Fall 2020 complete list with abstracts
Tuesday, December 15, 2020: Luke Miratrix (Harvard)
"Using national data and meta-analysis techniques to get a handle on how bad some biases might be in practice"
Discussant: Elizabeth Tipton (Northwestern University)
Abstract: Different designs come with different risks for bias and researchers and funding agencies have limited data on the magnitude of these different biases. How much do unobserved factors bias quasi-experimental designs in education evaluations? How problematic is attrition bias in randomized experiments? Across two projects we make use of a unique administrative dataset in England, linked to a large archive of RCTs. This allows us to complete two types of “within study comparison” analyses, one to see how well we can remove selection bias, and one to see how different attriters are from nonattriters. We then use tools from meta analysis to explore the magnitude of these biases in practice. Overall we find, for our context at least, somewhat reassuring news: quasi-experimental designs do tend to produce meaningful causal estimates with biases of around 0.03σ (effect size units), and attrition also tends to not be a massive concern, averaging at around 0.02σ. That being said, both concerns are nontrivial, especially for researchers and policymakers interested in low-cost interventions with smaller effects. We therefore discuss simple ways to revise uncertainty estimates in light of these findings.
Joint work with Ben Weidmann, Harvard Graduate School of Education
[Video] [Slides]Tuesday, December 8, 2020: Qingyuan Zhao (University of Cambridge)
"Selection bias in 2020"
Discussant: Louisa Smith (Harvard University)
Abstract: This talk will examine the selection bias that occurred in studying some most contentious problems in 2020. In the first case study, we will look at the estimation of the growth rate and incubation period of COVID-19 and demonstrate how early studies drastically misestimated them. In the second case study, we will review and hopefully clarify a recent debate on post-treatment selection in studying racial discrimination in policing. In the era of data, causal inference researchers are uniquely positioned to recognize selection bias, but the cost of no or slow action is high.
[Video] [Slides] [Discussant slides]Tuesday, December 1, 2020: Vanessa Didelez (University of Bremen)
"Causal reasoning in survival and time-to-event analyses"
Discussant: Els Goetghebeur (Ghent University)
Abstract: In this talk I will discuss why causal inference should pay special attention to survival and time-to-event settings. Even in an apparently simple case of a randomized point-treatment it is common that events other than the event of interest occur, sometimes called intercurrent events such as (semi-)competing events or time-varying mediators, and of course censoring. The choice of causal estimand in such situations should anticipate these issues and suitably represent the research question. Recently, we have proposed so-called ''separable effects'', which focus on contrasts relating to different components of treatment (or exposure) that can be manipulated separately. This approach provides practically relevant estimands for various applications faced with time-varying mediation or competing events.
Mostly, however, in time-to-event settings, treatments or exposures themselves are time-dependent (start / stop / switch treatment etc.); or we may, more generally, be interested in the causal relations among various types of events or processes. This entails potential sources of time-related biases such as time-dependent confounding or self-inflicted biases such as immortal-time bias. I will discuss a class of graphical models representing dynamic relations between processes which can help with causal reasoning in time-to-event settings and shed light onto some of the issues.
Examples from the field of cancer research will be given.
The presentation will focus on basic principles and concepts rather than technical details.
[Video] [Slides] [Paper 1] [Paper 2] [Paper 3] [Paper 4]Tuesday, November 24, 2020: Fan Li (Yale)
"Propensity score weighting for covariate adjustment in randomized clinical trials"
Discussant: Kari Lock Morgan (Penn State University)
Abstract: Chance imbalance in baseline characteristics is common in randomized clinical trials. Regression adjustment such as the analysis of covariance (ANCOVA) is often used to account for imbalance and increase precision of the treatment effect estimate. An objective alternative is through inverse probability weighting (IPW) of the propensity scores. Although IPW and ANCOVA are asymptotically equivalent, the former may demonstrate inferior performance in finite samples. In this article, we point out that IPW is a special case of the general class of balancing weights, and advocate to use overlap weighting (OW) for covariate adjustment. The OW method has a unique advantage of completely removing chance imbalance when the propensity score is estimated by logistic regression. We show that the OW estimator attains the same semiparametric variance lower bound as the most efficient ANCOVA estimator and the IPW estimator for a continuous outcome, and derive closed-form variance estimators for OW when estimating additive and ratio estimands. Through extensive simulations, we demonstrate OW consistently outperforms IPW in finite samples and improves the efficiency over ANCOVA and augmented IPW when the degree of treatment effect heterogeneity is moderate or when the outcome model is incorrectly specified. We apply the proposed OW estimator to the Best Apnea Interventions for Research (BestAIR) randomized trial to evaluate the effect of continuous positive airway pressure on patient health outcomes. All the discussed propensity score weighting methods are implemented in the R package PSweight.
