D’ya Like DAGs? A Survey on Structure Learning and Causal Discovery

Abstract

Causal reasoning is a crucial part of science and human intelligence. In order to discover causal relationships from data, we need structure discovery methods. We provide a review of background theory and a survey of methods for structure discovery. We primarily focus on modern, continuous optimization methods, and provide reference to further resources such as benchmark datasets and software packages. Finally, we discuss the assumptive leap required to take us from structure to causality.

Causal Machine Learning for Healthcare and Precision Medicine

“…Extensive background has been developed over the last 3 decades around discovering causal structures from observational data, as described in recent reviews of the subject [Peters et al, 2017, Glymour et al, 2019, Nogueira et al, 2021, Vowels et al, 2021. Most methods rely on conditional independence tests, combinatorial exploration over possible DAGs and/or assumptions about the data generation process' function class and noise distribution ( e.g.…”

A Causal Research Pipeline and Tutorial for Psychologists and Social Scientists

“…Application/Code Note : It is worth noting that many approaches to causal discovery exist, and readers are encouraged to consult surveys by Assaad, Devijver, and Gaussier [4], Glymour, Zhang, and Spirtes [45], and Vowels, Camgoz, and Bowden [154] for a discussion of various options. In this paper, we provide a working example using the Structural Agnostic Modeling (SAM) method by Kalainathan, Goudet, Guyon, Lopez-Paz, and Sebag [69].…”

“…Greenland, Pearl, and Robins [48], Heinze-Deml, Maathuis, and Meinshausen [55], and Vowels, Camgoz, and Bowden [154]. In general, approaches to causal discovery are split into a number of categories: Constraint based approaches, which test for conditional independence relationships; asymmetry based methods, which test for distributional asymmetries under different cause-effect directions; score-based approaches, which derive a structure based on the fit to the data; and finally, interventional approaches, which involve some kind of experimental manipulation.…”

“…In stage A we are concerned with theory development, and propose and iterated development and validation of an initial theory (one which may, for instance, have been informed by qualitative interviews) by accommodating techniques from the domain of causal discovery. This stage uses causal discovery techniques [154] to inform and augment our theories, by testing putative causal relationships in data.…”

“…One of the principal challenge with observational data, of course, is that causal effects are less readily derivable. Fortunately, in spite of the general reluctance to adopt them, there exist a wide variety of tools which facilitate the identification and estimation of causal quantities from observational data [45,51,67,95,154,156,157], and such tools are comparatively commonplace in other fields such as epidemiology [144] and economics [19,39]. Causal inference tools enable us to derive mathematical expressions for desired causal quantities in terms of observational quantities.…”