Spring 2022 complete list with abstracts

• Tuesday, June 28, 2022: Samuel Wang (Cornell University)
  - Uncertainty Quantification for Causal Discovery
  - Discussant: Daniel Malinsky (Columbia University)
  - Abstract: Causal discovery procedures are popular methods for discovering causal structure across the physical, biological, and social sciences. However, most procedures for causal discovery only output a single estimated causal model or single equivalence class of models. In this work, we propose a procedure for quantifying uncertainty in causal discovery. Specifically, we consider structural equation models where a unique graph can be identified and propose a procedure which returns a confidence sets of causal orderings which are not ruled out by the data. We show that asymptotically, a true causal ordering will be contained in the returned set with some user specified probability. In addition, the confidence set can be used to form conservative sets of ancestral relationships.
  [Video] [Slides] [Discussant slides] 

• Tuesday, June 21, 2022: Geneviève Lefebvre (Université du Québec à Montréal)
  - Bayesian joint modeling for causal mediation analysis with a binary outcome and a binary mediator
  - Discussant: Olli Saarela (University of Toronto)
  - Abstract: Mediation analysis with a binary outcome is notoriously more challenging than with a continuous outcome. In this talk, I will present a new approach, named t-link, to perform causal mediation with a binary outcome and a binary mediator. This approach relies on the Bayesian multivariate logistic regression model introduced by O'Brien and Dunson (Biometrics, 2004, 739-746, 60(3)) and its Student-t approximation. By re-expressing the mediation formula, I show how to use this multivariate latent model for estimating the natural direct and indirect effects of an exposure on an outcome in any measure scale of interest (e.g., odds or risk ratio, risk difference). The t-link mediation approach has several valuable features which, to the best of knowledge, are not found together in existing binary-binary mediation analysis approaches. In particular, it allows for sensitivity analyses regarding the impact of unmeasured mediator-outcome confounders on the natural effects estimates. The proposed mediation approach is evaluated and compared with two other benchmark approaches using simulated data, and is illustrated using pediatric cancer data.
  [Slides] [Discussant slides]

• Tuesday, June 14, 2022: AmirEmad Ghassami (Johns Hopkins University)
  - Combining Experimental and Observational Data for Identification and Estimation of Long-Term Causal Effects
  - Discussant: Guido Imbens
  - Abstract: We consider the task of identifying and estimating the causal effect of a treatment variable on a long-term outcome variable using data from an observational domain and an experimental domain. The observational domain is subject to unobserved confounding. Furthermore, subjects in the experiment are only followed for a short period of time; hence, long-term effects of treatment are unobserved but short-term effects will be observed. Therefore, data from neither domain alone suffices for causal inference about the effect of the treatment on the long-term outcome, and must be pooled in a principled way, instead. Athey et al. (2020) proposed a method for systematically combining such data for identifying the downstream causal effect in view. Their approach is based on the assumptions of internal and external validity of the experimental data, and an extra novel assumption called latent unconfoundedness. In this paper, we first review their proposed approach, and then we propose three alternative approaches for data fusion for the purpose of identifying and estimating average treatment effect as well as the effect of treatment on the treated. Our first approach is based on assuming equi-confounding bias for the short-term and long-term outcomes. Our second approach is based on a relaxed version of the equi-confounding bias assumption, where we assume the existence of an observed confounder such that the short-term and long-term potential outcome variables have the same partial additive association with that confounder. Our third approach is based on the proximal causal inference framework, in which we assume the existence of an extra variable in the system which is a proxy of the latent confounder of the treatment-outcome relation. We propose influence function-based estimation strategies for each of our data fusion frameworks and study the robustness properties of the proposed estimators.
  [Video] [Slides] [Discussant slides] [Paper]

• Tuesday, June 7, 2022: Mona Azadkia (ETH)
  - A Fast Non-parametric Approach for Causal Structure Learning in Polytrees
  - Discussant: Bryon Aragam (Chicago Booth)
  - Abstract: We study the problem of causal structure learning with no assumptions on the functional relationships and noise. We develop DAG-FOCI, a computationally fast algorithm for this setting that is based on the FOCI variable selection algorithm in (Azadkia 2021). DAG-FOCI requires no tuning parameter and outputs the parents and the Markov boundary of a response variable of interest. We provide high-dimensional guarantees of our procedure when the underlying graph is a polytree. Furthermore, we demonstrate the applicability of DAG-FOCI on real data from computational biology (Sachs et al., 2005) and illustrate the robustness of our methods to violations of assumptions.
  [Video] [Slides] [Paper]

