Summer 2020 complete list with abstracts
Tuesday, August 25, 2020: Cynthia Rudin, Alexander Volfovsky, Sudeepa Roy (Duke)
"Almost Matching Exactly"
Discussant: Guillaume Basse (Stanford)
[Video] [Almost Matching Exactly Website] [Speaker slides] [Discussant slides]
Abstract: We will present a matching framework for causal inference in the potential outcomes setting called Almost Matching Exactly. In this setting, the goal is to match treatment and control units on as many covariates as possible. We use machine learning on a hold-out training set to learn which variables are more important to match on; in essence, the method learns a distance metric for matching. This way, our methods retain the interpretability of matching, but also use distance metrics that are automated rather than being hand-designed, and are adaptive to the data rather than being fixed. The key constraint is that units are always matched on a set of covariates that together can predict the outcome well. Our techniques for discrete variables are called Fast Large-Scale Almost Matching Exactly (FLAME), Dynamic Almost Matching Exactly (DAME), Matching After Learning to Stretch (MALTS), and Adaptive Hyper-boxes. FLAME and DAME match units on a weighted Hamming distance for discrete variables using techniques that are natural for query processing in database management. FLAME rapidly produces high quality matched groups for discrete data, even for datasets that are too large to fit in memory. DAME produces higher quality matched groups than FLAME, but is slower. MALTS is useful for continuous variables and learns distance metrics that stretch the covariates in an interpretable way. Adaptive hyper-boxes can gracefully handle both discrete and continuous covariates flexibly and interpretably by optimizing a box around each treatment unit. These methods rival black box machine learning methods in their estimation accuracy but have the benefit of being interpretable and easier to troubleshoot.
This is joint work with several others in the Duke AME lab, including Tianyu Wang, Marco Morucci, Vittorio Orlandi, Harsh Parikh, Yameng Liu, Neha Gupta, Awa Dieng, Jerry Chang, and Usaid Awan.Tuesday, August 18, 2020: Judith Lok (Boston University)
"Causal organic indirect and direct effects: closer to Baron and Kenny, with a product method for binary mediators"
Discussant: Kosuke Imai (Harvard University)
Abstract: Baron and Kenny (1986, 92,251 Google Scholar citations) proposed estimators of indirect and direct effects: the part of a treatment effect that is mediated by a covariate and the part that is not. Subsequent work on natural indirect and direct effects provides a formal causal interpretation, based on cross-worlds counterfactuals: outcomes under treatment with the mediator “set” to its value without treatment. Organic indirect and direct effects (Lok 2016) avoid cross-worlds counterfactuals, using “organic” interventions on the mediator while keeping the initial treatment fixed at “treatment”. Organic indirect and direct effects apply also to settings where the mediator cannot be “set”. In linear models where the outcome model does not have treatment-mediator interaction, both organic and natural indirect and direct effects lead to the same estimators as in Baron and Kenny (1986). In this presentation, I propose organic interventions on the mediator that keep the initial treatment fixed at “no treatment”. I show that the product method, proposed in Baron and Kenny (1986), holds in linear models for these new indirect and direct effects even if there is treatment-mediator interaction. Moreover, I present a product method for binary mediators. Furthermore, I argue that this alternative organic indirect effect is more relevant for drug development. I illustrate the benefits of this new approach by estimating the organic indirect effect of curative HIV-treatments mediated by two HIV-persistence measures, using ART-interruption data without curative HIV-treatments combined with an estimated/hypothesized effect of the curative HIV-treatments on these mediators.
This is joint work with Ronald J. Bosch, Center for Biostatistics in AIDS Research, Harvard TH Chan School of Public Health.
