeneous interference as well as model misspecifications. We also apply our method to the Add Health dataset and find that smoking behavior exhibits interference on academic outcomes.
  [Video] [Slides] [Discussant Slides]

• Tuesday, December 7, 2021: Jann Spiess (Stanford University)
  - Improving Inference from Simple Instruments through Compliance Estimation
  - Discussant: Damian Kozbur (University of Zurich)
  - Abstract: Instrumental variables (IV) regression is widely used to estimate causal treatment effects in settings where receipt of treatment is not fully random, but there exists an instrument that generates exogenous variation in treatment exposure. While IV can recover consistent treatment effect estimates, they are often noisy. Building upon earlier work in biostatistics (Joffe and Brensinger, 2003) and relating to an evolving literature in econometrics (including Abadie et al., 2019; Huntington-Klein, 2020; Borusyak and Hull, 2020), we study how to improve the efficiency of IV estimates by exploiting the predictable variation in the strength of the instrument. In the case where both the treatment and instrument are binary and the instrument is independent of baseline covariates, we study weighting each observation according to its estimated compliance (that is, its conditional probability of being affected by the instrument), which we motivate from a (constrained) solution of the first-stage prediction problem implicit to IV. The resulting estimator can leverage machine learning to estimate compliance as a function of baseline covariates. We derive the large-sample properties of a specific implementation of a weighted IV estimator in the potential outcomes and local average treatment effect (LATE) frameworks, and provide tools for inference that remain valid even when the weights are estimated nonparametrically. With both theoretical results and a simulation study, we demonstrate that compliance weighting meaningfully reduces the variance of IV estimates when first-stage heterogeneity is present, and that this improvement often outweighs any difference between the compliance-weighted and unweighted IV estimands. These results suggest that in a variety of applied settings, the precision of IV estimates can be substantially improved by incorporating compliance estimation.
  [Video] [paper] [slides] [Discussant slides]

• Tuesday, November 30, 2021: Thomas Richardson (University of Washington)
  - Single World Intervention Graphs: A simple framework for unifying graphs and potential outcomes with applications to mediation analysis
  - Discussant: Mats Stensrud (EPFL)
  - Abstract: Causal models based on potential outcomes, also known as counterfactuals,  were introduced by Neyman (1923) and later popularized by Rubin. Causal Directed Acyclic Graphs (DAGs) are another approach, originally introduced by Wright (1921), but subsequently significantly generalized and extended by Spirtes and Pearl among others.
  In this talk I will first present a simple approach to unifying these two approaches via Single-World Intervention Graphs (SWIGs). The SWIG encodes the counterfactual independences associated with a specific hypothetical intervention on a set of treatment variables. The nodes on the SWIG are the corresponding counterfactual random variables. This represents a counterfactual model originally introduced by Robins (1986) using event trees.
  This synthesis permits a simplification of Pearl's do-calculus that clarifies and separates the underlying concepts. In turn this leads to a simple counterfactual formulation of a complete identification algorithm for causal effects in models with hidden variables.
  By expanding the graph, SWIGs may also be used to describe a novel interventionist approach to mediation analysis whereby treatment is decomposed into multiple separable components. This provides a means of discussing direct effects without reference to cross-world (nested) counterfactuals or interventions on the mediator. The theory preserves the dictum ``no causation without manipulation'' and makes questions of mediation empirically testable in future randomized controlled trials.

This is joint work with James M. Robins (Harvard) and Ilya Shpitser (Johns Hopkins).

[Video] [Slides] [Discussant slides]

• Tuesday, November 16, 2021: Linbo Wang (University of Toronto)
  - Causal inference on distribution functions
  - Discussant: Hongtu Zhu (University of North Carolina at Chapel Hill)
  - Abstract: Understanding causal relationships is one of the most important goals of modern science. So far, the causal inference literature has focused almost exclusively on outcomes coming from the Euclidean space Rp. However, it is increasingly common that complex datasets collected through electronic sources, such as wearable devices, cannot be represented as data points from Rp. In this paper, we present a novel framework of causal effects for outcomes from the Wasserstein space of cumulative distribution functions, which in contrast to the Euclidean space, is non-linear. We develop doubly robust estimators and associated asymptotic theory for these causal effects. As an illustration, we use our framework to quantify the causal effect of marriage on physical activity patterns using wearable device data collected through the National Health and Nutrition Examination Survey.
  [Paper] [Video] [Slides] [Discussant slides]

