Causality, Probability, and Medicine 1138829307, 9781315735542

Table of contents :
CONTENTS......Page 7
PREFACE......Page 15
INTRODUCTION......Page 17
Part I......Page 1
1 Action related theory of causality......Page 33
2 General discussion of AIM theories......Page 46
3 Koch on bacterial disease......Page 61
4 MECHANISTIC THEORIES OFCAUSALITY AND CAUSALTHEORIES OF MECHANISM......Page 87
5 TYPES OF EVIDENCE (I) EVIDENCE OF MECHANISM......Page 106
6 TYPES OF EVIDENCE (II) STATISTICAL EVIDENCE IN HUMAN POPULATIONS......Page 118
7 COMBINING STATISTICAL AND MECHANISM EVIDENCE......Page 144
8 THE RUSSO-WILLIAMSON THESIS......Page 149
9 THE RUSSO-WHILLIAMSON THESIS （II）......Page 166
10 OBJECTIONS TO THE RUSSO-WILLIAMSON THESIS......Page 179
PART III CAUSALITY AND PROBABILITY......Page 209
12 INDETERMINISTICAUSALITY......Page 211
13 CAUSAL NETWORKS......Page 220
14 HOW SHOULD PROBABILITIES BE INTERPRETED......Page 241
15 PEARL'S ALTERNATIVE APPROACH......Page 252
16 EXTENSIONO F THE ACTION RELATED THEORY......Page 268
Appendix 1......Page 275
Appendix_2_Mathematical_terminology......Page 278
Glossary_of_medical_terms......Page 280
References......Page 295
Index......Page 306

Citation preview

CAUSALITY, PROBABILITY, AND MEDICINE

Why is understanding causation so important in philosophy and the sciences? Should causation be defined in terms of probability? Whilst causation plays a major role in theories and concepts of medicine, little attempt has been made to connect causation and probability with medicine itself. Causality, Probability, and Medicine is one of the first books to apply philosophical reasoning about causality to important topics and debates in medicine. Donald Gillies provides a thorough introduction to and assessment of competing theories of causality in philosophy, including action-related theories, causality and mechanisms, and causality and probability. Throughout the book he applies them to important discoveries and theories within medicine, such as germ theory; tuberculosis and cholera; smoking and heart disease; the first ever randomized controlled trial designed to test the treatment of tuberculosis; the growing area of philosophy of evidence-based medicine; and philosophy of epidemiology. This book will be of great interest to students and researchers in philosophy of science and philosophy of medicine, as well as those working in medicine, nursing, and related health disciplines where a working knowledge of causality and probability is required. Donald Gillies is Emeritus Professor of Philosophy of Science and Mathematics at University College London, UK.

CAUSALITY, PROBABILITY, AND MEDICINE

Donald Gillies

First published 2019 by Routledge 2 Park Square, Milton Park, Abingdon, Oxon OX14 4RN and by Routledge 711 Third Avenue, New York, NY 10017 Routledge is an imprint of the Taylor & Francis Group, an informa business © 2019 Donald Gillies The right of Donald Gillies to be identified as author of this work has been asserted by him in accordance with sections 77 and 78 of the Copyright, Designs and Patents Act 1988. All rights reserved. No part of this book may be reprinted or reproduced or utilised in any form or by any electronic, mechanical, or other means, now known or hereafter invented, including photocopying and recording, or in any information storage or retrieval system, without permission in writing from the publishers. Trademark notice: Product or corporate names may be trademarks or registered trademarks, and are used only for identification and explanation without intent to infringe. British Library Cataloguing-in-Publication Data A catalogue record for this book is available from the British Library Library of Congress Cataloging-in-Publication Data A catalogue record for this book has been requested ISBN: 978-1-138-82928-2 (hbk) ISBN: 978-1-138-82930-5 (pbk) ISBN: 978-1-315-73554-2 (ebk) Typeset in Bembo by Deanta Global Publishing Services, Chennai, India

To Grazia, who has certainly benefited from scientific medicine

Vonka’s Thomist Maxim: Sublata causa, tollitur effectus. (If the cause is removed, the effect is taken away.)

CONTENTS

Acknowledgements xii Preface xv Introduction 1 0.1  Deterministic and indeterministic causality   2 0.2 AIM (Action, Intervention, Manipulation) theories of causality  4 0.3  Mechanistic theories of causality  5 0.4  Probabilistic theories of causality  8 Notes 13 PART I

Causality and action

15

1 An action-related theory of causality 1.1  Russell’s critique of the notion of cause  19 1.2  Collingwood’s AIM theory of causality  21 1.3  Productive actions and avoidance actions  24 Notes 29

17

2 General discussion of AIM theories of causality 2.1  Gasking’s contribution  30 2.2  Objection 1. Some causes cannot be manipulated  31 2.3  Objection 2. Causes exist independently of humans  37 2.4  Objection 3. Unavoidable circularity  38 2.5  Explanation of causal asymmetry  40

30

viii Contents

3 An example from medicine: Koch’s work on bacterial diseases and his postulates 45 3.1  The background to Koch’s work  45 3.2  Koch’s investigations of tuberculosis and cholera  47 3.3  Koch’s postulates  50 3.4 Modification of the postulates in the light of the action-related theory of causality  55 3.5 Koch establishes that the comma bacillus is the cause of cholera  57 Notes 67 PART II

Causality and mechanisms 4 Mechanistic theories of causality and causal theories of mechanism 4.1  The Dowe–Salmon theory of causality  71 4.2  Criticism of the Dowe–Salmon theory of causality  72 4.3  More general definitions of mechanism  73 4.4  A causal theory of mechanisms in medicine  79 4.5 The usefulness of postulating mechanisms for the confirmation of causal hypotheses and for discovering cures  81 4.6  Causes, activities, and Anscombe  82 Note 89 5 Types of evidence: (i) Evidence of mechanism 5.1 Confirmation and disconfirmation of causal hypotheses in medicine  90 5.2  Two kinds of evidence  91 5.3  Coronary heart disease (CHD)  92 5.4 An example of evidence of mechanism: Anitschkow’s study of experimental atherosclerosis in rabbits  95 Notes 101

69 71

90

6 Types of evidence: (ii) Statistical evidence in human populations 102 6.1  Ancel Keys and the dangers of saturated fat  102 6.2 An example of observational statistical evidence: The seven countries study  108 6.3 The problem of confounders, and the disconfirmation of causal hypotheses  115 6.4 Some further general points regarding the seven countries study  118

Contents  ix

6.5 Examples of interventional statistical evidence: Some clinical trials 119 Note 127   7 Combining statistical evidence with evidence of mechanism 7.1 Combining the results of Anitschkow, Dayton et al., and Keys et al.  128 7.2  Strength through combining  129 Note 132   8 The Russo–Williamson thesis: (i) Effects of smoking on health 8.1  Statement of the Russo–Williamson thesis  133 8.2 Some different views concerning the roles of statistical evidence and evidence of mechanism in medicine   134 8.3  Smoking and lung cancer  136 8.4 Smoking and heart disease: Is there a linking mechanism?   141 8.5  Research into atherosclerosis 1979–89  143 8.6  Research into atherosclerosis in the 1990s  145 8.7 Implications of our medical examples for the Russo–Williamson thesis   148 Note 149   9 The Russo–Williamson thesis: (ii) The evaluation of streptomycin and thalidomide 9.1  Evidence-based medicine  150 9.2  The trial of streptomycin against bed-rest  153 9.3  The investigation of the treatment mechanism  155 9.4  The trials of streptomycin and PAS  156 9.5  The streptomycin trials in relation to EBM and RWT  158 9.6 Generalizing from the streptomycin case, and the example of thalidomide 158 Notes 162 10 Objections to the Russo–Williamson thesis 10.1  Causal pluralism  163 10.2  McArdle disease  164 10.3  The Semmelweis case  168 10.4  Alternative medicine  174 10.5  Other statistical counter-examples  177

128

133

150

163

x Contents

11 Discovering cures in medicine and seeking for deeper explanations 11.1 The importance of mechanisms for discovering cures in medicine  185 11.2 Seeking deeper explanations in medicine and physics   188

185

PART III

Causality and probability

193

12 Indeterministic causality 12.1  The causation of cervical cancer  197 12.2  Probabilistic causality  200 Note 203

195

13 Causal networks 13.1  Conjunctive and interactive forks  204 13.2  Multi-causal forks  208 13.3 The rise of fast food, and the failure of the ‘Eat Well and Stay Well’ project  211 13.4  The Hesslow counter-example revisited  222 Notes 223

204

14 How should probabilities be interpreted? 14.1  Interpretations of probability  226 14.2  Interpreting the probabilities in causal models  229 14.3  Simpson’s paradox  232 14.4  Comparison with other suggested solutions  233 Notes 235

225

15 Pearl’s alternative approach to linking causality and probability 15.1  Debates about the Markov condition  236 15.2  Pearl’s alternative approach  243 15.3 Restoring the Markov condition by adjusting the model  246 Notes 251 16 Extension of the action-related theory to the indeterministic case 16.1  Mathematical formulation of the problem  252 16.2  Informal statement of Sudbury’s theorems  254

236

252

Contents  xi

16.3 Extension of the action-related theory to the indeterministic case  256 Notes 257 Appendix 1: Example of a simple medical intervention which is not an intervention in Woodward’s sense 259 Appendix 2: Mathematical terminology 262 Appendix 3: Sudbury’s theorems 264 Glossary of medical terms 268 References 283 Index 294

ACKNOWLEDGEMENTS

For almost a decade I have had the good fortune to be part of a research group whose other members are: Brendan Clarke, Phyllis McKay Illari, Federica Russo, and Jon Williamson. This group started on a purely informal basis when Brendan Clarke and I were at University College London (UCL) and the others at the University of Kent in Canterbury. In fact, we initially called ourselves the Kent-UCL group. Our first workshop (23 July 2008) was on ‘Causality and Linking Mechanisms’. In their 2007 work Federica Russo and Jon Williamson had formulated what became known as the Russo-Williamson thesis. This was much discussed within our group (and outside it as well), and the chapters of the present book dealing with this and related topics are, as will be specified in detail in the book’s main text, in part reports of joint research by the group as a whole. Even where I have been pursuing a theme on my own, I have greatly benefited by discussing my ideas with the group and getting their comments. The book itself was written as part of the project ‘Evaluating Evidence in Medicine’ (AH/M005917/1), and I am grateful to the UK’s Arts and Humanities Research Council (AHRC) for supporting this project. Many sections of the book were presented as papers at project meetings and I received very many useful comments from the project participants who included both the original KentUCL group, and in addition Mike Kelly, Charlie Norell, Veli-Pekka Parkkinen, Jan Vandenbroucke, Christian Wallmann, and Michael Wilde. I have also collaborated closely with Maria Carla Galavotti and Raffaella Campaner who are pursuing similar research at Bologna in Italy. They have given helpful comments on much of my material and also organized two workshops which I found very useful. These were on ‘Causal Inference in the Health Sciences’ (May 2011) and on ‘Causality in the Special Sciences’ (October 2015). Turning now to specific topics, the action-related theory of causality presented in Part I of the book was originally developed from the agency theory

Acknowledgements 

xiii

of Peter Menzies and Huw Price, and I have had very helpful discussions with Huw Price. Marco Buzzoni made several valuable points on this subject when we met at the Conference on Scientific Realism organized by Wenceslao Gonzalez at the University of A Coruña in September 2015. Regarding the question of mechanisms which is discussed in Part II of the book, I have had valuable discussions with Lindley Darden and Stuart Glennan, both when we met in Italy or London, and then by email exchanges. Regarding the problems of probability and causality discussed in Part III of the book, I had some very stimulating discussions with Carlo Berzuini, and Philip Dawid at the workshop in Bologna in May 2011. Their comments were quite critical, since they both supported the views of Judea Pearl against the tradition of “probabilistic causality” which I was trying to develop. Later Judea Pearl himself was kind enough to engage in quite an extended email correspondence with me on this subject. This did not end in agreement, but was very helpful for understanding Pearl’s position on the question. My work in this area was also read by Nancy Cartwright, David Corfield, and Jon Williamson, all of whom made useful comments. The development of a theory of how probability is related to causality gave rise to some mathematical questions, and, in dealing with these, I received invaluable help from two mathematicians who specialize in probability and statistics: Christian Hennig of UCL, and Aidan Sudbury of Monash University. Christian Hennig gave some crucial advice about how the problem should be formulated mathematically, and Aidan Sudbury, who was a fellow student of mine both at school and university, proved two theorems which showed how the problem could be resolved. On many issues in the book, I received valuable comments which led to improvements from fellow philosophers of science, who include Hasok Chang, Wenceslao Gonzalez, Ladislav Kvasz, Jun Otsuka, José Solano, and David Teira. Brian Simboli gave me a lot of help in tracing the history of the maxim: sublata causa, tollitur effectus (If the cause is removed, the effect is taken away.) Any philosopher of science dealing with medicine needs the help of medical researchers and I have been fortunate in that respect. I owe a particular debt of gratitude to Vladimir Vonka of Prague, whose own researches had an important role in clarifying the viral causation of cervical cancer. Reading his 2000 paper was one of the factors which got me interested in causality in medicine, and I was fortunate to be able to discuss this question with him on several visits to Prague. From Vladimir Vonka I learned the valuable maxim – sublata causa, tollitur effectus – which I have used as the motto for this book. More specifically, Vonka emphasized the importance of a successful vaccine in establishing causality. Nearer to home, a fellow student of mine at Cambridge, Robin Poston, who has spent his working life researching heart disease, was kind enough to take an interest in my project and gave me a lot of help. He explained the key points of modern ideas about atherosclerosis and heart disease, and suggested readings which would make me familiar with this area of medicine. He also read drafts of my account of how ideas on this subject had developed and suggested many improvements. For philosophers of science, the help of sympathetic researchers is really essential.

xiv Acknowledgements

Finally a word of thanks should be said to the students of the Department of Science and Technology Studies at UCL to whom I presented much of the material in this book in lectures and seminars. These included the students studying medicine at UCL, who did an intercalated year in Philosophy, Medicine, and Society in our Department. I can remember much stimulating discussion of the issues and involved, and much valuable feedback.

