Fall 2023 complete list with abstracts

• Tuesday, September 19, 2023: Andrew Gelman (Columbia University)
  - Discussants: Elizabeth Tipton (Northwestern), Avi Feller (Berkeley),  Jonathan Roth (Brown), Pedro Sant'Anna (Emory)
  - Title: Better Than Difference in Differences
  - Abstract: It is not always clear how to adjust for control data in causal inference, balancing the goals of reducing bias and variance. We show how, in a setting with repeated experiments, Bayesian hierarchical modeling yields an adaptive procedure that uses the data to determine how much adjustment to perform. The result is a novel analysis with increased statistical efficiency compared with the default analysis based on difference estimates. The increased efficiency can have real-world consequences in terms of the conclusions that can be drawn from the experiments. An open question is how to apply these ideas in the context of a single experiment or observational study, in which case the optimal adjustment cannot be estimated from the data; still, the principle holds that difference-in-differences can be extremely wasteful of data.
  The talk follows up on Andrew Gelman and Matthijs Vákár (2021), Slamming the sham: A Bayesian model for adaptive adjustment with noisy control data
  [Video] [Paper]

• Tuesday, September 26, 2023: Michael Hudgens and Chanhwa Lee (University of North Carolina at Chapel Hill)
  - Discussants: Hyunseung Kang (UW-Madison), Chris Harshaw (MIT, Berkeley)
  - Title: Efficient Nonparametric Estimation of Stochastic Policy Effects with Clustered Interference
  - Abstract: Interference occurs when a unit's treatment (or exposure) affects another unit's outcome. In some settings, units may be grouped into clusters such that it is reasonable to assume that interference, if present, only occurs between individuals in the same cluster, i.e., there is clustered interference. Various causal estimands have been proposed to quantify treatment effects under clustered interference from observational data, but these estimands either entail treatment policies lacking real-world relevance or are based on parametric propensity score models. Here, we propose new causal estimands based on modification of the propensity score distribution which may be more relevant in many contexts and are not based on parametric models. Nonparametric sample splitting estimators of the new estimands are constructed, which allow for flexible data-adaptive estimation of nuisance functions and are consistent, asymptotically normal, and efficient, converging at the usual parametric rate. Simulations show the finite sample performance of the proposed estimators. The proposed methods are applied to evaluate the effect of water, sanitation, and hygiene facilities on diarrhea among children in Senegal.
  [Video] [Slides] [Paper]

• Tuesday, October 3, 2023: Caleb Miles (Columbia University)
  - Discussant [standard form]: James Robins (Harvard University) and Thomas Richardson (University of Washington)
  - Title: Two fundamental problems in causal mediation analysis
  - Abstract: Scientists are often interested in understanding mediating mechanisms that can help explain causal effects. A vast body of literature on mediation analysis has accumulated since two foundational articles (Robins and Greenland, 1992; Pearl, 2001) formalized mediation in the language of causality. However, causal mediation analysis poses many fundamental, interesting, and unresolved difficulties in its causal interpretation, nonparametric identification assumptions (which are much stronger than most typical causal assumptions), and statistical inference. I will discuss some of my work in two of these categories: (i) the interpretation of alternative estimands that have been proposed to circumvent the standard identification assumptions, and (ii) the implicit assumptions embedded in the standard mediation identification assumptions and their shortcomings in handling the inherent time-varying nature of mediators.
  [Video] [Slides] [Discussant slides]

