Jiji Zhang

Causal models consisting of structural equations have proved a powerful formalism for approaching theoretical and applied questions on causal reasoning. The standard framework assumes that all variables that appear in a causal model must represent properties or events that are metaphysically independent, with no dependencies more intimate than causal relations, such as conceptual, constitutive, or other ontological dependencies. However, there are important contexts that call for lifting this restriction, such as the debate over high-level causation in philosophy and the task of causal representation learning in artificial intelligence. This paper proposes a natural generalization of structural causal models to accommodate (asymmetric) ontological dependence and provides an interpretation of causal counterfactuals based on such models. We discuss several implications of this generalization and present an axiomatization of the resulting counterfactual logic.

Assuming that the future of human relations will depend critically on whether AI technologies that already mediate and shape personal and professional interactions will be made trustworthy, I examine an apparently daunting challenge to making predictive AI trustworthy, which is the widely observed trade-off between the accuracy and fairness of predictive AI models. I argue that the significance of this trade-off has been misconceived, and the standard rhetoric of balancing accuracy and fairness indicates a misguided route to trustworthy AI. I suggest an alternative way to think about the alleged trade-off and discuss the relevant implications of the reconceptualization, especially the need to have a clear understanding of the intended use of AI, which in turn requires AI users to reflect on their goals and values.

In “Flagpoles anyone? Causal and explanatory asymmetries”, James Woodward supplements his celebrated interventionist account of causation and explanation with a set of new ideas about causal and explanatory asymmetries, which he extracts from some cutting-edge methods for causal discovery from observational data. Among other things, Woodward draws interesting connections between observational causal discovery and interventionist themes that are inspired in the first place by experimental causal discovery, alluding to a sort of unity between observational and experimental causal discovery. In this paper, I make explicit what I take to be the implicated unity. Like experimental causal discovery, observational causal discovery also relies on interventions (or exogenous variations, to be more accurate), albeit interventions that are not carried out by investigators and hence need to be detected as part of the inference. The observational patterns appealed to in observational causal discovery are not only surrogates for would-be interventions, as Woodward sometimes puts it; they also serve to mark relevant interventions that actually happen in the data generating process.

An important component in the interventionist account of causal explanation is an interpretation of counterfactual conditionals as statements about consequences of hypothetical interventions. The interpretation receives a formal treatment in the framework of functional causal models. In Judea Pearl’s influential formulation, functional causal models are assumed to satisfy a “unique-solution” property; this class of Pearlian causal models includes the ones called recursive. Joseph Halpern showed that every recursive causal model is Lewisian, in the sense that there is a possible worlds model in David Lewis’s semantics for counterfactuals that satisfies the exact same formulas as the causal model does in a certain language. Moreover, he demonstrated that some Pearlian (non-recursive) models are not Lewisian in this sense. This raises the question about the exact contour of Lewisian causal models. We address this question in this paper, not only regarding the class of Lewisian causal models, but also regarding an important subclass, that of Stalnakerian causal models, which is closer to Pearlian models, and an important superclass of Lewisian causal models, which results from weakening a constraint imposed by Lewis known as centering, and is in a way more natural than the class of Lewisian causal models. For each of the three classes, we provide a characterization and an axiomatization. Our results have philosophically interesting consequences, especially regarding cycles of counterfactual dependence.

Several of the most prominent theories of actual causation make use of a minimality condition to prevent irrelevant elements from being tacked onto a cause so that the conjunction or sum passes for a cause. Focusing on one theory in particular—the influential Halpern-Pearl definition of actual causation—we argue that either the minimality condition or its rationale ought to be revised. We produce proposals showing that both are live options and demonstrate their potential usefulness within the larger family of counterfactual approaches to causation.

Consider the problem of learning, from non-experimental data, the causal (Markov equivalence) structure of the true, unknown causal Bayesian network (CBN) on a given, fixed set of (categorical) variables. This learning problem is known to be very hard, so much so that there is no learning algorithm that converges to the truth for all possible CBNs (on the given set of variables). So the convergence property has to be sacrificed for some CBNs—but for which? In response, the standard practice has been to design and employ learning algorithms that secure the convergence property for at least all the CBNs that satisfy the famous faithfulness condition, which implies sacrificing the convergence property for some CBNs that violate the faithfulness condition (Spirtes, Glymour, and Scheines, 2000). This standard design practice can be justified by assuming—that is, accepting on faith—that the true, unknown CBN satisfies the faithfulness condition. But the real question is this: Is it possible to explain, without assuming the faithfulness condition or any of its weaker variants, why it is mandatory rather than optional to follow the standard design practice? This paper aims to answer the above question in the affirmative. We first define an array of modes of convergence to the truth as desiderata that might or might not be achieved by a causal learning algorithm. Those modes of convergence concern (i) how pervasive the domain of convergence is on the space of all possible CBNs and (ii) how uniformly the convergence happens. Then we prove a result to the following effect: for any learning algorithm that tackles the causal learning problem in question, if it achieves the best achievable mode of convergence (considered in this paper), then it must follow the standard design practice of converging to the truth for at least all CBNs that satisfy the faithfulness condition—it is a requirement, not an option.

