Miniworkshop on Causal Inference 2024

Abstracts

Niels Richard Hansen: Efficient adjustment and deconfounding 

Many causal parameters of interest, such as Average Treatment Effect (ATE), are identified and estimated from the observational distribution via adjustment. There is a substantial literature on semi-parametric efficient and doubly robust estimators, some of which use machine learning to avoid strong model assumptions. Despite this, adjusting for complex or high-dimensional covariates remains a practical challenge. To tackle this challenge we attempt to learn a covariate representation that is minimally sufficient for valid adjustment. We present general results characterizing such representations and specific results for the deconfounder, which is an adjustment algorithm based on a factor model of the covariates.

Negar Kyiavash: Triple Difference Estimator (for Targeted Policies)

The renowned difference-in-differences (DiD) estimator relies on the assumption of 'parallel trends,' which may not hold in many practical applications. To address this issue, economists are increasingly considering the triple difference estimator as a more credible alternative. Both DiD and triple difference are limited to assessing average effects exclusively. An alternative avenue is offered by the changes-in-changes (CiC) estimator, which provides an estimate of the entire counterfactual distribution by relying on assumptions imposed on the distribution of potential outcomes. In this talk, we discuss an extension of the triple difference estimator to accommodate the CiC framework, which we call the `triple changes estimator,' and its identification assumptions, thereby expanding the scope of the CiC paradigm. We further discuss the application of the proposed framework to a study examining the impact of Medicaid expansion on children's preventive care.

Martin Huber: Learning control variables and instruments for causal analysis in observational data

This study introduces a data-driven, machine learning-based method to detect suitable control variables and instruments for assessing the causal effect of a treatment on an outcome in observational data, if they exist. Our approach tests the joint existence of instruments, which are associated with the treatment but not directly with the outcome (at least conditional on observables), and suitable control variables, conditional on which the treatment is exogenous, and learns the partition of instruments and control variables from the observed data. The detection of sets of instruments and control variables relies on the condition that proper instruments are conditionally independent of the outcome given the treatment and suitable control variables. We establish the consistency of our method for detecting control variables and instruments under certain regularity conditions, investigate the finite sample performance through a simulation study, and provide an empirical application to labor market data from the Job Corps study.

Niklas Pfister: Extrapolation-Aware Nonparametric Statistical Inference

Extrapolation occurs in many data analysis applications and can invalidate the resulting conclusions if not taken into account. Formally, extrapolation refers to any type of statistical inference on a conditional function (e.g., a conditional expectation or conditional quantile) evaluated outside of the support of the conditioning variable. While extrapolation is straightforward in parametric models, it becomes challenging in nonparametric models. In this talk, we extend the nonparametric statistical model to explicitly allow for extrapolation and introduce a class of extrapolating assumptions that can be combined with existing inference techniques to draw extrapolation-aware conclusions. The proposed class of extrapolation assumptions stipulate that the conditional function of interest attains its minimal and maximal directional derivative in each direction within the observed support. We illustrate how the framework can be applied to several statistical applications including out-of-support prediction and extrapolation-aware uncertainty quantification.

Leonard Henckel: Adjustment Identification Distance: A gadjid for Causal Structure Learning

Benchmarking causal discovery algorithms is a challenging task, in part, because there are no commonly agreed upon performance metrics. We introduce a framework for developing causal distances between graphs which includes the structural intervention distance for directed acyclic graphs as a special case. We use this framework to develop improved adjustment-based distances as well as extensions to completed partially directed acyclic graphs and causal orders. We also develop new reachability algorithms to compute the distances efficiently and to prove their low polynomial time complexity. 

Jakob Runge: Causal inference for time series data

Many research questions across scientific disciplines are inherently causal. Causal inference provides the theoretical foundations to use data and qualitative domain knowledge to quantitatively answer these questions, complemented by statistical and machine learning techniques. Further, the underlying complex systems are often of a dynamical nature, leading to non-iid time series as the primary data source. The latter carries specific challenges as well as opportunities. In this talk I will focus on problem settings and recent advancements around causal inference for time series data. I will discuss challenges ahead and selected application scenarios to spark interest for collaborations on advancing causal inference for time series data.