• Tuesday, May 31, 2022: Bin Yu (UC Berkeley)
  - Predictability, stability, and causality with a case study to find genetic drivers of a heart disease
  - Discussant: Jas Sekhon (Yale University)
  - Abstract:  "A.I. is like nuclear energy -- both promising and dangerous" -- Bill Gates, 2019.
  Data Science is a pillar of A.I. and has driven most of recent cutting-edge discoveries in biomedical research and beyond.  Human judgement calls are ubiquitous at every step of a data science life cycle, e.g., in choosing data cleaning methods, predictive algorithms and data perturbations. Such judgment calls are often responsible for the "dangers" of A.I. To maximally mitigate these dangers, we developed a framework based on three core principles: Predictability, Computability and Stability (PCS). The PCS framework unifies and expands on the best practices of machine learning and statistics. It consists of a workflow and documentation and is supported by our software package v-flow.
  In this talk, we first illustrate the PCS framework through the development of iterative random forests (iRF) for predictable and stable non-linear interaction discovery (in collaboration with the Brown Lab at LBNL and Berkeley Statistics). In pursuit of genetic drivers of a heart disease called hypertrophic cardiomyopathy (HCM) as a CZ Biohub project in collaboration with the Ashley Lab at Stanford Medical School and others, we use iRF and UK Biobank data to recommend gene-gene interaction targets for knock-down experiments. We then analyze the experimental data to show promising findings about genetic drivers of HCM.
  [Slides] [Video]

• Tuesday, May 17, 2022: Mireille Schnitzer (University of Montreal)
  - Estimands and estimation of COVID-19 vaccine effectiveness under the test-negative design: connections to causal inference
  - Discussant: David Benkeser (Emory University)
  - Abstract: The test-negative design (TND) is routinely used for the monitoring of seasonal flu vaccine effectiveness. More recently, it has become integral to the estimation of COVID-19 vaccine effectiveness, in particular for more severe disease outcomes. Distinct from the case-control study, the design typically involves recruitment of participants with a common symptom presentation who are being tested for the infectious disease in question. Participants who test positive for the target infection are the “cases” and those who test negative are the “controls”. Logistic regression is the only statistical method that has been proposed to estimate vaccine effectiveness under the TND while adjusting for confounders. While under strong modeling assumptions it produces estimates of a causal risk ratio, it may be biased in the presence of effect modification by a confounder. I will present and justify an inverse probability of treatment weighting (IPTW) estimator for the marginal risk ratio, which is valid under effect modification.  I’ll discuss connections between the estimands targeted by these two methods and causal parameters under different interference assumptions.  I will then describe the results of a simulation study to illustrate and confirm the derivations and to evaluate the performance of the estimators. 
  [Slides] [Video]

• Tuesday, May 10, 2022: Tim Morrison (Stanford University); Harrison Li (Stanford University)
  - Talk 1: Optimality in multivariate tie-breaker designs
  - Abstract: Tie-breaker designs (TBDs), in which subjects with extreme values are assigned treatment deterministically and those in the middle are randomized, are intermediate between regression discontinuity designs (RDDs) and randomized controlled trials (RCTs). TBDs thus provide a convenient mechanism by which to trade off between the treatment benefit of an RDD and the statistical efficiency gains of an RCT. We study a model where the expected response is one multivariate regression for treated subjects and another one for control subjects. For a given set of subject data we show how to use convex optimization to choose treatment probabilities that optimize a prospective D-optimality condition (expected information gain). We can incorporate economically motivated linear constraints on those treatment probabilities as well as monotonicity constraints that have a strong ethical motivation. Our condition can be used in two scenarios: known covariates with random treatments, and random covariates with random treatments.
  [Video 1] [Speaker 1 slides] 
  - Talk 2: A general characterization of optimal tie-breaker designs
  - Abstract: Tie-breaker designs trade off between statistical efficiency and a preference for assigning a binary treatment to individuals with high values of a quantitative running variable x. Motivating examples include university scholarship programs and promotions for e-commerce companies. We explicitly characterize tie-breaker designs that optimize a D-optimality efficiency criterion under a two-line regression model, subject to equality constraints on the expected proportion of treated individuals and the covariance between x and the binary treatment indicator. Our results extend to any running variable distribution F with finite variance and any efficiency criterion depending continuously on the expected information matrix in the regression. If we additionally require treatment probabilities to be non-decreasing in x, an optimal design requires just two probability levels when the running variable distribution F is continuous. By contrast, the original tie-breaker design in Owen and Varian (2020) has three probability levels fixed at 0, 0.5, and 1. We find large efficiency gains for our optimal designs compared to using those three levels when fewer than half of the subjects are to be treated, or F is not symmetric. We illustrate these gains with a data example based on Head Start, a U.S. government early-childhood intervention program.
  [Video 2] [Speaker 2 slides]