PREFACE

The aim of this book is to develop a theory of causality for theoretical m ­ edicine. The method is that of history and philosophy of medicine. This means that any general views proposed are illustrated, and tested out, by applying them to historical case studies in medicine. History of medicine here is taken to include some very recent history. So considerations of contemporary medicine are not precluded. Part I of the book examines the relations between causality and action. An action-related theory of causality is proposed, and it is illustrated by considering Koch’s work in the 19th century on bacterial diseases, particularly cholera. Koch managed to convince the medical community of the time that a specific bacterium (vibrio cholerae) is the cause of cholera. Actions to prevent cholera epidemics could be based on this causal law. What was needed was to ensure that water for drinking had been filtered in such a way as to eliminate vibrio cholerae, and other pathogenic bacteria. This is a good example of how causality is connected to action. Part II of the book is concerned with the relations between causality and mechanisms. If a causal law such as A causes B is proposed in medicine, it is usually very important to try to discover a bodily mechanism M which links A to B. This leads to two different types of evidence for causal claims in medicine. Statistical evidence consists of the results of the study of the disease in human populations (assuming one is dealing with a human disease). Statistical evidence is obtained from epidemiological surveys and clinical trials. Evidence of mechanism is usually obtained in a different way from laboratory studies of animals, tissues, cells, microbes, etc. This is illustrated by considering 20th century investigations of the causes of coronary heart disease (CHD). It is now generally agreed that high blood cholesterol levels are a cause of CHD. The statistical evidence for this claim came from studying the incidence of CHD in the populations of

xvi Preface

seven different countries, and from a number of controlled trials. To this was added evidence of the mechanism by which high cholesterol levels in the blood lead to the formation of atherosclerotic plaques, and so to CHD. This evidence of mechanism came mainly from laboratory experiments of various kinds. This case illustrates the general principle of strength through combining, which states that the combination of evidence of different types results in much higher overall empirical confirmation than could be obtained from the same amount of evidence of a single type. Part III of the book deals with the relations between causality and probability. This was touched on in my earlier book: Philosophical Theories of Probability, and is dealt with in more detail in the present book. The problem is well illustrated by the example: smoking causes lung cancer, a causal law which was once highly controversial, but is now generally accepted. In fact, only about 5% of smokers ever get lung cancer. However, the probability of a smoker getting lung cancer is more than ten times that of a non-smoker getting lung cancer. This suggests a simple connection between causality and probability. Suppose A causes B, then the probability of B given that A occurs is greater than the probability of B given A does not occur. Unfortunately, and very surprisingly, this principle turns out to be false. There are cases in which A genuinely causes B, but in which the probability of B given A is actually less than the probability of B given not-A. As will be shown, this is related to another curious result in probability theory known as Simpson’s paradox. In Part III of the book I will try to develop an account of the connection between causality and probability which overcomes these difficulties. Considerations of probability, of course, quickly become involved in the mathematical aspects of the theory. In the main body of the text, however, I will try to present the main lines of argument on this subject with the minimum of mathematical detail. For those who are familiar with the mathematical theory of probability and statistics, however, I give a full mathematical treatment of the question in Appendices 2 and 3. This treatment involves some new and interesting theorems proved by my friend Aidan Sudbury. These are presented in Appendix 3 (‘Sudbury’s theorems’). Donald Gillies Emeritus Professor of Philosophy of Science and Mathematics Department of Science and Technology Studies University College London May 2018

INTRODUCTION

Causality is a key concept in medicine, but, to analyse this concept, it will be useful to begin by distinguishing between theoretical and clinical medicine. What could be called theoretical medicine consists of a body of laws and theories, many of them involving causality, which have been discovered and then confirmed by medical research. A typical accepted causal law is the following1:

The varicella zoster virus ( VZV ) causes chickenpox. (0.1)

This kind of causal claim is described as generic, because it covers many cases. In clinical medicine, however, a doctor examines a particular patient and has to find out what causes that patient’s symptoms. A doctor may, for example, decide that the rash of Miss Anne Smith, aged 4, is chickenpox and so caused by the varicella zoster virus. This is an instance of single-case, rather than generic causality. Some authors speak of type/token causality instead of generic/single-case causality, where type = generic, and token = single-case. In this book I will confine myself exclusively to generic causality, and will use the term causality only in this sense from now on. This is not of course to deny that single-case causality is very important. It is clearly of crucial importance for clinical medicine. However, I think it is easier to analyse generic causality first, and then use this analysis for a study of single-case causality. This procedure follows the normal course of medical education. Medical students generally learn theoretical medicine to begin with, and then, when they have mastered some current medical knowledge, they learn how to apply it to individual patients in clinical practice. So the aim of the book, as already stated in the preface, is to develop a theory of causality for theoretical medicine. To explain the plan in more detail, however, I need to make a further distinction concerning causality, which will be done in the next section.

只关注一般性因 果关系。

2 Introduction

0.1  Deterministic and indeterministic causality1 The distinction between deterministic and indeterministic causality is a most important one. A causal claim of the form A causes B is deterministic, only if, ceteris paribus, whenever A occurs, it is followed by B. Otherwise, the claim is indeterministic. To give a simple example of deterministic causality, let us consider a stretch of coastline where over-hanging cliffs descend directly into the sea without any beaches. In this stretch, the following causal claim is true:

Throwing a stone over the cliff edge causes it to fall into the sea. (0.2)

This is an example of deterministic causality, because each time a stone is thrown over the cliff edge, it will, ceteris paribus, fall into the sea. Deterministic causality is the concept of causality, which prevailed in the 18th and 19th centuries in what could be called the Newtonian era. Indeed, as an acknowledgement of this, my example (0.2) is a Newtonian one. For this reason, the concept of causality which was analysed by leading philosophers of that period, notably Hume and Kant, was, as we shall see in a moment, the concept of deterministic causality. 19th- and early 20th-century scientific medicine, as it was developed by Pasteur, Koch, and others, used a deterministic notion of causality. An attempt was made to show that each disease had a single cause, which was both necessary and sufficient for the occurrence of that disease (Carter, 2003). So, for example, tuberculosis was caused by a sufficiently large number of tubercle bacilli in a patient or experimental animal. Koch argued that whenever a suitable experimental animal was injected with a sufficiently large number of tubercle bacilli, it went on, ceteris paribus, to develop tuberculosis. This is deterministic causality. In the 20th century, however, a new concept of causality emerged in medicine, particularly in connection with medical epidemiology. If a casual claim of the form A causes B is indeterministic, then A can occur without always being followed by B. One of the first examples of indeterministic causality in medicine was the claim that

Smoking causes lung cancer. (0.3)

In 1947 when the hypothesis (0.3) was first seriously investigated in England, it was regarded as unlikely to be true. The standard view was that lung cancer was caused by general urban atmospheric pollution. Nowadays after much controversy, (0.3) is almost universally accepted as being true. Curiously, few participants in the heated discussion of the truth of (0.3) pointed out that it involved a new notion of causality. Yet this is clearly the case. In fact only about 5% of smokers get lung cancer. So, although smoking causes lung cancer, smoking is not always followed by lung cancer.

Hume and Kant研究的 是 deterministic causality

Introduction 

3

The notion of indeterministic causality has, since 1950, become ubiquitous in medicine. Eating a diet with a large percentage of animal fat is recognized as causing heart disease, having a gene or genes of a particular kind is recognized as causing Alzheimer’s, infections by certain viruses are recognized as causing cancers, and so on. In all these cases we are dealing with indeterministic causality. Unfortunately, indeterministic causality is harder to analyse than deterministic causality. A full account of indeterministic causality has to deal with the relation between causality and probability, and this is a complicated question which gives rise to several paradoxes. However, deterministic causality remains important in medicine today, since much of the work by Pasteur, Koch, and their successors on infectious diseases is still held to be largely valid. From the point of view of contemporary medicine, therefore, it is still well worth analysing the notion of deterministic causality. Let us now start this task by examining briefly some things, which Hume and Kant had to say on the subject. Hume’s A Treatise of Human Nature, published in 1738, contains a detailed discussion of the notion of cause in Book I, Part III. Hume mentions three conditions, which a causal law such as “A causes B” must satisfy. The first of these is that A and B must be contiguous. As Hume says (1738, p. 78): I find in the first place, that whatever objects are considered as causes or effects, are contiguous; and that nothing can operate in a time or place, which is ever so little removed from those of its existence. Hume’s second condition is that A must precede B in time. As he says (1738, p. 79): The second relation I shall observe as essential to causes and effects, is not so universally acknowledged, but is liable to some controversy. It is that of priority of time in the cause before the effect. Hume here seems to mean that A must occur at a time, which is strictly before that of B. The main opposing view, as we shall see later, is that A and B may occur simultaneously. Hume’s third condition is that A and B must be constantly conjoined. This is his most famous condition, and his overall theory is often referred to as the “constant conjunction” view of causality. Here is how Hume himself puts it (1738, p. 90): We remember to have had frequent instances of the existence of one species of objects; and also remember, that the individuals of another species of objects have always attended them, and have existed in a regular order of contiguity and succession with regard to them. Thus we remember to have seen that species of object we call flame, and to have felt that species of sensation we call heat. We likewise call to mind their constant conjunction in all past instances. Without any further ceremony, we call the one cause, and the other effect, and infer the existence of the one from that of the other. … 

4 Introduction

Thus, in advancing, we have insensibly discovered a new relation betwixt cause and effect when we least expected it, and were entirely employed upon another subject. This relation is their constant conjunction. So, according to Hume, if A causes B, then A and B are constantly conjoined, that is to say that, whenever A occurs, B follows. It is therefore a consequence of our earlier definition that Hume is dealing with deterministic causality. The same is true of Kant, as the following passages from the Critique of Pure Reason show. The schema of cause, and of the causality of a thing in general, is the real upon which, whenever posited, something else always follows. (Kant, 1781/7, A144/B183) the very concept of a cause …  manifestly contains the concept of a necessity of connection with an effect and of the strict universality of the rule. (Kant, 1787, B5) That concludes my discussion of the distinction between deterministic and indeterministic causality. We are now in a position to outline some of the principal positions in contemporary philosophy of causality. There are in fact a great number of different philosophical theories of causality in the philosophy of science of today. The situation is both complex and confused. I propose to classify these theories into three principal groups, and will briefly explain and outline each of these groups in turn in the next three sections. The groups are: (i) AIM (Action, Intervention, Manipulation) theories; (ii) Mechanistic theories; and (iii) Probabilistic theories. This classification is not exhaustive, since I will not consider, for example, counterfactual theories of causation. Nor would this particular classification be generally agreed. Other attempts to classify contemporary theories of causation are to be found in Psillos (2002), and Illari and Russo (2014).

0.2 AIM (Action, Intervention, Manipulation) theories of causality The basic idea of the AIM approach to causality is that causal laws have a close relation to actions, interventions and manipulations. This can be illustrated by the following example of an action, which is of a kind that will be familiar to readers of crime fiction. Let us start with the following true causal law.

