0% found this document useful (0 votes)

180 views45 pages

Causal Inference in Statistics, by Justin Belair

Uploaded by

Om Sai

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

180 views45 pages

Causal Inference in Statistics, by Justin Belair

Uploaded by

Om Sai

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 45

Causal Inference in Statistics

With Exercises, Practice Projects, and R Code Notebooks

(Draft - Introduction)

Justin Belair

2024
Contents

I Fundamentals of Causal Inference 13

1 The Causal Inference Revolution - Introduction and Examples 15
1.1 The Causal Inference Revolution : An Intellectual Landmark . . . . . . . . . . 15
1.2 The Battle for The Soul of Causal Inference . . . . . . . . . . . . . . . . . . . . 16
1.3 Lord’s Paradox–Decades of Statistical Debates Put To Bed? . . . . . . . . . . . 18
1.3.1 Lord’s Paradox In Brief . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.3.2 Statistician 1 - Change Score Approach . . . . . . . . . . . . . . . . . . 20
1.3.3 Statistician 2 - ANCOVA-style Approach . . . . . . . . . . . . . . . . . . 20
1.3.4 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.3.5 Duelling Statisticians and Their Competing Analyses . . . . . . . . . . 22
1.3.6 Potential Outcomes Approach . . . . . . . . . . . . . . . . . . . . . . . 23
1.3.7 Directed Acyclic Graph (DAG) Approach . . . . . . . . . . . . . . . . . 25
1.3.8 DAGs, SCMs, SEMs and Mediation Analysis . . . . . . . . . . . . . . . 27
1.4 To Adjust Or Not To Adjust? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
1.4.1 Simpson’s Paradox . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
1.4.2 Berkson’s Paradox, Selection Bias, or Collider Effects . . . . . . . . . . 34
1.5 When Genius Errs and Observational Studies Prevail–The Story of Smoking
and Lung Cancer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
1.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

2 Fundamentals of Potential Outcomes Theory 43

2.1 Historical Context, Development, and Motivation . . . . . . . . . . . . . . . . 43
2.2 Potential Outcomes and Causal Effects . . . . . . . . . . . . . . . . . . . . . . . 45
2.3 Gaining Strength From A Population . . . . . . . . . . . . . . . . . . . . . . . . 48
2.3.1 Randomization Inference–Does The Treatment Have An Effect At The
Individual Level? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
2.4 The Treatment Assignment Mechanism–How Experiments Differ From Obser-
vations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
2.4.1 Some Assignment Mechanisms Are Problematic, Some Are Useful . . 62
2.4.2 Classical Randomized Experiments . . . . . . . . . . . . . . . . . . . . 66
2.5 Average Treatment Effects (ATE) . . . . . . . . . . . . . . . . . . . . . . . . . . 70
2.5.1 Sampling From A Super-Population . . . . . . . . . . . . . . . . . . . . 71
2.5.2 Estimating Average Treatment Effects . . . . . . . . . . . . . . . . . . . 72
2.5.3 Average Treatment Effect for the Treated (ATT) and Average Treatment
Effect for the Controls (ATC) . . . . . . . . . . . . . . . . . . . . . . . . 76
2.6 Modes of Statistical Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

3
4 Contents

2.6.1Fisher’s Exact P-value (FEP) or Randomization-Based Tests of Sharp

Null Hypotheses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
2.6.2 Neyman’s Repeated Sampling Randomization-Based Inference . . . . 79
2.6.3 Repeated Sampling (Regression) Model-Based Inference . . . . . . . . 81
2.6.4 Bayesian Predictive Inference, or Full Probability Modelling . . . . . . 82
2.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
2.8 Case-Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
2.8.1 The Randomized Block Design in Practice . . . . . . . . . . . . . . . . . 91
2.9 Practice Project - Advanced . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
2.9.1 The Project . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
2.10 Further Readings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
2.11 Appendix - Algebraic Manipulations . . . . . . . . . . . . . . . . . . . . . . . . 94

3 Graphical Causal Models - A Picture Is Worth A Thousand Words 95

3.1 Historical Context, Development, and Motivation . . . . . . . . . . . . . . . . 95
3.2 Causal Diagrams–Translating Causal Hypotheses Into Observable Associations 96
3.2.1 Defining a DAG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
3.2.2 From DAGs to the More Familiar Territory of Probability . . . . . . . . 100
3.2.3 d-separation, Or How To Read Conditional Independencies Off A DAG 103
3.2.4 Spurious Associations Are Carried Along Backdoor Paths . . . . . . . 106
3.3 Common DAG Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
3.3.1 Confounding Bias . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
3.3.2 Colliders And Selection Bias . . . . . . . . . . . . . . . . . . . . . . . . 108
3.3.3 Mediator Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
3.3.4 M-Bias and Butterfly Bias . . . . . . . . . . . . . . . . . . . . . . . . . . 112
3.4 Identifying Causal Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
3.4.1 Causal Bayesian Networks, Or Interventions As Graph Surgery . . . . 115
3.4.2 Diagrams as Non-Parametric Structural Models . . . . . . . . . . . . . 119
3.4.3 [[observational equivalence?]] . . . . . . . . . . . . . . . . . . . . . . . 121
3.5 Selecting Adjustment Sets? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
3.6 Case-Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
3.7 Bibliographical Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
3.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121

4 Observational Studies 125

4.1 Experiments versus Observational studies . . . . . . . . . . . . . . . . . . . . . 126
4.2 The Role of Randomization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
4.3 Counterfactuals and DAGs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
4.4 Adjustment : Stratification, Regression Adjustment, Weighting Methods . . . 127
4.5 G-Computation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
4.6 Case-Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
4.7 Negative Controls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
4.8 Study Designs Widely Used in Epidemiology . . . . . . . . . . . . . . . . . . . 127
4.8.1 Cohort Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
4.8.2 Case-Control Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
4.9 Target Trial Emulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
4.10 Sensitivity Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
4.11 Reporting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
4.12 Bibliographical Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
4.13 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
© Justin Belair
Contents 5

II The Causal Inference Toolkit 129

5 Removing Confounding and Propensity Score Methods 131
5.1 The Propensity Score–Can We Understand The Assignment Mechanism By
Looking At The Covariates? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
5.2 Propensity Score in Observational Settings . . . . . . . . . . . . . . . . . . . . 134
5.3 Estimating the Propensity Score . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
5.4 Matching, IPW, Stratification, Covariate Adjustment with the Propensity Score 134
5.5 Sensitivity Analysis, Selection Bias, and Moderation Analysis . . . . . . . . . 134
5.6 Case-study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
5.7 Bibliographical Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
5.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134

6 Instrumental Variables 135

6.1 Randomization as an Instrument . . . . . . . . . . . . . . . . . . . . . . . . . . 135
6.2 Instrumental Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
6.2.1 Compliance in Clinical Trials : A Formalization . . . . . . . . . . . . . 136
6.2.2 Per-Protocol vs. As-Treated Analyses vs. Intention to treat (ITT) . . . . 137
6.3 Case-Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
6.4 Bibliographical Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
6.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139

7 Structural Equation Models and Path Analysis 141

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
7.1.1 History of randomization with Fisher . . . . . . . . . . . . . . . . . . . 141
7.1.2 Limits of randomization . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
7.2 DAGs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
7.2.1 D-separation tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
7.3 Path coefficients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
7.3.1 Correlated Errors in a DAG . . . . . . . . . . . . . . . . . . . . . . . . . 148
7.4 Structural Equations Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
7.5 Path Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
7.6 Case-Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
7.7 Bibliographical Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
7.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148

8 Mediation and Interaction Analysis 149

8.1 From Baron and Kenny to the modern causal inference approach . . . . . . . 149
8.2 Mediation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
8.3 Interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
8.4 Case-study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
8.5 Bibliographical Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
8.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149

9 Quasi-Experimental Designs 151

9.1 Apply differences-in-dfferences to real-world problems . . . . . . . . . . . . . 151
9.2 Apply Regression discontinuity to real-world problems . . . . . . . . . . . . . 151
9.3 Apply interrupted time-series to real-world problems . . . . . . . . . . . . . . 151
9.4 Synthetic Controls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
9.5 Bibliographical Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
9.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
© Justin Belair
III Advanced Topics in Causal Inference 153
10 Longitudinal Causal Inference 155
10.1 Hazard Ratios Have Built In Selection Bias . . . . . . . . . . . . . . . . . . . . . 155
10.2 Time-varying vs. Time-modified confounding . . . . . . . . . . . . . . . . . . 155
10.2.1 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
10.3 Marginal Structural Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
10.4 Robin’s G-formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
10.5 Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
10.6 Bibliographical Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
10.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158

11 Machine Learning and Causal AI 159

11.1 Targeted-Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
11.2 Meta-learners . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
11.3 DoubleML . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
11.4 Causal Discovery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159

IV Closing Up 161
12 Where To Go From Here? 163
12.1 Single-World Intervention Graphs . . . . . . . . . . . . . . . . . . . . . . . . . 163
12.2 Causal Inference with Interference . . . . . . . . . . . . . . . . . . . . . . . . . 163
12.3 Causal Inference with Cyclicality . . . . . . . . . . . . . . . . . . . . . . . . . . 163
12.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
Preface

Another Book on Causal Inference!?

This book stands on the shoulders of many great textbooks on causal inference.

A very gentle, short, and captivating non-technical book is [Rosenbaum, 2023]. A good

introductory book which eases into technical material is [Pearl et al., 2016], yet it has some

flaws in my opinion. Indeed, it does not go too far into the subject, as it is simply a primer

and takes around a third of the space to go over basic probability theory and graph theory

before really diving into causal inference. This short book thus leaves very little space to go

deeper in causal inference.

My aim has been to take the next step from this book. That is, bring you, dear reader,

up-to-date on current statistical methodology in causal inference! I made the book self-

contained, so we will go over all the necessary theory, but we assume you, dear reader, is

familiar with probability and statistics and leave all this basic theory in the Appendix 12.4,

for your reference.