• Tuesday, November 9, 2021: Jin Tian (Iowa State University)
  - Estimating Identifiable Causal Effects through Double Machine Learning - Graph-based & Data-driven Approaches
  - Discussant: Ilya  Shpitser (John Hopkins University)
  - Abstract: Inferring causal effects from observational data is a fundamental task throughout the empirical sciences. General methods have been developed to decide the identifiability of a target effect from a combination of observational data and the causal graph underlying the system. In practice, however, there are still challenges to estimating identifiable causal functionals from finite samples. We aim to fill this gap between causal identification and causal estimation. In this talk, I will discuss two versions of this problem. (1) Graph-based: For any identifiable causal functionals given a causal graph, we develop a general estimator with double/debiased machine learning (DML) properties enjoying doubly robustness against model misspecification and debiasedness against biases in nuisance function estimation permitting the use of machine learning techniques for estimating nuisances. This constitutes the first general result of identification with robustness guarantees given an arbitrary causal graph. (2) Data-driven: We study causal estimation from a Markov equivalence class (MEC) of the underlying causal graphs represented by a partial ancestral graph (PAG), which is learnable from observational data. In particular, we develop a general DML estimator for any identifiable causal effects in a PAG. The result provides an entirely data-driven solution to causal estimation, i.e., from observational data -> PAG by structure learning -> identifiability of target effect P(y|do(x)) -> estimating P(y|do(x)) from data.

Joint work with Yonghan Jung and Elias Bareinboim.
[Paper #1, #2] [Video] [Slides] [Discussant slides]

• Tuesday, November 2, 2021: Xinran Li (UIUC)
  - Randomization Inference beyond the Sharp Null: Bounded Null Hypotheses and Quantiles of Individual Treatment Effects
  - Discussant: Panos Toulis (Chicago Booth)
  - Abstract: Randomization (a.k.a. permutation) inference is typically interpreted as testing Fisher's "sharp" null hypothesis that all effects are exactly zero. This hypothesis is often criticized as uninteresting and implausible. We show, however, that many randomization tests are also valid for a "bounded" null hypothesis under which effects are all negative (or positive) for all units but otherwise heterogeneous. The bounded null is closely related to important concepts such as monotonicity and Pareto efficiency. Inverting tests of this hypothesis yields confidence intervals for the maximum (or minimum) individual treatment effect. We then extend randomization tests to infer other quantiles of individual effects, which can be used to infer the proportion of units with effects larger (or smaller) than any threshold. The proposed confidence intervals for all quantiles of individual effects are simultaneously valid, in the sense that no correction due to multiple analyses is needed. In sum, we provide a broader justification for Fisher randomization tests, and develop exact nonparametric inference for quantiles of heterogeneous individual effects. We illustrate our methods with simulations and applications, where we find that Stephenson rank statistics often provide the most informative results.
  [Video] [Slides] [Discussant slides]

• Tuesday, October 26, 2021: Carlos Cinelli (University of Washington)
  - Transparent and Robust Causal Inference in the Social and Health Sciences
  - Discussant: Guido Imbens (Stanford)
  - Abstract: The past few decades have witnessed rapid and unprecedented theoretical progress on the science of causal inference, ranging from the "credibility revolution” with the popularization of quasi-experimental designs, to the development of a complete solution to non-parametric identification with causal graphical models. Most of this theoretical progress, however, relies on strong, exact assumptions, such as the absence of unobserved common causes (ignorability assumptions), or the absence of certain direct effects (exclusion restrictions). Unfortunately, more often than not these assumptions are very hard to defend in practice. This leads to two undesirable consequences for applied quantitative work: (i) important research questions may be neglected, simply because they do not exactly match the requirements of current methods; or, (ii) researchers may succumb to making the required “identification assumptions” simply to justify the use of available methods, but not because these assumptions are truly believed (or understood).  In this talk, I will discuss new theory, methods, and software for permitting causal inferences under more flexible and realistic settings. In particular, I will focus on a flexible suite of sensitivity analysis tools for OLS (Cinelli and Hazlett, 2020) and instrumental variables (Cinelli and Hazlett, 2021) which can be immediately put to use to improve the robustness and transparency of current applied research. Notably, by building upon the familiar "omitted variable bias" framework, these tools: (i) do not require assumptions on the functional form of the treatment assignment mechanism nor on the distribution of the unobserved variables; (ii) naturally handle multiple unobserved variables, possibly acting non-linearly; (iii) exploit expert knowledge to bound sensitivity parameters; and, (iv) can be easily implemented using standard regression output. If time permits, I will also briefly discuss a graph-based approach for the algorithmic derivation of sensitivity curves for linear structural equation models (Cinelli et al, 2019).
  [Video] [Paper #1, #2, #3] [Slides] [Discussant slides]