Stephen Yablo’s notion of proportionality, despite controversies surrounding it, has played a significant role in philosophical discussions of mental causation and of high-level causation more generally. In particular, it is invoked in James Woodward’s interventionist account of high-level causation and explanation, and is implicit in a novel approach to constructing variables for causal modeling in the machine learning literature, known as causal feature learning (CFL). In this article, we articulate an account of proportionality inspired by both Yablo’s account of proportionality and the CFL account of variable construction. The resulting account has at least three merits. First, it illuminates an important feature of the notion of proportionality, when it is adapted to a probabilistic and interventionist framework. The feature is that at the center of the notion of proportionality lies the concept of “determinate intervention effects.” Second, it makes manifest a virtue of (common types of) high-level causal/explanatory statements over low-level ones, when relevant intervention effects are determinate. Third, it overcomes a limitation of the CFL framework and thereby also addresses a challenge to interventionist accounts of high-level causation.

Stephen Yablo’s notion of proportionality, despite controversies surrounding it, has played a significant role in philosophical discussions of mental causation and of high-level causation more generally. In particular, it is invoked in James Woodward’s interventionist account of high-level causation and explanation, and is implicit in a novel approach to constructing variables for causal modeling in the machine learning literature, known as causal feature learning (CFL). In this article, we articulate an account of proportionality inspired by both Yablo’s account of proportionality and the CFL account of variable construction. The resulting account has at least three merits. First, it illuminates an important feature of the notion of proportionality, when it is adapted to a probabilistic and interventionist framework. The feature is that at the center of the notion of proportionality lies the concept of “determinate intervention effects.” Second, it makes manifest a virtue of (common types of) high-level causal/explanatory statements over low-level ones, when relevant intervention effects are determinate. Third, it overcomes a limitation of the CFL framework and thereby also addresses a challenge to interventionist accounts of high-level causation.

We examine the performance of some standard causal discovery algorithms, both constraint-based and score-based, from the perspective of how robust they are against failures of the Causal Faithfulness Assumption. For this purpose, we make only the so-called Triangle-Faithfulness assumption, which is a fairly weak consequence of the Faithfulness assumption, and otherwise allows unfaithful distributions. In particular, we allow violations of Adjacency-Faithfulness and Orientation-Faithfulness. We show that the PC algorithm, a representative constraint-based method, can be made more robust against unfaithfulness by incorporating elements of the GES algorithm, a representative score-based method; similarly, the GES algorithm can be made less error-prone by incorporating elements of the conservative PC algorithm. As our simulations demonstrate, the increased robustness seems to matter even when faithfulness is not exactly violated, for with only finite sample, distributions that are not exactly unfaithful may be sufficiently close to being unfaithful to make trouble.

Compared to constraint-based causal discovery, causal discovery based on functional causal models is able to identify the whole causal model under appropriate assumptions [Shimizu et al. 2006; Hoyer et al. 2009; Zhang and Hyvärinen 2009b]. Functional causal models represent the effect as a function of the direct causes together with an independent noise term. Examples include the linear non-Gaussian acyclic model, nonlinear additive noise model, and post-nonlinear model. Currently, there are two ways to estimate the parameters in the models: dependence minimization and maximum likelihood. In this article, we show that for any acyclic functional causal model, minimizing the mutual information between the hypothetical cause and the noise term is equivalent to maximizing the data likelihood with a flexible model for the distribution of the noise term. We then focus on estimation of the PNL causal model and propose to estimate it with the warped Gaussian process with the noise modeled by the mixture of Gaussians. As a Bayesian nonparametric approach, it outperforms the previous one based on mutual information minimization with nonlinear functions represented by multilayer perceptrons; we also show that unlike the ordinary regression, estimation results of the PNL causal model are sensitive to the assumption on the noise distribution. Experimental results on both synthetic and real data support our theoretical claims.