I attribute [Pearl, 2009] with sparking my curiosity to go deeper in the fascinating sub-

ject of causal inference - it’s a true masterpiece and I recall fondly the intellectual awaken-

ing it sparked in me when I read it for the first time. It is a difficult book written with the

goal of rigorously defining Pearl’s approach to causal inference, which relies a lot on ad-

vanced mathematics–at least, more advanced than what a traditional applied (data) scientist

is equipped to handle. [Imbens and Rubin, 2021] is another very elegant textbook which I

loved reading, but it is also heavy on mathematics. [Morgan and Winship, 2014] is rich with

references to literature and real-world examples taken from decades of causal debates in so-

cial sciences, but it can be a bit heavy. We will also discuss epidemiology as it has become in-

7
8 Contents

creasingly technical, with much overlap with what was traditionally called biostatistics. Plus,

epidemiology relies almost uniquely on observational data, where causal inference method-

ology is indispensable. As we will see, a major intellectual breakthrough in causal inference

comes from our understanding of when and how we can analyze observational data as if

it was experimental, so-to-speak. Great references for epidemiology and more specifically

causal inference within the field, are [Rothman et al., 2008] and [Westreich, 2020]. There are

even more specialized books like [VanderWeele, 2015] and [Shipley, 2000a], an ecology book,

that will be referenced throughout. All this is without mentioning the 100s of papers I’ve

read over the years that you will find in the bibliography. This sprawling literature on causal

inference serves as a foundation to the book you’re about to read.

As a trained mathematician and statistician who’s taught dozens of university-level courses

to business students, data scientists, engineers, and researchers, I have the necessary back-

ground to decipher these complex textbooks and bring out the essence for non-mathematicians,

without having to go through long and rigorous mathematical proofs. I’ve also worked as

a statistical consultant with dozens of scientists from around the world in many biomedical

sciences, as well as researchers in psychology, psychiatry, agronomy, epidemiology, and even

Research and Development (R&D) in biotechnology industries. I’ve been applying causal

inference in applied scientific research for a long time now, through my services offered at a

company I founded and run on a daily basis, JB Statistical Consulting1 . I’m also a philoso-

phy dilettante, who’s been mostly interested in figuring out how this beautiful thing we call

Science actually works and how to do it well, through reading philosophy of science, history

of science, epistemology, and more. All this is to say that I (again, very humbly) believe I’m

uniquely positioned to take the complex mathematical and philosophical ideas behind causal

inference, decipher them and bring them to a level where an applied scientist can use them.

Plus, I absolutely love sharing knowledge I’m passionate about!

Don’t get me wrong, in this book, there will be some mathematics, and even some proofs,

as you will see by skimming through its contents, especially in the first few Chapters, where

we lay the conceptual foundation on which we will later build our causal inference toolbox.

This is simply because the causal inference revolution we present in Chapter 1 is mainly one

of finding a mathematical language to clearly and formally express ideas of causality. Rest
1 www.justinbelair.ca

assured, mathematics will not be the main goal and kept to a minimum. All the necessary

mathematical notations, definitions, and theorems will be stated and explained clearly. As

I just described in the previous paragraph, this book is NOT aimed at mathematicians, al-

though I sincerely hope they would still learn from it!

I don’t believe in the so-called dumbing down of content; I’d rather elevate the reader.

Learning anything worthwhile is usually difficult, requires patience and perseverance and

is not always pure pleasure. Causal inference is not different–it is admittedly a difficult sub-

ject, but oh so interesting! My promise to you is to do my part to make things as clear and

interesting as possible. And, believe me, there will definitely be lot’s of fun amidst the diffi-

culty, so don’t give up when it gets tough!

Rather than being submerged by complex theorems and their proofs, you will be treated

to elegant graphics representing the core ideas of the book, a suite of R notebooks with code

that you can download for free to play around and practice with data, as well as a series

of exercises to help you truly master the subject–something that is lacking in many causal

inference textbooks. I also like to incorporate historical nuggets–they bring out the people

and the stories behind these great ideas, giving them a bit more humanity. All the topics will

be accompanied by extensive references to the literature, if ever you, dear reader, wish to go

further!

Should I Read This Book? If So, Do I Have To Read Cover-to-

Cover?

Before answering the title-question of this section, let me rapidly go over the layout of the

book.

The first part of the book is formed by 4 chapters. The introductory chapter serves as a

guiding light through the journey we are about to embark on together. Chapters 2 and 3 form

the theoretical backbone of causal inference and Chapter 4 will discuss in detail the problem

of uncovering causal effects when we cannot perform experimental manipulations. It is here

that we will introduce the formal theories in sufficient detail for them to be understood clearly

and rigorously, especially to later apply them.

The second part of the book forms the bulk of the tools that are used by applied re-
© Justin Belair
10 Contents

searchers to study causal relationships in the data they care about, whether experimental

scientists in pharmacology, social scientists using observational data for program evaluation,

or psychologists trying to understand human behavior. We will tackle propensity score meth-

ods in Chapter 5, instrumental variables in Chapter 6, Structural Equation Models (SEM) in

Chapter 7, analysis of mediation and interactions in Chapter 8, and quasi-experimental meth-

ods in Chapter 9, such as differences-in-differences. Chances are that if you’re a scientist or

researcher who routinely uses data, you’ve heard some of these terms before. You might

even be familiar with some of these methods. In my humble opinion, this book can still be

extremely valuable owing to the fact that, as far as I know, no book combines a thorough

review of the theory with as much practical material through exercises and coding examples.

The final part of the book is dedicated to the frontier of causal inference in scientific re-

search, tackling advanced methods such as longitudinal causal inference and time-varying

confounding in Chapter 10. Methods related to machine learning and AI such as targeted

learning, double machine learning, meta-learners, and causal discovery in Chapter 11. These

recent and more advanced methods are extremely promising, as they build on the insights

that have been gleaned from a solid few decades of research in causal inference.

Obviously, the second and third parts of the book are more application-focused. Their

chapters can all be read individually and independently, even if they are loosely ordered by

conceptual difficulty. If you are strictly interested in learning how to apply methods and use

code to do so, you can tackle these chapters directly. On the other hand, the first part is mostly

focused on developing a rigorous conceptual approach to causal inference and might seem

difficult or even alien to you, depending on your background and your familiarity with math-

ematical ideas and abstract, conceptual thinking. In my opinion, I believe the first part of the

book is very valuable, even to readers who already understand both the potential outcomes

framework associated with the work of Donald Rubin and the graphical causal models (i.e.

DAGs) attributed to Judea Pearl, or readers that don’t have a particularly strong background

in mathematics. If I did not think them valuable, I would not have put them in the book, as

nobody has time to waste! Still, you might wish to skip over them, skim them rapidly, or go

back to them to help you apply the tools of the two latter parts of the book. Multiple different

approaches can work for you, depending on what kind of reader you are.

When writing the book, I was driven by the desire to make it easily navigable, both when
© Justin Belair
Contents 11

reading cover-to-cover or when simply picking and choosing subjects of interest. For this

reason, I also added an Appendix to cover some building blocks that you might not have

studied before, such as certain concepts in probability theory or statistics.

Fundamentals of Causal Inference

Chapter1

The Causal Inference Revolution -

Introduction and Examples

1.1 The Causal Inference Revolution : An Intellectual Land-

mark

We’re experiencing nothing short of an intellectual revolution1 . In a 2021 paper titled What are

the most important statistical ideas of the past 50 years?2 , counterfactual causal inference is the

first subject listed. Concepts of cause-and-effect, counterfactual storytelling, and learning

about the world through its manipulation (i.e. actions or interventions) are, as far as we

know, extremely intuitive and innate to human-beings. Consider the following examples.

• Cause and Effect. When it rains (sufficiently), the ground is wet.

• Manipulating the world through actions and interventions. When I drop something,

it falls. Yet, if I drop something and hold it by a string, I can prevent it from falling all

the way to the ground.

• Counterfactual storytelling. What if I hadn’t eaten my leftover seafood? I probably

wouldn’t be sick right now.

1 Ithas (almost) nothing to do with Large-Language Models (LLM), or chatGPT, a popular chatbot developed
by American tech company OpenAI. But causality is a very active area of reasearch in Artificial Intelligence (AI),
considered more broadly as the attempt to understand and emulate human intelligence. Indeed, there is no explicit
causal modelling behind chatGPT, and some researchers believe that achieving true Artifical General Intelligence
(AGI)–an ill-defined concept–will require some form of causality-based models.
2 [Gelman and Vehtari, 2021]

15
16 Chapter 1. The Causal Inference Revolution - Introduction and Examples

These sorts of causality-related utterances seem to be used routinely and with ease by

human beings. As a matter of fact, they seem to be indispensable to living as a human being

in the world. Children are known to go through phases of extreme playfulness in which it is

believed they learn about the world surrounding them through intervening on it and judging

the effects of their causal actions.

Yet, for all this ease and simplicity of causality in everyday life, the same cannot be said

scientifically and philosophically. It is only very recently that we have come to grips with

formal (i.e mathematical) language to discuss this sort of reasoning clearly and rigorously.

Computer scientist and mathematician Judea Pearl, a towering figure in the study of causality,

puts it like this:

In the last decade, owing partly to advances in graphical models, causality has

undergone a major transformation: from a concept shrouded in mystery into a

mathematical object with well-defined semantics and well-founded logic. Para-

doxes and controversies have been resolved, slippery concepts have been expli-

cated, and practical problems relying on causal information that long were re-

garded as either metaphysical or unmanageable can now be solve using elemen-

tary mathematics. Put simply, causality has been mathematized. [Pearl, 2009,

p.xiii]

This is the fascinating causal inference revolution I wish to present in this book!