• Tuesday, October 19, 2021: Juan Correa (Columbia University & Universidad Autónoma de Manizales) & Nicola Gnecco (University of Geneva)
  Talk1: Generalizing the Effect of Soft Interventions [Video] [Slides]
  Abstract: The challenge of generalizing causal knowledge across different environments is pervasive in scientific explorations, including in AI, ML, and Data Science. Experiments are usually performed in one environment/domain (e.g., in a lab, on Earth) with the intent, almost invariably, of being used elsewhere (e.g., outside the lab, on Mars), where the conditions are likely to be different. In the causal inference literature, this generalization task has been formalized under the rubric of transportability, for which several criteria and algorithms have been developed in the context of atomic, do-interventions. However, many real-world applications require more complex, stochastic interventions, often called "soft" interventions.
  In this work, we extend transportability theory to generalize the effect of soft interventions, which could appear both in the input and target distributions. Specifically, we develop a graphical condition that is both necessary and sufficient for deciding soft transportability. Second, we develop an algorithm to determine whether the effect of a soft intervention is computable from a combination of the distributions available across domains.
  Talk2: Causal discovery in heavy-tailed models [Video] [Slides]
  Abstract: Causal questions are omnipresent in many scientific problems. While much progress has been made in the analysis of causal relationships between random variables, these methods are not well suited if the causal mechanisms only manifest themselves in extremes. This work aims to connect the two fields of causal inference and extreme value theory. We define the causal tail coefficient that captures asymmetries in the extremal dependence of two random variables. In the population case, the causal tail coefficient is shown to reveal the causal structure if the distribution follows a linear structural causal model. This holds even in the presence of latent common causes that have the same tail index as the observed variables. Based on a consistent estimator of the causal tail coefficient, we propose a computationally highly efficient algorithm that estimates the causal structure. We prove that our method consistently recovers the causal order and we compare it to other well-established and nonextremal approaches in causal discovery on synthetic and real data. The code is available as an open-access R package.

• Tuesday, October 12, 2021: Colin Fogarty (MIT)
  - Prepivoting in Finite Population Causal Inference
  - Discussant:  Tirthanker Dasgupta (Rutgers)
  - Abstract: In finite population causal inference exact randomization tests can be constructed for sharp null hypotheses, hypotheses which fully impute the missing potential outcomes. Oftentimes inference is instead desired for the weak null that the sample average of the treatment effects takes on a particular value while leaving the subject-specific treatment effects unspecified. Without proper care, tests valid for sharp null hypotheses may be anti-conservative even asymptotically should only the weak null hold, creating the risk of misinterpretation when randomization tests are deployed in practice. We develop a general framework for unifying modes of inference for sharp and weak nulls, wherein a single procedure simultaneously delivers exact inference for sharp nulls and asymptotically valid inference for weak nulls. To do this, we employ randomization tests based upon prepivoted test statistics, wherein a test statistic is first transformed by a suitably constructed cumulative distribution function and its randomization distribution assuming the sharp null is then enumerated.  For a large class of test statistics common in practice, we show that prepivoting may be accomplished by employing a sample-based Gaussian measure governed by a suitably constructed covariance estimator. In essence, the approach enumerates the randomization distribution (assuming the sharp null) of a p-value for a large-sample test known to be valid under the weak null, and uses the resulting randomization distribution to perform inference. The versatility of the method is demonstrated through various examples, including inference for rerandomized experiments.
  [Video] [slides] [discussant slides]

• Tuesday, October 5, 2021: Eleanor Sanderson (University of Bristol)
  - Estimation of causal effects of an exposure at multiple time points through Multivariable Mendelian randomization
  - Discussant: Stephen Burgess (University of Cambridge)
  - Abstract: Mendelian Randomisation (MR) is a powerful tool in epidemiology which can be used to estimate the causal effect of an exposure on an outcome in the presence of unobserved confounding, by utilising genetic variants as instrumental variables (IVs) for the exposure. The effects obtained from MR studies are often interpreted as the lifetime effect of the exposure in question. However, the causal effects of many exposures are thought to vary throughout an individual’s lifetime and there may be periods that are more important for a particular outcome.  Multivariable MR (MVMR) is an extension of MR that allows for multiple, potentially highly related, exposures to be included in an MR estimation. We discuss the estimation and interpretation of MVMR models, including how the instrumental variable assumptions apply to models with multiple exposures. We then go on to explore the use of MVMR to estimate the effect of a single exposure at different time points in an individual’s lifetime on an outcome. We show that the direct effect of different time periods can be estimated through MVMR when the association between the genetic variants used as instruments and those time periods varies, however careful interpretation of the results obtained is important. Finally, we illustrate this approach through estimation of the causal effects of childhood and adult BMI on a range of different outcomes.  
  [Video] [Slides] [Discussant slides]

