Spring 2025 complete list with abstracts

See our homepage for future talks.

• Tuesday, June 17, 2025: Julie Josse (PreMeDICaL Inria-Inserm & University of Montpellier)
  - Title: Causal Alternatives to Meta-Analysis
  - Discussant: Larry Han (Northeastern University)
  - Abstract: Meta-analysis, by aggregating effect estimates from multiple studies conducted in diverse settings, stands at the top of the evidence-based medicine hierarchy. However, conventional approaches face key limitations: they often lack a clear causal interpretation, and practical constraints, such as data silos and regulatory barriers, frequently prevent access to individual-level data, limiting the feasibility of individual patient data meta-analyses.
  In this talk, we introduce three causal alternatives to classical meta-analysis. First, we present novel causal aggregation formulas—distinct from fixed- and random-effects models—that are compatible with standard meta-analysis practices and do not require individual-level data but only aggregated data. While these often align with traditional conclusions, they can expose important discrepancies, including cases where conventional methods suggest benefit when a causal approach reveals harm.
  Second, we connect federated learning with causal inference to enable treatment effect estimation from decentralized data sources. Finally, we discuss generalization methods aimed at predicting treatment effects in new target populations. Across these approaches, we focus on both absolute and relative measures, such as risk ratios and odds ratios, which pose unique challenges due to their non-linearity.
  [Slides][Papers: #1, #2, #3, #4][Video][Discussant slides]

• Tuesday, June 10, 2025: Matias Cattaneo (Princeton University)
  - Title: Estimation and Inference in Boundary Discontinuity Designs
  - Discussants: Alberto Abadie (MIT) and Kosuke Imai (Harvard University)
  - Abstract: Boundary Discontinuity Designs are used to learn about treatment effects along a continuous boundary that splits units into control and treatment groups according to a bivariate score variable. These research designs are also called Multi-Score Regression Discontinuity Designs, a leading special case being Geographic Regression Discontinuity Designs. We study the statistical properties of commonly used local polynomial treatment effects estimators along the continuous treatment assignment boundary. We consider two distinct approaches: one based explicitly on the bivariate score variable for each unit, and the other based on their univariate distance to the boundary. For each approach, we present pointwise and uniform estimation and inference methods for the treatment effect function over the assignment boundary. Notably, we show that methods based on univariate distance to the boundary exhibit an irreducible large misspecification bias when the assignment boundary has kinks or other irregularities, making the distance-based approach unsuitable for empirical work in those settings. In contrast, methods based on the bivariate score variable do not suffer from that drawback. We illustrate our methods with an empirical application. Companion general-purpose software is provided.
  [Slides][Paper][Video][Discussant 1 Slides][Discussant 2 Slides]

• Tuesday, June 03, 2025: Edward Kennedy (Carnegie Mellon University)
  - Title: Causal inference with high-dimensional treatments
  -  Discussant: Iván Díaz (New York University)
  - Abstract: In this work we consider causal inference when the number of treatment levels is comparable to or larger than the number of observations. This setting brings two unique challenges: first, the treatment effects of interest are a high-dimensional vector rather than a low-dimensional scalar and, second, positivity violations are unavoidable. We first discuss fundamental limits of estimating effects in such settings, showing consistent estimation is impossible without further assumptions. We go on to propose new doubly robust thresholding/penalized estimators, and analyze them under exact and approximate sparsity assumptions. We then derive minimax lower bounds to characterize optimal rates of convergence. Finally we illustrate the methods in an education application studying school effects on test scores, where the number of treatments (schools) is in the thousands. This is joint work with Patrick Kramer (CMU) and Isaac Opper (RAND).