We study the identifiability and estimation of functional causal models under selection bias, with a focus on the situation where the selection depends solely on the effect variable, which is known as outcome-dependent selection. We address two questions of identifiability: the identifiability of the causal direction between two variables in the presence of selection bias, and, given the causal direction, the identifiability of the model with outcome-dependent selection. Regarding the first, we show that in the framework of post-nonlinear causal models, once outcome-dependent selection is properly modeled, the causal direction between two variables is generically identifiable; regarding the second, we identify some mild conditions under which an additive noise causal model with outcome-dependent selection is to a large extent identifiable. We also propose two methods for estimating an additive noise model from data that are generated with outcome-dependent selection.

A primary object of causal reasoning concerns what would happen to a system under certain interventions. Specifically, we are often interested in estimating the probability distribution of some random variables that would result from forcing some other variables to take certain values. The renowned do-calculus gives a set of rules that govern the identification of such post-intervention probabilities in terms of pre-intervention probabilities, assuming available a directed acyclic graph that represents the underlying causal structure. However, a DAG causal structure is seldom fully testable given preintervention, observational data, since many competing DAG structures are equally compatible with the data. In this paper we extend the do-calculus to cover cases where the available causal information is summarized in a so-called partial ancestral graph that represents an equivalence class of DAG structures. The causal assumptions encoded by a PAG are significantly weaker than those encoded by a full-blown DAG causal structure, and are in principle fully testable by observed conditional independence relations.

We clarify the status of the so-called causal minimality condition in the theory of causal Bayesian networks, which has received much attention in the recent literature on the epistemology of causation. In doing so, we argue that the condition is well motivated in the interventionist (or manipulability) account of causation, assuming the causal Markov condition which is essential to the semantics of causal Bayesian networks. Our argument has two parts. First, we show that the causal minimality condition, rather than an add-on methodological assumption of simplicity, necessarily follows from the substantive interventionist theses, provided that the actual probability distribution is strictly positive. Second, we demonstrate that the causal minimality condition can fail when the actual probability distribution is not positive, as is the case in the presence of deterministic relationships. But we argue that the interventionist account still entails a pragmatic justification of the causal minimality condition. Our argument in the second part exemplifies a general perspective that we think commendable: when evaluating methods for inferring causal structures and their underlying assumptions, it is relevant to consider how the inferred causal structure will be subsequently used for counterfactual reasoning

Learning equivalence classes of acyclic models with latent and selection variables from multiple datasets with overlapping variables is discussed. The problem of inferring the presence of latent variables, their relation to the observables, and the relation among themselves, is considered. A different approach for identifying causal structures, one that results in much simpler equivalence classes, is provided. It is found that the computational cost is much higher than the procedure implemented, but if datasets are individually of modest dimensionality, it might be doable in practice. From the point of view of search algorithms for optimizing structure, much of the machinery of combinatorial optimization could optimize the penalized composite likelihood score by enforcing constraints such that the independence models over different subsets of variables agree on the overlapping sets.

Different directed acyclic graphs may be Markov equivalent in the sense that they entail the same conditional independence relations among the observed variables. Chickering provided a transformational characterization of Markov equivalence for DAGs, which is useful in deriving properties shared by Markov equivalent DAGs, and, with certain generalization, is needed to prove the asymptotic correctness of a search procedure over Markov equivalence classes, known as the GES algorithm. For DAG models with latent variables, maximal ancestral graphs provide a neat representation that facilitates model search. However, no transformational characterization -- analogous to Chickering's -- of Markov equivalent MAGs is yet available. This paper establishes such a characterization for directed MAGs, which we expect will have similar uses as it does for DAGs.

It is commonplace to encounter nonstationary or heterogeneous data, of which the underlying generating process changes over time or across data sets. Such a distribution shift feature presents both challenges and opportunities for causal discovery. In this paper we develop a principled framework for causal discovery from such data, called Constraint-based causal Discovery from Nonstationary/heterogeneous Data, which addresses two important questions. First, we propose an enhanced constraint-based procedure to detect variables whose local mechanisms change and recover the skeleton of the causal structure over observed variables. Second, we present a way to determine causal orientations by making use of independence changes in the data distribution implied by the underlying causal model, benefiting from information carried by changing distributions. Experimental results on various synthetic and real-world data sets are presented to demonstrate the efficacy of our methods.

The account of causal regularities in the influential INUS theory of causation has been refined in the recent developments of the regularity approach to causation and of the Boolean methods for inference of deterministic causal structures. A key element in the refinement is to strengthen the minimality or non-redundancy condition in the original INUS account. In this paper, we argue that the Boolean framework warrants a further strengthening of the minimality condition. We motivate our stronger condition by showing, first, that a rationale for strengthening the original minimality condition in the INUS theory is also applicable to our proposal to go further, and second, that the new element of the stronger condition is a counterpart to a well-established minimality condition for probabilistic causal models. We also compare the various minimality conditions in terms of the difference-making criteria they imply and argue for the criterion implied by our condition. Finally, we show that putative counterexamples to our proposal can be addressed in the same way that the Boolean theorists defend the current minimality conditions in their framework.