The ingestion of a sufficiently large dose of arsenic causes death. (0.4)

A villain (Mr V) wishes to murder his aunt in order to inherit her money. He knows (0.4), and so places a sufficiently large dose of arsenic into his aunt’s tea. Unsuspectingly she drinks the tea and expires. Here Mr V has carried out his

Introduction 

5

wicked action on the basis of the causal law (0.4). He has intervened in the situation, and manipulated the tea in his aunt’s cup. So causal laws form the basis of actions, interventions, and manipulations, and this, according to AIM theories of causality, is precisely what characterizes the notion of causality. This approach to causality can perhaps be traced back to the following famous quotation from Bacon (1620, III, p. 259): Human knowledge and human power meet in one; for where the cause is not known the effect cannot be produced. Nature to be commanded must be obeyed; and that which in contemplation is as the cause is in operation as the rule. A connection between cause and action is also mentioned by Kant, who says (1781/7, A204/B249): “Causality leads to the concept of action.” However, this remark is not followed up in any detail, and Kant’s overall position on causality cannot be described as an AIM theory. In fact, AIM theories of causality have been mainly developed since the 1930s. In this book, I will advocate an AIM theory of causality, which is developed from a version, which was originally published in my 2005a paper. I call this “an actionrelated theory of causality”. It will be expounded in Part I of the book, but only for deterministic causality. I have adopted this plan, because, as already remarked, indeterministic causality is considerably more complicated than deterministic causality. It is therefore easier to deal with the deterministic case first, and extend the analysis later to indeterministic causality. This will be done in Part III of the book. In Part I, as well as expounding my own version of the AIM approach to causality, I will compare it in detail to other versions of this approach. I will discuss the main objections to AIM theories and show how they might be overcome, as well as pointing out the advantages of this approach. The aim of this book is to develop a theory of causality for theoretical medicine. So I will try at every stage to show that the views adopted apply satisfactorily to important medical examples. For deterministic causality, the obvious example to choose is the work of Robert Koch in developing the germ theory of disease. Koch used a deterministic notion of causality, and applied it in his discoveries concerning the bacteriological causes of significant diseases such as anthrax 2, tuberculosis2, and cholera 2. His results are still largely accepted today. Koch also formulated his famous postulates, designed to show what evidence was needed to show that a specific bacterium 2 was indeed the cause of a particular disease. In Part I, I will examine in detail whether, and to what extent, the action-related theory of causality gives a satisfactory account of Koch’s researches.

0.3  Mechanistic theories of causality The next important group are the mechanistic theories of causality. To explain the basic idea behind these theories, I will use a striking example introduced by one of the leading advocates of the mechanistic approach – Wesley Salmon.

6 Introduction

This example concerns a dog whistle. We observe a dog owner who from time to time blows a whistle. No sound comes from the whistle, and yet each time it is blown the owner’s dog comes. So we observe a constant conjunction between the blowing of the whistle and the dog’s going back to its owner. However, according to Salmon, this constant conjunction is not sufficient to infer causality. In addition to the constant conjunction, we have to satisfy ourselves that there is some mechanism linking the blowing of the dog whistle to the dog’s return. If there were no such mechanism, it could be that the dog was returning, not because of the blowing of the whistle, but because the dog spotted some gesture which the owner was making at the same time. Of course, there is such a mechanism in this case, because dog whistles emit sounds which are too high for humans to hear, but which can be heard by dogs. This is how Salmon puts it (1993, p. 22): Suppose, for example, that a person has a whistle and a dog. When the person blows the whistle, we hear nothing but in every case the dog comes. We learn inductively that blowing the whistle is a sufficient condition for the appearance of the dog. But we do not understand the relation between the cause and effect unless we know that there is a sound wave that we cannot hear because the frequency is too high for humans but not too high for dogs. The dog comes because it hears the whistle of its master. The wave is the causal process that provides the connection. Salmon explicitly contrasts his view with that of Hume. He states Hume’s position as follows (1993, p. 15): Hume concludes …  that, in situations in which we believe that there is a causal relation, we perceive three features: (1) the temporal priority of the cause to the effect; (2) the spatiotemporal contiguity of the cause to the effect; and (3) the fact that on every occasion on which the cause occurs, the effect follows (constant conjunction). However, according to Hume, we cannot find a physical connection between the cause and the effect; On the contrary, Salmon replies, we can find a physical connection between the cause and the effect. In the case of the dog whistle this is the sound wave, which travels from the whistle to the dog’s ear. The general mechanistic approach to causality can therefore be stated as follows. A causes B if and only if there is a mechanism which links A to B. In Part II of the book, I will discuss the various mechanistic theories of causality, starting with the last version of Salmon’s own theory, into which he introduced some ideas from Dowe. This theory was criticized for having too limited a notion of mechanism. So this investigation led to attempts to define mechanism in more general terms. The problem, however, with these more general definitions is that they all seem to involve either the word “cause” or some words which are more

Introduction 

7

or less equivalent, such as “produce” or “bring about”. So to try to characterize “cause” in terms of “mechanism” appears to involve a vicious circle. In this situation, my suggestion is that instead of trying to characterize “cause” in terms of “mechanism”, we should move in the opposite direction and try to define “mechanism” in terms of “cause”. So, the general idea is to characterize “cause” using the action-related theory, and then define mechanism using this notion of cause. In this sense, therefore, I reject mechanistic theories of causality, but, at the same time, it is still possible, from the causal definition of mechanism, to obtain many of the results of the mechanistic theories of causality. Specifically, if “A causes B” is a causal law in medicine, it is frequently desirable, and indeed often essential, to seek for a mechanism linking A to B. In Part II, as in Part I, I will relate the philosophical discussion to important medical examples. One of these in Part II is the causal claim, now generally accepted, that smoking causes heart disease. This is, in many ways, quite like Salmon’s dog whistle example. Statistical surveys showed a strong correlation between smoking and heart disease, but initially no one was sure whether this correlation was causal in character. There was at that time no obvious mechanism linking smoking to heart disease. However, as a result of some excellent scientific work involving laboratory experiments, a mechanism linking smoking and heart disease was discovered, and the causal connection between smoking and heart disease was established. In general, it has been claimed that to establish a causal law, such as “A causes B” in medicine, we need not only statistical evidence relating to human populations, but also evidence of a mechanism linking A to B. This is one form of a thesis, known as the Russo-Williamson thesis, which will be defended in Part II. In addition to the need for mechanisms in order to establish causal laws in medicine, there is another powerful reason for seeking the mechanisms leading to diseases. Medical researchers are constantly looking for ways of curing or preventing diseases. If the details of the mechanisms leading to a disease are brought to light, these mechanisms may contain many steps. Blocking any one of these steps may cure or prevent the disease in question, and so the more steps that are known, the more possibilities there are of finding an intervention, which cures or prevents the disease. The search for ever more detailed mechanisms is also the search for ever greater depth in the understanding of the disease. I will thus try to show in Part II that, starting from the action-related theory of causality, we are led to assigning a very important role to mechanisms in medicine, even though the mechanistic theories of causality in medicine are rejected. It should be added that this is not the only way of looking at the relation between AIM theories of causality, and mechanistic theories of causality in medicine. A different approach is to be found in Campaner and Galavotti (2007). They argue that the two types of theory lead to different notions of causality, namely the manipulative notion of causality and the mechanistic notion of causality. They analyse a particular medical example, namely the deep brain stimulation used for treating Parkinson’s disease, and try to show that there is an interplay in the

8 Introduction

use of these two types of causality. This is how they summarize their position (2007, p. 195): The example used from medicine suggests that the mechanical and manipulative notions of causality are both useful, and should be seen as complementary rather than conflicting. This is a prominent sense in which we regard causality as a pluralistic notion. Despite the differences between the approach of Campaner and Galavotti and the one advocated in this book, there is a point in common. Both sides would agree that a full analysis of causality in medicine should involve both manipulations and mechanisms. The difference lies in how the pieces of this jigsaw are fitted together. There would also be consensus that there is another piece of the jigsaw, namely probability, and we turn to this in the next section.

0.4  Probabilistic theories of causality If “A causes B” is an example of deterministic causality, then, ceteris paribus, whenever A occurs, it will be followed by B. Obviously there is no room for probability here. So, the search for a link between causality and probability will surely only begin with the introduction of indeterministic causality. Now, as I argued earlier, indeterministic causality appeared in medicine for the first time in the early 1950s, when the question of whether smoking caused lung cancer was being considered. It is to be expected therefore that the first probabilistic theories of causality would appear in the middle to late 1950s. This turns out to be indeed the case. Williamson in his 2009 work gives an admirable account of probabilistic theories of causality. He claims that the first such theories were due to Reichenbach (1956) and Good (1959). I will deal with Reichenbach’s views in more detail in Part III of the book. So to introduce probabilistic theories of causation here, I will consider Good (1959). Before doing so, however, a word of caution is needed. Probabilistic theories of causation were devised following the appearance of indeterministic causes in medicine. However, it is possible in some cases to use indeterministic causes in medicine in a completely satisfactory fashion without introducing probabilities. In the first chapter of Part III (Chapter 12, section 12.1), I give, as an example of this, a recent striking advance in medicine. This is the discovery, for which zur Hausen was awarded the Nobel Prize in 2008, that cervical cancer is caused by a preceding infection by a papilloma virus. This is clearly an example of indeterministic causation, since not all infections by the papilloma virus are followed by cervical cancer. However, it does appear that an infection by a papilloma virus is a necessary condition for cervical cancer to develop. It follows that a vaccination, which prevents infections by a papilloma virus, will prevent cervical cancer. The reasoning here led to a very effective preventative measure, but does not depend in any way on probability.

Introduction 

9

Still it remains true that most uses of indeterministic causality do involve a consideration of probabilities. Let us therefore turn to Good’s 1959 paper, which was one of the first to introduce a probabilistic theory of causality. Good begins by pointing out that he is modifying the classical view of causality (1959, p. 307): The view of causality that I shall describe here is a modification of the classical view. The classical view has strong overtones of strict determinism. …  There is, however, a more modern view of causality in which determinism is not taken for granted. The modern or modified view seems to me to be a more useful one than the classical one. What Good here describes as “the classical view” is what I have called: deterministic causality; and what he describes as “the modern or modified view” is what I have called: indeterministic causality. Good goes on to say that he reached the conclusion that causality could be defined in terms of probability as a result of a conversation in 1955 with L. J. Savage (1959, p. 307): In a conversation in Chicago in 1955, L. J. Savage remarked that the understanding of probability had reached a point where it should be possible to give a definition of causality. He adds (1959, p. 307): I have recently given more thought to the question and have decided, in agreement with Popper, that the definition should be formulated in terms of what he calls ‘propensity’. Good here refers in a footnote to Popper (1957a) – the paper in which Popper introduces a new interpretation of probability, the propensity interpretation. I will now explain the basic idea of the propensity interpretation of probability in terms of a simple example. Let us suppose that we are rolling a loaded die. We do so by placing the die in a dice box, shaking the box thoroughly and then turning it over so that the die falls onto a table top. Let us refer to these conditions for rolling this particular die as S. S is a set of conditions, which can be repeated as often as we like. Let us now consider what is meant by saying that the probability of getting a 5 relative to this set S of repeatable conditions is p [P(5 | S) = p]. According to the propensity interpretation of probability, the meaning is this: the repeatable conditions S are endowed with a propensity (or tendency or disposition) to give the result 5 on any repetition of S. The magnitude of this propensity is given by p. On the propensity interpretation, probabilities are objective, and we can measure their magnitudes by using statistical frequencies. The preceding paragraph of course gives only a very rough sketch of the propensity theory of probability.3 However, it is obvious already that there might

10 Introduction

be a connection between propensity and causality. Could we not regard the propensity of the conditions S to produce a particular result, such as 5, as a kind of causal power associated with these conditions? This causal power would be an instance of indeterministic causality, since 5 does not appear on every repetition, but only on some. Returning to Good, his plan was to define causality in terms of probability, where probability is interpreted as propensity. I will not here give the full details of Good’s definition. In fact, he modified some of the details in subsequent papers. However, I will mention one of the clauses of his definition, which became a central principle in probabilistic theories of causality. Let F be a cause and E its effect, and let H be some background assumptions. The principle is (Good, 1959, p. 309):

P ( E|F.H ) >P ( E|not-F.H ) (0.5)

Leaving aside a consideration of the background assumptions H, this says that the probability of the effect E, given the cause F [P(E | F )] is greater than the probability of the effect E, given that the cause F does not occur [P(E | not-F)]. An example would be that the probability of lung cancer for a smoker is greater than the probability of lung cancer for a non-smoker. This certainly seems a very plausible principle, and could be called the principle that causes raise the probability of their effects. Good adds the following observation (1959, p. 309):

If we had P ( E|F.H ) = 1 we would be trying to define old-fashioned ( = strict ) causality.



In other words, if the probability of the effect E given the cause F is 1, then we have deterministic causality, which thus turns out to be a special case of indeterministic causality. Popper seems to have accepted Good’s general approach here, because Popper writes (1990, p. 20): Causation is just a special case of propensity: the case of propensityy equal to 1. Good’s plan for defining causality in terms of probability, interpreted as propensity, must have seemed a very promising one in 1959. Unfortunately though, as happens so often in philosophy, attempts to carry it out led to some quite unexpected difficulties. In Part III, I will discuss these difficulties, and what might be done to overcome them. Here, by way of introduction, I will give a rough sketch of two of the most important problems. The first problem is what is known as Humphreys’ Paradox. It was discovered by Paul Humphreys, but first published, with an acknowledgement of the author, by Salmon in1979. Humphreys’ own account of the paradox appeared in 1985.

