Accepted Abstracts

Contributed Talks 

Authors: Daria Bystrova, Emilie Devijver, Vardan Manucharian, Julie Mondet, Pascal Mossuz
Title: Difference graph over two populations: Implicit Difference Inference algorithm
Abstract:  Comparing causal relations across different populations is essential in various fields, including medicine and ecology. Recently, several methods have been developed to directly infer difference graphs from observational data. These methods rely on semi-parametric assumptions and suppose  that the data is continuous.  We propose a new approach for discovering causal difference graphs without any semi-parametric assumption and can be applied on continuous or discrete or mixed data.  We provide theoretical guarantees of the new method and test it on simulated data.

Authors: Aurelien Bibaut, Nathan Kallus, Apoorva Lal
Title: Nonparametric Jackknife Instrumental Variable Estimation and Confounding Robust Surrogate Indices
Abstract:   Jackknife instrumental variable estimation (JIVE) is a classic method to leverage many weak instrumental variables (IVs) to estimate linear structural models, overcoming the bias of standard methods like two-stage least squares. In this paper, we extend the jackknife approach to nonparametric IV (NPIV) models with many weak IVs. Since NPIV characterizes the structural regression as having residuals projected onto the IV being zero, existing approaches minimize an estimate of the average squared projected residuals, but their estimates are biased under many weak IVs. We introduce an IV splitting device inspired by JIVE to remove this bias, and by carefully studying this split-IV empirical process we establish learning rates that depend on generic complexity measures of the nonparametric hypothesis class. We then turn to leveraging this for semiparametric inference on average treatment effects (ATEs) on unobserved long-term outcomes predicted from short-term surrogates, using historical experiments as IVs to learn this nonparametric predictive relationship even in the presence of confounding between short- and long-term observations. Using split-IV estimates of a debiasing nuisance, we develop asymptotically normal estimates for predicted ATEs, enabling inference.

Authors: Yuhao Wang, Arnab Bhattacharyya, Jin Tian, N. V. Vinodchandran
Title: PAC Style Guarantees for Doubly Robust Generalized Front-Door Estimator
Abstract:  Doubly robust estimators present a promising methodology for estimating treatment effects in observational studies. This paper provides a finite sample analysis of the doubly robust estimators for  both the back-door model (where treatment, outcome, and covariates are observed) and the generalized front-door model (which includes unmeasured confounding). Our approach establishes PAC-style guarantees of the deviation of the estimators in term of the divergence of probability distributions. These bounds demonstrate that minimizing the estimation error of the treatment effect in terms of Chi-square distance is crucial for minimizing the variance between true and estimated model.

Authors: Johan de Aguas, Sebastian Krumscheid, Johan Pensar, Guido Biele
Title: Partial identification and efficient estimation for the stratum-specific probability of benefit with thresholds on a continuous outcome
Abstract:  We define the probability of benefit for a scenario involving a binary exposure, a continuous outcome, and a partition of the outcome support with $K$ fixed thresholds. As with other counterfactual queries, this parameter is often not $g$-identifiable, and we show that monotonicity assumption is not sufficient when $K>1$. We introduce a partial identification strategy by adapting existing bounds. Conducting asymptotic inference and uncertainty quantification for estimates of these bounds is challenging due to potential nonregularity and the lack of differentiability of the involved functionals. Moreover, results might be sensitive to model specification. To address this, we reformulate the problem in terms of individualized rules, adapting the available online one-step estimator with stabilizing weights. We show the connection with solutions based on conservative optimal transport and illustrate the advantages over surrogate bounds derived from smooth approximations. We present an application aimed at estimating the probability of benefit from pharmacological treatment for ADHD upon school performance using observational data.