1.2 The Battle for The Soul of Causal Inference

To make things even more intellectually captivating, there is a full-scale battle being waged

for the soul of causal inference. On one hand, there is the potential outcomes framework

which was revitalized and developed through the work of statistician Donald Rubin begin-

ning in the 1970s (sometimes called the Neyman-Rubin potential outcomes frameork, or the

Rubin Causal Model (RCM)), which we will thoroughly discuss in Chapter 2. On the other

hand, there is the theory of Structural Causal Models (SCM), a very visual graphical ap-

proach to causality, most strongly associated with Pearl. A major tool of this approach is the

Directed Acyclic Graph (DAG). He developed this work through his long and productive

career beginning also in the 1970s, for which he was even awarded the Turing prize
© Justin Belair
1.2. The Battle for The Soul of Causal Inference 17

For fundamental contributions to artificial intelligence through the development

of a calculus for probabilistic and causal reasoning3 .

We will learn about this “calculus for probabilistic and causal reasoning”, the so-called

do-calculus, in Chapter 3.

For all these fascinating advances in causal inference, Pearl and Rubin disagree. Pearl

has been more keen to engage with the ideas from the opposing camp, most notably dis-

cussing them at different places in his work4 . Rubin, on the other hand, seems to want

nothing to do with pearlian ideas related to causality. In his monograph on causal infer-

ence, [Imbens and Rubin, 2021], a major reference for the book you are about to read, there

is exactly one paragraph, totalling around 100 words, devoted to Pearl’s approach to causal

inference.

Pearl’s work is interesting [...] In our work, [...] we have not found this approach

to aid drawing of causal inferences, and we do not discuss it further in this text5 .

Rubin was also interviewed in 2024 to discuss his famous paper on the propensity score,

an early landmark in causal inference (i.e. [Rosenbaum and Rubin, 1983]), a subject we tackle

in Chapter 5. At a certain point in the conversation, Rubin is expressing his thoughts what he

calls a “big idea”, by which he more or less means mathematical simplicity and beauty. To

him, the (Neyman-)Rubin Causal Model, the subject of Chapter 2, is in competition with the

graphical approach of Chapter 3:

For a while, I think there was this competition. I think the competition for RCM,

for this Rubin Causal Model potential outcomes approach, was the graphical stuff.

You know, it was the DAG stuff of Judea Pearl. And Pearl’s a great character, but

I always found just the DAG stuff to be clutter, because it added something that

wasn’t essential at all. [...] As soon as you get to something with experimental

design, for example, split-plot models or things like that, complicated, not just

simple randomization, the whole things falls apart. The whole simplicity of DAGs

goes away. So why keep pursuing it6 ?

3 [Russell, 2011]
4 For example, in [Pearl,
2009], the entry for Rubin in the Name Index list 26 places where he is discussed by Pearl.
5 [Imbensand Rubin, 2021, p.22]
6 https://youtu.be/eF12bVsvq2Q?si=n0LUEB3xoFLL5wWV&t=770

While there is, in my opinion, an incredible beauty behind the potential outcomes ap-

proach, especially when it comes to discussing concepts like experimental design, treatment

assignment mechanisms, and randomization, I argue from a purely subjective point of view

that the graphical approach of Pearl contains itself unmistakable elegance, especially when

trying to uncover causal effects from observational studies that contain complex relation-

ships between many variables. We will discuss observational studies throughout, but espe-

cially in Chapter 4. Of course, these are both subjective opinions.

Despite this competition, so-to-speak, there is a growing number of enthusiastic researchers,

in economics, sociology, political science, epidemiology, and more, which have embraced both

causal inference traditions and have taken a pragmatic approach to choosing the best concep-

tual tools from each framework to answer their scientific questions with as much rigour and

ease as possible. This is not merely a subjective appreciation of mathematical elegance, but

rather pragmatic evidence that both approaches are useful, at least to certain camps of em-

pirical researchers.

The overarching goal of this book will be to explore both these frameworks in depth

(Chapters 2 and 3), but even more so to unite and compare their ideas when possible. It

is my point of view that any user of causal inference methods is confronted with answer-

ing scientific research questions, rather than debating foundational issues. We can draw a

parallel to the unsettled debates for the foundations of probability theory (Frequentism vs.

Bayesianism), which do not bother any empirical researcher, even as they rely consistently on

probability theory to do research and advance science7 . This pragmatic approach to causal

inference has led researchers to develop ideas, most notably through the notion of a counter-

factual, that allow for a partial unification of these frameworks. We will devote a good deal

of space to discussing this in Chapter 4.

1.3 Lord’s Paradox–Decades of Statistical Debates Put To Bed?

Let’s move on to important intellectual contributions that causal inference have brought to

the table - solving of statistical paradoxes.

7 If we believe Linkedin and Twitter, these debates are extremely important, as they garner much engagement and

make for easy-to-publish content. I argue that in the real scientific literature, empirical researchers don’t care, for
better and for worse.

In 1967, Frederic M. Lord, considered by the National Council on Measurement in Edu-

cation to be “the father of modern testing”, published a famous paper titled A Paradox in the

Interpretation of Group Comparisons8 in which two hypothetical statisticians using two differ-

ent, but defensible, analysis strategies come to contradictory conclusions. Lord also discussed

the same type of problem in two other hypothetical examples in his 1968 paper Statistical Ad-

justments When Comparing Preexisting Groups 9 . This problem has come to be known as Lord’s

Paradox.

After explaining the nature of the paradox, we will simulate data to get a good grasp

of the purported problem we have to face. We will then draw on modern insights to solve

the paradox, leveraging causal inference approaches, namely those of Holland & Rubin in

On Lord’s Paradox, from a 1982 technical report [Holland and Rubin, 1982] and from Pearl’s

Lord’s Paradox Revisited - (Oh Lord! Kumbaya!), published in 2016 [Pearl, 2016].

This example sits, perhaps due to coincidences, at the confluence of both the intellectual

battle for the soul of causal inference being waged between the Neyman-Rubin potential out-

comes framework10 and Pearl’s approach11 , and as a pivotal example at the heart of statistics

history.

A brief note : throughout, we assume familiarity with Analysis-of-Covariance, (ANCOVA)12 .

1.3.1 Lord’s Paradox In Brief

Lord’s example is concerned with comparing preexisting groups (e.g. biological sex). He

acknowledges that

it is widely recognized that ideally the research worker should assign cases or

individuals at random to the groups that are to be studied ... In behavioral re-

search and in many other areas, such random assignment is usually difficult or

impossible 13 .

Precisely, the example is concerned with assessing

8 [Lord, 1967]
9 [Lord, 1968]
10 See Chapter 2
11 See Chapter 3
12 https://en.wikipedia.org/wiki/Analysis_of_covariance
13 [Lord, 1967, p.304]

the effects on the students of the diet provided in the university dining halls and

any sex differences in these effects14 .

The available data concerns the weight of each student gathered at the beginning of the

school year in September (for which we will use X), and at the end, the following June (for

which we will use Y ). The sex of each student is recorded in our variable G.

The crux of the matter lies in analyzing either the so-called change scores, Y − X in both

sex groups, or rather regressing Y on X in both groups, through an ANCOVA-style approach.

1.3.2 Statistician 1 - Change Score Approach

In Lord’s example, the average change for BOTH groups is 0, that is Ȳ − X̄ = 0. Therefore,

the first statistician concludes that as far as these data are concerned, there is no

evidence of any interesting effect of the school diet (or of anything else) on student

weight. In particular, there is no evidence of any differential effect on the two

sexes15 .

1.3.3 Statistician 2 - ANCOVA-style Approach

Lord states that the second statistician finds that

the slope of the regression line of final weight [i.e. Y ] on initial weight [i.e. X] is

essentially the same for the two sexes. [Yet,] the difference between the intercepts

is statistically highly significant. The second statistician concludes, as is custom-

ary in such cases, that the boys showed significantly more gain in weight than the

girls when proper allowance is made for differences in initial weight between the

two sexes16 .

Obviously, the statisticians cannot be both right. There cannot be no effect at all of diet on

weight, and a significant effect of diet on weight. These statements are plainly contradictory!

14 [Lord, 1967, p.304]

15 [Lord, 1967, p.305]
16 [Lord, 1967, p.305]

1.3.4 Simulation

Let’s simulate data reproducing the main features of Lord’s example. Here, G ∈ {A, B}, A

for girls and B for boys, using Lord’s terminology17 .

Next, Lord states

although the weight of individual boys and girls has usually changed during the

course of the year, [...] the group of girls considered as a whole has not changed

in weight, nor has the group of boys18 .

Drawing the scatterplot of Y and X in Figure 1.1, we have in grey the girls and in black the

boys. The large dots represent the mean values for each group, and the diamond represents

the overall mean. The means all fall on the dashed 45◦ line, representing equal X and Y

values, or no changes in the means from September to July, in both groups. Obviously, the

data varies around this mean, as would be typical in a real-life setting19 .

Figure 1.1: The weight in September (Y) in relation to the weight in June (X) for Boys (black)
and Girls (grey)–group means are large dots and overall mean is diamond

17 The code is omitted for clarity, but available for download on Github. As a matter of fact, you can find the code

used to generate every single Figure and Table from the book.
18 [Lord, 1967, p.304]
19 The change scores would almost never be exactly 0, but omitting this consideration simplifies the analysis with-

out Lord’s arguments losing any force.

1.3.5 Duelling Statisticians and Their Competing Analyses

Statistician 1

As stated earlier, the Statistician 1, presumably after plotting the figure above, would compute

change scores in both groups and presumably used a paired t-test to discern if the effect is

significantly different from 0, as shown in Table 1.1.

group estimate statistic p.value CI

Girls 0.11 0.16 0.87 [-1.23 ; 1.45]
Boys 0.30 0.49 0.63 [-0.91 ; 1.51]

Table 1.1: Pairwise Two-Sided T-Tests Comparing Weigth in September (Y) with Weigth in
June (X), for Girls and Boys

Hurray! The results of the t-test confirm that there are no significant weight differences

on average in both groups.