Francesco Locatello: Causal representation learning with the invariance principle

Machine learning and AI have the potential to transform data-driven scientific discovery, enabling not only accurate predictions for several scientific phenomena but also causal understanding. Toward this, I will present a new framework for causal representation learning based on the invariance principle that generalizes most existing methodologies. Thanks to the increased flexibility, we show improved performance on our ISTAnt data set, the first real-world benchmark for estimating causal effects from high-dimensional observations in experimental ecology. Further, I will connect causal representation learning with recent advances in dynamical systems discovery that, when combined, enable learning scalable and controllable models with identifiable trajectory-specific parameters, which we apply to real-world climate data. 

Isabel Valera: Causethical ML -  from theory to practice 

In this talk I will give an overview of the role of causality in ethical machine learning, and in particular, in fair and explainable ML. In particular, I will first detail how to use causal reasoning to study fairness and interpretability problems in algorithmic decision making, stressing the main limitations that we encounter when aiming to address these problems in practice.  Then, I will provide some hints about how to solve some of these practical limitations by using causal generative models. A novel class of deep generative models that do not only accurately fit observational data but can also provide accurate estimates to interventional and counterfactual queries. I will finally discussed the open challenges of designing such causal generative models.

Sara Magliacane: Causal representation learning in temporal settings 

Causal inference reasons about the effect of unseen interventions or external manipulations on a system. Similar to classic approaches to AI, it typically assumes that the causal variables of interest are given from the outset. However, real-world data often comprises high-dimensional, low-level observations (e.g., pixels in a video) and is thus usually not structured into such meaningful causal units. Causal representation learning aims to address this gap by learning high-level causal variables along with their causal relations directly from raw, unstructured data, e.g., images or videos.  In this talk, I will focus on learning causal representations in temporal sequences, e.g., sequences of images. In particular, I will present some of our recent work on causal representation learning in environments in which we can perform interventions or actions. I will start by presenting CITRIS, where we leveraged the knowledge of which variables are intervened in each timestep to learn a provably disentangled representation of the potentially multidimensional ground truth causal variables, as well as a Dynamic Bayesian Network representing the causal relations between these variables. I will then show iCITRIS, an extension that allows for instantaneous effects between variables. Finally, I will focus on our most recent method, BISCUIT, which overcomes one of the biggest limitations of our previous methods: knowing which variables are intervened. In BISCUIT we instead leverage actions with unknown effects on an environment. Assuming that each causal variable has exactly two distinct causal mechanisms, we prove that we can recover each ground truth variable from a sequence of images and actions up to permutation and element-wise transformations. This allows us to apply BISCUIT to realistic simulated environments for embodied AI, where we can learn a latent representation that allows us to identify and manipulate each causal variable, as well as a mapping between each high-level action and its effects on the latent causal variables.

Qingyuan Zhao: Acyclic Directed Mixed graphs: matrix algebra, statistical models, confounder selection

Directed mixed graphs (DMGs) permit directed and bidirected edges between any two vertices and
play an essential role in statistical modeling. This talk discusses the role of such graphs in causal inference and is divided into three parts:

• First, I will introduce a matrix algebra for walks on DMGs, motivated by Wright’s path analysis, that allows its user to easily describe and visualize complex graphical concepts.
• Second, I will discuss various interpretations of acyclic DMGs and their relations. Acyclic DMGs are often interpreted as directed acyclic graphs (DAGs) with latent variables. I will introduce a slight modification of this interpretation and argue that it should be used as the default interpretation.
• Third, I will discuss a new approach to confounder selection for causal inference that does not require the full causal graph. Instead, this approach solicits partial information about the graph in an iterative and interactive fashion.

The last part of the talk is based on joint work with F Richard Guo.

Jalal Etesami: Causal effect identification under Uncertainty

Causal identification is at the core of the causal inference literature, where complete algorithms have been proposed to identify causal queries of interest. The validity of these algorithms hinges on the restrictive assumption of having access to a correctly specified causal structure. In this talk, I explore the settings where a probabilistic model of the causal structure is available. Specifically, the edges in a causal graph exist with uncertainties which may, for example, represent degree of belief from domain experts. The question that naturally arises in this setting is: Given such a probabilistic graph and a specific causal effect of interest, what is the subgraph which has the highest plausibility and for which the causal effect is identifiable?