Introduction 

11

If A causes B, then in general it is not true that B causes A. To use an example of Pearl’s (2000, p. 134), it is true that rain causes mud, but not that mud causes rain. However, if P(A) >  0 and P(B) >  0, then whenever P(B | A) is defined, then P(A | B) is defined. Let us take A = rain and B = mud. By filling in a few details, we can make these the outcomes of a chance set up. For example, we could consider a field at a particular season of the year, where “mud” is evaluated at sunset, and “rain” means that it rained during the day. Relative to such a set up, P(mud | rain) and P(rain | mud) are both well defined. They are objective probabilities whose values can be measured by statistical frequencies collected by observation. Yet only one of these two probabilities, namely P(mud | rain) is causal in character, since rain causes mud, while the other P(rain | mud) is not causal in character since mud does not cause rain. Another way of putting the point is to say that causality is asymmetrical, while probability is symmetrical. These differences between causality and probability cast doubt on the project of defining causality in terms of probability. Reactions to this situation are quite varied.4 Humphreys himself holds that propensities should be considered to be causal in character, from which it follows that not all probabilities are propensities. Another approach, which is the one I favour, is to maintain that propensities are indeed probabilities in the sense of satisfying the usual axioms of probability. From this position it follows that propensities are not always causal in character. At all events, Humphreys’ paradox shows that “causality” and “probability” are very different concepts. Yet there must be some connection between these two concepts, even if we cannot define causality in terms of probability. This brings us to the second of our two problems. The second problem concerns what we have called the principle that causes raise the probabilities of their effects. This was adopted by Good, and we have given his formulation of the principle as 0.5. Now, even if we abandon the project of defining causality in terms of probability, we might still use a principle like 0.5 to provide a connection between causality and probability. Unfortunately, despite its great plausibility, the principle that causes raise the probabilities of their effects turns out to be untrue. Cases were discovered in which A undoubtedly causes B, and yet the probability of B given A is actually lower than the probability of B given non-A. The most striking such counter-example was published by Hesslow in 1976, and is stated in section 12.2. Hesslow’s own reaction to his counter-example is of great interest. He suggests that indeterministic causality should be abandoned, and that there should be a return to deterministic causality. By this time, however, indeterministic causality had become so widely used in medicine that its elimination was no longer a practical possibility. There was nothing for it but to try to resolve the problems of connecting causality and probability. Luckily in the 1980s there was a new development, which made this task a little easier. As we have seen, indeterministic causality first appeared in medicine in the early 1950s with the investigation of the hypothesis that smoking caused lung cancer. In the late 1950s the first probabilistic theories of causality were introduced,

12 Introduction

FIGURE 0.1 

Diagrammatic representation of a causal network.

and these were developed in the 1960s and early 1970s. Then in the 1970s, difficulties in these theories emerged, such as Humphreys’ paradox and Hesslow’s counter-example. In the 1980s there appeared the new theories of causal networks and Bayesian networks. Important contributions to this theoretical development are Pearl (1988) and Lauritzen and Spiegelhalter (1988). Neapolitan (1990) gave an account of these new theories, which made them generally accessible. The theory of causal networks uses diagrams, like the one shown in Figure 0.1. Here the arrows joining two nodes of the network such as A and C mean that A causes C (more strictly that A has a causal influence on C). So, in Figure 0.1, A causes C, and B causes C. C in turn causes E, and D is also a cause of E. Diagrams of this sort can exhibit complicated groups of causal relations in a clear and comprehensible fashion. Another advantage of this way of representing causes is that a probability distribution can be given to the set of variables involved, which, in this case is {A, B, C, D, E}. Thus probabilities are added to the causal relations such as A causes B, etc., though these probabilities are usually stated separately rather than made part of the diagram. Now of course the theory of causal networks becomes quite mathematically complicated when developed in detail. As the aim of this book is to develop the philosophical aspects of these questions, in the main text I will use only the diagrams associated with causal networks. These diagrams form a kind of visual language, which clarifies the underlying problems. In an appendix, however, designed for those with mathematical knowledge and interests, I will give a full mathematical treatment of the question, which includes some new theorems proved by my friend, Aidan Sudbury. Sudbury’s theorems are of considerable interest in themselves, and suggest new ways in which the mathematical theory could be developed. The main aim of Part III of the book is really to rework Good’s original suggestion of 1959, using the theory of causal networks and taking account of the difficulties in the proposal which have emerged since 1959. Because of Humphreys’

Introduction 

13

paradox and other problems, it is no longer feasible to seek a definition of causality in terms of probability. Instead we should seek for a link between causality and probability, what I call a causality probability connection principle (or CPCP). However, we can use a version of Good’s principle 0.5, reformulated in terms of causal networks, for CPCP. The most straightforward formulations of CPCP are still liable to objections such as Hesslow’s counter-example, which can also be formulated in terms of causal networks (see section 13.4). However, the causal network formulation also suggests a way in which this difficulty can be overcome. The next problem is to consider how the probabilities which are linked to generic causes should be interpreted. Here I once again follow Good in giving these probabilities a propensity interpretation. With this interpretation it can be shown that Hesslow’s counter-example is essentially the same as another wellknown problem – Simpson’s paradox, which is stated in section 14.3. Interpreting probabilities as propensities, and hence as objective probabilities whose values can be measured using statistical frequencies, is not accepted by everyone. Pearl always interprets probabilities subjectively as degrees of belief (see his 1988 and 2000 works). Pearl also rejects any principle connecting causality and probability along the lines of Good’s 0.5, and he tries to establish this connection in a different way. In Part III of the book, the position advocated will be compared to and contrasted with Pearl’s. The views presented in Part III of the book constitute an extension of the action-related theory of probability given in Part I to the case of indeterministic causality. Part III thus completes the book’s general project. That concludes my outline of the book, and we can now start on the detail of Part I.

Notes 1 Statements used as examples, and equations, will be numbered by chapter, so that, e.g. 3.2 means the second statement or equation of Chapter 3. For convenience 0 will be used for the introduction. The same system of numbering will be used for sections of the chapters or introduction. 2 The meanings of medical terms, which may be unfamiliar to some readers, are explained in the Glossary of medical terms at the end of the book. 3 I give a more detailed account of the propensity theory in my 2000 work in Chapters 6 and 7, pp. 113–168. This account has been updated in some respects in my more recent 2016a work. These references provide a survey of how the propensity theory has been developed, and give details of its various versions, and their respective merits, as well as of the objections which have been raised to the propensity approach. 4 A fuller account of the Humphreys paradox, which discusses the various reactions to it, is to be found in my 2000 work, pp. 129–136.

PART I

Causality and action

1 AN ACTION-RELATED THEORY OF CAUSALITY

In the introduction I have described an important group of philosophical ­theories of causality, which were called AIM theories, where AIM means Action, Intervention, Manipulation. In this chapter I will discuss the historical development of AIM theories, and put forward a particular theory of this type, which will be called “an action-related theory of causality”. Then in the next chapter, I will give a general discussion of AIM theories of causality, showing how they attempt to overcome the difficulties in this approach, and what are the points in their favour. This discussion will give a comparison of the action-related theory advocated in this book with other contemporary versions of AIM theories of causality. In the introduction I mentioned some ancestors of the AIM approach to causality, most notably Bacon. However, AIM theories of causality really only begin in the late 1920s and 1930s. Perhaps the first sketch of an AIM theory of causality is to be found in a paper by Frank Ramsey, ‘General Propositions and Causality’, which was written in 1929, and published after his death in 1931. This sketch consists of only three sentences which run as follows (Ramsey, 1931, p. 250)1: Again from the situation when we are deliberating seems to me to arise the general difference of cause and effect. We are then engaged not on disinterested knowledge or classification (to which this difference is utterly foreign), but on tracing the different consequences of our possible actions, which we naturally do in sequence forward in time, proceeding from cause to effect not from effect to cause. We can produce A or A’ which produces B or B’ which etc. … ; the probabilities of A, B are mutually dependent, but we come to A first from our present volition. Here Ramsey argues that “the general difference of cause and effect” (in say “A causes B”) arises not from “disinterested knowledge” but from a consideration of

18  An action-related theory of causality

“our possible actions”. So we can produce A which produces B. He contrasts the situation with probability by saying that “the probabilities of A, B are mutually dependent”. What Ramsey seems to mean here is that if A is probabilistically dependent on B, then B is probabilistically dependent on A, i.e. the relation is symmetric. By contrast, if A causes B, B does not in general cause A. This is because “we come to A first from our present volition”, which is to say that we use A to produce B, but not vice versa. As can be seen, this remarkable passage anticipates not only AIM theories of causality, but also the Humphreys paradox, which was described in the introduction. However, the passage is so short and cryptic that it must surely be described as a precursor rather than an exposition of these ideas. The first to develop a detailed version of an AIM theory of causality in the 20th century were Collingwood, in a 1938 paper, and subsequent book (1940), and Dingler in his 1938 book.2 The AIM approach to causality was then espoused and developed by Gasking (1955), and von Wright in his 1973 paper and 1974 book. More recently, there have been a number of developments of AIM theories of causality by philosophers of science. Menzies and Price have developed one of these theories (see Price [1992] and Menzies and Price [1993]). They refer to their theory as an agency theory of causality. Woodward (2003) has developed a theory which he calls: “a manipulationist or interventionist account of …  causation” (p. v). I refer to my own version (Gillies 2005a) as an action-related theory of causality. Pearl’s 2000 book should also be mentioned here. Pearl is not committed exclusively to an AIM account of causality, and introduces other conceptions of causality into his scheme. However, intervention still plays an important role for him. He introduces the variable X3 which stands for a sprinkler which can be on or off (2000, p. 23). He urges the reader: “Note the difference between the action do(X3 = On) and the observation X3 = On”. Later in the book, Pearl introduces a do-calculus, which is a mathematical way of representing interventions. AIM theories of causality are thus one of the leading trends in contemporary philosophy of causality. In this book, I will naturally follow my own AIM theory, which differs from some of the others. It is, however, perhaps the closest to Collingwood’s original version, and I will expound it by first stating Collingwood’s theory and then explaining how I think this needs to be changed and developed. Before coming to Collingwood, however, I want to discuss briefly one famous paper on causality, even though this does not present, or even mention, an action, intervention, manipulation approach. The paper in question is Russell’s 1913 ‘On the Notion of Cause’, which was his presidential address to the Aristotelian Society in November 1912. This paper has a particular interest for me since I believed its conclusions for many years, and even though I now reject Russell’s overall position, I still think that some of his ideas are valid, and will attempt to incorporate these ideas into my action-related theory of causality. Moreover, Russell’s 1913 paper on causality was, as we shall see, very important also for Collingwood.

An action-related theory of causality  19

1.1  Russell’s critique of the notion of cause Russell begins his paper in a dramatic fashion by claiming the word ‘cause’ should be altogether banished from philosophy (1913, p. 173): In the following paper I wish, first, to maintain that the word ‘cause’ is so inextricably bound up with misleading associations as to make its complete extrusion from the philosophical vocabulary desirable; …  Russell’s principal reason for this recommendation is that the word ‘cause’ has already disappeared from the advanced sciences. As he says: All philosophers, of every school, imagine that causation is one of the fundamental axioms or postulates of science, yet, oddly enough, in advanced sciences such as gravitational astronomy, the word ‘cause’ never occurs. Russell, writing just before the First World War, concludes that causality, like the British monarchy, is something no longer appropriate in the modern age. With his customary wit, he puts the point as follows (1913, p. 173): The law of causality, I believe, like much that passes muster among philosophers, is a relic of a bygone age, surviving, like the monarchy, only because it is erroneously supposed to do no harm. In these quotations Russell has perhaps rather exaggerated his own position in the interest of some striking turns of phrase. If we read on, we discover that he does after all allow a weak notion of causality. However, he claims that this notion of causality is useful only in everyday life and the infancy of science. His main thesis, therefore, is that the concept of cause disappears from science as it advances. Regarding the weak notion of causality which he in fact allows, Russell gives an analysis similar to a Humean “constant conjunction” account, except that Russell claims that the sequences are of frequent rather than absolutely constant conjunction, and that they yield no more than probability (1913, p. 185): the sequence, in any hitherto unobserved instance, is no more than probable, whereas the relation of cause and effect was supposed to be necessary. … Thus in our present sense, A may be the cause of B even if there actually are cases where B does not follow A. Striking a match will be the cause of its igniting, in spite of the fact that some matches are damp and fail to ignite. Thus causal laws for Russell are laws of probable sequence. He points out one curious consequence of this neo-empiricist or neo-Humean account, namely that it turns out to be correct to say that night causes day (1913, p. 185). Intuitively however the statement that night causes day does not seem to be true.