• Tuesday, May 3, 2022: Tyler VanderWeele (Harvard University)
  - Title: Causal Inference and Measure Construction: Towards a New Model of Measurement
  - Discussant: Fredrik Sävje (Yale University)
  - Abstract. Psychosocial constructs can only be assessed indirectly, and measures are typically formed by a combination of indicators that are thought to relate to the construct. Reflective and formative measurement models offer different conceptualizations of the relation between the indicators and what is sometimes conceived of as a univariate latent variable supposedly corresponding to the construct. I argue that the empirical implications of these models will often be violated by data since the causally relevant constituents will generally be multivariate, not univariate. In fact, the assumption of an underlying univariate structural latent variable is so strong that it has empirically testable implications, even though the latent is unobserved. Formal statistical tests can be developed to reject this assumption, but factor analysis, as typically practiced, is not adequate to do so. Factor analysis also suffers from the inability to distinguish associations arising from causal versus conceptual relations. I put forward an outline for a new model of the process of measure construction and propose a causal interpretation of associations between constructed measures and subsequent outcomes that is applicable even if the usual assumptions of reflective and formative models fail. I discuss the practical implications of these observations and proposals for the provision of definitions, the selection of items, item-by-item analyses, the construction of measures, and the causal interpretation of regression analyses.
  [Video] [Slides] [Discussant slides]

• Tuesday, April 26, 2022: Shu Yang (NCSU)
  - Test-based integrative analysis for heterogeneous treatment effects combining randomized trial and real-world data
  - Discussant: Issa Dahabreh (Harvard University)
  - Abstract: Parallel randomized trial (RT) and real-world (RW) data are becoming increasingly available for treatment evaluation. Given the complementary features of the RT and RW data, we propose a test-based elastic integrative analysis of RT and RW data for accurate and robust estimation of the heterogeneity of treatment effect (HTE), which lies at the heart of precision medicine. When the RW data are not subject to bias, e.g., due to hidden confounding, our approach combines the RT and RW data for optimal estimation by exploiting semiparametric efficiency theory. Utilizing the design advantage of RTs, we construct a built-in test procedure to gauge the reliability of the RW data and decide whether or not to use RW data in an integrative analysis. A data-adaptive procedure is proposed to select the threshold of the test statistic that promises the smallest mean square error of the proposed estimator of the HTE. Lastly, we construct an adaptive confidence interval that has a good finite-sample coverage property. We apply the proposed method to characterize who can benefit from adjuvant chemotherapy in patients with stage IB non-small cell lung cancer. If time permits, I will cover other approaches such as using the notion of the confounding function to improve inference for HTEs using RT and RW data.
  Keywords: Bias function; Least favorable confidence interval; Nonregularity; Pre-test estimator; Semiparametric efficient score.
  [Video] [Slides] [Discussant slides] [Paper] 

• Tuesday, April 19, 2022: Alex Luedtke (University of Washington)
  - Adversarial Monte Carlo Meta-Learning of Conditional Average Treatment Effects
  - Discussant:  Jonas Metzger (Stanford University)
  - Abstract: We frame the meta-learning of conditional average treatment effect estimators as a search for an optimal strategy in a two-player game. In this game, Nature selects a prior over distributions that generate labeled data consisting of covariates, treatment, and an associated outcome, and the Estimator observes data sampled from a distribution drawn from this prior. The Estimator's objective is to learn a function that maps from a new feature to an estimate of the conditional average treatment effect. We establish that, under reasonable conditions, the Estimator's has an optimal strategy that is equivariant to shifts and rescalings of the outcome and is invariant to permutations of the observations and to shifts, rescalings, and permutations of the features. We introduce a neural network architecture that satisfies these properties.
  [Video] [Discussant slides] [Slides]

• Tuesday, April 12, 2022: Neil Davies (University of Bristol)
  - Average causal effect estimation via instrumental variables: the no simultaneous heterogeneity assumption
  - Discussant: Eric Tchetgen Techetgen
  - Abstract: Instrumental variables (IVs) can be used to provide evidence as to whether a treatment X has a causal effect on Y. Z is a valid instrument if it satisfies the three core IV assumptions of relevance, independence and the exclusion restriction. Even if the instrument satisfies these assumptions, further assumptions are required to estimate the average causal effect (ACE) of X on Y. Sufficient assumptions for this include: homogeneity in the causal effect of X on Y; homogeneity in the association of Z with X; and No Effect Modification (NEM). Here, we describe the NO Simultaneous Heterogeneity (NOSH) assumption, which requires the heterogeneity in the X-Y causal effect to be independent of both Z and heterogeneity in the Z-X association. We describe the necessary conditions for NOSH to hold, in which case conventional IV methods are consistent for the ACE even if both homogeneity assumptions and NEM are violated. We illustrate these ideas using simulations and by re-examining selected published studies.
  [Video] [Slides] [Discussant slides] [Paper]