• Tuesday, September 28, 2021: Youjin Lee (Brown University)
  - Evidence factors from multiple, possibly invalid, instrumental variables
  - Discussant: Jose Zubizarreta (Harvard University)
  - Abstract: Instrumental variables have been widely used to estimate the causal effect of a treatment on an outcome in the presence of unmeasured confounders. When several instrumental variables are available and the instruments are subject to possible biases that do not completely overlap, a careful analysis based on these several instruments can produce orthogonal pieces of evidence (i.e., evidence factors) that would strengthen causal conclusions when combined. We develop several strategies, including stratification, to construct evidence factors from multiple candidate instrumental variables when invalid instruments may be present. Our proposed methods deliver nearly independent inferential results each from candidate instruments under the more liberally defined exclusion restriction than the previously proposed reinforced design.  We apply our stratification method to evaluate the causal effect of malaria on stunting among children in Western Kenya using three nested instruments that are converted from a single ordinal variable. Our proposed stratification method is particularly useful when we have an ordinal instrument of which validity depends on different values of the instrument.
  This is based on joint work with Anqi Zhao, Dylan Small, and Bikram Karmarkar.
  [Video] [Slides] [Discussant slides]

• Tuesday, September 21, 2021: Ted Westling (University of Massachusetts, Amherst)
  - Title: Nonparametric tests of the causal null with non-discrete exposures
  - Discussant: Oliver Dukes (University of Pennsylvania)
  - Abstract: Many methods have been developed to test for the presence of a causal effect of a discrete exposure on an outcome when there are no unobserved confounders. In this talk, we introduce a class of nonparametric tests of the null hypothesis that there is no average causal effect of an arbitrary univariate exposure on an outcome when there are no unobserved confounders. Our tests apply to discrete, continuous, and mixed discrete-continuous exposures. We demonstrate that our proposed tests are doubly-robust consistent, that they have correct asymptotic type I error if both nuisance parameters involved in the problem are estimated at fast enough rates, and that they have power to detect local alternatives approaching the null at the rate $n^{-1/2}$. We study the performance of our tests in numerical studies, and use them to test for the presence of a causal effect of BMI on immune response in early-phase vaccine trials.
  [Video] [Paper] [Slides] [Discussant slides]

• Tuesday, September 14, 2021: Daniel Malinsky (Columbia University)
  - Title: Explaining the Behavior of Black-Box Prediction Algorithms with Causal Learning
  - Discussant: Joshua  Loftus (LSE)
  - Abstract: We propose to explain the behavior of black-box prediction methods (e.g., deep neural networks trained on image pixel data) using causal graphical models. Specifically, we explore learning the structure of a causal graph where the nodes represent prediction outcomes along with a set of macro-level “interpretable” features, while allowing for arbitrary unmeasured confounding among these variables. The resulting graph may indicate which of the interpretable features, if any, are possible causes of the prediction outcome and which may be merely associated with prediction outcomes due to confounding. The approach is motivated by a counterfactual theory of causal explanation wherein good explanations point to factors that are “difference-makers” in an interventionist sense. The resulting analysis may be useful in algorithm auditing and evaluation, by identifying features which make a causal difference to the algorithm’s output.
  [Video] [Paper] [Slides] [Discussant slides]

• Tuesday, September 7, 2021: Joseph Antonelli (University of Florida)
  Title: Heterogeneous causal effects of neighborhood policing in New York City with staggered adoption of the policy
  Discussant: Matthew Cefalu (RAND Corporation)
  Abstract: Communities often self select into implementing a regulatory policy, and adopt the policy at different time points. In New York City, neighborhood policing was adopted at the police precinct level over the years 2015-2018, and it is of interest to both (1) evaluate the impact of the policy, and (2) understand what types of communities are most impacted by the policy, raising questions of heterogeneous treatment effects. We develop novel statistical approaches that are robust to unmeasured confounding bias to study the causal effect of policies implemented at the community level. Using techniques from high-dimensional Bayesian time-series modeling, we estimate treatment effects by predicting counterfactual values of what would have happened in the absence of neighborhood policing. We couple the posterior predictive distribution of the treatment effect with flexible modeling to identify how the impact of the policy varies across time and community characteristics. Using pre-treatment data from New York City, we show our approach produces unbiased estimates of treatment effects with valid measures of uncertainty. Lastly, we find that neighborhood policing decreases discretionary arrests, but has little effect on crime or racial disparities in arrest rates.
  [Video] [Paper] [Slides]

• Tuesday, August 31, 2021: Susan Athey and Stefan Wager (Stanford)
  Title: Estimating heterogeneous treatment effects in R
  Abstract: This tutorial will survey recent advances in machine learning based estimation of conditional average treatment effects under unconfoundedness. We will also discuss methods for validating and interpreting estimates of treatment heterogeneity. Methods will be illustrated using numerical examples in R.
  [Video] [Athey Slides] [Wager Slides]