Posters 

Authors: Jakob Zeitler, Raul Astudillo
Title: Causal Elicitation for Bayesian Optimization
Abstract: Causal Inference allows scientists and businesses to draw causal conclusions about e.g. their drug- development or marketing campaign. Causal Entropy Search Branchini et al. [2023] was introduced as a way to learn both the causal graph as well as optimise an intervention of interest at the same time. It combines Bayesian Optimisation with the Causal Inference Framework to identify the right molecule or marketing tagline. Here, we present initial work on a crucial extension of CEO, namely the introduction of preference elicitation, an increasingly popular technique in Bayesian Optimisation to elicit crucial causal knowledge from subject matter experts. We introduce the problem of Causal Elicitation for Bayesian Optimisation, discuss elicitation strategies and initial work on empirical evaluation.

Authors: Dhurim Cakiqi, Max A Little
Title: Algorithmic syntactic causal identification
Abstract:  Causal identification in causal Bayes nets (CBNs) is an important tool in causal inference allowing the derivation of interventional distributions from observational distributions where this is possible in principle. However, most existing formulations of causal identification using techniques such as d-separation and do-calculus are expressed within the mathematical language of classical probability theory on CBNs. However, there are many causal settings where probability theory and hence current causal identification techniques are inapplicable such as relational databases, dataflow programs such as hardware description languages, distributed systems and most modern machine learning algorithms. We show that this restriction can be lifted by replacing the use of classical probability theory with the alternative axiomatic foundation of symmetric monoidal categories. In this alternative axiomatization, we show how an unambiguous and clean distinction can be drawn between the general syntax of causal models and any specific semantic implementation of that causal model. This allows a purely syntactic algorithmic description of general causal identification by a translation of recent formulations of the general ID algorithm through fixing. Our description is given entirely in terms of the non-parametric ADMG structure specifying a causal model and the algebraic signature of the corresponding monoidal category, to which a sequence of manipulations is then applied so as to arrive at a modified monoidal category in which the desired, purely syntactic interventional causal model, is obtained. We use this idea to derive purely syntactic analogues of classical back-door and front-door causal adjustment, and illustrate an application to a more complex causal model.

Authors: Kenneth Lee, Bruno Ribeiro, Murat Kocaoglu
Title: Constraint-based Causal Discovery from a Collection of Conditioning Sets
Abstract:   In constraint-based causal discovery, existing algorithms systematically use a series of conditional independence (CI) relations observed in the data to recover an equivalence class of causal graphs in the large sample limit. One limitation of these algorithms, such as the PC algorithm, is the reliance on CI tests, which can quickly lose statistical power due to finite samples as the conditioning set size increases or the support of the conditioning set is large. The idea of bounding the size of conditioning sets has been proposed for robust causal discovery. However, the existing algorithms require exhaustive testing of all CI relations with conditioning set sizes up to a certain integer $k$. To further relax this restriction, we propose using CI tests where the conditioning sets are restricted to a given set of conditioning sets including the empty set. We call this set a conditionally closed set $\mathcal{C}$. We define the notion of $\mathcal{C}$-Markov equivalence. We propose a graphical representation to characterize $\mathcal{C}$-Markov equivalence between two causal graphs. We propose a sound constraint-based algorithm called the $\mathcal{C}$-PC algorithm for learning the $\mathcal{C}$-Markov equivalence class. We demonstrate the utility of the proposed algorithm via experiments in scenarios where high-dimensional variables and spurious correlations are present in the data. 

Authors: Klaus-Rudolf Kladny, Julius von Kügelgen, Bernhard Schölkopf, Michael Muehlebach
Title: Backtracking Counterfactuals for Deep Structural Causal Models
Abstract:  Counterfactuals answer questions of what would have been observed under altered circumstances and can therefore offer valuable insights. Whereas the classical interventional interpretation of counterfactuals has been studied extensively, backtracking constitutes a less studied alternative where all causal laws are kept intact. In the present work, we introduce a practical method called deep backtracking counterfactuals (DeepBC) for computing backtracking counterfactuals in structural causal models that consist of deep generative components. We employ constrained optimization to generate counterfactuals for high-dimensional data and conduct experiments on a modified version of MNIST. 