[Video] [Paper] [Speaker slides] [Discussant slides]Tuesday, August 11, 2020: Panos Toulis (Chicago Booth)
"Randomization tests for spillovers under general interference: A graph-theoretic approach"
Discussant: Peng Ding (Berkeley)
Abstract: Interference exists when a unit’s outcome depends on another unit’s treatment assignment. For example, intensive policing on one street could have a spillover effect on neighboring streets. Classical randomization tests typically break down in this setting because many null hypotheses of interest are no longer sharp under interference. A promising alternative is to instead construct a conditional randomization test on a subset of units and assignments for which a given null hypothesis is sharp. Finding these subsets is challenging, however, and existing methods either have low power or are limited to special cases. In this paper, we propose valid, powerful, and easy to-implement randomization tests for a general class of null hypotheses under arbitrary interference between units. Our key idea is to represent the hypothesis of interest as a bipartite graph between units and assignments, and to find a biclique of this graph. Importantly, the null hypothesis is sharp for the units and assignments in this biclique, enabling randomization-based tests conditional on the biclique. We can apply off-the-shelf graph clustering methods to find such bicliques efficiently and at scale. We illustrate this approach in settings with clustered interference and show advantages over methods designed specifically for that setting. We then apply our method to a large-scale policing experiment in Medellín, Colombia, where interference has a spatial structure.
Joint work with David Puelz (U Chicago, Booth), Guillaume Basse (Stanford) and Avi Feller (UC Berkeley)
[Video] [Paper] [Speaker slides] [Discussant slides]Tuesday, August 4, 2020: Andrew Gelman (Columbia University)
"100 Stories of Causal Inference"
Abstract: In social science we learn from stories. The best stories are anomalous and immutable (see http://www.stat.columbia.edu/~gelman/research/published/storytelling.pdf). We shall briefly discuss the theory of stories, the paradoxical nature of how we learn from them, and how this relates to forward and reverse causal inference. Then we will go through some stories of applied causal inference and see what lessons we can draw from them. We hope this talk will be useful as a model for how you can better learn from own experiences as participants and consumers of causal inference.
[Video] [Paper]Tuesday, July 28, 2020: Georgia Papadogeorgou (University of Florida) and Lihua Lei (Stanford University)
Talk 1: "Causal inference with spatio-temporal data: estimating the effects of airstrikes on insurgent violence in Iraq" (Georgia Papadogeorgou)
Abstract: Many causal processes have spatial and temporal dimensions. Yet the classic causal inference framework is not directly applicable when the treatment and outcome variables are generated by spatio-temporal processes with an infinite number of possible event locations at each point in time. We take up the challenge of extending the potential outcomes framework to these settings by formulating the treatment point process as stochastic intervention. Our causal estimands include the expected number of outcome events in a specified area of interest under a particular stochastic treatment assignment strategy. We develop an estimation technique that applies the inverse probability of treatment weighting method to spatially-smoothed outcome surfaces. We demonstrate that the proposed estimator is consistent and asymptotically normal as the number of time period approaches infinity. A primary advantage of our methodology is its ability to avoid structural assumptions about spatial spillover and temporal carryover effects. We use the proposed methods to estimate the effects of American airstrikes on insurgent violence in Iraq (February 2007 – July 2008). We find that increasing the average number of daily airstrikes for up to one month increases insurgent attacks across Iraq and within Baghdad. We also find evidence that airstrikes can displace attacks from Baghdad to new locations up to 400 kilometers away.
Talk 2: "Conformal Inference of Counterfactuals and Individual Treatment Effects" (Lihua Lei)
Abstract: Evaluating treatment effect heterogeneity widely informs treatment decision making. At the moment, much emphasis is placed on the estimation of the conditional average treatment effect via flexible machine learning algorithms. While these methods enjoy some theoretical appeal in terms of consistency and convergence rates, they generally perform poorly in terms of uncertainty quantification. This is troubling since assessing risk is crucial for reliable decision-making in sensitive and uncertain environments. In this work, we propose a conformal inference-based approach that can produce reliable interval estimates for counterfactuals and individual treatment effects under the potential outcome framework. For completely randomized or stratified randomized experiments with perfect compliance, the intervals have guaranteed average coverage in finite samples regardless of the unknown data generating mechanism. For randomized experiments with ignorable compliance and general observational studies obeying the strong ignorability assumption, the intervals satisfy a doubly robust property which states the following: the average coverage is approximately controlled if either the propensity score or the conditional quantiles of potential outcomes can be estimated accurately. Numerical studies on both synthetic and real datasets empirically demonstrate that existing methods suffer from a significant coverage deficit even in simple models. In contrast, our methods achieve the desired coverage with reasonably short intervals.