[Video] [Paper] [Slides] [Discussant slides]Tuesday, November 17, 2020: Interview with Judea Pearl (UCLA)
- Data versus Science: Contesting the Soul of Data-Science
- Radical Empiricism and Machine Learning Research
- What Statisticians Want to Know about Causal Inference and The Book of Why (Interview by David Hand)
- The Seven Tools of Causal Inference with Reflections on Machine Learning
[Video]Tuesday, November 10, 2020: Justin Grimmer (Stanford), Dean Knox (Wharton), Brandon Stewart (Princeton)
"Naïve regression requires weaker assumptions than factor models to adjust for multiple cause confounding"
Discussants: Ilya Shpitser (Johns Hopkins), Betsy Ogburn (Johns Hopkins), Eric Tchetgen Tchetgen (Wharton)
Abstract: The empirical practice of using factor models to adjust for shared, unobserved confounders, Z, in observational settings with multiple treatments, A, is widespread in fields including genetics, networks, medicine, and politics. Wang and Blei (2019, WB) formalizes these procedures and develops the “deconfounder,” a causal inference method using factor models of A to estimate “substitute confounders,” Ẑ, then estimating treatment effects—regressing the outcome, Y , on part of A while adjusting for Ẑ. WB claim the deconfounder is unbiased when there are no single-cause confounders and Ẑ is “pinpointed.” We clarify pinpointing requires each confounder to affect infinitely many treatments. We prove under these assumptions, a naïve semiparametric regression of Y on A is asymptotically unbiased. Deconfounder variants nesting this regression are therefore also asymptotically unbiased, but variants using Ẑ and subsets of causes require further untestable assumptions. We replicate every deconfounder analysis with available data and find it fails to consistently outperform naïve regression. In practice, the deconfounder produces implausible estimates in WB’s case study to movie earnings: estimates suggest comic author Stan Lee’s cameo appearances causally contributed $15.5 billion, most of Marvel movie revenue. We conclude neither approach is a viable substitute for careful research design in real-world applications.
[Video] [Paper] [Slides] [Discussant slides]Tuesday, November 3, 2020: Interview with Donald Rubin (Harvard)
[Video]Tuesday, October 27, 2020: David Blei (Columbia University)
"The Deconfounder: What is it? What is its theory? Is it useful?"
Discussant: Guido Imbens (Stanford)
Abstract: I will discuss the deconfounder algorithm and the assumptions it requires. Several refinements have been suggested around the theory of the deconfounder. Among these, Imai and Jiang clarified the assumption of "no unobserved single-cause confounders." Using their assumption, I will clarify the theory. Finally, I will discuss whether the deconfounder is useful in practice. This talk will largely follow Wang and Blei (2020).
[Paper] [Slides] [Discussant slides]Tuesday, October 20: Ismael Mourifie (University of Toronto)
"Testing Identification assumptions in Fuzzy Regression Discontinuity Designs"
Discussant: Zhuan Pei (Cornell)
Abstract: We propose a new specification test for assessing the validity of fuzzy regression discontinuity designs (FRD-validity). We derive a new set of testable implications, characterized by a set of inequality restrictions on the joint distribution of observed outcomes and treatment status at the cut-off. We show that this new characterization exploits all the information in the data useful for detecting violations of FRD-validity. Our approach differs from, and complements existing approaches that test continuity of the distributions of running variables and baseline covariates at the cut-off since ours focuses on the distribution of the observed outcome and treatment status. We show that the proposed test has appealing statistical properties. It controls size in large sample uniformly over a large class of distributions, is consistent against all fixed alternatives, and has non-trivial power against some local alternatives. We apply our test to evaluate the validity of two FRD designs. The test does not reject the FRD-validity in the class size design studied by Angrist and Lavy (1999) and rejects in the insurance subsidy design for poor households in Colombia studied by Miller, Pinto, and Vera-Hernández (2013) for some outcome variables, while existing density tests suggest the opposite in each of the cases.
Joint with Y. Arai, Y-C Hsu, T. Kitagawa and Y. Wan.
