See our homepage for future talks.
Tuesday, Mar 25, 2025: Guido Imbens (Stanford University)
- Title: Identification of nonparametric factor models for average treatment effects
- Discussant: Bryan Graham (UC Berkeley)
- Abstract: There is a growing literature on methods for estimating causal effects in settings with panel or longitudinal data using two-way-fixed-effect, linear factor, and synthetic control methods. These methods attempt to adjust for unobserved differences between units as well as unobserved differences over time. Many of these methods partly rely on functional form assumptions to allow for such adjustments. Here we propose a set up that does not involve functional form assumptions. We show that by matching units on a similarity distance we can find units that are comparable in terms of expected outcomes and use those units to consistently estimate average causal effects under a matrix version of standard ignorability conditions.
[Slides][Video]
Tuesday, Mar 18, 2025: Alberto Abadie (MIT)
- Title: Synthetic Controls for Experimental Design
- Discussant: Dmitry Arkhangelsky (CEMFI)
- Abstract: This article studies experimental design in settings where the experimental units are large aggregate entities (e.g., markets), and only one or a small number of units can be exposed to the treatment. In such settings, randomization of the treatment may result in treated and control groups with very different characteristics at baseline, inducing biases. We propose a variety of experimental non-randomized synthetic control designs (Abadie, Diamond and Hainmueller, 2010, Abadie and Gardeazabal, 2003) that select the units to be treated, as well as the untreated units to be used as a control group. Average potential outcomes are estimated as weighted averages of the outcomes of treated units for potential outcomes with treatment, and weighted averages the outcomes of control units for potential outcomes without treatment. We analyze the properties of estimators based on synthetic control designs and propose new inferential techniques. We show that in experimental settings with aggregate units, synthetic control designs can substantially reduce estimation biases in comparison to randomization of the treatment.
[Paper][Slides][Video][Discussant slides]
Tuesday, Mar 11, 2025: Katherine A. Keith (Williams College)
- Title: Proximal Causal Inference with Text Data
- Discussant: Naoki Egami (Columbia University)
- Abstract: Recent text-based causal methods attempt to mitigate confounding bias by estimating proxies of confounding variables that are partially or imperfectly measured from unstructured text data. These approaches, however, assume analysts have supervised labels of the confounders given text for a subset of instances, a constraint that is sometimes infeasible due to data privacy or annotation costs. In this work, we address settings in which an important confounding variable is completely unobserved. We propose a new causal inference method that uses two instances of pre-treatment text data, infers two proxies using two zero-shot models on the separate instances, and applies these proxies in the proximal g-formula. We prove, under certain assumptions about the instances of text and accuracy of the zero-shot predictions, that our method of inferring text-based proxies satisfies identification conditions of the proximal g-formula while other seemingly reasonable proposals do not. To address untestable assumptions associated with our method and the proximal g-formula, we further propose an odds ratio falsification heuristic that flags when to proceed with downstream effect estimation using the inferred proxies. We evaluate our method in synthetic and semi-synthetic settings—the latter with real-world clinical notes from MIMIC-III and open large language models for zero-shot prediction—and find that our method produces estimates with low bias. We believe that this text-based design of proxies allows for the use of proximal causal inference in a wider range of scenarios, particularly those for which obtaining suitable proxies from structured data is difficult.
[Paper][Video][Slides]
Tuesday, Mar 04, 2025: Vasilis Syrgkanis (Stanford University)
- Title: Detecting clinician implicit biases in diagnoses using proximal causal inference
- Discussant: Ilya Shpitser (Johns Hopkins University)
- Abstract: Clinical decisions to treat and diagnose patients are affected by implicit biases formed by racism, ableism, sexism, and other stereotypes. These biases reflect broader systemic discrimination in healthcare and risk marginalizing already disadvantaged groups. Existing methods for measuring implicit biases require controlled randomized testing and only capture individual attitudes rather than outcomes. However, the "big-data" revolution has led to the availability of large observational medical datasets, like EHRs and biobanks, that provide the opportunity to investigate discrepancies in patient health outcomes. In this work, we propose a causal inference approach to detect the effect of clinician implicit biases on patient outcomes in large-scale medical data. Specifically, our method uses proximal mediation to disentangle pathway-specific effects of a patient’s sociodemographic attribute on a clinician’s diagnosis decision. We test our method on real-world data from the UK Biobank. Our work can serve as a tool that initiates conversation and brings awareness to unequal health outcomes caused by implicit biases.