Authors: Shashaank Khanna, Marina Maciel Ansanelli, Mathew F. Pusey, Elie Wolfe
Title: Classifying Causal Structures: Ascertaining Inequality Constraints
Abstract:  The causal relations between a set of variables, some observed and some latent, can induce both equality constraints (typically conditional independencies) as well as inequality constraints (Instrumental and Bell inequalities being prototypical examples) on their compatible distribution over the observed variables. Enumerating a causal structure’s implied inequality constraints is generally far more difficult than enumerating its equalities. Furthermore, only inequality constraints ever admit violation by quantum correlations. For both those reasons, it is important to classify causal scenarios into those which impose inequality constraints versus those which do not. Here we develop methods for detecting such scenarios by appealing to d-separation, e-separation, and incompatible supports. Many (perhaps all?) scenarios with exclusively equality constraints can be detected via a condition articulated by Henson, Lal and Pusey (HLP). Considering all scenarios with up to 4 observed variables, which number in the thousands, we are able to resolve all but three causal scenarios, providing evidence that the HLP condition is, in fact, exhaustive. 

Authors: Marina Maciel Ansanelli, Elie Wolfe, Robert Spekkens
Title: Everything that can be learned about a causal structure with latent variables by observational and interventional probing schemes
Abstract:  When is it impossible to distinguish between two causal structures with latent variables from statistical data obtained by probing each visible variable? If we simply passively observe each visible variable, then it is well-known that many different causal structures can realize the same joint probability distributions. Even for the simplest case of two visible variables, for instance, one cannot distinguish between one variable being a causal parent of the other and the two variables being confounded by a latent common cause. However, it is possible to distinguish between these two causal structures if we have recourse to more powerful probing schemes, such as the possibility of intervening on one of the variables and observing the other. Herein, we address the question of which causal structures remain indistinguishable even given the most informative types of probing schemes on the visible variables. We find that two causal structures remain indistinguishable if and only if they are both associated with the same mDAG structure (as defined in [Evans, 2016]). We also investigate to what extent one can weaken the probing schemes implemented on the visible variables, such as allowing only for do-interventions that can fix a variable to one of its possible values affects, and still have the same discrimination power as a maximally informative probing scheme. 

Authors: Nadja Rutsch, Sara Magliacane, Stéphanie L. van der Pas
Title: MSE-optimal adjustment sets in linear Gaussian causal models with finite sample size
Abstract:  Covariate selection for causal inference based on the causal graph commonly aims for unbiasedness and asymptotic efficiency of the causal effect estimator. When the sample size is finite, these approaches can lead to results that are suboptimal in terms of the Mean Squared Error (MSE). We aim to find the adjustment set that is optimal in terms of MSE, taking into account the joint distribution of the causal variables and the sample size. We present examples where the MSE-optimal adjustment set differs from the optimal adjustment set, depending on the sample size. To find the MSE-optimal adjustment set, we introduce a sample size criterion that compares two adjustment sets in linear Gaussian models. We develop graphical criteria to reduce the search space for this adjustment set based on the causal graph. In preliminary experiments, we show that the estimated MSE-optimal adjustment set can outperform the optimal adjustment set in finite sample size settings, and performs competitively in larger sample size settings. 

Authors: Danru Xu, Dingling Yao, Sebastien Lachapelle, Perouz Taslakian, Julius von Kügelgen, Francesco Locatello, Sara Magliacane
Title: A Sparsity Principle for Partially Observable Causal Representation Learning
Abstract: Causal representation learning aims at identifying high-level causal variables from perceptual data. Most methods assume that all latent causal variables are captured in the high-dimensional observations. We instead consider a partially observed setting, in which each measurement only provides information about a subset of the underlying causal state. Prior work has studied this setting with multiple domains or views, each depending on a fixed subset of latents. Here we focus on learning from unpaired observations from a dataset with an instance-dependent partial observability pattern. Our main contribution is to establish two identifiability results for this setting: one for linear mixing functions without parametric assumptions on the underlying causal model, and one for piecewise linear mixing functions with Gaussian latent causal variables. Based on these insights, we estimate the underlying causal variables by enforcing sparsity in the inferred representation. Based on these insights, we propose two methods for estimating the underlying causal variables by enforcing sparsity in the inferred representation. 