In the causal inference framework of Spirtes, Glymour, and Scheines, inferences about causal relationships are made from samples from probability distributions and a number of assumptions relating causal relations to probability distributions. The most controversial of these assumptions is the Causal Faithfulness Assumption, which roughly states that if a conditional independence statement is true of a probability distribution generated by a causal structure, it is entailed by the causal structure and not just for particular parameter values. In this paper we show that the addition of the Causal Faithfulness Assumption plays three quite different roles in the SGS framework: it reduces the degree of underdetermination of causal structure by probability distribution; computationally, it justifies reliable causal inference algorithms that would otherwise have to be slower in order to be reliable; and statistically, it implies that those algorithms reliably obtain the correct answer at smaller sample sizes than would otherwise be the case. We also consider a number of variations on the Causal Faithfulness Assumption, and show how they affect each of these three roles

One conception of underdetermination is that it corresponds to the impossibility of reliable inquiry. In other words, underdetermination is defined to be the situation where, given a set of background assumptions and a space of hypotheses, it is logically impossible for any hypothesis selection method to meet a given reliability standard. From this perspective, underdetermination in a given subject of inquiry is a matter of interplay between background assumptions and reliability or success criteria. In this paper I discuss underdetermination in causal inference along this line. In particular I will analyze several success criteria that can be applied to causal inference from statistical regularities. The criteria center on the notions of consistency in mathematical statistics. For each criterion I present its epistemic implication in terms of simple conditions under which the criterion cannot possibly be met. I then investigate which of the familiar principles and their variants in the literature, if adopted as background assumptions, are sufficient to overcome different levels of underdetermination induced by the success criteria. “亚决定性”是知识论和科学哲学中一个重要的概念。对这个概念的一种阐释是把它对应于可靠探索的不可能性。就是说，在一个（经验）问题中，给定一些公设和一些供选择的理论或假说，如果逻辑上不可能找到一种理论选择的方法能满足一定的可靠或成功标准，那么相对于这个标准就存在亚决定性。从这个观点看，亚决定性总是相对于一个问题设定，尤其是公设和成功标准而言的。本文从这个角度对近来的统计因果推理研究作一番梳理。首先，基于数理统计中的一致性概念，我会讨论和分析一系列可应用于因果推理的成功标准。对每一个标准，我会用一个相对简单的条件来刻画它对应的亚决定性。然后我对文献里一部分重要的结果作一个综述，以澄清什么样的公设可以消除什么样的亚决定性。

A fundamental question in causal inference is whether it is possible to reliably infer the manipulation effects from observational data. There are a variety of senses of asymptotic reliability in the statistical literature, among which the most commonly discussed frequentist notions are pointwise consistency and uniform consistency (see, e.g. Bickel, Doksum [2001]). Uniform consistency is in general preferred to pointwise consistency because the former allows us to control the worst case error bounds with a finite sample size. In the sense of pointwise consistency, several reliable causal inference algorithms have been established under the Markov and Faithfulness assumptions [Pearl 2000, Spirtes et al. 2001]. In the sense of uniform consistency, however, reliable causal inference is impossible under the two assumptions when time order is unknown and/or latent confounders are present [Robins et al. 2000]. In this paper we present two natural generalizations of the Faithfulness assumption in the context of structural equation models, under which we show that the typical algorithms in the literature are uniformly consistent with or without modifications even when the time order is unknown. We also discuss the situation where latent confounders may be present and the sense in which the Faithfulness assumption is a limiting case of the stronger assumptions.

Using the flexibility of recently developed methods for causal discovery based on Boolean satisfiability solvers, we encode a variety of assumptions that weaken the Faithfulness assumption. The encoding results in a number of SAT-based algorithms whose asymptotic correctness relies on weaker conditions than are standardly assumed. This implementation of a whole set of assumptions in the same platform enables us to systematically explore the effect of weakening the Faithfulness assumption on causal discovery. An important effect, suggested by simulation results, is that adopting weaker assumptions greatly alleviates the problem of conflicting constraints and substantially shortens solving time. As a result, SAT-based causal discovery is potentially more scalable under weaker assumptions.