• Tuesday, October 10, 2023: Ruoqi Yu (UIUC) Q&A moderator: Peng Ding (UC Berkeley)
  - Discussant: José Zubizarreta (Harvard) and Luke Keele (UPenn) [new format]
  - Title: How to learn more from observational factorial studies
  - Abstract: Many scientific questions in biomedical research, environmental sciences, and psychology involve understanding the impact of multiple factors on an outcome of interest. Randomized factorial experiments are a popular tool for evaluating the causal effects of multiple factors and their interactions simultaneously. However, randomization is often infeasible in many empirical studies, and drawing reliable causal inferences for multiple factors in observational studies remains challenging. As the number of treatment combinations grows exponentially with the number of factors, some treatment combinations can be rare or even missing in the data due to chance, posing additional difficulties in factorial effects estimation. To address these challenges, we propose a novel weighting approach tailored for observational studies with multiple factors. Our approach uses weighted observational data to approximate a randomized factorial experiment, enabling us to estimate the effects of multiple factors and their interactions simultaneously using the same set of weights. Our investigations reveal a crucial nuance that achieving balance among observed covariates, as conventionally practiced with single-factor scenarios, is necessary but insufficient for unbiased estimation of the factorial effects of interest. Notably, our findings suggest that balancing the treatments for each contrast is also essential in multi-factor settings. Moreover, we discuss how to extend the proposed weighting method when some treatment combinations are missing in the data. Finally, we study the asymptotic behavior of the new weighting estimators and propose a consistent variance estimator, allowing researchers to conduct inferences for the factorial effects in observational studies. Our approach is applicable to various observational studies, providing a valuable tool for investigators interested in estimating the causal effects of multiple factors.
  [Video]

• Tuesday, October 17, 2023: Ricardo Silva (University College London)
  - Discussant: Anish Agarwal (Columbia University)
  - Title: Intervention Generalization: A View from Factor Graph Models
  - Abstract: One of the goals of causal inference is to generalize from past experiments and observational data to novel conditions. While it is in principle possible to eventually learn a mapping from a novel experimental condition to an outcome of interest, provided a sufficient variety of experiments is available in the training data, coping with a large combinatorial space of possible interventions is hard. Under a typical sparse experimental design, this mapping is ill-posed without relying on heavy regularization or prior distributions. Such assumptions may or may not be reliable, and can be hard to defend or test. In this paper, we take a close look at how to warrant a leap from past experiments to novel conditions based on minimal assumptions about the factorization of the distribution of the manipulated system, communicated in the well-understood language of factor graph models. A postulated interventional factor model (IFM) may not always be informative, but it conveniently abstracts away a need for explicit unmeasured confounding and feedback mechanisms, leading to directly testable claims. We derive necessary and sufficient conditions for causal effect identifiability with IFMs using data from a collection of experimental settings, and implement practical algorithms for generalizing expected outcomes to novel conditions never observed in the data. This is joint work with Gecia Bravo-Hermsdorff, David S. Watson, Jialin Yu and Jakob Zeitler.
  [Video] [Slides] [Discussant slides] [Paper]

• Tuesday, October 24, 2023 (Young researcher seminar)
  - Speaker1: Michael Celentano (UC Berkeley)
  - Title: Challenges of the inconsistency regime: Novel debiasing methods for missing data models
  - Abstract: In this talk, I will discuss semi-parametric estimation of the population mean when data is observed missing at random (MAR) in the  "inconsistency regime," in which neither the outcome model nor the propensity/missingness model can be estimated consistently. I focus on a high-dimensional linear-GLM specification in which the number of confounders is proportional to the sample size. In the case n > p, past work has developed theory for the classical AIPW estimator in this model and established its variance inflation and asymptotic normality when the outcome model is fit by ordinary least squares. Ordinary least squares is no longer feasible in the case studied here, and I will demonstrate that a number of classical debiasing procedures become inconsistent. This challenge motivates the development and analysis of a novel procedure: we establish that it is consistent for the population mean under proportional asymptotics allowing for n < p. Providing such guarantees in the inconsistency regime requires a new debiasing approach that combines penalized M-estimates of both the outcome and propensity/missingness models in a non-standard way.
  [Video] [Slides]
  - Speaker 2: Chris Harshaw (UC Berkeley + MIT)
  - Title: Clip-OGD: An Experimental Design for Adaptive Neyman Allocation in Sequential Experiments
  - Abstract: From clinical trials and public health to development economics and political science, randomized experiments stand out as one of the most reliable methodological tools, as they require the fewest assumptions to estimate causal effects. Adaptive experiment designs – where experimental subjects arrive sequentially and the probability of treatment assignment can depend on previously observed outcomes – are becoming an increasingly popular method for causal inference, as they offer the possibility of improved precision over their non-adaptive counterparts. However, in simple settings (e.g. two treatments) the extent to which adaptive designs can improve precision is not sufficiently well understood. In this talk, I present my recent work on the problem of Adaptive Neyman Allocation, where the experimenter seeks to construct an adaptive design which is nearly as efficient as the optimal (but infeasible) non-adaptive Neyman design which has access to all potential outcomes. I will show that the experimental design problem is equivalent to an adversarial online convex optimization problem, suggesting that any solution must exhibit some amount of algorithmic sophistication. Next, I present Clip-OGD, an experimental design that combines the online gradient descent principle with a new time-varying probability-clipping technique. I will show that the Neyman variance is attained in large samples by showing that the expected regret of the online optimization problem is bounded by O(\sqrt{T}), up to sub-polynomial factors. Even though the design is adaptive, we construct a consistent (conservative) estimator for the variance, which facilitates the development of valid confidence intervals. Finally, we demonstrate the method on data collected from a micro-economic experiment. Joint work with Jessica Dai and Paula Gradu.
  [Video] [Slides] [Paper]