20  An action-related theory of causality

In terms of the analysis of causal laws as laws of probable sequence, Russell is in a position to state his main thesis which he does as follows (1913, p. 186): such laws of probable sequence, though useful in daily life and in the infancy of science, tend to be displaced by quite different laws as soon as a science is successful. The law of gravitation will illustrate what occurs in any advanced science. In the motions of mutually gravitating bodies, there is nothing that can be called a cause and nothing than can be called an effect; there is merely a formula. Certain differential equations can be found, which hold at every instant for every particle of the system, and which, given the configuration and velocities at one instant, or the configurations at two instants, render the configuration at any other earlier or later instant theoretically calculable. That is to say, the configuration at any instant is a function of that instant and the configurations at two given instants. This statement holds throughout physics, and not only in the special case of gravitation. But there is nothing that could be properly called ‘cause’ and nothing that could be properly called ‘effect’ in such a system. Russell’s position depends on distinguishing between different types of scientific law. Causal laws are those of the form: “A causes B”. However, there are other types of scientific law. Russell mentions functional laws relating variables, and laws expressed by differential equations. He could have added probabilistic or statistical laws, such as the law that radioactive emissions follow a Poisson distribution. Russell’s thesis is that causal laws are useful only in daily life and in the infancy of a science, and are not to be found in an advanced science. I believed this thesis for many years, and I still think it is correct for theoretical physics. However, Jon Williamson persuaded me that it cannot be true for all advanced sciences, since it is plainly false for medicine. Considerations of causality arise at every step in medicine. Consider, for example, a doctor who carries out a medical diagnosis is attempting to ascertain the cause of a patient’s symptoms. If the patient suffers from pains in the chest, it is most important to know whether these are caused by lung cancer, angina, a bacterial infection of the bronchi, or something else altogether. The treatment given will be quite different for different causes. We see from this that the notion of cause and causal laws play a crucial role in medical practice. The same applies to medical research, as can be illustrated by one of Pasteur’s famous medical discoveries (cf. Debré , 1994, pp. 330–40). In June 1879, Pasteur collected pus from the boil of one of his assistants, and discovered that it contained a bacterium which was later named staphylococcus. In February 1880 Pasteur took pus samples from deep in the bone of a little girl aged 12 who was being operated on for the bone disease osteomyelitis. He discovered that the pus contained a bacterium of the same type. This led him to conclude that boils and osteomyelitis are both caused by the same bacterium. This was a very significant discovery since it showed that a serious disease located deep in the tissues had the same cause as a superficial and generally slight illness. The result eliminated the difference

An action-related theory of causality  21

between internal and external pathology. Note that this most impressive discovery was the discovery of a causal law. Discoveries in medicine are used for either the prevention or the cure of diseases. In the case of staphylococcal infections, preventative measures were relatively easy. It was a matter of preventing the pathogenic bacterium entering the body through greater care about hygiene and antisepsis. The discovery of a cure proved much harder. It required finding a substance (an antibiotic) which would kill staphylococci in a patient’s body without harming the patients. Fleming discovered penicillin in 1928, but it required a lot of development work both by him, and later by the Oxford team of Florey, Chain, and others before penicillin could become an effective antibiotic. (For details see Macfarlane, 1984, especially pp. 165–186.) In fact, it was not until May 1941 that the first patient was cured of a staphylococcal infection (a four-inch carbuncle on the back) using penicillin. Hence 62 years elapsed between the discovery of the cause of a group of illnesses and the development of a successful cure. However, it is worth noting that the discovery of the cause was a precondition for developing a cure. Returning now to Russell, we can see that medicine completely refutes his claim that causal laws belong only to the infancy of science. Medicine has been highly successful and produced cures which would have been regarded as miraculous in a former age. Thus medicine has every claim to be an advanced science, and yet makes essential use of the notion of cause and of causal laws. Russell, however, was correct in his claim that non-causal mathematical laws have replaced causal laws in theoretical physics. His mistake was to assume that the same applied to every advanced science. This is an instance of a saying of Wittgenstein’s (1953, § 593): A main cause of philosophical disease – a one-sided diet: one nourishes one’s thinking with only one kind of example. It also shows that one must study history and philosophy of science, and not philosophy of science without the history. Only the history of science can provide the variety of examples needed for philosophy of science. To sum up then: Russell’s thesis appears to be true of some advanced sciences, e.g. theoretical physics, but false of others, e.g. medicine. This situation poses the following question: why is it that some advanced sciences can dispense with the use of causal laws, whereas such laws continue to play a central role in other advanced sciences? We will consider this question further later in the chapter, but now let us turn to a consideration of Collingwood’s views on causality.

1.2  Collingwood’s AIM theory of causality In Collingwood’s 1938 paper on causality, he refers to Russell’s earlier 1913 paper on two occasions. In the first, he describes Russell’s 1913 paper as (1938, p. 101): a paper of very great importance, to which I shall have to refer again; but I want here and now to express my great admiration for it and my great

22  An action-related theory of causality

indebtedness to it. …  But I find myself, very reluctantly, unable to accept all of what I take Mr. Russell to mean. The second runs as follows (1938, p. 108): Among more recent writers I will first refer to Mr. Russell, whose brilliant paper “on the notion of cause,” mentioned above, is worth everything else put together that has been written on the subject during the present century. The neglect of this paper by the crowd of subsequent writers on the same subject is to my mind a very disquieting symptom of the state of English philosophy. Collingwood’s admiration for Russell is undoubtedly real, but it should not blind us to the fact that Collingwood is attempting to answer Russell, and to rehabilitate the notion of causality in the face of Russell’s critique. The truth is that Collingwood and Russell belonged to two different and opposed schools of philosophy. In the last decades of the 19th century and the early 20th century, philosophy in Britain was strongly influenced by a school of idealism which was to some extent based on Kant and Hegel. The leading representatives of this school were F.H. Bradley and T.H. Green at Oxford and J.M.G. McTaggart at Cambridge. Collingwood at Oxford and Russell at Cambridge were both initially brought up in this school, but their reactions to it were very different. Collingwood largely continued the tradition by developing a form of historical idealism. Russell, however, with his friend G.E. Moore, criticized idealism and developed a new approach to philosophy – analytic philosophy. A key tool of analytic philosophy was the new formal logic which Russell had helped to develop in his work Principia Mathematica, written with Whitehead. As far as the English-speaking world was concerned, analytic philosophy triumphed and the earlier school of British idealism was largely abandoned. This put Collingwood in an awkward position. In 1935 he was appointed to one of the leading chairs in philosophy at Oxford – the Waynflete Professorship of Metaphysical Philosophy – but this was at a time when most British philosophers were abandoning Collingwood’s approach in favour of the new analytic philosophy of Russell. In his autobiography, Collingwood says (1939, p. 33) that he “parted company with what I called propositional logic”, that is to say with the logic which Russell advocated as the basis of the analytic approach to philosophy. Collingwood goes on to say (1939, pp. 36-37): “For a logic of propositions I wanted to substitute what I called a logic of question and answer”. This logic of question and answer is a kind of dialectical logic. Regarding the development of the opposing propositional logic, Collingwood writes (1939, pp. 35–36): Hence that numerous and frightful offspring of propositional logic …  the various attempts at a ‘logical language’, beginning with the pedantry of the text-books …  and ending, for the present, in the typographical jargon of Principia Mathematica.

An action-related theory of causality  23

It has to be said that Collingwood’s attempts to criticize Russell’s logic did not meet with much success, but he was more fortunate when he criticized Russell’s views on causality. Collingwood distinguishes three senses of causality, and his AIM (Action, Intervention, Manipulation) account is applied only to causality in sense II. In this sense, what is caused is an event in nature. However, Collingwood goes on to say (1938, p. 89): In sense II …  the word cause expresses an idea relative to human action; but the action …  is an action intended to control …  things in “nature”, or “physical” things. In this sense, the “cause” of an event in nature is the handle, so to speak, by which we can manipulate it. …  This sense of the word may be defined as follows. A cause is an event or state of things which it is in our power to produce or prevent, and by producing or preventing which we can produce or prevent that whose cause it is said to be. Here Collingwood relates cause to action, and introduces the striking comparison of a cause to a handle by which we can manipulate the effect of the cause. Collingwood goes on to say that causes in his sense II occur in what he calls “practical sciences”, and he adds that, (1938, p. 90): “A conspicuous example of practical natural science is medicine”. Collingwood’s emphasis on the connection between causality and action and manipulation leads him to criticize Hume’s “constant conjunction” view. In a passage quoted in the introduction, Hume puts this position as follows (1738, p. 90): Thus we remember to have seen that species of object we call flame, and to have felt that species of sensation we call heat. We likewise call to mind their constant conjunction in all past instances. Without any further ceremony, we call the one cause, and the other effect, and infer the existence of the one from that of the other. Collingwood comments on this as follows (1938, pp. 93–4): For a mere spectator there are no causes. When Hume tried to show how the mere act of spectation could in time generate the idea of a cause, where “cause” meant the cause of empirical science, he was trying to explain how something happens which in fact does not happen. If sciences are constructed consisting of causal propositions in sense II of the word cause, … Their constituent propositions will be … experimental …  In calling them experimental I mean that their assertion will depend on “experiment.” No amount of “observation” will serve to establish such a proposition; for any such proposition is a declaration of our ability to produce or prevent a certain state of things by the use of certain means, and no one knows what he can do, or how he can do it, until he tries.

24  An action-related theory of causality

Collingwood here makes explicit a very important consequence of the AIM approach to causality. Any AIM theory of causality stresses the connection between causal laws and interventions. A causal law has to support an intervention. It follows from this, that, in establishing a causal law, it is desirable to have interventional evidence as well as observational evidence. By interventional evidence, I mean evidence obtained by making an intervention and recording its results. Note that “establishing a causal law” here does not mean “establishing such a law with complete certainty or beyond doubt” because scientific claims can never be established in this sense. “Establishing” should rather be understood in something like the following sense. Suppose a scientific claim has become so well confirmed by the available evidence that it can be accepted for the time being as the basis for action. Then this claim can be said to be established. However, this by no means excludes the possibility that the further advance of science will lead us to modify, or even reject, the claim. These considerations suggest the following principle, which I will call the Principle of Interventional Evidence: A causal law cannot be taken as established unless it has been confirmed by some interventional evidence. This Principle of Interventional Evidence does not apply universally since there are some areas of the social sciences where it is not possible to gather interventional evidence (see Russo [2009]). But in medicine it is always possible to collect interventional evidence, since in addition to statistical evidence in human populations, there is also laboratory evidence using experimental animals, tissues, or cells. So for establishing causality in medicine, I would argue that the Principle of Interventional Evidence should be accepted. That concludes what I want to say about Collingwood’s views on causality. I am now in a position to state my own version of the AIM approach to causality, which I call “an action-related theory of causality”. This theory is a development of Collingwood’s, and incorporates some ideas from Russell’s 1913 paper which was discussed in section 1.1.

1.3  Productive actions and avoidance actions The basic idea behind the action-related theory of causality is quite simple. It is that causal laws are useful and appropriate in situations where there is a close link between the law and action based on the law. The concept of cause has evolved in such a way that the transition between a causal law and carrying out actions based on that law is particularly easy and straightforward. Let us consider a causal law of the form A causes B. Collingwood stresses that causality is closely related to human action, and he mentions two types of action. One kind of action is designed to produce B, and I will call such an action a productive action. The other kind of action is designed to eliminate B, or to prevent B from occurring. I will call such an action

An action-related theory of causality  25

an avoidance action. We can illustrate these two types of action using Pearl’s example of turning a sprinkler on or off. The corresponding causal law is:

Turning on the sprinkler causes the grass to become wet. (1.1)

We can imagine that the sprinkler is operated by a handle, so that Collingwood’s analogy between a cause and a handle becomes literally the case in this instance. I will now make a few remarks about (1.1) which will obviously apply to causal laws of the same form. If the sprinkler is turned on, then the grass will always, ceteris paribus, become wet. For causal laws like (1.1) only apply “other things being equal”. Collingwood has an illuminating analysis of the ceteris paribus clause in terms of what he calls (1938, p. 91): “conditiones sine quibus non”, that is to say conditions which must hold if the causal law is to apply. In the case of (1.1) these conditions might be something like the following: the handle is properly attached to the rest of the sprinkler, the sprinkler is connected to the water mains, the water mains are functioning, etc. The point is that these conditions are never explicitly spelt out, but they are tacitly assumed as part of the background to the causal law. If the conditiones sine quibus non do in fact hold, then turning on the sprinkler is a sufficient condition for the grass to become wet. This is the standard case of a productive action, where instantiating the cause A produces the effect B. Let us now turn from productive actions to avoidance actions. Collingwood argues in the passage quoted above that, for a causal law of the form A causes B, we can prevent B from occurring by preventing A from occurring. However, this implies that A is a necessary condition for B, and, even when, as in this part of the book, we confine ourselves to deterministic causality, there are many cases in which this is not the case. Consider our example (1.1). Even if we keep the sprinkler off, the grass may still get wet. There are other causes apart from the sprinkler which can make the grass wet. For example,