[Paper][Slides][Video]
Tuesday, Feb 25, 2025: Young researchers' seminar
Speaker 1: Julius von Kügelgen (ETH Zürich)
- Title: Multi-Domain Causal Representation Learning
- Abstract: A key obstacle to more widespread use of causal models is requiring the relevant variables to be specified a priori. Yet, the causal relations of interest often do not occur at the level of raw observations such as pixels, but instead play out among abstract high-level latent concepts. Machine learning (ML) has proven successful at automatically extracting useful and compact representations of such complex data. Causal representation learning (CRL) aims to combine core strengths of ML and causality by learning representations in the form of latent variables endowed with causal structure and interventional semantics.
In this talk, I will present two identifiability studies [1,2] for CRL from multiple domains or environments arising from interventions on the latent causal variables. In particular, we focus on the continuous, nonparametric case in which both the latent causal model and the mixing function are completely unconstrained. First, we introduce Causal Component Analysis (CauCA) [1], a generalization of independent component analysis (ICA), in which the causal graph is known but non-trivial. As a special case of CRL, CauCA provides insights into necessary assumptions; it also yields novel identifiability results for ICA. We then study the full CRL setting [2] with unknown graph and intervention targets and prove identifiability subject to a suitable notion of genericity or given coupled interventions.
[1] L Wendong, A Kekić, J von Kügelgen, S Buchholz, M Besserve, L Gresele, B Schölkopf. Causal Component Analysis. In: Advances in Neural Information Processing Systems, 2023.
[2] J von Kügelgen, M Besserve, L Wendong, L Gresele, A Kekić, E Bareinboim, DM Blei, B Schölkopf. Nonparametric Identifiability of Causal Representations from Unknown Interventions. In: Advances in Neural Information Processing Systems, 2023.
[Slides][Video]
Speaker 2: David Bruns-Smith (Stanford University)
- Title: Augmented balancing weights as linear regression
- Abstract: We provide a novel characterization of augmented balancing weights, also known as automatic debiased machine learning (AutoDML). These popular doubly robust or de-biased machine learning estimators combine outcome modeling with balancing weights — weights that achieve covariate balance directly in lieu of estimating and inverting the propensity score. When the outcome and weighting models are both linear in some (possibly infinite) basis, we show that the augmented estimator is equivalent to a single linear model with coefficients that combine the coefficients from the original outcome model and coefficients from an unpenalized ordinary least squares (OLS) fit on the same data. We see that, under certain choices of regularization parameters, the augmented estimator often collapses to the OLS estimator alone; this occurs for example in a re-analysis of the LaLonde (1986) dataset. We then extend these results to specific choices of outcome and weighting models. We first show that the augmented estimator that uses (kernel) ridge regression for both outcome and weighting models is equivalent to a single, undersmoothed (kernel) ridge regression. This holds numerically in finite samples and lays the groundwork for a novel analysis of undersmoothing and asymptotic rates of convergence. When the weighting model is instead lasso-penalized regression, we give closed-form expressions for special cases and demonstrate a “double selection” property. Our framework opens the black box on this increasingly popular class of estimators, bridges the gap between existing results on the semiparametric efficiency of undersmoothed and doubly robust estimators, and provides new insights into the performance of augmented balancing weights.
[Paper][Slides][Video]
[SCI-OCIS joint webinar] Wednesday, February 19, 2025 @ 12:00 PM EST: Elizabeth A. Stuart (Johns Hopkins) [Registration link]
- Title: Lessons in "causality" from National Academies consensus panels
- Abstract: Quantitative researchers working in causal inference generally have a broadly common understanding about what we mean by “causal inference,” at least with respect to estimating causal effects. Many statistical methods have been developed to estimate causal effects in individual studies, and there is a growing literature on methods for combining (or “integrating”) multiple data sources together. However, it is unclear how these advances and frameworks fit in terms of broader discussions of “causality” in science, especially for broad scientific questions that require synthesis of a wide variety of types of evidence, ranging from biological mechanistic knowledge to narrow randomized experiments to large-scale non-experimental studies, and even medical case histories. This talk will discuss lessons learned about “causality” from serving on National Academies panels, in particular one assessing a framework for “causality” used by the Environmental Protection Agency to establish potential links between exposures and health and ecological outcomes, and another that aimed to assess the literature on possible links between antimalarial exposure and long-term psychiatric symptoms among Veterans. The talk will describe the scientific contexts and lessons for us as statisticians to ensure our work is relevant and useful for such broad scientific questions.