Authors: William Chang, Haohan Wang
Title: Understanding Domain Adaptation in the Lens of Causality
Abstract: Domain adaptation has been of great interest as it allows us to apply models developed from training to unseen data from a different distribution. To evaluate the success of these models, various generalization bounds have been derived. In parallel, causality has recently gained traction as it allows us to more rigorously understand how input variables affect output variables. In this paper, we provide an alternative perspective to domain adaptation using causality to show the importance of understanding the causal structure when training. More specifically, we look at the accuracy of our model when we train a model using some input variables on the training distributions, but fail to consider latent variables. We assume that the latent variables are both present in the generation of the training and testing datasets. Using the causal language \cite{pearl2009causal}, we are able to derive generalization bounds. Denoting the resulting discrepancy measure as $c(\theta, P_s)$ we then compare $c(\theta, P_s)$ with other commonly studied discrepancy measures such as  $\mathcal{H} \Delta \mathcal{H}$ divergence, Wasserstein distance, and Rademacher complexity. Finally, we run experiments under various parameters to corroborate the bounds provided in our paper.

Authors: Kenneth Lee, Murat Kocaoglu
Title: RCPC: A Sound Causal Discovery Algorithm under Orientation Unfaithfulness
Abstract:  In causal discovery, the constraint-based approaches often rely on an assumption known as faithfulness/stability, only the variables that are d-separated in a directed acyclic graph will be statistically independent. This assumption can be partitioned into two subconditions: orientation faithfulness and adjacency faithfulness. Under adjacency faithfulness, a sound algorithm known as CPC, a conservative version of PC algorithm, has been developed and is conjectured to be complete. In this work, we show that the CPC algorithm is not complete and propose two new sound orientation rules as part of a sound causal discovery algorithm called revised CPC (RCPC) under orientation unfaithfulness. 

Authors: Luka Kovačević, Izzy Newsham, Sach Mukherjee, John C Whittaker
Title: Simulation-based Benchmarking for Causal Structure Learning in Gene Perturbation Experiments
Abstract:  Causal structure learning (CSL) refers to the task of learning causal relationships from data. Advances in CSL now allow learning of causal graphs in diverse application domains, which has the potential to facilitate data-driven causal decision-making. Real-world CSL performance depends on a number of *context-specific* factors, including context-specific data distributions and non-linear dependencies, that are important in practical use-cases. However, our understanding of how to assess and select CSL methods in specific contexts remains limited. To address this gap, we present *CausalRegNet*, a multiplicative effect structural causal model that allows for generating observational and interventional data incorporating context-specific properties, with a focus on the setting of gene perturbation experiments. Using real-world gene perturbation data, we show that CausalRegNet generates accurate distributions and scales far better than current simulation frameworks. We illustrate the use of CausalRegNet in assessing CSL methods in the context of interventional experiments in biology. 

Authors: Inwoo Hwang, Yesong Choe, Yeahoon Kwon, Sanghack Lee
Title: Causal Identification with Relaxed Positivity
Abstract:   Identifying and estimating a causal effect is a fundamental task when researchers want to infer a causal effect using an observational study without experiments. A conventional assumption is the strict positivity of the given distribution, or so called positivity (or overlap) under the unconfounded assumption that the probabilities of treatments are positive. However, there exist many environments where neither observational data exhibits strict positivity nor unconfounded assumption holds. In this work, we examine the graphical counterpart of the conventional positivity condition so as to license the use of an identification formula without strict positivity. In particular, we explore various approaches, including analysis in a post-hoc manner, do-calculus, $Q$-decomposition, and algorithmic, to yielding a positivity condition for an identification formula. We relate these approaches, providing a comprehensive view. 