Statistician 2

The second statistician, perhaps feeling a bit more sophisticated, is of the opinion that the

proper way to analyze this is through ANCOVA, so she first plots the ANCOVA regression

lines in Figure 1.2.

Then, seeing the magnitude of the difference between intercepts, she proceeds to compute

the proper estimates to test this group difference. She indeed finds that both the group coeffi-

cient (Table 1.2) and the F-test (Table 1.3) on the group parameter in the ANCOVA are highly

significant, contradicting statistician 1! [[Difference between beta parameter and F-test]]

term estimate std.error statistic p.value

(Intercept) 10.85 2.22 4.90 <0.01
X 0.48 0.10 4.57 <0.01
group 7.59 1.69 4.48 <0.01

Table 1.2: ANCOVA Model Coefficients for the Weight of Students in June (Y) regressed on
the Weight of Students in September (X) and Group

What gives–how can data like this exist!? As the concrete simulated data shows, this is not

simply a theoretical idea20 . This problem has been at the core of decades of debate and dis-

cussions by statisticians, with the most recent breakthroughs coming from developments in
20 Real-life examples can be found in [Ross, 2004], [Charig et al., 1986], and [Bickel et al., 1975].

Figure 1.2: The weight in September (Y) in relation to the weight in June (X), with ANCOVA
regression lines for both Boys (black) and Girls (grey)

term sumsq df statistic p.value

X 551.98 1.00 20.85 <0.01
group 531.51 1.00 20.08 <0.01
Residuals 3891.93 147.00

Table 1.3: ANCOVA Model F-Tests for the Weight of Students in June (Y) regressed on the
Weight of Students in September (X) and Group

causal inference. Can we solve this paradox and crown victorious the statistician that emerges

unscathed from this statistical duel?

1.3.6 Potential Outcomes Approach

21
In Holland and Rubin , the problem is attacked using the Neyman-Rubin potential out-

comes framework, which will be discussed in depth in Chapter 2. Concretely, they define

• Population (P ) : The students at the university in the specified school

• Treatment (T ) : The dining hall diet

• Control (C) : ??? (No control!)

21 [[Holland Rubin citation]]

• Treatment indicator (S) : S = t for all units in P

• Gender (G) : Student Gender (A = Girls, B = boys)

• Weight of Student in December (X)

• Weight of Student in June (Y )

The first problem raised is that of the absence of a control group. It is a fundamental

idea in experimental science that to compare the effect of a treatment, or an experimental

manipulation, there must an alternative called a control treatment, presumably identical in

all aspects to the so-called active treatment, but inactive. Indeed, the question “Did a certain

diet lead to weight gain?” is ill-posed in this framework. The question can only be rigorously

asked by qualifying the statement like such : “Did a certain diet (T ) lead to weight gain, when

compared to another diet (C)?”. This leads to defining two alternative potential outcomes

for Y , one under the treatment Y (T ) and one under the control Y (C), which we wish to

compare. Ideally, we would then randomize a sample of units either to T or C. But Lord never

mentioned an alternative diet. Rather, he was explicit that the two groups to be compared

were based on G, the gender:

It is widely recognized that ideally the research worker should assign cases or

individuals at random to the groups that are to be studied by analysis of covari-

ance. In behavioral research and in many other areas, such random assignment

is usually difficult or impossible22 .

He proceeds to give the example of comparing racial groups, but he clearly places Gender

(G) in the same category as that of a variable representing preexisting groups to be compared.

In Lord’s pessimistic conclusion, he says

The researcher wants to know how the groups would have compared if there had

been no preexisting uncontrolled differences23 .

The role of “if” in the previous statement, is what philosophers call “counterfactuals”,

or “counterfactual language”. Counterfactuals describe “alternative worlds” so-to-speak,

22 [Lord, 1967, p.304]
23 [Lord, 1967, p.305, emphasis mine]

where things have gone differently than in our world, which is assumed to be real and unique24 .

This counterfactual language, if placed in an experimental framework, would presuppose

that either we could randomize students to a given initial weight (X) at the experiment’s

onset, or that we could assign them to a given gender (G), which would in turn partially

determine their initial weight (X). Of course, this is not possible in a true experiment.

Therefore, Rubin and Holland famously quipped “no causation without manipulation”,

although Holland later softened his position 25 . There is debate around what counterfactuals

mean, if anything, due to their abstract, or even metaphysical nature. Indeed, if it’s possible

to imagine alternative worlds where some events happen differently, why wouldn’t it be pos-

sible, at least in an abstract sense, or even in a thought experiment, to imagine what would

have been different had a participant had a different initial weight (X)? Or even a different

gender (G) altogether?

We will discuss Rubin and Holland’s paper in more depth once we’ve put in place the

proper notation and formal language required to discuss potential outcomes.

1.3.7 Directed Acyclic Graph (DAG) Approach

In 2016, Pearl published a paper in which he takes the position of having decisively put Lord’s

Paradox to bed26 . Essentially, he stands as having solved this longstanding problem, a van-

tage point from which he is purportedly able to point out the shortcomings and missteps in

all previous discussions around the issue. In his own words :

The purpose of this paper is to trace back Lord’s paradox from its original for-

mulation, resolve it using modern tools of causal analysis, explain why it resisted

prior attempts at resolution and, finally, address the general methodological issue

of whether adjustments for preexisting conditions is justified in group compari-

son applications27 .

The first step in any Pearlian approach to causal problems is establishing a graphical

model representing the causal relationships between the variables under study. We will for-
24 We will not go in depth in philosophical debates surrounding counterfactuals and causality, but will rather refer

the reader to appropriate literature when warranted. We will assume a somewhat naı̈ve view of causality, which we
will describe a bit later.
25 [Holland, 2003]
26 [Pearl, 2016]
27 [Pearl, 2016, Abstract]

mally denote and define all these concepts later in Chapter 3, but one of the main advantages

of DAGs as conceptual tools is their (superficially) simple and intuitive nature.

The first step is to represent all variables under study as nodes in a graph. Then, using

unidirectional arrows, causal links between variables can be drawn. Indeed, if changes in A

cause changes in B, we draw an arrow from A to B. The real challenge lies in establishing

where to place arrows - if we knew all the causal links between variables with certainty, our

study would be of less relevance. Still, for sake of exposition, we will assume the DAG Pearl

posits as being appropriate for the situation at hand. We draw it in Figure 1.3.

Figure 1.3: A Directed Acyclic Graph (DAG) Of Lord’s Paradox

Pearl introduces a new variable, Z which is simply the change score Z = Y −X and points

two arrows into Z, one from X and one from Y . This is obvious, since Z is defined precisely

equal to the difference of Y and X. Then, he posits that gender is a direct cause of both X

and Y , a sensible assumption. Finally, he affirms that the initial weight X is a cause of the

final weight Y 28 . His final assumption is that all the relationships in his diagram are uncon-

founded, which for now simply means that the estimates of causal effects are unbiased by
28 Some readers might already feel uneasy of treating Gender as a cause of a person’s weight, yet this is often done

in research, based on solid pragmatic arguments. Pearl simply shrugs off the problem by reassuring us that if we
wanted, the Gender variable could be replaced by a nexus of hormone variables that distinguish between biological
sexes. This opens up a philosophical can-of-worms on the ontological status of social-construct study variables. We
will not go down this rabbit-hole for now.

unobserved variables not included in the model. This is obviously a very strong assumption,

as there could always be hidden confounders29 .

Recall that Pearl is interested in the following counterfactual statement, or at least some-

thing close to it : “What would the subject’s weight be in September if their Gender had been dif-

ferent?.” Obviously, this is abstract language but it shouldn’t be too concerning as humans

routinely make and use these kinds of counterfactual statements to understand and navigate

the world30 .

1.3.8 DAGs, SCMs, SEMs and Mediation Analysis

At this stage, a few notes are in order. As Pearl insists in multiple different places throughout

his work, a DAG causal model is non-parametric, i.e. the postulated causal relationships are

not assumed to follow any specific functional form that could be reduced to a handful of

parameters. These non-parametric models are sometimes referred to as Structural Causal

Models (SCM). More informally, by non-parametric we mean that an arrow in a DAG is

indeed a function mapping between two variables, but that the function is not necessarily

described using a few symbols (i.e. parameters)–the function can be any complex function,

even one that doesn’t have neat equations to describe it31 . On the contrary, a parametric

function is usually defined by a small collection of values, called parameters. In statistics, we

often used parametric models to simplify analyses. For example, the function describing the

gaussian distribution can be boiled down to two parameters, the mean µ and variance σ 2 .

The next step taken to make SCMs more useful, is imposing an additional fully parametric

model structure. This essentially has the effect of restricting the functions linking variables

together to have predefined forms which can then be estimated by estimating a handle of
29 In this book, we will discuss what confounding means for causal estimation and how this phenomenon is for-

mally represented in both the potential outcomes approach and the DAG approach to causal inference. We will get
back to the notion of confounding especially in Chapters 2 and 3 but it will be discussed throughout. It is a central
idea in modern science that has been defined and re-defined many times. In my opinion, this notion only makes
sense when viewed through the lens of causal models, which is why it has had a tortuous existence.
30 [Shpitser, 2012] says :

Human beings reach an intuitive consensus on the meaning of many causal utterances and there have
been numerous attempts to formalize causality in a way that is faithful to this consensus.
See also the references cited therein.
31 A reminder that a function is a concept that tells us how to transform, or map, an input value to an output value.

