When are graphical causal models not good models?

Cybernetics and Systems, 2008

In this paper we describe an important structure used to model causal theories and a related problem of great interest to semi-empirical scientists. A causal Bayesian network is a pair consisting of a directed acyclic graph (called a causal graph) that represents causal relationships and a set of probability tables, that together with the graph specify the joint probability of the variables represented as nodes in the graph. We briefly describe the probabilistic semantics of causality proposed by Pearl for this graphical probabilistic model, and how unobservable variables greatly complicate models and their application. A common question about causal Bayesian networks is the problem of identifying casual effects from nonexperimental data, which is called the identifability problem. In the basic version of this problem, a semi-empirical scientist postulates a set of causal mechanisms and uses them, together with a probability distribution on the observable set of variables in a domain of interest, to predict the effect of a manipulation on some variable of interest. We explain this problem, provide several examples, and direct the readers to recent work that provides a solution to the problem and some of its extensions. We assume that the Bayesian network structure is given to us and do not address the problem of learning it from data and the related statistical inference and testing issues.

Journal of Economic Methodology, 2005

A graphical model is a graph that represents a set of conditional independence relations among the vertices (random variables). The graph is often given a causal interpretation as well. I describe how graphical causal models can be used in an algorithm for constructing ...

arXiv (Cornell University), 2017

2005

Discrete Bayesian Networks (BNs) have been very successful as a framework both for inference and for expressing certain causal hypotheses. In this paper we present a class of graphical models called the chain event graph (CEG) models, that generalises the class of discrete BN models. This class is suited for representing conditional independence and sample space structures of asymmetric models. It retains many useful properties of discrete BNs, in particular admitting conjugate estimation. It provides a flexible and expressive framework for representing and analysing the implications of causal hypotheses, expressed in terms of the effects of a manipulation of the generating underlying system. We prove that, as for a BN, identifiability analyses of causal effects can be performed through examining the topology of the CEG graph, leading to theorems analogous to the Backdoor theorem for the BN.

2013

Scientists often use directed acyclic graphs (days) to model the qualitative structure of causal theories, allowing the parameters to be estimated from observational data. Two causal models are equivalent if there is no experiment which could distinguish one from the other. A canonical representation for causal models is presented which yields an efficient graphical criterion for deciding equivalence, and provides a theoretical basis for extracting causal structures from empirical data. This representation is then extended to the more general case of an embedded causal model, that is, a dag in which only a subset of the variables are observable. The canonical representation presented here yields an efficient algorithm for determining when two embedded causal models reflect the same dependency information. This algorithm leads to a model theoretic definition of causation in terms of statistical dependencies.

Innovations in Bayesian Networks, 2008

Several paradigms exist for modeling causal graphical models for discrete variables that can handle latent variables without explicitly modeling them quantitatively. Applying them to a problem domain consists of different steps: structure learning, parameter learning and using them for probabilistic or causal inference. We discuss two well-known formalisms, namely semi-Markovian causal models and maximal ancestral graphs and indicate their strengths and limitations. Previously an algorithm has been constructed that by combining elements from both techniques allows to learn a semi-Markovian causal models from a mixture of observational and experimental data. The goal of this paper is to recapitulate the integral learning process from observational and experimental data and to demonstrate how different types of inference can be performed efficiently in the learned models. We will do this by proposing an alternative representation for semi-Markovian causal models.

Journal of Machine Learning Research

The goal of many sciences is to understand the mechanisms by which variables came to take on the values they have (that is, to find a generative model), and to predict what the values of those variables would be if the naturally occurring mechanisms were subject to outside manipulations. The past 30 years has seen a number of conceptual developments that are partial solutions to the problem of causal inference from observational sample data or a mixture of observational sample and experimental data, particularly in the area of graphical causal modeling. However, in many domains, problems such as the large numbers of variables, small samples sizes, and possible presence of unmeasured causes, remain serious impediments to practical applications of these developments. The articles in the Special Topic on Causality address these and other problems in applying graphical causal modeling algorithms. This introduction to the Special Topic on Causality provides a brief introduction to graphical causal modeling, places the articles in a broader context, and describes the differences between causal inference and ordinary machine learning classification and prediction problems.

EPSA Epistemology and Methodology of Science, 2009

National Conference on Artificial Intelligence, 2006

This paper addresses the problem of identifying causal effects from nonexperimental data in a causal Bayesian network, i.e., a directed acyclic graph that represents causal relationships. The identifiability question asks whether it is possible to com- pute the probability of some set of (effect) variables given intervention on another set of (intervention) variables, in the presence of non-observable (i.e., hidden

amstat.org

This paper develops a method of design-based causal inference from a sample to a finite population. Causal effects are defined as finite population parameters based on the causal structure of the population. Only a non-parametric model for the joint distribution of the causal structure, which can be expressed as a directed acyclic graphical model, must be assumed. The finite population causal effects are expressed in two equivalent forms: (1) as stratified sampling type estimators with the strata determined by the confounders, and (2) as Horvitz-Thompson type estimators with the selection probabilities replaced by the propensity score. Both forms lead to sampling estimators that are subtle variations on common estimators, and we provide a delta method for computing their approximate variances.