Authors: Roger Pros, Jordi Vitria
Title: Preventing spurious interactions in tree-based metalearners
Abstract:   In recent years, various insights have been employed to enhance causal machine learning methods by refining estimation techniques and introducing robust algorithms that account for causal structures and dependencies within the data. Building on this trend, we propose a novel method to improve the estimation of the Conditional Average Treatment Effect (CATE). A common approach in CATE estimation involves the use of metalearners, which can estimate CATE if certain identification properties are met. However, this approach employs causal knowledge only for identifying the estimand, not for the estimation process itself. We present a new method that utilizes causal knowledge in the estimation phase by imposing variable interaction constraints during model training. These constraints are based on total or partial knowledge about the underlying data-generating process. By applying these constraints to traditional tree-based estimation algorithms, we show that models trained in this manner achieve improved performance and reduced variability in estimating CATE.

Authors: Simon Rittel, Sebastian Tschiatschek
Title: On Differentiable Bayesian Causal Structure Learning
Abstract:  This extended abstract reviews differentiable Bayesian causal structure learning (CSL) and discusses why recent works on Bayesian causal discovery published in top-tier conference do not yet meet important desiderata. In particular, we advocate against the current trend of global regularization via prior terms. 

Authors: Mátyás Schubert, Tom Claassen, Sara Magliacane
Title: SNAP: Sequential Non-Ancestor Pruning for Targeted Causal Effect Estimation With an Unknown Graph
Abstract:  Causal discovery often serves as a precursor to causal effect estimation, but it can be computationally demanding due to the number of conditional independence tests involved. If we are interested in estimating only the causal effects on a small subset of the measured variables, many of these tests may be unnecessary. Existing methods addressing this issue often have strong assumptions about the causal relations between variables. In this paper, we consider targeted causal effect estimation with an unknown graph, a task that focuses on identifying the causal effect between multiple target variables. This task combines causal discovery and effect estimation, aligning the discovery objective with the effects to be estimated. We show that the non-ancestors of the target variables are unnecessary to estimate the causal effects between the targets. We sequentially identify and prune these non-ancestors during the process of existing algorithms. Our results show that our approach substantially reduces the number of tests without compromising the quality of causal effect estimations.

Authors: Tianzhu Zhang, Davide Rinaldi, Fabio Pianese, Armen Aghasaryan
Title: Evaluating the Robustness of Causal Discovery Algorithms with Observations and Interventions in VNF Deployments
Abstract:   Causal discovery (CD) incorporates a large collection of interdisciplinary research endeavors from statistics, computer science, and philosophy to uncover the true causal relationship from data and move beyond mere correlations to expose the underlying data generation mechanism. Despite the rich set of causal discovery algorithms, they also bear some common limitations, including demanding assumptions and lack of validation using real-world data, making their applicability in real systems questionable. This paper explores the practical challenges of performing causal discovery in real systems. We construct a controllable Network Function Virtualization (NFV) system that allows the deployment and perturbation of interconnected topologies of high-performance Virtual Network Functions (VNFs). Our contribution is a comparison of the ability of state-of-the-art CD algorithms to reconstruct the correct causal configuration from data in observational and interventional settings.

Authors: Stelios Triantafyllou, Aleksa Sukovic, Yasaman Zolfimoselo, Goran Radanovic
Title: Counterfactual Effect Decomposition in Multi-Agent Sequential Decision Making
Abstract:   We address the challenge of explaining counterfactual outcomes in multi-agent Markov decision processes. In particular, we aim to explain the total counterfactual effect of an agent's action to some realized outcome through its influence on the environment dynamics and the agents' behavior. To achieve this, we introduce a novel causal explanation formula that decomposes the counterfactual effect of an agent's action by attributing to each agent and state variable a score reflecting its respective contribution to the effect.