[Video] [Paper (supplement)] [Speaker slides] [Discussant slides]Monday, October 12: Interview with Esther Duflo
[Video]Tuesday, October 7, 2020: Peng Ding (UC Berkeley)
"Randomization and Regression Adjustment"
Discussant: Tirthankar DasGupta (Rutgers)
[Video] [Paper] [Speaker Slides] [Discussant Slides]
Abstract: Randomization is a basis for the statistical inference of treatment effects without strong assumptions on the outcome-generating process. Appropriately using covariates further yields more precise estimators in randomized experiments. R. A. Fisher suggested blocking on discrete covariates in the design stage or conducting analysis of covariance in the analysis stage.We can embed blocking in a wider class of experimental design called rerandomization, and extend the classical analysis of covariance to more general regression adjustment. Rerandomization trumps complete randomization in the design stage, and regression adjustment trumps the simple difference-in-means estimator in the analysis stage. It is then intuitive to use both rerandomization and regression adjustment. Under the randomization inference framework, we establish a unified theory allowing the designer and analyser to have access to different sets of covariates.We find that asymptotically, for any given estimator with or without regression adjustment, rerandomization never hurts either the sampling precision or the estimated precision, and, for any given design with or without rerandomization, our regression-adjusted estimator never hurts the estimated precision. Therefore, combining rerandomization and regression adjustment yields better coverage properties and thus improves statistical inference. To quantify these statements theoretically, we discuss optimal regression-adjusted estimators in terms of the sampling precision and the estimated precision, and then measure the additional gains of the designer and the analyser. We finally suggest the use of rerandomization in the design and regression adjustment in the analysis followed by the Huber–White robust standard error.Tuesday, September 29, 2020: Emilija Perkovic (University of Washington)
"Causal effects in maximally oriented partially directed acyclic graphs (MPDAGs): Identification and efficient estimation"
Discussant: Thomas Richardson (University of Washington)
[Video] [Paper 1] [Paper 2] [Speaker slides] [Discussant slides]
Abstract: We present a necessary and sufficient causal identification criterion for maximally oriented partially directed acyclic graphs (MPDAGs). MPDAGs as a class of graphs include directed acyclic graphs (DAGs), completed partially directed acyclic graphs (CPDAGs), and CPDAGs with added background knowledge. As such, they represent the type of graph that can be learned from observational data and background knowledge under the assumption of no latent variables. Our causal identification criterion can be seen as a generalization of the g-formula of Robins (1986), or the truncated factorization formula (Pearl, 2009).
In the second part of the talk, we further assume that data is generated by a linear structural causal model that can be represented by a causal DAG. We still assume we only have access to an MPDAG representing the true causal DAG. In this case, the causal identification formula leads to a simple estimator based on recursive least squares and graphically based on path analysis. We show that our estimator is the most asymptotically efficient among all regular estimators that are based on the sample covariance. This class of esimators includes covariate adjustment (Shpitser et al, 2010, Perković et al, 2017) and the estimators employed by the joint-IDA algorithm (Nandy et al, 2017). Notably, our results hold without assuming Gaussian errors.Tuesday, September 23, 2020: Falco Bargagli Stoffi (Harvard); Eli Ben-Michael (UC Berkeley)
Talk 1: "Causal Rule Ensemble: Interpretable Inference of Heterogeneous Treatment Effects" (Falco Bargagli Stoffi)
Abstract: In environmental epidemiology, it is critically important to identify subpopulations that are most vulnerable to the adverse effects of air pollution so we can develop targeted interventions. In recent years, there have been many methodological developments for addressing heterogeneity of treatment effects in causal inference. A common approach is to estimate the conditional average treatment effect (CATE) for a pre-specified covariate set. However, this approach does not provide an easy-to-interpret tool for identifying susceptible subpopulations or discover new subpopulations that are not defined a priori by the researchers. In this paper, we propose a new causal rule ensemble (CRE) method with two features simultaneously: 1) ensuring interpretability by revealing heterogeneous treatment effect structures in terms of decision rules, and 2) providing CATE estimates with high statistical precision. We provide theoretical results that guarantee consistency of the estimated causal effects for the newly discovered causal rules. Furthermore, via simulations, we show that the CRE method has competitive performance on its ability to discover subpopulations and then accurately estimate the causal effects. We also develop a new sensitivity analysis method to examine robustness to unmeasured confounding bias. Lastly, we apply the CRE method to the study of the effects of long-term exposure to air pollution on the 5-year mortality rate of the New England Medicare-enrolled population in United States.
Joint work with Kwonsang Lee (Sungkyunkwan University) and Francesca Dominici (Harvard University)
Talk 2: "Synthetic Controls with Staggered Adoption" (Eli Ben-Michael)
Staggered adoption of policies by different units at different times creates promising opportunities for observational causal inference. The synthetic control method (SCM) is a recent addition to the evaluation toolkit but is designed to study a single treated unit and does not easily accommodate staggered adoption. In this paper, we generalize SCM to the staggered adoption setting. Current practice involves fitting SCM separately for each treated unit and then averaging. We show that the average of separate SCM fits does not necessarily achieve good balance for the average of the treated units, leading to possible bias in the estimated effect. We propose "partially pooled" SCM weights that instead minimize both average and state-specific imbalance, and show that the resulting estimator controls bias under a linear factor model. We also extend our proposal to balance auxiliary covariates and to include separate intercepts for each SCM problem. We assess the performance of the proposed method via extensive simulations and apply our results to the question of whether teacher collective bargaining leads to higher school spending, finding minimal impacts. We implement the proposed method in the augsynth R package.