• Tuesday, May 27, 2025: Francesco Locatello (Institute of Science and Technology Austria)
  - Title: Powering causality with ML: Discovery, Representations, and Inference
  - Discussant: Jason Hartford (University of Manchester & Valence Labs)
  - Abstract: Machine learning and AI have the potential to transform data-driven scientific discovery, enabling not only accurate predictions for several scientific phenomena but also accelerating causal understanding. In this talk, I will show how machine learning has the potential to power causal analysis across 3 pillars: discovery, representations, and inference. I will discuss first how causal structure can be discovered using score matching approaches. Next, I will turn to a new interpretation of causal representation learning, introducing the measurement model as a way to conceptualize causal representations and the invariance principle as a unifying methodology. Finally, I will focus on powering causal inferences from high-dimensional observations. Using our real-world ISTAnt benchmark in experimental ecology as a motivating example, I will discuss challenges from selection bias during transfer learning and encoder bias due to pre-training and generalization across populations (e.g., multiple experiments with different treatments).
  [Slides][Video][Discussion slides]

• Tuesday, May 20, 2025: Luke Keele  (University of Pennsylvania)
  - Title: Clustered Observational Studies: A Review of Concepts and Methods
  - Discussant: Eli Ben-Michael (Carnegie Mellon University)
  - Abstract: The clustered observational study (COS) design is the observational counterpart to the clustered randomized trial. In a COS, a treatment is assigned to intact groups, and all units within the group are exposed to the treatment. However, the treatment is non-randomly assigned. COSs are common in both education and health services research. In education, treatments may be given to all students within some schools but withheld from all students in other schools. In health studies, treatments may be applied to clusters such as hospitals or groups of patients treated by the same physician. In this presentation, I provide a review of research on the COS design.  I start with identification issues for COS designs.  Specifically, I outline how this design is subject to differential selection, which occurs when the units' cluster selections depend on the clusters' treatment assignments. This can occur when units switch clusters in response to how the cluster was assigned to treatment. Early work on COSs made an implicit assumption that rules out the presence of differential selection. I review how causal effects can be identified in the presence of differential selection and how to reason about when it is likely.  Next, I review how COS designs are often subject to common support violations. Such violations are likely to occur in a COS, especially with a small number of treated clusters. I outline how trimming to ensure common support holds can result in extremely narrow causal estimands that are unlikely to generalize to other contexts. I then review methods for covariate adjustment in COS design. Extant methods for covariate adjustment include specialized forms of matching and weighting. I review these different methods and outline current research that compares the performance of these different methods. I conclude with more recent work that has developed key results on the nature of omitted confounders. Throughout I motivate concepts and methods with a number of COS applications from education and health services research.
  [Slides][Video][Discussant slides]

• Tuesday, May 13, 2025: Sam Pimentel (UC Berkeley)
  - Title: Design Sensitivity and Its Implications for Weighted Observational Studies
  - Discussant: Jacob Dorn (University of Pennsylvania)
  - Abstract: Sensitivity to unmeasured confounding is not typically a primary consideration in designing weighted treated-control comparisons in observational studies. We introduce a framework allowing researchers to optimize robustness to omitted variable bias at the design stage using a measure called design sensitivity. Design sensitivity, which describes the asymptotic power of a sensitivity analysis, allows transparent assessment of the impact of different weighted estimation strategies on sensitivity. We apply this general framework to two commonly-used sensitivity models, the marginal sensitivity model and the variance-based sensitivity model. By comparing design sensitivities, we interrogate how key features of weighted designs --- including definition of treatment, inclusion criteria, and handling of extreme weights --- impact robustness to unmeasured confounding, and how these impacts may differ for the two different sensitivity models. We illustrate the proposed framework on a study examining drivers of support for the 2016 Colombian peace agreement.
  [Paper][Slides][Video][Discussant slides]