• Tuesday, April 5, 2022: Zijian Guo (Rutgers University)
  - Two Stage Curvature Identification with Machine Learning: Causal Inference with Possibly Invalid Instrumental Variables
  - Discussant:  Frank Windmeijer (University of Oxford)
  - Abstract: Instrumental variables regression is a popular causal inference method for endogenous treatment. A significant concern in practical applications is the validity and strength of instrumental variables. This paper aims to perform causal inference when all instruments are possibly invalid. To do this, we propose a novel methodology called two stage curvature identification (TSCI) together with a generalized concept to measure the strengths of possibly invalid instruments: such invalid instruments can still be used for inference in our framework. We fit the treatment model with a general machine learning method and propose a novel bias correction method to remove the overfitting bias from machine learning methods. Among a collection of spaces of violation functions, we choose the best one by evaluating invalid instrumental variables' strength. We demonstrate our proposed TSCI methodology in a large-scale simulation study and revisit the important economics question on the effect of education on earnings. This is a joint work with Dr. Peter Bühlmann.
  [Video] [Slides] [Discussant slides] [Paper]

• Tuesday, March 29, 2022
  - Speaker 1: Shuangning Li (Stanford University)
  - Random Graph Asymptotics for Treatment Effect Estimation under Network Interference
  - Abstract: The network interference model for causal inference places all experimental units at the vertices of an undirected exposure graph, such that treatment assigned to one unit may affect the outcome of another unit if and only if these two units are connected by an edge. This model has recently gained popularity as means of incorporating interference effects into the Neyman--Rubin potential outcomes framework; and several authors have considered estimation of various causal targets, including the direct and indirect effects of treatment. In this paper, we consider large-sample asymptotics for treatment effect estimation under network interference in a setting where the exposure graph is a random draw from a graphon. When targeting the direct effect, we show that -- in our setting -- popular estimators are considerably more accurate than existing results suggest, and provide a central limit theorem in terms of moments of the graphon. Meanwhile, when targeting the indirect effect, we leverage our generative assumptions to propose a consistent estimator in a setting where no other consistent estimators are currently available. We also show how our results can be used to conduct a practical assessment of the sensitivity of randomized study inference to potential interference effects. Overall, our results highlight the promise of random graph asymptotics in understanding the practicality and limits of causal inference under network interference.
  [Video] [Slides]
  - Speaker 2: Michael Oberst  (MIT)
  - Regularizing towards Causal Invariance: Linear Models with Proxies
  - Abstract: We propose a method for learning linear models whose predictive performance is robust to causal interventions on unobserved variables, when noisy proxies of those variables are available. Our approach takes the form of a regularization term that trades off between in-distribution performance and robustness to interventions. Under the assumption of a linear structural causal model, we show that a single proxy can be used to create estimators that are prediction optimal under interventions of bounded strength. This strength depends on the magnitude of the measurement noise in the proxy, which is, in general, not identifiable. In the case of two proxy variables, we propose a modified estimator that is prediction optimal under interventions up to a known strength. We further show how to extend these estimators to scenarios where additional information about the "test time" intervention is available during training. We evaluate our theoretical findings in synthetic experiments and using real data of hourly pollution levels across several cities in China.
  [Video] [Slides]

• Tuesday, March 22, 2022: Mathias Drton  (Technical University of Munich)
  - Half-Trek Criterion for Identifiability of Latent Variable Models
  - Discussant: Robin Evans (University of Oxford)
  - Abstract: We consider linear structural equation models with latent variables and develop a criterion to certify whether the direct causal effects between the observable variables are identifiable based on the observed covariance matrix. Linear structural equation models assume that both observed and latent variables solve a linear equation system featuring stochastic noise terms. Each model corresponds to a directed graph whose edges represent the direct effects that appear as coefficients in the equation system. Prior research has developed a variety of methods to decide identifiability of direct effects in a latent projection framework, in which the confounding effects of the latent variables are represented by correlation among noise terms. This approach is effective when the confounding is sparse and effects only small subsets of the observed variables. In contrast, the new latent-factor half-trek criterion (LF-HTC) we develop in this paper operates on the original unprojected latent variable model and is able to certify identifiability in settings, where some latent variables may also have dense effects on many or even all of the observables. Our LF-HTC is an effective sufficient criterion for rational identifiability, under which the direct effects can be uniquely recovered as rational functions of the joint covariance matrix of the observed random variables. When restricting the search steps in the LF-HTC to consider subsets of latent variables of bounded size, the criterion can be verified in time that is polynomial in the size of the graph.
  [Video] [Slides] [Discussant slides] [paper]