Rain causes the grass to become wet. (1.2)

Now a causal law such as (1.2) poses some problems because we cannot manipulate the cause like a handle. We cannot turn the rain on or off at will. What we can do, however, is manipulate some of the conditiones sine quibus non, which are implicitly assumed by (1.2). (1.2) only holds if the grass is not covered by a waterproof sheet. Now in cricket matches, it is desirable to keep the grass of the wicket dry when it rains. So, in the event of rain, covers are put over the wicket. This is a successful avoidance action based on the causal law (1.2), but it consists in operating not on the cause, but on one of the conditiones sine quibus non, which are implicitly assumed by the causal law. I will call an action of this sort a blocking action. In this case the rain continues, but its usual effects are blocked. Having distinguished between productive and avoidance actions, it is now important to stress that avoidance actions are much more important in the case of 在医学中避免性行为比产生性行为更重要

26  An action-related theory of causality

medicine. Thus, to take the Pasteur example given in section 1.1, if we want to avoid having boils, we must take care to avoid being infected by staphylococci. If we have a boil, we should seek to eliminate it by taking an antibiotic which kills staphylococci. In general, the aim of medicine is to cure and prevent diseases, and not to produce diseases. I use the term “avoidance” rather than Collingwood “prevention”, because curing a disease is one way of avoiding the disease, but it is not usually considered to be “prevention”. A very important way of preventing a disease is vaccination, and it is clear from the above analysis that vaccination is a blocking action. Infection by a particular bacterium or virus may normally lead to the development of a disease, but, if the immune system is primed through vaccination to deal with that disease, this usual consequence of the infection is blocked. Vonka emphasizes the importance of avoidance actions in medicine, and he gives a maxim of St Thomas Aquinas which expresses this idea very well. It runs as follows (Vonka, 2000, p. 1835): Sublata causa, tollitur effectus (if the cause is removed, the effect is taken away) I will refer to this as Vonka’s Thomist maxim,3 since the maxim was used by St Thomas, and Vonka introduced it into 21st century discussions of causality in medicine. It might be objected that this maxim does not apply to blocking actions where the cause, e.g. rain or infective bacteria, remains, but its usual effects are blocked. However, once a cause has its usual effects blocked, it really ceases to be a cause. So, in this case, we can still say that the cause has been removed, and Vonka’s Thomist maxim applies. Note, however, that the maxim only applies if the cause is a necessary condition for the effect, which, as remarked above, is not always the case. The action-related theory of causality explains the value of causal laws both in everyday life and in medicine. Everyday life imposes on us the constant necessity for taking action, and so it is obviously convenient to cast our common-sense knowledge in a form which is closely linked to action, that is to say in the form of causal laws. In medicine too, a doctor has to try to cure his or her patients, and so is under an obligation to act. Thus, medical knowledge, even though it involves advanced scientific considerations going far beyond common sense, must still be cast in a form which is closely linked to action, that is to say in the form of causal laws. Turning now to the laws of theoretical physics, I do not of course want to deny that these too are linked to practical applications and actions. Thus, Maxwell’s equations can be used for making radio transmitters, or quantum mechanics for making nuclear weapons. However, the connection between the laws and the corresponding practical actions is much more remote in such cases than it is in everyday life or medicine. In the examples from physics, long mathematical calculations are needed, approximations must be made, and, usually, additional empirical assumptions introduced, before a theoretical scheme such as Maxwell’s

An action-related theory of causality  27

equations can be connected to a practical problem such as radio transmission. As the link between theoretical knowledge and practical action is here not at all close, laws other than causal laws can make their appearance, and indeed prove more convenient than causal laws. This constitutes a partial justification for Russell’s claim that as a science advances it abandons the use of causal laws, since this claim is in fact true of theoretical physics, even though it is false of medicine. The action-related theory of causality has, however, something more to say about the case of physics. Suppose we start with very theoretical laws of physics such as Maxwell’s equations or Schrö d inger’s equation, and then through mathematical calculations, the construction of more detailed models, etc. gradually move towards some practical application. The action-related theory of causality suggests that, as we approach practical actions, the causal laws which were absent at the very theoretical level might well reappear. This does indeed seem to be the case as is illustrated by the following simple example. Let us consider the ideal gas law PV = RT. This is a law of functional form relating the three variables P = pressure, V = volume, and T = temperature. The law is clearly not of causal form since we cannot say that any one of the variables, e.g. V, is a cause, and another, e.g. P, is an effect. The variables are all on a par. If, however, we apply the law to a particular concrete situation where a gas is being manipulated in some way, causal notions do reappear. However, what counts as a cause may differ in different applications. The law as we know holds approximately for some gases under certain conditions. Suppose we are handling such a gas under these conditions. Let us first suppose (see Figure 1.1) that this gas is in a cylinder one end of which is closed by a piston which can be moved in or out. The apparatus is held at constant temperature. It follows from the ideal gas law that if the piston is moved in, the pressure of the gas increases; whereas if it is moved out, the gas’s pressure decreases. But now we can consider the change in volume as the cause and the change in pressure as the effect. This is because, by means of the piston, we can manipulate the volume in order to produce changes in the pressure. Since there is now a close positive connection between the law as applied in this specific case, and the corresponding action, causality has reappeared just as the action-related theory would suggest that it should. Let us contrast this with a different application of the ideal gas law. Suppose our gas is now in a sealed container made of such a material that it can be regarded as having constant volume for the range of temperatures which occur. Suppose

FIGURE 1.1 

Gas in cylinder with piston.

28  An action-related theory of causality

FIGURE 1.2 

Sealed container over Bunsen flame.

this sealed container can be heated by a Bunsen flame as shown in Figure 1.2, or cooled by being placed in a refrigerator. It follows from the ideal gas law that if the temperature increases, the pressure of the gas will increase; whereas if the temperature is decreased, the gas’s pressure will decrease. Once again causality has reappeared in a specific case, but now it is the change in temperature rather than the change in volume which acts as the cause, since manipulating the temperature produces changes in the pressure. This simple example suggests a reason why, in the development of theoretical physics, it may have proved advantageous to replace causal laws with laws of other kinds. A single functional law such as the ideal gas law yields different 函数定律比因 果定律概括了 causal laws when applied in different situations. This functional law thus in a 更多的内容 sense summarizes a number of different causal laws and thus produces greater economy than could have been achieved by the exclusive use of causal laws. The advantage of non-causal laws may therefore be that they allow greater economy; their disadvantage that they create a greater gap between the law and action based on the law. This is a gap which may need to be filled by mathematical calculations, approximations, the construction of models, and the other devices mentioned above. That concludes my account of the action-related theory of causality as it applies in the case of generic deterministic causality. In the next chapter I will consider objections to this theory and how they might be overcome, as well as a point in favour of the theory. These objections, however, apply to any other AIM (Action, Intervention, Manipulation) theory of causality. So, this consideration of the objections (and the point in favour) is a good opportunity to compare the particular version of the AIM approach advocated here (the action-related theory) to other versions of this approach. The aim of this comparison will be to examine how successful the various versions of the AIM approach are in overcoming the objections to that approach, and how much they benefit from the point in favour of the AIM approach.

An action-related theory of causality  29

Notes 1 I am grateful to Maria Carla Galavotti and Huw Price for drawing my attention to this passage. I had helpful suggestions regarding its significance from them, and from Jon Williamson. 2 I owe the reference to Dingler to Marco Buzzoni. See his interesting 2014 paper. 3 The maxim: sublata causa, tollitur effectus, has a rather fascinating history. I discussed the matter with Brian Simboli, and he managed to locate a passage where St Thomas Aquinas uses the maxim. The reference is Summa Theologiae II–II, q. 20, a. 2 ad 1, where Aquinas says: “Ad primum ergo dicendum quod effectus tollitur non solum sublata causa prima, sed etiam sublata causa secunda.” (Reply to Objection 1. The effect is taken away not only if the first cause is removed, but also if the second cause is removed.) Here Aquinas seems to be using what he regards as a well-known principle of causality. It is unclear therefore whether he introduced the maxim, or whether it had become a well-known maxim in his day either introduced by an earlier scholastic or perhaps by either an Islamic philosopher, or by a philosopher of the ancient world. I have not been able to find an earlier use of the maxim. Brian Simboli also discovered that the maxim is cited by Robert Burton in his Anatomy of Melancholy, first published in 1621. Burton says (1621, p. 92): Fernelius puts a kind of necessity in the knowledge of causes …  without which it is impossible to cure or prevent any manner of disease, Empirics may ease, and sometimes help, but not thoroughly root out; sublata causa tollitur effectus as the saying is, if the cause be removed, the effect is likewise vanquished. Kant endorses the maxim in his Vorlesungen ü ber die Metaphysik (Lectures on Metaphysics). These were given from the 1760s to the 1790s, and published after Kant’s death in 1821. In the chapter ‘Von der Ursache und Wirkung’ (Of Cause and Effect), Kant, after mentioning another causal maxim, says (1821, p. 72): “Aber sublata causa tollitur effectus ist eben so gewiss.” (But sublata causa tollitur effectus is just as certain.) The maxim is also cited by Claude Bernard in his Introduction to the Study of Experimental Medicine of 1865. Bernard says (1865, pp. 55–56): Counterproof decides whether the relation of cause to effect, which we seek in phenomena, has been found. To do this, it removes the accepted cause, to see if the effect persists, relying on that old and absolutely true adage: sublata causa, tollitur effectus. This is what we still call the experimentum crucis. It is clear that the maxim: sublata causa, tollitur effectus has been often cited and used from the 13th century to the present. However, the exact origins of the maxim are still not known, though perhaps they will be discovered by further research.

2 GENERAL DISCUSSION OF AIM THEORIES OF CAUSALITY

2.1  Gasking’s contribution

Copyright © 2018. Routledge. All rights reserved.

The next contribution to the AIM approach to causality after that of Collingwood was made by Gasking in 1955. Gasking’s version of the AIM theory largely follows Collingwood, but Gasking discusses two things which are not mentioned by Collingwood. The first is an objection to the AIM approach which concerns causal claims in which the cause cannot be manipulated by humans. How can such claims be dealt with in an approach which links causality to human actions and manipulations? This is how Gasking puts it (1955, p. 483): one can sometimes properly say of some particular happening, A, that it caused some other particular event, B, even when no-one could have produced A, by manipulation, as a means of producing B. For example, one may say that the rise in mean sea-level at a certain geological epoch was due to the melting of the Polar ice-cap. Gasking’s example is one of singular causation and falls outside the range of this book which deals only with generic causation. However, one can easily give similar examples which involve generic causation (causal laws). One such is the following.

The Moon causes the tides. (2.1)

This is generally accepted as correct, and yet humans cannot manipulate the position of the Moon in order to create higher or lower tides. This objection (that some causes cannot be manipulated) will be discussed in section 2.2. Then in sections 2.3 and 2.4 I will discuss two further objections to AIM theories of causality which are not mentioned by Gasking, though they are listed and discussed by Menzies and Price (1993, sections 4 and 6).

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:47:55.

General discussion of AIM theories of causality  31

As well as introducing an important objection to the AIM approach to causality, Gasking also introduces a point in its favour. An important feature of causality is that it is normally asymmetric. If A causes B, then usually B does not cause A. In the example given earlier, although rain causes mud, it is not true that mud causes rain. Causality differs from probability in this respect. If we are dealing with nonzero probabilities, then, if the probability of A given B [P(A | B)] is well-defined, so is the probability of B given A [(P(B | A)]. In our example, both P(mud | rain) and P(rain | mud) are well-defined. This difference between causality and probability is what is referred to as the Humphreys’ paradox. But how is this asymmetry of causality to be explained? Gasking points out that if a cause A strictly precedes a cause B in time, then we can easily explain the asymmetry of causality on the AIM approach. Gasking writes (1955, p. 483): This account fits in with the principle that an event, A, at time t2 cannot be the cause of an event B at an earlier time, t1. It is a logical truth that one cannot alter the past. One cannot, therefore, by manipulations at t2 which produce A at t2 also produce B retrospectively at t1. The question of the asymmetry of causation becomes a bit more complicated if one introduces, as I think one should, simultaneous causation as well as causation in which the cause strictly precedes the effect in time. In section 2.5 I will discuss the explanation of causal asymmetry in a more general setting which includes simultaneous causation. First, however, I will discuss the three principal objections to the AIM approach to causality in the next three sections. In each case I will compare the answers to these objections given by the action-related theory of causality to those given other versions of the AIM theory of causality, namely that of Menzies and Price, and that of Woodward.

Copyright © 2018. Routledge. All rights reserved.