Authors: Ana Esponera Gómez, Giovanni Cinà
Title: Interchange Intervention Training Applied to Post-meal Glucose Prediction for Type 1 Diabetes Mellitus Patients
Abstract:   This research explores the application of Interchange Intervention Training (IIT) in predicting blood glucose levels in Type 1 Diabetes Mellitus (T1DM) patients by leveraging expert knowledge encoded in causal models. The study utilizes an acyclic version of the simglucose simulator approved by the FDA to train a Multi-Layer Perceptron (MLP) model, employing IIT to abstract its causal internal structure. Results show that the model trained with IIT effectively abstracted the causal structure and it outperformed the standardly trained one in terms of predictive performance across different prediction horizons (PHs) post-meal. This technique also allows us to measure the extent to which the causal structure has been abstracted, promoting the interpretability of the black-box model. These preliminary results with the acyclic model suggest the potential of IIT in enhancing predictive models in healthcare by effectively complying with expert knowledge. 

Authors: Vera Kvisgaard, Johan Pensar
Title: Bayesian estimation of causal effects from observational categorical data
Abstract:   We develop a scaleable Bayesian method for estimation of all pairwise causal effects in a system from observational data, under the assumption that the underlying causal model is an unknown discrete Bayesian network and that there are no latent confounders. Specifically, we build upon the the Bayesian IDA (BIDA) and extend this method to the categorical setting. The key-idea is to combine Bayesian estimation of intervention distributions through the so-called backdoor formula with Bayesian model averaging. The main motivation of the method is to inform future experiments about plausible strong relationships, and we demonstrate by numerical experiments that our Bayesian modeling averaging approach can be highly relevant for this task.

Authors: Joshua R. Loftus, Lucius E.J. Bynum, Sakina Hansen
Title: Causal Dependence Plots for Interpretable Machine Learning
Abstract:   To use artificial intelligence and machine learning models wisely we must understand how they interact with the world, including how they depend causally on data inputs. In this work we develop Causal Dependence Plots (CDPs) to visualize how a model's predicted outcome depends on changes in a given predictor along with consequent causal changes in other predictor variables. Crucially, this differs from standard methods based on independence or holding other predictors constant, such as regression coefficients or Partial Dependence Plots (PDPs). Our explanatory framework generalizes PDPs, including them as a special case, as well as a variety of other interpretive plots that show, for example, the total, direct, and indirect effects of causal mediation. Since people often think causally about input-output dependence, CDPs can be powerful tools in the xAI or interpretable machine learning toolkit and contribute to applications like scientific machine learning and algorithmic fairness.

Authors: Marina Maciel Ansanelli, Daniel Centeno, Elie Wolfe, Matt Jones, Matthew F. Pusey
Title: Causal Criterion for Multivariate Correlation under Postselection

Authors: Mouad El Bouchattaoui, Myriam Tami, BENOIT LEPETIT, Paul-Henry Cournède
Title: Toward a more transparent causal representation learning
Abstract:  This work addresses the challenge of causal representation learning (CRL) for complex, high-dimensional, time-varying data. We enhance transparency and confidence in learned causal abstractions by linking them to observational space. The existing literature rarely explores the association between latent causal variables and observed ones, with only one notable work imposing a simplistic single-latent-factor decoding constraint. Our approach, in contrast, allows for a flexible entangling of latent factors, reflecting the complexity of real-world datasets. We introduce a structural sparsity pattern over generative functions and leverage induced grouping structures over observed variables for better model understanding. Our regularization technique, based on sparse subspace clustering over the Jacobian matrix of the decoder, promotes the sparsity and readability of model results. We apply our model to real-world datasets, including Saint-Gobain purchase data and MIMIC III medical data.

Authors: Sahil Satish Kumar, Claudia Soares
Title: Causal Inference Explanations for Graph Neural Networks
Abstract:  Explainable Artificial Intelligence has emerged, aiming to enhance the trustworthiness of black box models by devising explanation methods that clarify their inner workings. However, prevalent explanation techniques predominantly leverage correlation and association rather than employing causality, a significant aspect of human comprehension. We propose a novel explanation method grounded in causal inference tailored specifically for Graph Neural Networks. Our approach seeks to illuminate the decision-making process of Graph Neural Networks, thereby augmenting their transparency and trustworthiness. We apply our method to the medical referral problem in healthcare.