For example, y = f (x) means that given input value x, the function f returns the output value y. A function that
is not defined by an equation is sometimes said to have no analytical form. This idea is abstract, but an easier way to
grasp the idea is to think of such a function as a huge lookup table. Given an input, say x again, we know how to
find the output, here y, by looking in the table and finding x. Yet, this table has no clear, concise equation to describe
it.

parameters. This tool is known as the Structural Equation Model (SEM) and is widely used

in many areas of research, most notably in social sciences and biology32 .

Conceptually, the easiest fully parametric model to work with is one where we impose

functions to be linear, in addition to giving every variable it’s own independent gaussian

error term. Linear, gaussian SEMs enjoy the property of being simply multivariate gaussian

distributions with a given correlational structure. This essentially conceptualizes a DAG as

a series of interwoven linear models. The mathematical simplicity afforded by this (strong)

set of assumptions even allows the analyst to work with latent variables, that is, unobserved

variables that we know are part of the causal story we wish to unravel. We will discuss this

in much detail later in Chapter 7. Still, without taking the extra-step of moving from a DAG

to and SEM, we can still say interesting things about the relationships between the variables

in Lord’s paradox, when represented using Pearl’s DAG. This is one of the major strengths of

the DAG approach–taken at its core, it relies on few assumptions yet still uncovers important

truths about the system of variables being studied.

When following arrows in a DAG in their proper direction, we get directed paths. On the

other hand, undirected paths, where we follow arrows without concern for their direction,

will also be important when discussing confounding and the back-door criterion in Chapter

3. For example, G → X → Y → Z is a directed path from G to the outcome of interest,

Z = Y − X. We thus say that X is a mediator between G and Z; it is affected by Gender and

in turn affects the change score in Weight. It is also a mediator between G and Y . Stated in

words, we say that the weight in September mediates the relationship between Gender and

the outcome (either weight in June or change score in weights).

We also have a path G → Y, which is of interest if we don’t look at the change score Z, but

rather the final weight in September Y. This path is unmediated, and represents a direct effect.

It might already becoming clear that DAGs give us a neat conceptual tool to look at effects

of variables and how they travel along networks of cause-effect relationships. We will see in

Chapter 7 that what was traditionally considered simply as an effect of a variable on another,

can in fact be looked at more specifically. This is sometimes called “mediation analysis”. We

will define notions of total effect, controlled direct effect, natural direct effect, and natural

32 Social sciences and Biology encompass so many disciplines, that these alone account for a huge chunk of empir-

ical research being done today.

indirect effect. We will then discuss the extremely important concept of identification: given

our Structural Causal Model and the data on hand, what kind of effects can be estimated

from the data? Path analysis, that is estimating effects along paths and using sets of rules

for combining them, is a very old technique dating back to [Wright, 1921] that has seen a

resurgence in applications. It will also be discussed in Chapter 7. Now, back to Lord’s Paradox

and its interpretation by Pearl.

Total Effects, Direct Effects, and Indirect Effects

The total effect of G on Z is akin to varying G and letting the effect of this variation propagate

through the graph and looking how this affects the change score Z. Formally, this amounts

to estimating

Total Effect = E(Z|G = Boy) − E(Z|G = Girl)

= E(Y − X|G = Boy) − E(Y − X|G = Girl)

= E(Y |G = Boy) − E(X|G = Boy) − E(Y |G = Girl) − E(X|G = Girl,

| {z } | {z }
average change in weight for boys average change in weight for girls

where the last equation is given by linearity of expectation. In the equation above, the ex-

pression we called average change in weight for boys can be estimated by taking the boys and

computing Ȳ − X̄. Equivalently, the average change in weight for girls can be estimated by doing

the same computation using the group of girls33 . This is of course what Statistician 1 did in

Section 1.3.2 when he computed Ȳ − X̄ in both groups and found them both to be 0.

33 In this book, we will focus on causality and identification of causal effects of interest. That is, we will be more

interested in determining if an estimation procedure can be reasonably assumed to estimate the true causal effect
of interest. Obviously, estimating a causal effect using a plausible causal modeling strategy will not absolve us of
the traditional duty that statisticians must observe. That is, by adding a layer of causal inference to an analysis, a
statistician is not suddenly absolved of all traditional issues that statistics have been developed to tackle : sampling
variability, research design, measurement error, inferring from a sample to a population, maximizing use of infor-
mation in choosing estimators, etc. If anything, causal inference makes the duty of the statistician greater when
conducting an analysis.

On the other hand, the direct effect34 is rather

Z
Direct Effect = [E(Z|G = Boy, X = x) − E(Z|G = Girl, X = x)] P(X = x|G = Girl)dx
Zx
= [E(Y − X|G = Boy, X = x) − E(Y − X|G = Girl, X = x)] P(X = x|G = Girl)dx
x
Z
= [E(Y |G = Boy, X = x) − E(Y |G = Girl, X = x)] P(X = x|G = Girl)dx.
x | {z }
(∗)

To jump from the second line to the third in the above derivation, we simply used the lin-

earity of expectation along with that fact that by conditioning on X = x, there is no more

randomness involved in X. This yields

E(X|G = Boy, X = x) = x

and

E(X|G = Girl, X = x) = x,

so both of them cancel out.

Recall, Statistician 2 using ANCOVA in Section 1.3.3. Indeed, ANCOVA proceeds by esti-

mating a linear model for the conditional expected value of the final weight Y as such

E(Y |G, X) = β0 + β1 G + β2 X.

Therefore, the expression (∗) becomes

β0 + β2 X − (β0 + β1 + β2 X) = β1 , (1.1)

if we set the value of G for boys to be 1 and for girls to be 0, as is traditionally done for binary

variables in regression models. Putting Equation 1.1 back into the expression for the direct

34 This expression is given in [Pearl, 2016]. We will also rely heavily on [Pearl, 2014] for Chapters 7 and 8 where

we will formally define all these concepts and discuss when we can identify such effects. That is, when the effects
can be parsed out from the data we have on hand.

effect, we get

Z
Direct Effect = β1 P(X = x|G = Girl)dx
x

= β1 ,

R
since β1 does not depend on X and can be moved out of the integral and x
P(X = x|G =

Girl)dx = 1, since we are integrating a probability density function. Thus, Statistician 2, by

looking at the beta coefficient associated with the Gender variable in an ANCOVA, is estimat-

ing the direct effect under the assumption of a linear model for the expected value–which

was found to be significantly different from 0. No wonder Statistician 1 and Statistician 2 do

not agree, they are not estimating the same thing!

The conceptual clarity provided by the concepts, tools, and language associated with

causal inference allows us to move beyond the nebulous so-called paradox illustrated by Lord

in his famous paper. When everybody speaks the same language, the paradox dissolves. The

idea of dissolving paradoxes is exciting! How about we attack other longstanding so-called

paradoxes that have caused much confusion and debate in statistics?

1.4 To Adjust Or Not To Adjust?

It is a basic scientific principle that is repeated ad nauseam among many scientists that proper

adjustment, or control, must be incorporated in statistical models. This is usually done by

including variables such as Age, Sex, etc. that we believe are related somehow to the causal

effect we are trying to estimate. Usually, there is an incentive to include all available variables

in the adjustment–the more the better. Yet, without proper causal formalism, this can turn

out to be ill-advised and even further bias the effect of interest. We will also learn in Chapter 3

that it is possible to eliminate bias resulting from confounding without necessarily including

all confounders! We will also add a simple and intuitive online tool to our toolkit that, given

a causal DAG, selects the smallest set of confounders needed to properly control for when

estimating causal effects. But first, some paradoxical examples!

1.4.1 Simpson’s Paradox

In [Simpson, 1951], popularized a curious and apparently paradoxical aspect of data anal-

ysis. Indeed, Simpson presents a data table in which a hypothetical experiment involving a

treatment given to male and female patients. The death status of each patient is then clas-

sified according to the patient’s sex, and whether the patient received the treatment or not.

The Table 1.4 is a scaled-up version of the one given in [Simpson, 1951, p.241]35 .

Male Untreated Treated Female Untreated Treated

Alive 800 1600 Alive 400 2400
Dead 600 1000 Dead 600 3000

Table 1.4: Patients’ Survival Rates in Simpson’s Original Example

An elementary procedure learned in most Statistics 101 classes is that this sort of 2 × 2

contingency tables involving counts can be used to test the probabilistic independence of the

row and column variables through the chi-square test, as in Table 1.5.

Sex statistic p.value parameter method

Male 7.33 <0.01 1 Pearson’s Chi-squared test
Female 6.77 <0.01 1 Pearson’s Chi-squared test

Table 1.5: Chi-Square Tests For Simpson’s Original Example

The results of the chi-squared tests applied to the male and female tables separately are

both highly significant, which strongly suggests that there is a significant positive effect of the

treatment on death–for both males and females, the treatment reduces the chance of death.

Yet, if we look at the data on aggregate, as in Table 1.6 the chi-square statistic is perfectly

equal to 0 - the data fits exactly what we would expect to observe if treatment status and death

were completely independent of one another, as we see in Table 1.7.

Aggreagated Untreated Treated

Alive 1200 4000
Dead 1200 4000

Table 1.6: Aggregated Patients’ Survival Rates in Simpson’s Original Example

We could go even further. It is easy to construct similar examples where the aggregate
35 We scale up the values from Simpson’s original table to increase the power of the chi-square tests of indepen-

dence as to render the observed differences statistically significant, to further reinforce the paradoxical element of
this data.

statistic p.value parameter method

0.00 1.00 1 Pearson’s Chi-squared test

Table 1.7: Chi-Square Test For Simpson’s Original Example (Aggregated)

data does not collapse to a null effect but is rather reversed as compared to the sign of the effect

in the subgroups. This means, for example, that one could observe that a treatment is positive

for both males and females, but negative for the overall population. [Pearl et al., 2016, p.2]

eloquently state how disconcerting this paradoxical result should be :

The data seem to say that if we know the patient’s gender–male or female–we can

prescribe the drug, but if the gender is unknown we should not! Obviously, that

conclusion is ridiculous. If the drug helps men and women, it must help anyone;

our lack of knowledge of the patient’s gender cannot make the drug harmful.