2.2  Objection 1. Some causes cannot be manipulated Our problem then is to see how the action-related theory of causality can be applied to causal laws of the form “A causes B” where A cannot be manipulated by human actions. In fact, we already considered a case of this sort in Chapter 1, namely “Rain causes the grass to get wet”. The technique was to consider an avoidance action based on this law. What needed to be done was not to manipulate the cause (rain) which cannot be turned on and off by human actions, but rather to manipulate the ceteris paribus conditions presupposed by the law. One of these ceteris paribus conditions is that the grass is not covered, and so by altering this condition and covering the grass with a waterproof sheet, we can carry out an avoidance action based on the law. This was called “a blocking action”. This then is one general method of dealing with causes which cannot be manipulated. We make use of avoidance actions – particularly blocking actions. I will now give another example of the use of this technique, and then describe a second method which covers a different set of cases.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:47:55.

32  General discussion of AIM theories of causality

Menzies and Price (1993, p. 195) give the following example of an ­unmanipulable cause: “the 1989 San Francisco earthquake was caused by friction between continental plates.” They comment (p. 195): Obviously, no normal agent has the capacity to bring about friction between continental plates. The agency approach, then, appears to be seriously defective in not being able to accommodate such straightforward causal claims. Now this example is a singular causal claim while in this book I am dealing with generic causal laws. So, I will change the example to turn it into the following general law.

Copyright © 2018. Routledge. All rights reserved.



Earthquakes are caused by friction between continental plates. (2.2)

Here we cannot manipulate continental plates to cause or prevent earthquakes, but, we can certainly use (2.2) as the basis of an avoidance action. We need only refrain from going into areas which are on the boundary between continental plates, and we can be sure of avoiding earthquakes. Moreover, those who need to stay near a boundary between continental plates can, and do, take actions to block the harmful effects of earthquakes. There is, however, a second strategy for dealing with unmanipulable causes, and that is to use Russell’s approach of eliminating causal laws in favour of other sorts of law such as functional laws. As we argued in Chapter 1, Russell’s plan for eliminating cause and causal laws from all advanced science certainly cannot be carried out, for the use of the concept of cause is indispensable in medicine which is clearly an advanced science. However, Russell’s approach does seem to work well for theoretical physics, and, in particular, for his own example of celestial mechanics. We can therefore use Russellian elimination to deal with unmanipulable causes in areas such a theoretical physics and cosmology. An example for which this treatment seems appropriate is the following: “The change of seasons is caused by the movement of the Earth in its orbit round the Sun.” This apparently causal law can easily be replaced by a functional law. The seasons depend on the angle of elevation (θ  say) of the sun at midday, and on the direction of change dθ /dt of this angle. A relatively simple functional law relates θ , dθ /dt to the position φ  say of the Earth in its orbit round the Sun. An example like “the background radiation in the universe was caused by the big bang” can be dealt with in the same way, although the non-causal laws involved will obviously be more complicated. The example, given earlier as 2.1, of the Moon causing the tides can also be dealt with by this Russellian technique. So, we can sum up as follows. The existence of some causes which cannot be manipulated is a difficulty for any AIM theory of causality. Within the context of the action-related theory, this can be resolved by the use of one or other of two strategies. The first strategy involves avoidance actions, particularly blocking

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:47:55.

General discussion of AIM theories of causality  33

actions, rather than productive actions. The second strategy employs Russellian elimination of causal laws in favour of functional laws. The second strategy is useful particularly in cases of theoretical physics and cosmology. Let us now turn to considering how the same difficulty about unmanipulable causes is dealt with in other versions of the AIM (Action, Intervention, Manipulation) approach. One of the most important alternative versions is the “agency theory of causality” developed by Price (1992) and Menzies and Price (1993). I will next give two general statements of this position, one from each of the two papers just cited. The first comes from Price (1992, p. 514): Roughly, to think of A as a cause of B is to think of A as a potential means for achieving (or making more likely) B as an end. The second is from Menzies and Price (1993, p. 189):

Copyright © 2018. Routledge. All rights reserved.

the common idea to agency accounts of causation is that an event A is a cause of a distinct event B just in case bringing about the occurrence of A would be an effective means by which a free agent could bring about the occurrence of B. These passages show that the agency theory of causality is closely related to the action-related theory, but that there is a difference. The two characterizations of agency theories of causality only mention what we have called productive actions based on a causal law, and not avoidance actions based on such a law. Yet avoidance actions are very important, particularly in medicine. It is true that smoking causes lung cancer, but this does not mean that smoking is an effective means by which a free agent could give himself or herself lung cancer. It is rather that refraining from smoking is a sensible strategy for trying to avoid getting lung cancer. The term “agency” does indeed suggest the productive action, and this is why I have introduced the term “action-related” which suggests that it is possible to base different types of action (both productive and avoidance) on causal laws. Now one of our strategies for dealing with unmanipulable causes involved avoidance actions. As Menzies and Price do not consider such actions, this suggests that they will use another method for handling unmanipulable causes, and this is indeed the case. They suggest that the agency theory should be weakened in a way they describe as follows (1993, pp. 197–8): In its weakened form, the agency account states that a pair of events are causally related just in case the situation involving them possesses intrinsic features that either support a means-end relation between the events as is, or are identical with (or closely similar to) those of another situation involving an analogous pair of means-end related events. Clearly, the agency account, so weakened, allows us to make causal claims about unmanipulable events such as the claim that the 1989 San

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:47:55.

34  General discussion of AIM theories of causality

Francisco earthquake was caused by friction between continental plates. We can make such causal claims because we believe that there is another situation that models the circumstances surrounding the earthquake in the essential respects and does support a means-end relation between an appropriate pair of events. The paradigm example of such a situation would be that created by seismologists in their artificial simulations of the movement of continental plates. Gasking, who first drew attention to the problem of unmanipulable causes, proposed a rather similar solution. Gasking’s example was that of the melting of the Polar ice-cap causing the rise in mean sea-level at a certain geological epoch. Gasking comments (1955, p. 483): But when one can properly say this sort of thing it is always the case that people can produce events of the first sort as a means to producing events of the second sort. For example, one can melt ice in order to raise the level of water in a certain area.

Copyright © 2018. Routledge. All rights reserved.

模拟

Here again, the idea seems to be to create a simulation. Because we cannot melt the Polar ice-cap at a remote geological epoch, we place some ice in a bucket of sea water and see if the melting of the ice raises the level of the sea water in the bucket. This use of simulations does not seem to me a satisfactory solution to the problem. The whole point of an agency or action-related theory of causality is to link causal laws with practical human actions in the real world. Simulations are better considered as a theoretical preliminary to practical action rather than as part of such action. Let me next turn to considering the way in which Woodward deals with the problem in the version of the AIM approach which he develops in his 2003 book. Woodward’s theory hinges on the unusual account he presents of the notion of intervention. Woodward gives a formal definition of an intervention on X with respect to Y (2003, p. 98). This definition is long and complicated. The first step is to define an intervention variable or IV. This requires four clauses. Then, in a further clause an intervention or IN is defined in terms of IV. These definitions are rather technical, and so we will consider them in more detail only in Appendix 1. As far as the main text of the book is concerned, we can get an idea of Woodward’s approach by considering some of the informal remarks about his notion of intervention, which Woodward makes in the following pages. The first point is that interventions in Woodward’s sense need not be human actions. Woodward is very clear about this, writing (2003, p. 103): although there will be realistic cases in which manipulations carried out by human beings will qualify as interventions in virtue of satisfying IN, the conditions in IN make no reference to human activities or to what human

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:47:55.

General discussion of AIM theories of causality  35

beings can or can’t do. Notions such as “human agency” and “freely ­chosen action” do not occur as primitives in IN. Instead, the conditions in IN are characterized purely in terms of notions such as “cause” and (statistical) “independence.” An event or process not involving human action at any point will qualify as an intervention on X as long as it satisfies IN. (It is this possibility that scientists have in mind when they speak of “natural experiments.”) In this respect, a manipulability theory that appeals to IN is quite different from traditional agency theories (such as those of von Wright …  and Menzies and Price 1993). We can add that Woodward’s theory also differs in this respect from the other theories in the AIM tradition which we mentioned above, that is to say Collingwood 1938 and 1940, Gasking 1955, and Gillies 2005. This is also true of more recent work in the AIM tradition. For example, Buzzoni writes in 2014 (p. 377): In spite of the unquestionable merits of Woodward’s theory, I cannot agree with its separating the notion of causality from that of human intervention. On this point, I side with von Wright, Price, and Menzies: the close link between intervention and causality cannot be understood without reference to the free agency of human beings. In fact, in contrast to Woodward, all the other theories in the AIM tradition regard interventions and manipulations as always human actions. Even more surprisingly, Woodward allows interventions, which are not physically possible. Once again, he is very clear about this, writing (2003, p. 129):

Copyright © 2018. Routledge. All rights reserved.

It is of considerable interest that there appear to be cases in which X causes Y but interventions on X are not physically possible even in the weak sense. The example, which Woodward considers, is the one which we gave as (2.1), namely that the Moon causes the tides. In his discussion, Woodward begins by stating this rather more precisely as follows (2003, p. 129): Changes in the position of the moon with respect to the earth and corresponding changes in the gravitational attraction exerted by the moon on various points on the earth’s surface cause changes in the motion of the tides. He then goes on to say (2003, pp. 129–130): All physically possible processes that would change the position of the moon might be …  “too ham-fisted” to satisfy all the conditions in IN. …  Perhaps with sufficient ingenuity we may be able to think up some physically possible intervention that changes the position of the moon and

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:47:55.

除了Woodward 之外，其他AIM 理论家都将干预 视为人类行为。

36  General discussion of AIM theories of causality

Copyright © 2018. Routledge. All rights reserved.

that satisfies the conditions IN, but I hope to have made it plausible that ­nothing in the truth of the original causal claim (3.5.1) guarantees that this will be possible. What we need for such an intervention is a physically possible process that is sufficiently fine-grained and surgical that it does not have any other effects on the tides besides those that occur through the change that it produces in the position of the moon, and it may well be that the laws of nature guarantee that all real causal processes will have such additional effects. For this reason, Woodward allows interventions, which may not be physically possible, but only “logically or conceptually possible” (2003, p. 132). If such interventions violate a law of nature, he is prepared to refer to them as “miracles” (2003, p. 133). So far, we have seen that Woodward’s definition of intervention goes beyond the usual sense of intervention by allowing interventions, which would not normally be considered as such. However, because of the complexity of his definitions of IV and IN, with their total of five clauses which need to be satisfied, it may well be that there are many interventions in the normal sense which are not interventions in Woodward’s sense – making Woodward’s notion of intervention more restrictive than the usual one. Woodward himself does not consider this possibility, but it turns out that it does in fact occur. In fact, very simple medical interventions fail to satisfy Woodward’s formal definition of an intervention. For example, suppose a doctor is treating a patient who has high blood pressure. The doctor prescribes a pill. The patient takes the pill and his blood pressure drops. This would naturally be regarded as a successful medical intervention, but it turns out not to be an intervention in Woodward’s sense. This is fairly easy to show, but, because of the technical complexity of Woodward’s definition, I give the demonstration of this result in Appendix 1. In this appendix, I also explain why, in my view, the complexity of Woodward’s definition of intervention is quite unnecessary, and that, for the discussion of causality in medicine, we can take the meaning of “intervention” to be given by the way it is ordinarily used in medical practice. Let us return to our example: “The Moon causes the tides” (2.1). The problem with this case for AIM theories of causality is that human actions cannot change the position of the Moon in order to produce higher or lower tides. Woodward, as we have seen, solves this problem by introducing miraculous interventions which change the position of the Moon in a way which contradicts the laws of physics. The obvious objection to this solution is that (2.1) is a fairly standard scientific law, and yet its explication on Woodward’s account involves highly metaphysical assumptions concerning miraculous interventions. Would it not be better to explicate (2.1) in a way which does not stray outside the boundaries of scientific empiricism into the regions of dubious metaphysics? In fact, as I will now show, if we adopt the solution suggested earlier which involves Russell’s eliminationism, there is no need to make any dubious metaphysical assumptions.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:47:55.

存在简单的医 学干预不符合 Woodward的 定义。

General discussion of AIM theories of causality  37

Copyright © 2018. Routledge. All rights reserved.