• Tuesday, October 31, 2023: Richard Guo (University of Cambridge)
  - Discussant: Ilya Shpitser (Johns Hopkins University)
  - Title: Confounder selection via iterative graph expansion
  - Abstract: Confounder selection, namely choosing a set of covariates to control for confounding between a treatment and an outcome, is arguably the most important step in the design of observational studies. Previous methods, such as Pearl's celebrated back-door criterion, typically require pre-specifying a causal graph, which can often be difficult in practice. We propose an interactive procedure for confounder selection that does not require pre-specifying the graph or the set of observed variables. This procedure iteratively expands the causal graph by finding what we call "primary adjustment sets" for a pair of possibly confounded variables. This can be viewed as inverting a sequence of latent projections of the underlying causal graph. Structural information in the form of primary adjustment sets is elicited from the user, bit by bit, until either a set of covariates are found to control for confounding or it can be determined that no such set exists. We show that if the user correctly specifies the primary adjustment sets in every step, our procedure is both sound and complete.
  [Video] [Slides] [Discussant slides] [Paper]

• Tuesday, November 7, 2023: Anish Agarwal (Columbia University) 
  - Discussant: Iavor Bojinov (Harvard University) and Ashesh Rambachan (MIT)
  - Title: On Causal Inference with Temporal and Spatial Spillovers in Panel Data
  - Abstract: Panel data is a ubiquitous setting where one collects multiple measurements over time of a collection of heterogeneous units (e.g., individuals, firms, geographic entities). Two pervasive challenges in doing causal inference in such settings are: (1) temporal spillovers - interventions in the past affect current outcomes; (2) spatial spillovers - interventions a unit receives affects the outcomes of its neighboring units. We develop a causal framework to tackle these two challenges. We do so by proposing a novel latent factor model that allows one to share information across units, measurements, and interventions, and incorporates both types of spillovers. The model subsumes linear time-varying dynamical systems and autoregressive processes as a special case, and previously studied models for spatial spillovers. We show how a variety of causal parameters are identified and estimated despite unobserved confounding.
  [Video] [Slides] [Paper #1, #2]

• Tuesday, November 14, 2023: Maya Mathur (Stanford University)
  - Discussant: Eric Tchetgen Tchetgen (University of Pennsylvania) and Nan Laird (Harvard University) [new format]
  - Title: A common-cause principle for eliminating selection bias in causal estimands through covariate adjustment
  - Abstract: Average treatment effects (ATEs) may be subject to selection bias when they are estimated among only a non-representative subset of the target population. Selection bias can sometimes be eliminated by conditioning on a “sufficient adjustment set” of covariates, even for some forms of missingness not at random (MNAR). Without requiring full specification of the causal structure, we consider sufficient adjustment sets to allow nonparametric identification of conditional ATEs in the target population. Covariates in the sufficient set may be collected among only the selected sample. We establish that if a sufficient set exists, then the set consisting of common causes of the outcome and selection, excluding the exposure and its descendants, also suffices. We establish simple graphical criteria for when a sufficient set will not exist, which could help indicate whether this is plausible for a given study. Simulations considering selection due to missing data indicated that sufficiently-adjusted complete-case analysis (CCA) can considerably outperform multiple imputation under MNAR and, if the sample size is not large, sometimes even under missingness at random. Analogous to the common-cause principle for confounding, these sufficiency results clarify when and how selection bias can be eliminated through covariate adjustment.
  [Video] [Slides] [Paper]