What is the analyst left to do? Should their be an adjustment for sex of should the data be

considered in aggregate?

Answering this question satisfactorily cannot be done by statistical arguments alone. The

only satisfactory answer relies on causal inference concepts, namely understanding what

kind of causal relationship holds between our Sex, Treatment, and Death variables. More

precisely, the adjustment for Sex should only be considered if Sex is a confounder of the

causal relationship between treatment and death.

The most basic confounding that sex S can have on the relationship between a treatment

D and an outcome Y can be represented by the following DAG.

The (undirected) path X ← S → Y is called a backdoor path. As we will see in Chapter

3, we can properly identify and estimate causal effects on DAGs by using a criteria known as

d-separation, that entails blocking all backdoor paths. Here, we would block the path, or

d-separate the cause from the effect, by conditioning (or adjusting for, or controlling for) sex

S when we estimate the association between D and Y . It is a common belief that adjusting

for each and every variable is a valid strategy, perhaps after pruning those which don’t pass

significance tests in our model36

36 We will discuss variable selection in a causal context, where we will look at when and how to include variables

that could potentially confound our causal effect of interest. We will also learn how to use the propensity score
estimates in Chapter 5, a very useful technique when the number of confounders is large, which can cause practical
challenges.
In any case, it is strongly advised NOT to use stepwise (forward or backward) selection algorithms. Unfortunately,

Figure 1.4: A DAG of Basic Confounding of X on Y by S

If the Sex variable turns out to be a collider in our causal model, adjusting for it would be

catastrophic. Lets consider another example. You guessed it, it is also labelled a paradox.

1.4.2 Berkson’s Paradox, Selection Bias, or Collider Effects

In [Berkson, 1946], numerical examples for 2 × 2 contingency tables are given, much in the

same spirit than [Simpson, 1951], as discussed in Section 1.4.1. Berkson was worried that

in case-control studies37 , selecting hospitalized patients could lead to a spurious association

between exposure and disease (or between disease and disease). Indeed, as he shows, if

the disease and exposure of interest causally affect the risk of being hospitalized, selecting

these tools are common, as they can be implemented automatically without much knowledge. This is precisely why
they are wrong. I suggest reading Chapter 4, most notably Section 4.3 of [Harrell, 2015] for an excellent discussion
of this topic. A free version is available online at [Harrell, 2024]. Harrell also gives workshops related to the book.
37 A case-control study is an important epidemiological design used in observational studies that is usually learned

in Epidemiology 101. As we will see in Section 1.5, Berkson thought this argument weakened the possibility that
smoking tobacco had a causal effect on development on lung cancer. The paradox was contentious for a while in the
research community, due to its peculiar nature. Indeed, the case-control design consists in finding all the diseased,
so-called cases, in a certain population, called the study base. To ascertain whether there is an association between
the case disease and an exposure of interest, another pool of disease-free units is selected from the study base as
controls. Then, the exposure-disease association is measured. This may seem like a hopelessly biased approach,
as we oversampled all the diseased from the population–the strategy seems to violate a fundamental intuition of
selecting representative samples by, say, using simple random sampling from a well-defined population. Yet, as it
turns out, from this design we can estimate odds-ratios without bias. Still, the design is extremely tricky. To my
knowledge, there is no better discussion of case-control studies than in [Rothman et al., 2008], a must-read textbook
for anyone interested in epidemiology.

hospitalized patients can induce a spurious exposure-disease association. This came to be

known as Berkson’s Paradox, Berkson’s Fallacy or Berkson’s Bias. It is a form of selection

bias, a common term that represents many different forms of bias that result when a certain

selection mechanism for units of observation misrepresents the true causal effect of interest.

Consider the following famous example, where we assume that intellectual ability and

athletic ability are uncorrelated. The scatter-plot represents a population of units for which

we have some measure of both these variables. It is apparent in the data-cloud that the two

variables are not associated. Indeed, the regression line has slope 0, and no other patterns

are detected.

Figure 1.5: Intellectual ability and athletic ability are completely independent in the popula-
tion, as attested by the flat regression slope

Now, consider we select a portion of the units for study, say graduates from the most pres-

tigious university in the world38 . Assume further that prestigious universities select their stu-

dents based on their intellectual ability and/or their athletic ability39 This can be represented

by the following DAG.

This procedure will tend to select students in the upper-right corner of the data cloud :

38 I will let the reader imagine the most prestigious university without committing to any!
39 Especially in North America, college sports teams are vital to branding efforts by the universities, and thus there

are complex selection processes used to find the best athletes possible.

Figure 1.6: A DAG showing collider bias - Controlling or adjusting for college admission
status, a collider, can create a spurious correlation between intellectual and athletic abilities

students with high intellectual ability are further on the right of the X−axis, and students

with high athletic ability are further up on the Y −axis.

Figure 1.7: When selected into college, a (biased) negative correlation between intellectual
ability and athletic ability is created in the selected group

When we isolate these students, we create a spurious negative association between in-

tellectual and athletic abilities. Intuitively, this is simply because a student selected on high

intellectual ability will, on average, have lower athletic ability, and vice-versa. Thus, when

looking only at the units selected, the regression slope is negative, an artificial association

that we’ve created through selection bias, or conditioning on a collider. That is, when two

arrows in a DAG point to a common variable, this variable is called a collider–the causes

collide at a common effect. It might first seem counter-intuitive that conditioning on a vari-

able can create a spurious association between two of its causes, but hopefully the university

graduates example gives you a good intuition of why this bias occurs.

We will learn to correctly identify confounders and colliders in DAGs, while taking care

to not mix them up and properly address their related issues in Chapter 3. We will also learn

about a slew of so-called selection biases and represent them with DAGs. As we will see, all

forms of selection bias are induced by conditioning on colliders.

1.5 When Genius Errs and Observational Studies Prevail–The

Story of Smoking and Lung Cancer

In 1959, Sir Ronald A. Fisher, a British statistician and geneticist widely considered to be one

of the most influential scientists of all-time, wrote the following in a Letter to the Editors of

Nature, a prestigious scientific journal :

The association observable between the practice of cigarette smoking and the in-

cidence of cancer of the lung [...] has been interpreted [...] almost as though it

demonstrated a causal connection between these variables[Fisher, 1958d, emphasis

mine].

He added in another paper :

Unfortunately, considerable propaganda is now being developed to convince the

public that cigarette smoking is dangerous [...] [Fisher, 1958b]

Yet, in an ironic twist of fate, it has been well established that the propaganda machine

was working in exactly the opposite camp. Tobacco companies were found to be running
© Justin Belair
38 Chapter 1. The Causal Inference Revolution - Introduction and Examples

vast and secretive campaigns to influence public opinion in favor of tobacco smoke, despite

having themselves conducted much research on the topic, that all pointed towards tobacco

smoke being dangerous to human health40 . Was Fisher a victim of this propaganda?

The overwhelming evidence for the association between smoking and lung cancer was not

considered controversial to Fisher. Rather, he refused to acknowledge the straightforward

causal explanation for this association. He knew very well that once an association between

two variables (e.g., X : smoking, and Y : lung cancer) has been established with a high degree

of certainty, we find ourselves in one of three positions as pertains to their causal connection:

1. X causes Y ;

2. Y causes X, sometimes called reverse causation;

3. Common unmeasured variables cause both X and Y, a phenomenon known as con-

founding.

In the quote above, it is clear that Fisher excluded the first possibility listed above. He

even suggested the second possibility, that lung cancer could indeed cause smoking habits,

a rather odd point of view:

Is it possible, then, that lung cancer—that is to say, the precancerous condition

which must exist and is known to exist for years in those who are going to show

overt lung cancer—is one of the causes of smoking cigarettes? I don’t think it can

be excluded [Fisher, 1958a, p.162]41 .

40 There is now mounting evidence that the playbook used by tobacco companies to obscure scientific facts by sewing

seeds of doubt in the population is being replayed when a large industry feels threatened by scientific evidence of the
harms said industry is causing. The most distressing case comes from climate change deniers and greenwashing, but
the same story is found over and over, in debates around the health concerns of asbestos, the impacts of insecticides
on bee pollination and biodiversity, and more. See [Oreskes and Conway, 2019] for an excellent discussion on so-
called Merchants of Doubt.
41 Fisher goes on to further justify this point of view with curious anecdotal evidence, if we can even call it that :

I don’t think we know enough to say that it is such a cause. But the pre-cancerous condition is one
involving a certain amount of slight chronic inflammation. The causes of smoking cigarettes may be
studied among your friends, to some extent, and I think you will agree that a slight cause of irritation—a
slight disappointment, an unexpected delay, some sort of a mild rebuff, a frustration—are commonly
accompanied by pulling out a cigarette and getting a little compensation for life’s minor ills in that way.
And so, anyone suffering from a chronic inflammation in part of the body (something that does not
give rise to conscious pain) is not unlikely to be associated with smoking more frequently, or smoking
rather than not smoking. It is the kind of comfort that might be a real solace to anyone in the fifteen
approaching lung cancer. And to take the poor chap’s cigarrettes away from him would be rather like
taking away his white stick from a blind man. It would make an already unhappy person a little more
unhappy than he need be.

Of course, he had no data to back this claim up, as it is quite ludicrous. He did in fact pro-

vide data to back up his suggestion of the third explanation, that is that there are unmeasured

confounders that could explain the causal effect, specifically, genetic factors. We illustrate his

argument in the following very simple DAG.