[Video] [Papdogeorgou slides] [Lei slides] [Lei Paper] [Papadogeorgou Paper]Tuesday, July 21, 2020: Yiqing Xu (Stanford) and Xun Pang (Tsinghua University)
"A Bayesian Alternative to Synthetic Control for Comparative Case Studies: A Dynamic Multilevel Latent Factor Model with Hierarchical Shrinkage"
Discussant: Dmitry Arkhangelsky (CEMFI, Madrid)
Abstract: This paper proposes a Bayesian alternative to the synthetic control method (SCM) for comparative case studies based on a posterior predictive approach to Rubin's causal model. Our counterfactual imputation method generalizes the SCM by assigning observation-specific parameters to covariates of treated units and exploiting high-order relationships between treated and control time series. The model includes a dynamic latent factor term to correct biases induced by unit-specific time trends and other unobserved time-varying confounders. To reduce model dependence, we develop a Bayesian hierarchical shrinkage strategy for factor selection and model specification search. It allows researchers to make causal inferences about both individual and average treatment effects based on empirical posterior distributions of treated counterfactuals. We apply this method to simulated data and two empirical examples and show that, compared to existing approaches, our method has smaller biases, higher efficiency, and more flexibility.
[Video] [Paper] [Speaker slides] [Discussant slides]Tuesday, July 14, 2020: Michal Kolesár (Princeton)
"Finite-Sample Optimal Estimation and Inference on Average Treatment Effects Under Unconfoundedness"
Discussant: Luke Miratrix (Harvard)
Abstract: We consider estimation and inference on average treatment effects under unconfoundedness conditional on the realizations of the treatment variable and covariates. Given nonparametric smoothness and/or shape restrictions on the conditional mean of the outcome variable, we derive estimators and confidence intervals (CIs) that are optimal in finite samples when the regression errors are normal with known variance. In contrast to conventional CIs, our CIs use a larger critical value that explicitly takes into account the potential bias of the estimator. When the error distribution is unknown, feasible versions of our CIs are valid asymptotically, even when √n-inference is not possible due to lack of overlap, or low smoothness of the conditional mean. We also derive the minimum smoothness conditions on the conditional mean that are necessary for √n-inference. When the conditional mean is restricted to be Lipschitz with a large enough bound on the Lipschitz constant, the optimal estimator reduces to a matching estimator with the number of matches set to one. We illustrate our methods in an application to the National Supported Work Demonstration.
[Video] [Paper] [Speaker slides] [Discussant Slides]Tuesday, July 7, 2020: Caroline Uhler (MIT)
"Causal Inference in the Light of Drug Repurposing for SARS-CoV-2"
Abstract: Massive data collection holds the promise of a better understanding of complex phenomena and ultimately, of better decisions. An exciting opportunity in this regard stems from the growing availability of perturbation / intervention data (drugs, knockouts, overexpression, etc.) in biology. In order to obtain mechanistic insights from such data, a major challenge is the development of a framework that integrates observational and interventional data and allows predicting the effect of yet unseen interventions or transporting the effect of interventions observed in one context to another. I will present a framework for causal structure discovery based on such data and characterize the causal relationships that are identifiable in this setting. We end by demonstrating how these ideas can be applied for drug repurposing in the current SARS-CoV-2 crisis.
[Video] [Speaker slides]Tuesday, June 30, 2020: Marloes Maathuis (ETH Zürich)
"Total causal effect estimation by combining causal structure learning and covariate adjustment"
Discussant: Daniel Malinsky (Columbia)
Abstract: I will discuss a line of work that combines causal structure learning and covariate adjustment to estimate causal effects from observational data. In particular, I will discuss the IDA algorithm and some of its variations, the generalized backdoor criterion, the generalized adjustment criterion, and a graphical characterization of efficient adjustment sets. Throughout, examples will be used to illustrate the concepts.