• Tuesday, November 28, 2023: Yuqi Gu (Columbia University)
  - Discussant: Qingyuan Zhao (University of Cambridge)
  - Title: Identifiable Deep Generative Models for Rich Data Types with Discrete Latent Layers
  - Abstract: We propose a class of identifiable deep generative models for very flexible data types. The key features of the proposed models include (a) discrete latent layers and (b) a shrinking pyramid- or ladder-shaped deep architecture. We establish model identifiability by developing transparent conditions on the sparsity structure of the deep generative graph. The proposed identifiability conditions can ensure estimation consistency in both the Bayesian and frequentist senses. As an illustration, we consider the two-latent-layer model and propose shrinkage estimation methods to recover the latent structure and model parameters. Simulation results corroborate the identifiability theory and also demonstrates the nice empirical performance of our estimation algorithm. Applications of the methodology to a DNA nucleotide sequence dataset and an educational test response time dataset both give interpretable results. The proposed framework provides a recipe for identifiable, interpretable, and reliable deep generative modeling. In the future, it would be interesting to explore the implications on causal structure discovery and causal representation learning of the proposed identifiability results.
  [Video] [Slides] 

• Tuesday, December 5, 2023: Erica Moodie (McGill University)
  - Discussant: Yu Cheng (University of Pittsburgh), Peter Thall (MD Anderson)
  - Title: Flexible modeling of adaptive treatment strategies for censored outcomes
  - Abstract: To achieve the goal of providing the optimal care to each patients, physicians must customize treatments. Making decisions at multiple stages as a disease progresses can be formalized as an adaptive treatment strategy (ATS). To be able to recommend an optimal treatment, an understanding of the causal effect of treatment is required. In this talk, I will discuss an extension of a general Bayesian machine learning framework for the popular estimation approach of Q-learning, adapted to censored outcomes using Bayesian additive regression trees (BART) for each stage under an accelerated failure time modeling framework. The developments are motivated by and applied to an analysis aimed at estimating optimal immunosuppressant treatment strategies to maximize the disease-free survival time in a cohort of patients who underwent allogeneic hematopoietic cell transplant to treat myeloid leukemia. Joint work with Xiao Li, Brent R Logan, and S M Ferdous Hossain.
  [Video] [Slides]

• Tuesday, December 12, 2023: Mats Stensrud (EPFL) and Aaron Sarvet (EPFL)
  - Discussant: Kerollos Wanis (Western University) and Vanessa Didelez (BIPS) . Q&A moderators: Lan Wen (U Waterloo)
  - Title: Interpretational errors in causal inference and how to avoid them
  - Abstract: Pioneering works in causal inference were explicitly grounded in practical disciplines, aiming at formalizing real questions with mathematical definitions. Now, causal inference methods provide an architecture that profoundly regulates what practical questions get asked and how they get answered. Here we consider subtly different approaches for causal inference research, and their implications for theory development and practice. In this process, we formalize an interpretational error that is increasingly apparent in the causal literature, which we coin "identity slippage".  This formalization can be used for error detection whenever policy decisions depend on the accurate interpretation of statistical results, which is nearly always the case. Therefore, broad awareness of identity slippage will aid in the successful translation of data into public good. As an illustration, we describe how this error appears in the methodological literature on mediation analysis, and how the error has propagated to applied disciplines.
  [Video] [Slides] [Paper]