Figure 1.8: A DAG showing Fisher’s argument for a hidden confounder - According to Fisher,
a common cause for both tendency to smoke and lung cancer could explain the observed
association between the two

The gist of the argument is that there could be genetic factors that influenced both smok-

ing habits and the probability of developing lung cancer, creating a confounding bias in the

estimates of the effect. While it was less controversial to assume that genetic factors were a

direct cause of cancer, the argument for the cause-and-effect relationship between the same

genetic factor and the proclivity to become a tobacco smoker were less evident. His argument

was based on data that showed that in a few dozen pairs of monozygotic twins, their smok-

ing behaviour was much more likely to be identical than in dizygotic twins, [Fisher, 1958c]42 .

Let’s look again at Fisher’s argument using our DAG, this time highlighting the fact of the

uncertainty of genetic factors directly causing tendency to smoke.

42 A simplistic version of his argument says that people with the same genes are more likely to have similar be-

haviors than people with different genes. The use of twins helps to remove confounding from other factors, namely
that people with different genes are also more likely to live in different environments, which in turn impacts their
behavior.

Figure 1.9: A DAG showing Fisher’s argument for a hidden confounder - with dashed lines
representing unconclusive evidence

In a very limited sense, Fisher was right, as his argument was based on the absence of ran-

domization. By randomizing X, we automatically exclude the second and third possibilities

listed above. Why? Because, by definition, the only cause of X is the randomization scheme!

Fisher understood this better than anybody, since he was the first to truly develop the theory

of the randomized experiment. Yet, if we believed Fisher had the last word in causal infer-

ence, many scientific problems would forever remain insoluble. The causal effect of cigarette

smoking on lung cancer is a case in point: it is both unethical and unfeasible to conduct a

randomized experiment to settle this matter. Fisher said it himself :

[The proponents of the causal effect of cigarette smoking on lung cancer] cannot

produce evidence in which a thousand children of teen age have been laid under a

ban that they shall never smoke, and a thousand more chosen at random from the

same age group have been under compulsion to smoke at least thirty cigarettes

a day. If that type of experiment could be done, there would be no difficulty,

[Fisher, 1958a, p.155].

There are other cases where we can never uncover causal mechanisms through random-

ization. Take as established that smoking causes lung cancer, which we will symbolize by
© Justin Belair
1.5. When Genius Errs and Observational Studies Prevail–The Story of Smoking and Lung Cancer 41

X → Y . A more complete description would involve a nexus of physiological processes that

we choose to leave unexplained, as a black-box. We could write this X → B → Y, where

B denotes the unexplained black-box phenomena by which physiological mechanisms turn

cigarette smoking into lung cancer. The issue is that many sciences focus specifically on these

sort of mechanisms, sometimes at cellular levels. Obviously, there are many cases where it is

impossible to conduct randomized studies at this level of detail. What are we to do?

At the time of publication of these criticisms, Fisher was closing in on 70 years of age, and

his best scientific days were behind him. It took a new generation of researchers to establish

the idea that even without randomization, it is possible to make causal claims. Aren’t we glad

they did! Among other factors, these debates spurred researchers to discover when and why

observational evidence could be trusted. The use of observational data to discover causal effects

is often what is meant by causal inference, as experiments were widely considered the only

way to detect causality.

Today, it is one of the most well-established scientific facts in Public Health and Epidemi-

ology that cigarette is not only harmful to health, but it is one of the worst possible behaviors

one can have in terms of impact on health outcomes. It not only causes lung cancer, but also

a great deal of deadly health conditions43 . Of course, there were also many prominent sci-

entists, such as Sir Austin Bradford Hill, who held the position that the association found

over-and-over again between tobacco smoke and lung cancer was indeed causal.

How did we go from a debate about hidden confounders to having scientifically estab-

lished a clear scientific consensus on the causal effect of smoking on lung cancer through

strictly observational means? A major step was the introduction of sensitivity analysis, a

framework which allows to measure the amount of bias hidden confounding must bring in

order to eliminate the effects measured44 .

43 See [Doll, 1998], who claims that among nearly 40 diseases or causes of death which are positively associated

with cigarette smoke, a great majority have been found to be causal. The paper also serves as a fascinating overview
of the history of changing social and political attitudes towards the mounting scientific evidence of the harms caused
by tobacco.
44 Paul Rosenbaum, a student and collaborator of Rubin who has made tremendous contributions to causal infer-

ence, says in [Rosenbaum, 2023, p.64]

There are always covariates that were not measured. Always.

1.6 Summary

I hope this Chapter has given you a nice overview of what we’re about to learn together! We

end with some exercises. This is the only Chapter where we forego a case-study - expect one

at the end of every other Chapter.

Good luck!

1.7 Exercises

1. Go to Github and set up a free account. Star the repository (repo) associated with the

book, if you like it!

2. For each of the following scenarios, identify the cause and the effect. Explain your

reasoning.

(a) A farmer applies fertilizer to a crop and observes an increase in yield.

(b) A child plays outside in the rain without a coat and catches a cold.

decrease.

3. Explain why the causal conclusions stated below are wrong.

(a) On a daily basis, there is a strong positive correlation between ice cream sales

and temperature. Therefore, to help with global warming, we should prohibit

ice cream sales.

(b) There is a strong association between money spent on lawyer fees and probability

of divorce. Therefore, to avoid divorce, couples should be mindful of their lawyer

expenses and try to keep them at minimum.

4. For each situation, write a counterfactual statement that explores an alternative out-

come.

(b) “Since the city has built the new park, many families have been moving into the

area.”

(c) “I took the early bus at arrived on time for my meeting.”

5. Describe a real-world scenario (not related to the examples in the chapter) where Berk-

son’s Paradox might occur. Explain how the selection of a specific population could

create a spurious association between two variables.

6. Select a recent study or news article that discusses a cause-and-effect relationship (e.g.,

the impact of a public health intervention). Analyze the following:

• Identify the causal relationship being studied.

• Discuss the potential outcomes framework as it applies to this study.

• Suggest how a DAG could visually represent the causal relationships in this sce-

nario.

7. For each scenario below, identify potential confounding variables and discuss how they

might affect the causal interpretation:

• A study finds that students who participate in after-school tutoring perform better

on standardized tests.

• A new traffic law is implemented, and accidents decrease at intersections.

• A company notices that employees who attend wellness programs report higher

job satisfaction.

8. What key lessons did you learn about the importance of understanding the causal re-

lationships between variables in statistical analysis?

9. How do the concepts of confounding and selection bias impact the interpretation of

observational data in public health research?

Fundamentals of Potential Outcomes

Theory

2.1 Historical Context, Development, and Motivation

It is now well-known and almost universally applied in practice that experiments with ran-

domization of treatment assignment can help uncover causality in empirical science, indus-

trial research and development (R&D), data science, engineering, pharmaceutical research

and drug development, online experimentation (e.g. A/B testing), and more. Scientists have

always been experimenting, but it wasn’t until relatively recently that the role of randomiza-

tion was properly understood–a role we will cut and dissect in many different ways, as it is

truly one of the greatest achievements of humanity’s scientific enterprise. [Fisher, 1925] first

made the explicit case for randomization for uncovering causality, although use of chance

mechanisms by scientists can be found earlier in historical records1 . Most of the traditional

textbooks on design of experiments focused on industrial applications in manufacturing set-

tings and quality control, for example [Box et al., 2005], a masterpiece we will discuss in

this Chapter. With the advent of digital technology, there is now a large and growing de-

mand for expertise in experimentation and analysis of online randomized experiments that

is now added to the demand that already existed in more traditional sectors (i.e. pharma-

ceuticals, manufacturing, engineering, etc.). These experiments are done in marketing, web

development, application optimization, etc. and are often called A/B tests. According to