[Video]
Tuesday, Feb 18, 2025: Eric Tchetgen Tchetgen (University of Pennsylvania)
- Title: Revisiting Identification in the Binary Instrumental Variable Model: the NATE and Beyond
- Abstract: This talk revisits the identification problem in the canonical binary instrumental variable model. The work reveals new conditions for the classical Wald ratio estimand (WR) to be endowed with a nonparametric causal interpretation. Specifically, we describe a straightforward set of conditions under which the Wald Ratio point identifies the Nudge Average Treatment Effect (NATE), defined as the average causal effect for the subgroup of units whose treatment can be manipulated by the instrument, a sub-group referred to as Nudge-able. Crucially, the nudge-able may include both compliers and defiers therefore obviating the need for the standard no-defier IV condition known to identify the Local Average Treatment Effect. Our key identification condition for the NATE is that any variability of the treatment effect induced by a hidden counfounder must be is uncorrelated with corresponding variability in the share of compliers among the Nudge-able. An important and easily interpretable sufficient condition for this assumption is that, although a priori unrestricted, the share of compliers within the subgroup of nudge-able units is balanced across strata of the unmeasured confounders. We will also describe various generalizations of our results, including new straightforward conditions for identification of the average treatment effect for the treated by a generalized Wald ratio estimand, together with new quasi-IV identification results with an imperfect instrument which violates the exclusion restriction assumption.
[Paper][Slides][Video]
Tuesday, Feb 11, 2025: Corwin Zigler (Brown University)
- Title: Causal health impacts of power plant emission controls under modeled and uncertain physical process interference
- Discussant: Fredrik Sävje (Uppsala University)
- Abstract: Causal inference with spatial environmental data is often challenging due to the presence of interference: outcomes for observational units depend on some combination of local and nonlocal treatment. This is especially relevant when estimating the effect of power plant emissions controls on population health, as pollution exposure is dictated by: (i) the location of point-source emissions as well as (ii) the transport of pollutants across space via dynamic physical-chemical processes. In this work we estimate the effectiveness of air quality interventions at coal-fired power plants in reducing two adverse health outcomes in Texas in 2016: pediatric asthma ED visits and Medicare all-cause mortality. We develop methods for causal inference with interference when the underlying network structure is not known with certainty and instead must be estimated from ancillary data. Notably, uncertainty in the interference structure is propagated to the resulting causal effect estimates. We offer a Bayesian, spatial mechanistic model for the interference mapping, which we combine with a flexible nonparametric outcome model to marginalize estimates of causal effects over uncertainty in the structure of interference. Our analysis finds some evidence that emissions controls at upwind power plants reduce asthma ED visits and all-cause mortality; however, accounting for uncertainty in the interference renders the results largely inconclusive.
[Slides][Paper][Video][Discussant slides]
Tuesday, Feb 4, 2025: Michal Kolesár (Princeton University)
- Title: Evaluating Counterfactual Policies Using Instruments
- Discussant: Edward Vytlacil (Yale University)
- Abstract: In settings with instrumental variables, the TSLS estimator is the most popular way of summarizing causal evidence. Yet in many settings, the instrument monotonicity assumption needed for its causal interpretation is refuted. A prominent example are designs using the (quasi-)random assignment of defendants to judges as an instrument for incarceration. But ultimately, we may not be interested in the TSLS estimand itself, but rather in the impact of some counterfactual policy intervention (e.g. an encouragement to release more defendants). In this paper, we derive tractable sharp bounds on the impact of such counterfactual policies under reasonable sets of assumptions. We show that for a variety of common policy exercises, the bounds do not depend on whether one imposes instrument monotonicity, and thus one can drop this often-tenuous assumption without loss of information. We explore other restrictions that can help to tighten the bounds, including the policy invariance assumption commonly used in applications of the marginal treatment effects framework and its relaxations. We illustrate the usefulness of this approach in an application involving the quasi-random assignment of prosecutors to defendants in Massachusetts.