Joint work with Avi Feller and Jesse Rothstein.
[Video] [Bargagli Stoffi slides] [Ben-Michael slides]Tuesday, September 15, 2020: Nathan Kallus and Xiaojie Mao (Cornell University)
"Localized Debiased Machine Learning: Efficient Inference on Quantile Treatment Effects and Beyond"
Discussant: Alexandre Belloni (Duke)
Abstract: We consider the efficient estimation of a low-dimensional parameter in an estimating equation involving high-dimensional nuisances that depend on the parameter of interest. An important example is the (local) quantile treatment effect ((L)QTE) in causal inference, where the efficient estimating equation involves as a nuisance the covariate-conditional cumulative distribution function evaluated at the quantile to be estimated. Debiased machine learning (DML) is a data-splitting approach to address the need to estimate nuisances using flexible machine learning methods that may not satisfy strong metric entropy conditions, but applying it to problems with parameter-dependent nuisances is impractical. For (L)QTE estimation, DML requires we learn the whole conditional cumulative distribution function, conditioned on potentially high-dimensional covariates, which is far more challenging than the standard supervised regression task in machine learning. We instead propose localized debiased machine learning (LDML), a new data-splitting approach that avoids this burdensome step and needs only estimate the nuisances at a single initial rough guess for the parameter. For (L)QTE estimation, this involves just learning two binary regression (i.e., classification) models, for which many standard, time-tested machine learning methods exist, and the initial rough guess may be given by inverse propensity weighting. We prove that under lax rate conditions on nuisances, our estimator has the same favorable asymptotic behavior as the infeasible oracle estimator that solves the estimating equation with the unknown true nuisance functions. Thus, our proposed approach uniquely enables practically-feasible and theoretically-grounded efficient estimation of important quantities in causal inference such as (L)QTEs and in other coarsened data settings.
[Video] [Paper] [Speaker slides]Tuesday, September 8, 2020: Joris Mooij (University of Amsterdam)
"Joint Causal Inference: A Unifying Perspective on Causal Discovery"
Discussant: Philip Dawid (University of Cambridge)
Abstract: Many questions in science, policy making and everyday life are of a causal nature: how would a change of A affect B? An important research topic is therefore how cause-effect relationships can be discovered from data and how these can be used for making predictions in situations where a system has been perturbed by an external intervention. In this talk, I will introduce the basics of two, apparently quite different, approaches to causal discovery. I will discuss how both approaches can be elegantly combined in Joint Causal Inference (JCI), a novel constraint-based framework for causal discovery from multiple data sets. This perspective has inspired novel causal discovery algorithms that lead to a significant increase in the accuracy and identifiability of the predicted causal relations.
[Video] [Paper] [Speaker slides] [Discussant slides]Tuesday, September 2, 2020: Karthika Mohan (Berkeley); David Hirshberg (Stanford)
Talk 1: "Causal Graphical Models for Handling Missing Data" (Karthika Mohan)
Abstract: “Missingness Graphs” (m-graphs) are causal graphical models used for processing missing data. They portray the causal mechanisms responsible for missingness and thus encode knowledge about the underlying process that generates data. Using m-graphs, we develop methods to determine if there exists a consistent estimator for a given quantity of interest such as joint distributions, conditional distributions and causal effects. Our methods apply to all types of missing data including the notorious and relatively unexplored NMAR (Not Missing At Random) category. We further address the question of testability i.e. if and how an assumed model can be subjected to statistical tests, considering the missingness in the data. Viewing the missing data problem from a causal perspective has ushered in several surprises such as recoverability when variables are causes of their own missingness, testability of MAR models and the indispensability of causal assumptions for handling missing data problems.
Talk 2: "Balance in Causal Inference: From Poststratification to Regularized Riesz Representers" (David Hirshberg)
Abstract: In causal inference, various notions of comparability between samples are used to justify the interpretation of observed differences as causal. Collectively, these are called balance. By reweighting units, we can establish balance between otherwise incomparable samples. We will discuss this tradition, and recent generalizations that allow the estimation of a variety of causal summaries.
[Hirshberg video] [Hirshberg slides]