1 See for example [Peirce and Jastrow, 1884].

Causal Inference in Statistics - Solution M - Judea Pearl
No ratings yet
Causal Inference in Statistics - Solution M - Judea Pearl
75 pages
European E Democracy in Practice
No ratings yet
European E Democracy in Practice
359 pages
KPMG - MITSDE - Operational Anlaytics
No ratings yet
KPMG - MITSDE - Operational Anlaytics
11 pages
Transportation Engg
No ratings yet
Transportation Engg
56 pages
5374 21346 2 PB
No ratings yet
5374 21346 2 PB
70 pages
Computational Petroleum Laboratory Lab Manual JCT COLLEGE OF ENGINEERING
No ratings yet
Computational Petroleum Laboratory Lab Manual JCT COLLEGE OF ENGINEERING
65 pages
Causal Inference Book Part I-Ifqdve
No ratings yet
Causal Inference Book Part I-Ifqdve
158 pages
Introduction To Machine Learning and Data Science: by Myself and Slidedeck Ai:)
No ratings yet
Introduction To Machine Learning and Data Science: by Myself and Slidedeck Ai:)
6 pages
Mastering LLM Applications With LangChain and Hugging Face
No ratings yet
Mastering LLM Applications With LangChain and Hugging Face
526 pages
Regional Sustainability: Tewodros Abuhay, Endalkachew Teshome, Gashaw Mulu
No ratings yet
Regional Sustainability: Tewodros Abuhay, Endalkachew Teshome, Gashaw Mulu
12 pages
A Beginner - S Guide To Conversion Rate Optimization
No ratings yet
A Beginner - S Guide To Conversion Rate Optimization
8 pages
Critical Success Factors in Managing Modular Production Design Six Company Case Studies in Hong Kong, China, and Singapore - Lau - 2011
No ratings yet
Critical Success Factors in Managing Modular Production Design Six Company Case Studies in Hong Kong, China, and Singapore - Lau - 2011
16 pages
Chapter III - Revised No.2
No ratings yet
Chapter III - Revised No.2
10 pages
Probabilistic Seismic Loss Estimation of Aging Highway Bridges Subjected To Multiple Earthquake Events
No ratings yet
Probabilistic Seismic Loss Estimation of Aging Highway Bridges Subjected To Multiple Earthquake Events
21 pages
04 Activity 1 - STRATEGIC BUSINESS ANALYSIS
No ratings yet
04 Activity 1 - STRATEGIC BUSINESS ANALYSIS
2 pages
Operation Research Notes
100% (1)
Operation Research Notes
124 pages
Slime Journal For Students
No ratings yet
Slime Journal For Students
22 pages
DS100-2-Grp#4 Chapter 6 Advanced Analytical Theory and Methods Regression (CADAY, CASTOR, CRUZ, SANORIA, TAN)
No ratings yet
DS100-2-Grp#4 Chapter 6 Advanced Analytical Theory and Methods Regression (CADAY, CASTOR, CRUZ, SANORIA, TAN)
4 pages
(SpringerBriefs in Mathematics) Qi He, Le Yi Wang, George G. Yin - System Identification Using Regular and Quantized Observations - Applications of Large Deviations Principles-Springer (2013)
No ratings yet
(SpringerBriefs in Mathematics) Qi He, Le Yi Wang, George G. Yin - System Identification Using Regular and Quantized Observations - Applications of Large Deviations Principles-Springer (2013)
108 pages
Advances in Applied Mathematics and Global Optimization in Honor
No ratings yet
Advances in Applied Mathematics and Global Optimization in Honor
542 pages
Time Series Forecasting
100% (1)
Time Series Forecasting
52 pages
!!!discourse Analysis
No ratings yet
!!!discourse Analysis
12 pages
ST2195 Complete
No ratings yet
ST2195 Complete
430 pages
Chapter 2 Methology
No ratings yet
Chapter 2 Methology
17 pages
2017 Hu & Zhang - A Pathway To Learner Autonomy A Self-Determination Theory Perspective
No ratings yet
2017 Hu & Zhang - A Pathway To Learner Autonomy A Self-Determination Theory Perspective
11 pages
Project
No ratings yet
Project
10 pages
PyTorch - Advanced Deep Learning
No ratings yet
PyTorch - Advanced Deep Learning
237 pages
An Analysis On The English Teachers Strategies in Teaching Reading Comprehension SMP 1 Wonomulyo
No ratings yet
An Analysis On The English Teachers Strategies in Teaching Reading Comprehension SMP 1 Wonomulyo
1 page
Functional Data Analysis With R and Matlab
No ratings yet
Functional Data Analysis With R and Matlab
220 pages
Teachers' Viewpoint: Translanguaging in English-Instructed Classes
No ratings yet
Teachers' Viewpoint: Translanguaging in English-Instructed Classes
6 pages
AI in Software Testing
No ratings yet
AI in Software Testing
15 pages
Ppue Book
No ratings yet
Ppue Book
275 pages
Stat 331 Course Notes
No ratings yet
Stat 331 Course Notes
79 pages
Ace The Data Science Interview-1
No ratings yet
Ace The Data Science Interview-1
5 pages
Questions: Temesgen - Worku@aau - Edu.et
100% (1)
Questions: Temesgen - Worku@aau - Edu.et
3 pages
Information Entropy Fundamentals
No ratings yet
Information Entropy Fundamentals
19 pages
13 Colonies Lesson Plan
No ratings yet
13 Colonies Lesson Plan
4 pages
Introduction To Statistics in Python
100% (2)
Introduction To Statistics in Python
211 pages
STAT 650 - Foundations of Data Science Syllabus
No ratings yet
STAT 650 - Foundations of Data Science Syllabus
13 pages
Nnu-Net: A Self-Configuring Method For Deep Learning-Based Biomedical Image Segmentation
No ratings yet
Nnu-Net: A Self-Configuring Method For Deep Learning-Based Biomedical Image Segmentation
14 pages
Mathematical Foundations For Data Science
No ratings yet
Mathematical Foundations For Data Science
2 pages
OHSMS Manual v1.0
No ratings yet
OHSMS Manual v1.0
9 pages
Stock Market Analysis 0th Review
No ratings yet
Stock Market Analysis 0th Review
27 pages
DK Critical Appraisal EBM Harm
No ratings yet
DK Critical Appraisal EBM Harm
17 pages
Evaluation of Personnel
100% (2)
Evaluation of Personnel
7 pages
Use of Calculators in Examinations
No ratings yet
Use of Calculators in Examinations
3 pages
Won Fong K Lau Examining A Brief Measure of Parent Involvement in Children's Education
No ratings yet
Won Fong K Lau Examining A Brief Measure of Parent Involvement in Children's Education
12 pages
Research Perspectives Overview
100% (1)
Research Perspectives Overview
15 pages
Sanet ST
No ratings yet
Sanet ST
385 pages
2021 - Nature - Bayesian Statistics and Modelling
100% (1)
2021 - Nature - Bayesian Statistics and Modelling
26 pages
K - PPT - Simple Regression and Correlation
No ratings yet
K - PPT - Simple Regression and Correlation
21 pages
Sequence Spaces Kamthan Gupta
100% (1)
Sequence Spaces Kamthan Gupta
382 pages
List of Books - PDF
100% (1)
List of Books - PDF
335 pages
Bayesian Learning Methods
No ratings yet
Bayesian Learning Methods
57 pages
Binomial Distribution
No ratings yet
Binomial Distribution
16 pages
Fundamentals of Statistical Signal Processing Volume I Estimation Theory PDFDrive
No ratings yet
Fundamentals of Statistical Signal Processing Volume I Estimation Theory PDFDrive
603 pages
Exploratory Causal Analysis With Time Series Data - James M. McCracken (Morgan & Claypool, 2016)
100% (1)
Exploratory Causal Analysis With Time Series Data - James M. McCracken (Morgan & Claypool, 2016)
134 pages
Mathematical Statistics With Applications Solution Manual
No ratings yet
Mathematical Statistics With Applications Solution Manual
5 pages
Communication Systems and Techniques PDF
50% (2)
Communication Systems and Techniques PDF
2 pages
(Alan Holt) Network Performance Analysis Using TH PDF
No ratings yet
(Alan Holt) Network Performance Analysis Using TH PDF
228 pages
(Image Processing Series) Luciano Da Fona Costa, Roberto Marcond Cesar Jr. - Shape Classification and Analysis - Theory and Practice-CRC Press (2009) PDF
No ratings yet
(Image Processing Series) Luciano Da Fona Costa, Roberto Marcond Cesar Jr. - Shape Classification and Analysis - Theory and Practice-CRC Press (2009) PDF
674 pages
Entropy: Explainable AI: A Review of Machine Learning Interpretability Methods
No ratings yet
Entropy: Explainable AI: A Review of Machine Learning Interpretability Methods
45 pages
Notes
No ratings yet
Notes
422 pages
Bayesian Lecture Notes
No ratings yet
Bayesian Lecture Notes
28 pages
Horaratnam of Srimanmishra Balabhadra (Vol. 2) - Hindi PDF
No ratings yet
Horaratnam of Srimanmishra Balabhadra (Vol. 2) - Hindi PDF
114 pages
Numerical Methods - B. Ram
No ratings yet
Numerical Methods - B. Ram
236 pages
Formal Language Theory
No ratings yet
Formal Language Theory
69 pages
An Introduction To Deep Reinforcement Learning PDF
No ratings yet
An Introduction To Deep Reinforcement Learning PDF
140 pages
Solution 2
0% (1)
Solution 2
4 pages
STAT 0202 Outlier Syllabus - Spring 2020 - Regular Term (14-Wk) - 12-2
No ratings yet
STAT 0202 Outlier Syllabus - Spring 2020 - Regular Term (14-Wk) - 12-2
7 pages
Combinatorial Optimization: Alexander Schrijver
No ratings yet
Combinatorial Optimization: Alexander Schrijver
34 pages
Jyotish Kalpdrum Bhasha
100% (1)
Jyotish Kalpdrum Bhasha
64 pages
2015 406105 Satpanchasika
No ratings yet
2015 406105 Satpanchasika
115 pages
STAT613
No ratings yet
STAT613
295 pages
Pattern Recognition and Machine Learning Errata and Additional Comments
0% (1)
Pattern Recognition and Machine Learning Errata and Additional Comments
7 pages
Business Statistics: A First Course: Simple Linear Regression
No ratings yet
Business Statistics: A First Course: Simple Linear Regression
65 pages
Databook PDF
No ratings yet
Databook PDF
64 pages
Principles of Data Analysis - Prasenjit Saha (2003) PDF
No ratings yet
Principles of Data Analysis - Prasenjit Saha (2003) PDF
86 pages
MATH 115: Lecture IV Notes
No ratings yet
MATH 115: Lecture IV Notes
5 pages
Support Vector Machines: The Interface To Libsvm in Package E1071 by David Meyer FH Technikum Wien, Austria
No ratings yet
Support Vector Machines: The Interface To Libsvm in Package E1071 by David Meyer FH Technikum Wien, Austria
8 pages
Vemanasatakam
No ratings yet
Vemanasatakam
13 pages
COmputing and Combinatoria
No ratings yet
COmputing and Combinatoria
490 pages
Cheet Sheet
No ratings yet
Cheet Sheet
47 pages
Numpy Ref
No ratings yet
Numpy Ref
1,128 pages
Statistical Learning Theory
No ratings yet
Statistical Learning Theory
4 pages
Automated Reasoning With PROLOG
100% (1)
Automated Reasoning With PROLOG
5 pages
Hawk Eye Technology in Depth
No ratings yet
Hawk Eye Technology in Depth
35 pages
Parallel
No ratings yet
Parallel
13 pages
45 Genetic Algorithms
No ratings yet
45 Genetic Algorithms
20 pages
Teaching Bayesian Method
No ratings yet
Teaching Bayesian Method
20 pages

Note: This service is not intended for secure transactions such as banking, social media, email, or purchasing. Use at your own risk. We assume no liability whatsoever for broken pages.