[Slides][Video][Discussant slides]
Tuesday, Jan 28, 2025: Anne Helby Petersen (University of Copenhagen)
- Title: What are we discovering? Two perspectives on interpretable evaluation of causal discovery algorithms
- Discussant: Vanessa Didelez (Leibniz Institute for Prevention Research and Epidemiology)
- Abstract: Causal discovery algorithms aim to recover (parts) of causal data generating mechanism by analyzing empirical data they generated. Methodological research on causal discovery has been immensely productive, but a standard practice for evaluating performance has still not been established. It is therefore difficult to understand what algorithms may be most suited for what tasks, and what performance we can expect on real data. Evaluating performance of causal discovery algorithms can be divided into two questions: First, how well do the algorithms perform in an (artificial) controlled environment, where the true data generating mechanism is assumed to be known? And secondly, do they provide useful information in real world applications?
As an example of how the second question may be adressed, I present a case-based study that compares graphs found using causal discovery algorithms with DAGs constructed by experts in a life coure epidemiologic study [1]. For the first question, I propose using random guessing as a common, interpretable baseline for evaluating causal discovery algorithms [2]. I provide mathematical insights into expected behavior of commonly used evaluation approaches under random guessing, as well as simulation-based generalizations. I furthermore showcase how use of negative controls enhance interpretability of classic causal discovery evaluation designs, including the study in [1].
[Slides][Video][Discussant slides]
Tuesday, Jan 21, 2025: Davide Viviano (Harvard University)
- Title: Program Evaluation with Remotely Sensed Outcomes
- Discussant: Hyunseung Kang (University of Wisconsin-Madison)
- Abstract: While traditional program evaluations typically rely on surveys to measure outcomes, certain economic outcomes such as living standards or environmental quality may be infeasible or costly to collect. As a result, recent empirical work estimates treatment effects using remotely sensed variables (RSVs), such mobile phone activity or satellite images, instead of ground-truth outcome measurements. Common practice predicts the economic outcome from the RSV, using an auxiliary sample of labeled RSVs, and then uses such predictions as the outcome in the experiment. We prove that this approach leads to biased estimates of treatment effects when the RSV is a post-outcome variable. We nonparametrically identify the treatment effect, using an assumption that reflects the logic of recent empirical research: the conditional distribution of the RSV remains stable across both samples, given the outcome and treatment. Our results do not require researchers to know or consistently estimate the relationship between the RSV, outcome, and treatment, which is typically mis-specified with unstructured data. We form a representation of the RSV for downstream causal inference by predicting the outcome and predicting the treatment, with better predictions leading to more precise causal estimates. We re-evaluate the efficacy of a large-scale public program in India, showing that the program’s measured effects on local consumption and poverty can be replicated using satellite imagery.
[Slides][Paper][Video][Discussant slides]
Tuesday, Jan 14, 2025: Qingyuan Zhao (University of Cambridge)
- Title: On statistical and causal models associated with acyclic directed mixed graphs
- Discussant: Thomas Richardson (University of Washington)
- Abstract: Causal models in statistics are often described using acyclic directed mixed graphs (ADMGs), which contain directed and bidirected edges and no directed cycles. This article surveys various interpretations of ADMGs, discusses their relations in different sub-classes of ADMGs, and argues that one of them -- nonparametric equation system (the E model below) -- should be used as the default interpretation. The E model is closely related to but different from the interpretation of ADMGs as directed acyclic graphs (DAGs) with latent variables that is commonly found in the literature. Our endorsement of the E model is based on two observations. First, in a subclass of ADMGs called unconfounded graphs (which retain most of the good properties of directed acyclic graphs and bidirected graphs), the E model is equivalent to many other interpretations including the global Markov and nested Markov models. Second, the E model for an arbitrary ADMG is exactly the union of that for all unconfounded expansions of that graph. This property is referred to as completeness, as it shows that the model does not commit to any specific latent variable explanation. In proving that the E model is nested Markov, we also develop an ADMG-based theory for causality. Finally, we criticize the common interpretation of ADMG as a convenient shortcut to represent some unspecified large causal DAG that generate the data. We argue that the "latent DAG" interpretation is mathematically unnecessary, makes obscure ontological assumptions, and discourages practitioners from deliberating over important structural assumptions.
[Slides][Paper][Video][Discussant slides]