[Video] [Paper 1] [Paper 2] [Paper 3] [Paper 4 (supplement)] [Paper 5] [Speaker slides] [Discussant slides]Tuesday, June 23, 2020: Eytan Bakshy (Facebook)
"Efficient Experimentation and Inference for Large Decision Spaces"
Discussant: Dean Eckles (MIT)
Abstract: Internet service providers routinely leverage randomized experiments to optimize products and improve decision making. I will describe recent directions in adaptive experimentation, meta-analysis, and causal inference that draw on large-scale experiments at Facebook. I will first describe the general problem of experimenting with large action spaces, and a simple solution to this problem: Bayesian optimization. I will then describe how meta-analysis can be used to improve decision quality, and may ultimately be used to further accelerate experimentation via statistical surrogacy. Finally, I will discuss the increasing role of causal inference in practical decision-making problems faced by internal organizations.
[Video] [Speaker slides] [Discussant Slides]Tuesday, June 16, 2020: Jonas Peters (University of Copenhagen)
"Causality and distribution generalization"
Discussant: Yuansi Chen (ETH Zürich)
Abstract: Purely predictive methods do not perform well when the test distribution changes too much from the training distribution. Causal models are known to be stable with respect to distributional shifts such as arbitrarily strong interventions on the covariates, but may not perform well when the test distribution differs only mildly from the training distribution. As a result, methods have been proposed that provide a trade off between causal and predictive models. We provide conditions under which such methods can be proved to generalize well to unseen distributions discuss theoretical limitations of this idea and show an example for inferring metabolic networks.
[Video] [Main paper] [Paper 1] [Paper 2] [Paper 3] [Speaker slides] [Discussant slides]Tuesday, June 9, 2020: Dean Eckles (MIT)
"Noise induced randomization in regression discontinuity designs"
Discussant: Michal Kolesár (Princeton)
Abstract: Joint work with Nikolaos Ignatiadis, Stefan Wager & Han Wu.
Regression discontinuity designs are used to estimate causal effects in settings where treatment is determined by whether an observed running variable crosses a pre-specified threshold. While the resulting sampling design is sometimes described as akin to a locally randomized experiment in a neighborhood of the threshold, standard formal analyses do not make reference to probabilistic treatment assignment and instead identify treatment effects via continuity arguments. Here we propose a new approach to identification, estimation, and inference in regression discontinuity designs that exploits measurement error in the running variable for identification. Under an assumption that the measurement error is exogenous, we show how to consistently estimate causal effects using a class of linear estimators that weight treated and control units so as to balance a latent variable of which the running variable is a noisy measure. We find this approach to facilitate identification of both familiar estimands from the literature, as well as policy-relevant estimands that correspond to the effects of realistic changes to the existing treatment assignment rule. Regression discontinuity designs with estimable measurement error arise in many settings, including medicine, education, and meta-science. We demonstrate the method with a study of retention of HIV patients and evaluate its performance using simulated data and a regression discontinuity design artificially constructed from tests scores in early childhood.
[Video] [Speaker slides]Tuesday, June 2, 2020: Paul Rosenbaum (Wharton)
"Replication and Evidence Factors in Observational Studies"
Abstract: Observational studies are often biased by failure to adjust for a covariate that was not measured. A series of studies may replicate an association because the bias that produced this association has been replicated, not because a treatment effect has been demonstrated. To be of value, a replication should remove, or reduce, or at least vary a potential source of bias that resulted in uncertainty in earlier studies. Having defined the goal of replication in this way, we may ask: Can one observational study replicate itself? Can it provide two statistically independent tests of one hypothesis about treatment effects such that the two tests are susceptible to different unobserved biases? Can the sensitivity analyses for these two tests be combined using meta-analytic techniques as if they came from unrelated studies, despite using the same data twice? When this is possible, the study is said to possess two evidence factors. The talk is divided into two parts, a brief, largely conceptual discussion of replication in observational studies, followed by a longer, more technical discussion with results about and practical examples of evidence factors.
[Video] [Handout] [Speaker slides]