Recent advancements in causal inference have made it possible to gain profound insight about our world and the complex systems which operate in it. While industry professionals and academics in every domain ask questions of their data, traditional statistical methods often fall short of providing conclusive answers. This is where causality can help.

This book gives readers the tools necessary to use causal inference in applied settings by building from theoretical foundations all the way to hands-on case studies in Python. We wrote this book primarily for the practitioner who knows how to work with data but may not be familiar with causal inference concepts, or how to apply those concepts to real-world problems.

Contents by Chapter

Part 1 begins by motivating why causality is a promising resource and laying the foundation of the necessary concepts from causal inference, culminating in an understanding of the potential outcomes framework. After this, we explore the full causal estimation process, providing the tools necessary to go from an initial question and a dataset to the creation and evaluation of causal estimates. We then provide an overview of causal discovery, which allows us to learn causal structures from observational data. The contents of Part 1 are as follows:

• Chapter 1 first seeks to motivate the reader by answering the question “why should I use causality?” It does this by providing illustrative examples of Simpson’s paradox and spurious correlation, and introducing how causal methods address these problems. The chapter then describes the paradigm shift between the traditional statistical or machine learning workflow to that of a causal inference setting. The chapter closes by giving a brief history of causal inference – from Judea Pearl’s early work to modern applications in industry – and providing a list of existing resources for those wishing to learn more.
• Chapter 2 introduces the potential outcome framework to explain concepts and theoretical foundations of causal effects. By exploring the key ideas and theories, this chapter aims to deepen readers’ understanding of cause-and-effect relationships. A comprehensive case study is presented, utilizing a well-known causal study conducted with an educational television program. This practical example aims to give readers a firsthand understanding of the discussed concepts. The chapter concludes by addressing various challenges from the Gelman and Hill book, enabling readers to develop the necessary skills to contemplate, model, and apply the theory discussed throughout the chapter.
• Chapter 3 delves further into the world of causal modeling, building upon the foundational knowledge established with the potential outcome framework and fundamental causal concepts. This chapter explores causal graphs as a practical approach for inferring causal relationships. We introduce causal graphs, discuss the high-level process of causal inference, examine various methods and techniques, and present a comprehensive case study with the Lalonde dataset for comparative analysis.
• Chapter 4 probes into the challenges of constructing causal models in practice and highlights the emergence of causal discovery techniques based on observational data as an alternative. It introduces a range of techniques developed for this purpose and presents an array of causal discovery algorithms, explaining their relative strengths and limitations. The chapter concludes with a real-world case study, showcasing the practical utility of these algorithms in uncovering causal relationships.

We shift gears in Part 2 to discuss how causal inference is currently being used within other sub-domains of machine learning, including computer vision, natural language processing, and in time-dependent settings. These are the chapters in Part 2:

• Chapter 5 focuses on how to apply methods of causal inference natural language processing, specifically for data that includes text. We consider how to compute causal effect sizes when the treatment and/or outcome is text, with or without the presence of confounding text. We include a case study analyzing film revenue data.
• Chapter 6 details the intersection of causality and computer vision. It introduces the ever-present issues of spurious correlation and confounding in image data – problems well-suited for applications of causal methods. The chapter showcases research efforts to apply causal inference techniques in specific areas of computer vision, including image classification and visual question-answering. The chapter concludes with a case study of causal methods designed to improve robustness, using an adversarial transfer dataset.
• Chapter 7 explores a recent method for time-dependent causal inference that is able to not only determine causation, but the temporal delay between the cause and effect variables. At present, this chapter does not describe time-dependent causal inference in the presence of confounding associations. We include a case study using an open-source bike sharing dataset.

In Part 3, we discuss some advanced topics within the field of causality:

• Chapter 8 explores a special case of causal inference: assessing model fairness. The chapter provides a high-level introduction to the issues of algorithmic bias, describes existing non-causal approaches of measuring unfair bias, and presents an argument in favor of using causal model fairness techniques. It details causal approaches and additional considerations for the confounders that may exist in fairness settings. The chapter closes with a case study comparing causal and non-causal methods on the infamous COMPAS dataset.
• Chapter 9 contains an overview of cutting-edge applications of causality in reinforcement learning, including techniques to improve world models in model-based RL, merge online and offline data, improve sample efficiency, and explain agent incentives. This chapter closes with a discussion of the challenges preventing large-scale adoption of causal RL technique.