• Tuesday, May 06, 2025: Zijian Guo (Rutgers University)
  - Title: Multi-Source Learning with Minimax Optimization: From Adversarial Robustness to Causal Invariance
  - Discussant: Kaizheng Wang (Columbia University)
  - Abstract: Empirical risk minimization often fails to deliver reliable predictions when the target distribution deviates from source populations. This talk explores leveraging multi-source data to build models that generalize and transfer effectively to target domains. We propose a distributionally robust optimization framework that maximizes worst-case performance over a set of potential target distributions. We first introduce the definition of a distributionally robust prediction model and show that it can be expressed as a weighted average of the conditional outcome models from the source populations.
  We then extend this framework to the transfer learning regime by incorporating a small number of labeled outcomes from the target domain to guide the robust optimization. This approach achieves faster convergence rates than models trained solely on labeled target data, while mitigating negative transfer.
  Finally, we connect causal invariance learning to distributionally robust learning with negative weights. We show that allowing negative weights over environment-specific risks enables the identification of invariant causal models, despite introducing nonconvexity. We establish a necessary and sufficient condition for identifying the causal invariance model under additive interventions. Under this causal identification condition, we show that standard gradient-based methods can efficiently solve the resulting nonconvex optimization problem, with provable convergence rates depending on the sample size and number of iterations.
  [Slides][Video]

• Tuesday, Apr 29, 2025: Angela Zhou  (USC Marshall)
  - Title: Robust Fitted-Q-Evaluation and Iteration under Sequentially Exogenous Unobserved Confounders
  - Discussant: Qingyuan Zhao (University of Cambridge)
  - Abstract: Offline reinforcement learning is important in domains such as medicine, economics, and e-commerce where online experimentation is costly, dangerous or unethical, and where the true model is unknown. However, most methods assume all covariates used in the behavior policy’s action decisions are observed. Though this assumption, sequential ignorability/unconfoundedness, likely does not hold in observational data, most of the data that accounts for selection into treat- ment may be observed, motivating sensitivity analysis. We study robust policy evaluation and policy optimization in the presence of sequentially-exogenous unobserved confounders under a sensitivity model. We propose and analyze orthogonalized robust fitted-Q-iteration that uses closed-form solutions of the robust Bellman operator to derive a loss minimization problem for the robust Q function, and adds a bias-correction to quantile estimation. Our algorithm enjoys the computational ease of fitted-Q-iteration and statistical improvements (reduced dependence on quantile estimation error) from orthogonalization. We provide sample complexity bounds, insights, and show effectiveness both in simulations and on real-world longitudinal healthcare data of treating sepsis. In particular, our model of sequential unobserved confounders yields an online Markov decision process, rather than partially observed Markov decision process: we illustrate how this can enable warm-starting optimistic reinforcement learning algorithms with valid robust bounds from observational data.
  [Paper][Slides][Video][Discussant slides]

• Tuesday, Apr 22, 2025: Young researchers' seminar
  Speaker 1:  Drago Plečko (Columbia University)
  - Title: Monotonicity in graphical causal models: an algorithmic approach
  - Abstract: A longstanding debate between the proponents of the potential outcomes (PO) and the graphical models (GM) approach to causal inference concerns the identification of causal effects under shape constraints (such as monotonicity). Scholars in the PO framework have developed seminal results leveraging monotonicity constraints in practical applications, such as in identification of local average treatment effects (LATE). Various assertions have been made in the PO literature that the graphical approach may be inherently limited for encoding shape constraints. At the same time, the GM approach has successfully provided complete algorithms for non-parametric identification of causal effects. Some special cases, such as linear or additive-noise models, are also well understood within this approach. However, incorporating shape constraints such as monotonicity, has so far eluded its scope. In this talk, we discuss a recent advance which allows one to encode monotonicity constraints in a graphical model, and leverage monotonicity for identification.
  [Slides][Video]
  Speaker 2: Daiqi Gao (Harvard University)
  - Title: Harnessing Causality in Online Reinforcement Learning for Digital Interventions
  - Abstract: Mobile health leverages mobile devices to monitor and manage patient health through digital interventions. To improve the efficacy of interventions given user heterogeneity, we use online reinforcement learning (RL) to learn personalized intervention policies. Online learning in mobile health faces challenges due to noisy user behavior and limited user-specific data. To improve sample efficiency, we embed established behavioral theories via an expert-provided causal directed acyclic graph (DAG), which forms the assumptions in our RL algorithms. The causal DAG guides both state construction and reward design. Specifically, we consider bagged decision times, where each bag represents a finite sequence of consecutive decisions, and the transition dynamics are non-Markovian and non-stationary within the bag. Leveraging the conditional independencies implied by the DAG, we construct an optimal state that ensures Markov transitions within and across bags, maximizes the optimal value function across all state constructions, and minimizes variance in learning. When the DAG informs mediators between actions and rewards, we learn a surrogate reward online using these mediators and demonstrate that optimizing the surrogate reward improves the regret bound.
  [Slides][Video]