This paper has two main parts. In the first part, we motivate a kind of indeterminate, suppositional credences by discussing the prospect for a subjective interpretation of a causal Bayesian network, an important tool for causal reasoning in artificial intelligence. A CBN consists of a causal graph and a collection of interventional probabilities. The subjective interpretation in question would take the causal graph in a CBN to represent the causal structure that is believed by an agent, and interventional probabilities in a CBN to represent suppositional credences. We review a difficulty noted in the literature with such an interpretation, and suggest that a natural way to address the challenge is to go for a generalization of CBN that allows indeterminate credences. In the second part, we develop a decision-theoretic foundation for such indeterminate suppositional credences, by generalizing a theory of coherent choice functions to accommodate some form of act-state dependence. The upshot is a decision-theoretic framework that is not only rich enough to, so to speak, ground the probabilities in a subjectively interpreted causal network, but also interesting in its own right, in that it accommodates both act-state dependence and imprecise probabilities.

Causal discovery becomes especially challenging when the possibility of latent confounding and/or selection bias is not assumed away. For this task, ancestral graph models are particularly useful in that they can represent the presence of latent confounding and selection effect, without explicitly invoking unobserved variables. Based on the machinery of ancestral graphs, there is a provably sound causal discovery algorithm, known as the FCI algorithm, that allows the possibility of latent confounders and selection bias. However, the orientation rules used in the algorithm are not complete. In this paper, we provide additional orientation rules, augmented by which the FCI algorithm is shown to be complete, in the sense that it can, under standard assumptions, discover all aspects of the causal structure that are uniquely determined by facts of probabilistic dependence and independence. The result is useful for developing any causal discovery and reasoning system based on ancestral graph models

In “Flagpoles anyone? Causal and explanatory asymmetries”, James Woodward supplements his celebrated interventionist account of causation and explanation with a set of new ideas about causal and explanatory asymmetries, which he extracts from some cutting-edge methods for causal discovery from observational data. Among other things, Woodward draws interesting connections between observational causal discovery and interventionist themes that are inspired in the first place by experimental causal discovery, alluding to a sort of unity between observational and experimental causal discovery. In this paper, I make explicit what I take to be the implicated unity. Like experimental causal discovery, observational causal discovery also relies on interventions, albeit interventions that are not carried out by investigators and hence need to be detected as part of the inference. The observational patterns appealed to in observational causal discovery are not only surrogates for would-be interventions, as Woodward sometimes puts it; they also serve to mark relevant interventions that actually happen in the data generating process.

Forster presented some interesting examples having to do with distinguishing the direction of causal influence between two variables, which he argued are counterexamples to the likelihood theory of evidence. In this paper, we refute Forster's arguments by carefully examining one of the alleged counterexamples. We argue that the example is not convincing as it relies on dubious intuitions that likelihoodists have forcefully criticized. More importantly, we show that contrary to Forster's contention, the consilience-based methodology he favored is accountable within the framework of the LTE.

Different directed acyclic graphs may be Markov equivalent in the sense that they entail the same conditional indepen- dence relations among the observed variables. Meek characterizes Markov equiva- lence classes for DAGs by presenting a set of orientation rules that can correctly identify all arrow orienta- tions shared by all DAGs in a Markov equiv- alence class, given a member of that class. For DAG models with latent variables, maxi- mal ancestral graphs provide a neat representation that facilitates model search. Earlier work has identified a set of orientation rules sufficient to con- struct all arrowheads common to a Markov equivalence class of MAGs. In this paper, we provide extra rules sufficient to construct all common tails as well. We end up with a set of orientation rules sound and complete for identifying commonalities across a Markov equivalence class of MAGs, which is partic- ularly useful for causal inference.

Much of the recent work on the epistemology of causation has centered on two assumptions, known as the Causal Markov Condition and the Causal Faithfulness Condition. Philosophical discussions of the latter condition have exhibited situations in which it is likely to fail. This paper studies the Causal Faithfulness Condition as a conjunction of weaker conditions. We show that some of the weaker conjuncts can be empirically tested, and hence do not have to be assumed a priori. Our results lead to two methodologically significant observations: (1) some common types of counterexamples to the Faithfulness condition constitute objections only to the empirically testable part of the condition; and (2) some common defenses of the Faithfulness condition do not provide justification or evidence for the testable parts of the condition. It is thus worthwhile to study the possibility of reliable causal inference under weaker Faithfulness conditions. As it turns out, the modification needed to make standard procedures work under a weaker version of the Faithfulness condition also has the practical effect of making them more robust when the standard Faithfulness condition actually holds. This, we argue, is related to the possibility of controlling error probabilities with finite sample size (“uniform consistency”) in causal inference.