Authors: Rezaur Rashid, Gabriel Terejanu
Title: Graph Neural Networks for Probabilistic Causal Discovery
Abstract:  Conventional causal discovery algorithms face significant challenges in dealing with large-scale observational datasets and in capturing global structural information. ⁤⁤To address these limitations, we introduce a novel graph neural network (GNN)--based probabilistic framework for causal structure learning that generates a probability distribution over the entire graph space. By encoding the node and edge attributes into a unified graph representation, our framework enables the GNN to learn the complex causal structure directly from the data augmented with statistical and information-theoretic measures, which exploit the local and global data properties. ⁤⁤Our approach outperforms benchmark methods, both traditional and recent non-GNN-based, in terms of accuracy and scalability on synthetic and real-world datasets. Notably, our framework advances the causal discovery paradigm by generating a probability distribution over the causal graphs, rather than learning a single causal graph.

Authors: Vijay Keswani
Title: On the Role of Control in Auditing Risk Prediction Tools
Abstract:  This paper explores the role of individual agency in algorithmic risk predictions. By comparing the relative control individuals have over various features used for risk prediction to the predictive relevance of these features, we formalize an audit procedure to assess the usability of risk prediction in practice.

Authors: Oscar Clivio, David Bruns-Smith, Avi Feller, Christopher C. Holmes
Title: Towards Principled Representation Learning to Improve Overlap in Treatment Effect Estimation
Abstract:  A common approach to mitigate undesirable effects of poor overlap is to use well-crafted representations of covariates as adjustment sets. In this abstract, we motivate quantifying the overlap induced by a representation using the $\chi^2$-divergence, show that the overlap improvement under this metric is precisely how much the representation does not predict the propensity score, which confirms intuitions in previous work, and discuss next steps.

Authors: Anna K Raichev, Jin Tian, Alexander Ihler, Rina Dechter
Title: Estimating Causal Effects from Learned Causal Networks
Abstract:   The standard approach to answering an identifiable causal-effect query (e.g., P(Y |do(X)) when given a causal diagram and observational data is to first generate an estimand, or probabilistic expression over the observable variables, which is then evaluated using the observational data. In this paper, we propose an alternative paradigm for answering causal-effect queries over discrete observable variables. We instead learn the causal Bayesian network and its confounding latent variables directly from the data. Then, efficient probabilistic graphical model (PGM) algorithms can be applied to the learned model to answer queries. Surprisingly, we show that this model completion learning approach can be more effective than estimand approaches, particularly for larger models in which the estimand expressions become computationally difficult. We illustrate our method’s potential using a benchmark collection of Bayesian networks and synthetically generated causal models.

Authors: Caleb H Miles
Title: The central role of the mediator process in mediation analysis
Abstract:   Causal mediation has traditionally been framed as the effect of an exposure on an outcome through some intermediate variable, where each variable is measured at three sequential time points. However, definitions of mediated effects and their corresponding identification assumptions generally ignore the fact that the mediator of interest is, in many if not most circumstances, a stochastic process indexed by time from baseline to follow-up. I demonstrate that the failure to account for the mediator process has profound implications for defining the relevant causal estimand of interest as well as its identification and estimation. Additionally, I introduce novel versions of direct and indirect effect definitions that account for the entire mediator process.

Authors: Razieh Nabi, David Benkeser
Title: Fair Risk Minimization under Causal Path-Specific Effect Constraints
Abstract:   Causal mediation has traditionally been framed as the effect of an exposure on an outcome through some intermediate variable, where each variable is measured at three sequential time points. However, definitions of mediated effects and their corresponding identification assumptions generally ignore the fact that the mediator of interest is, in many if not most circumstances, a stochastic process indexed by time from baseline to follow-up. I demonstrate that the failure to account for the mediator process has profound implications for defining the relevant causal estimand of interest as well as its identification and estimation. Additionally, I introduce novel versions of direct and indirect effect definitions that account for the entire mediator process.