• Tuesday, Apr 15, 2025 [SCI+OCIS]: Gabriel Loewinger (NIH), Emre Kiciman (Microsoft), Natalie Levy (Aetion)Ga
  - Title:  Roundtable Panel – Exploring Career Paths in Pharma, Government, and Technology
  [Video]

• Tuesday, Apr 08, 2025: Nathan Kallus (Cornell University)
  - Title: Learning Surrogate Indices from Historical A/Bs: Adversarial ML for Debiased Inference on Functionals of Ill-Posed Inverses
  - Discussant: Rahul Singh (Harvard University)
  - Abstract: Experimentation on digital platforms often faces a dilemma: we want rapid innovation but we also want to make decisions based on long-term impact. Usually one resorts to looking at indices (i.e., scalar-valued functions) that combine multiple short-term surrogate outcomes. Constructing indices by regressing long-term metrics on short-term ones is easy with off-the-self ML but suffers bias from confounding and direct (i.e., unmediated) effects. I will discuss how to instead leverage past experiments as instrumental variables (IVs) and some surrogates as negative-control outcomes, with real-world examples from Netflix. There are two key technical challenges to surmount to make this possible. First, past experiments characterize the right surrogate index as a solution to an ill-posed system of moment equations: it does not uniquely identify an index, and approximately solving it does not translate to approximating any solution. We tackle this by developing a novel debiasing method for inference on linear functionals of solutions to ill-posed problems (as average long-term effects are such functionals of the index) and adversarial ML estimators for the solution admitting flexible hypothesis classes, such as neural nets and reproducing kernel Hilbert spaces. Second, even as we observe more past experiments, we have non-vanishing bias in estimating the moment equation implied by each one, since each experiment has a bounded size that is often just barely powered to detect effects. We tackle this by incorporating an instrument-splitting technique into our estimators, leading to a ML analogue of the classic (linear) jackknife IV estimator (JIVE) with guarantees for flexible function classes in terms of generic complexity measures.
  [Video][Discussant slides]

• Tuesday, Apr 01, 2025: Dylan Small (University of Pennsylvania)
  - Title:  Exploration, Confirmation, and Replication in the Same Observational Study: A Two Team Cross-Screening Approach to Studying the Effect of Unwanted Pregnancy on Mothers’ Later Life Outcomes
  - Discussant:  Ying Jin (Harvard University)
  - Abstract: The long term consequences of unwanted pregnancies carried to term on mothers have not been much explored. We use data from the Wisconsin Longitudinal Study (WLS) and propose a novel approach, namely two team cross-screening, to study the possible effects of unwanted pregnancies carried to term on various aspects of mothers’ later-life mental health, physical health, economic well-being and life satisfaction. Our method, unlike existing approaches to observational studies, enables the investigators to perform exploratory data analysis, confirmatory data analysis and replication in the same study– this is a valuable property when there is only a single data set available with unique strengths to perform exploratory, confirmatory and replication analysis. In two team cross-screening, the investigators split themselves into two teams and the data is split as well according to a meaningful covariate. Each team then performs exploratory data analysis on its part of the data to design an analysis plan for the other part of the data. The complete freedom of the teams in designing the analysis has the potential to generate new unanticipated hypotheses in addition to a prefixed set of hypotheses. Moreover, only the hypotheses that looked promising in the data each team explored are forwarded for analysis (thus alleviating the multiple testing problem). These advantages are demonstrated in our study of the effects of unwanted pregnancies on mothers’ later life outcomes.  This is joint work with Samrat Roy, Marina Bogomolov, Ruth Heller, Tishra Beeson and Amy Claridge.
  [Paper][Slides][Discussant slides][Video]