Russell suggested that the concept of cause and causal law should be eliminated in all advanced sciences in favour of the use of other types of law such as functional laws. I argued that this was not a reasonable proposal for medicine, but that for theoretical physics and Russell’s own example of celestial mechanics, it was entirely acceptable. To apply this approach to our example: “The Moon causes the tides”, I will simplify by using only classical mechanics and ignoring relativistic refinements. Newton’s theory can be formulated in terms of the three laws of motion and the law of gravity, where all four of these laws can be stated in functional rather than causal form. This theory gained strong empirical confirmation by deriving from it many results which could be checked empirically. Some of these concerned the solar system, such as Kepler’s laws as well as deviations from these laws. Others concerned terrestrial mechanics, such as Galileo’s law of falling bodies and connected results concerned with projectiles and pendulums. Newton’s well-confirmed theory was applied to calculate the gravitational forces which the Moon exerted on the waters of the Earth’s oceans. In this way a theory (T say) of the relation between the Moon and the tides was produced, and T could in its turn be tested against empirical observations.T of course did not use the concept of cause, and the laws it involved were functional rather than causal in character. Using Russell’s approach, we eliminate (2.1) in favour of  T. In this way, the problems created by (2.1) for AIM theories of causality are resolved without going beyond scientific empiricism into any dubious metaphysics. So, my criticism of Woodward’s approach is that it involves unacceptable metaphysics. What is curious here is that Woodward in a recent paper (2014) declares that he is very opposed to metaphysics. He claims in this paper that he is giving a functional account of causation, and, to quote part of the title of the paper, presenting a “defense of the legitimacy of causal thinking by reference to the only standard that matters – usefulness …  as opposed to metaphysics … ”. Yet his functional account includes appeals to miraculous interventions on the Moon’s orbit which contradict the laws of physics. This is surely a somewhat metaphysical notion!

2.3  Objection 2. Causes exist independently of humans Any AIM theory of causality relates causes to actions, interventions, and manipulations. Thus it would seem on this approach, there can be no causes without human beings to carry out these actions, interventions, and manipulations. Yet many causes seem to exist quite independently of human beings. For example, the law that rain causes mud, surely held at the time of the dinosaurs, long before there were any human beings. This objection is sometimes expressed by saying that AIM theories of causality are unacceptably anthropocentric. This is certainly not a problem for Woodward, since for him interventions need not be human actions. Indeed, there is a section of his 2003 book (pp. 103–4) entitled nonanthropomorphism. But can the difficulty be overcome in agency or action-related theories? I believe that it can, and my answer to the difficulty falls into two parts.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:47:55.

38  General discussion of AIM theories of causality

I will first argue that agency or action-related theories do not prevent causal laws being objective and potentially human-independent. However, I will then qualify this by arguing that such theories do give the concept of cause an anthropocentric aspect, and that this is desirable rather than harmful. The first line of argument runs as follows. The natural world is to be considered as highly complicated and subject to many variations, while at the same time obeying quite a number of laws and regularities. From this profusion of laws, humans, naturally enough, pick out those, which are useful to them in carrying out actions either to achieve desirable goals or to avoid undesirable situations.This is why there is, in particular, a search for laws, which are of causal form, and so can be related to human action in a simple and straightforward way. These laws have the property of being closely linked to human action, but they hold objectively quite apart from this property. Indeed, if they did not hold objectively, they would not be useful as a basis of action.This is why we can suppose that so many causal laws would continue to operate in the absence of humans. The point could be put in this way. The fact that a causal law is a law of nature means that it holds objectively, independent of humans. Yet the fact that it has been picked out from the complexity of the natural world by humans as a causal law shows that it is closely related to human action. This is why causal laws can at the same time be objective and potentially human independent, while having an anthropocentric aspect.

2.4  Objection 3. Unavoidable circularity

Copyright © 2018. Routledge. All rights reserved.

Let us turn to the third objection to AIM theories, which is that they appear to involve a vicious circularity. This problem is stated so clearly by Menzies and Price in the context of their agency theory that I will quote their formulation of it (1993, p. 193): The apparent circularity is plain enough in our informal statement of the agency approach, according to which A is a cause of a distinct event B just in case bringing about A would be an effective means by which a free agent could bring about B. This statement contains two references to ‘bringing about’, which seems on the face of it to be a causal notion: doesn’t an agent bring about some event just in case she causes it to occur? It would appear, then, that agency accounts are vitiated by the fact that they employ as part of their analyses the very concept which they are trying to analyse. Obviously, the same problem arises in the action-related theory of causality, since any human action can be considered as a human agent causing something to occur, and so the notion of action involves that of causality. On this point, Woodward writes (2003, p. 124): However, unlike the view I have defended, Menzies and Price attempt to appeal to the notion of agency to provide a noncircular, reductive analysis of causation. Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:47:55.

General discussion of AIM theories of causality  39

Here Woodward claims that Menzies and Price are reductionists, i.e. that their project is to define causation in terms of agency, and so to reduce causality in general to human actions in particular. Woodward himself is not a reductionist. He writes regarding any manipulability theory similar to his own (2003, p. 106):

Copyright © 2018. Routledge. All rights reserved.

a manipulability theory can provide a nontrivial constraint on what it is for a relationship to be causal without providing a reductive analysis of causality. I won’t here go into the details of the dispute between Menzies and Price, and Woodward on this point since this is dealt with by Price himself (2017). Instead I will argue that Woodward’s anti-reductionist approach can be applied also to an agency or action-related theory, and this approach is actually more plausible for such theories than it is for Woodward’s own theory. Philosophers often try to analyse a concept (C say) in order to make C’s nature clearer and to enable it to be used better. One approach to such an analysis is to try to define C in terms of another concept, D, which is taken to be clearer and less problematic than C. This is reductionism. C is reduced to D. This approach was popular in the past, but it never led to any very successful reductions. So, as Woodward points out (2003, p. 106) a non-reductionist approach to philosophical analysis is now often preferred. On the non-reductionist approach, no attempt is made to define C. Instead C is accepted as a primitive, undefined concept. However, an attempt is made to clarify C by showing how C is related to other concepts such as D. So, in the case of agency and action-related theories, cause need not be defined in terms of agency and human actions, but an attempt could be made to show how cause relates to human actions. Now human actions, as we have pointed out, do involve some notion of causality. If we were trying to reduce causality to human action, this would be a vicious circle. Since we are only trying to relate causality to human actions, the circle is not vicious. However, there is still a circle here. Does this mean that this attempt to clarify the concept of causality is useless? To answer this question, let us take a particular example. One area of medical research is concerned with investigating possible viral causes of cancers. However, the sense of cause in which a virus might cause a cancer is not at all clear, and is certainly in need of clarification. Might relating the concept of cause here to human action help in this clarification? I certainly think that it might, and the fact that there is some notion of cause involved in the concept of human action is not a problem. Cause in the context of human action is more familiar and less problematic than in those parts of medical theory which deal with possible viral causes of cancers. We all learn a great deal about human actions from an early age, both by acting ourselves and by observing the actions of others. This is an area about which we know a great deal, and so it makes complete sense to use it in trying to clarify causality in more problematic areas. Consequently, an anti-reductionist strategy is quite in order here.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:47:55.

40  General discussion of AIM theories of causality

Woodward’s use of anti-reductionism is, however, much more problematic. Woodward defines his concept of intervention (IN) using cause. In fact, the word “cause(s)” appears in every one of the five clauses which he uses in the definition of IN. He then goes on to use IN to analyse causality. The circularity of his anti-reductionism is just as pronounced as it is in the case of agency or action-related theories. However, now comes the difference. The agency or action-related theories analyse causality in general using a concept (human action) which does indeed involve causality, but which is a concept that is very familiar and about which a lot is known. Does the same apply to Woodward’s concept IN? Clearly it does not. Woodward’s concept IN is very obscure, complicated and unfamiliar. It is actually very hard to tell whether some putative intervention is really an example of IN. Some examples of IN are definitely not interventions in the usual sense of the word, while some interventions in the usual sense of the word are not interventions in the sense of IN. These then are my reasons for holding that the non-reductionist strategy, even though it involves some circularity in both cases, is much more plausible in agency and action-related theories than it is in Woodward’s manipulationist theory.

2.5  Explanation of causal asymmetry Let us now turn from objections to AIM theories to something in their favour. As we showed in section 2.1, Gasking in 1955 pointed out that an AIM theory of causality provides a convincing explanation of the asymmetry of causation. This is also stressed by Price who writes (1992, p. 515):

Copyright © 2018. Routledge. All rights reserved.

the agency account of causation …  has a significant and largely unrecognized advantage; it is particularly well placed to explain the nature of causal asymmetry, and its prevailing orientation in time. I will show in this section that both the action-related and agency theories of causality do indeed explain the asymmetry of causality, but that Woodward’s theory cannot explain the asymmetry. This is another considerable disadvantage of Woodward’s approach. First, however, we must look more closely at causation in relation to time. Consider the law “A causes B”. As we saw in the introduction, Hume argued that a cause A always precedes its effect in time, but he also added (1738, p. 79) that this principle “is not so universally acknowledged, but is liable to some controversy.” In fact, other authors have claimed that A can occur at the same time as B. Kant mentions simultaneous causation, although it raises some difficulties for his views. He writes (1781/7: A202/B248): At this point a difficulty arises with which we must at once deal. The principle of the causal connection among appearances is limited in our formula Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:47:55.

General discussion of AIM theories of causality  41

to their serial succession, whereas it applies also to their coexistence, when cause and effect are simultaneous. For instance, a room is warm while the outer air is cool. I look around for the cause, and find a heated stove. Now the stove, as cause, is simultaneous with its effect, the heat of the room. Here there is no serial succession in time between cause and effect. They are simultaneous …  Medicine affords many examples of simultaneous causation. A tumour may be the同时因果关系：肿 cause of a patient’s pain, although tumour and pain occur simultaneously. In such瘤导致疼痛。 cases of simultaneous causation, we are no longer dealing with a situation in which one event causes another event, but with a situation in which two processes are involved, one of which causes the other. Thus, one process (the burning of fuel in the stove) causes another process (the temperature of the room remaining above that of the outside air). In the medical case, we can suppose that the tumour and pain begin and develop together. The more the tumour grows, the worse the pain. It seems to me that examples of this type do definitely establish the existence of simultaneous causation. In section 2.1 it was shown how Gasking derived the asymmetry of causation from his agency theory of causality in the case where a cause A is earlier in time than its effect B. We must now see whether, and to what extent, this can be extended to cover the case of simultaneous causation if we adopt the action-related theory of causality. We must begin by assuming the following principle:

Copyright © 2018. Routledge. All rights reserved.



Human action cannot change the past. (2.3)

Gasking, in the quotation given earlier, states that this is a logical truth. Perhaps this claim is too strong for some philosophers have questioned (2.3). However, (2.3) certainly applies in nearly all cases, and I think, therefore that we are justified in adopting it. From (2.3) and the action-related theory of causality, a second principle follows immediately, namely:

“A causes B” cannot hold, if B occurs earlier in time than A. (2.4)

According to the action-related theory, if “A causes B” holds, then we can either produce B by manipulating A, or avoid B by manipulating A, or by blocking its usual effects. If, however, B occurred earlier in time than A, such manipulations, or blocking actions, would involve altering the past, which, by (2.3) is not possible. Hence (2.4) is true. From (2.4) it follows in turn that if A causes B, then either B occurs later than A, or A and B are simultaneous. Let us consider these two cases in turn. If B is later than A, then it is immediate by (2.4) that B cannot cause A. So, the asymmetry of causation is established in this case. The case in which A and B are simultaneous is more complicated and requires more careful consideration.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:47:55.

42  General discussion of AIM theories of causality

If A and B are simultaneous, then the asymmetry of causation continues to hold in many cases, but there are some cases of simultaneous causation in which the asymmetry breaks down, and we have that A causes B and B causes A, i.e. an interactive process. However, proceeding on a case by case basis, we can still use the action-related theory to explain why asymmetry applies or fails to apply. In the two cases of simultaneous causation considered earlier, asymmetry continues to apply. The stove causes the elevated temperature of the room, but not vice versa. The tumour causes the pain, but the pain does not cause the tumour. The action-related theory of causation explains these asymmetries as follows. If we want to heat the room, we can light the stove, but we cannot ensure that the stove is burning by altering the temperature of the room, assuming that we have only 18th-century technology at our disposal, and there are no thermostats and heating control systems. Similarly, it may be possible to eliminate the patient’s pain by manipulating the tumour, for example by excising it surgically. However, it is clearly not possible to alter the tumour by acting upon the pain in some way. Let us now consider an example of simultaneous causation where the asymmetry of causation breaks down and we have an interactive process. Kistler provides an excellent example of this kind (2013). This is a device called a magnetic stirrer, which is used in chemistry for mixing substances. For such a stirrer if L = the angular momentum, and μ  = the magnetic moment, these obey the law L = μ ( 2m/e), where m is the mass of the object, and e its electric charge. Kistler remarks (2013, p. 69):

Copyright © 2018. Routledge. All rights reserved.

The law can be tested either by manipulating L by exerting a mechanical force F on the rotating object while holding fixed its mass m and charge e, and observing the change in μ , or by manipulating μ . The latter can be done by increasing the strength of the magnetic field B, which accelerates the rotation of the charge. In the case of a magnetic stirrer, therefore, we can either produce changes in μ by manipulating L, or produce changes in L by manipulating μ . So, it follows from any AIM theory of causality that L causes μ , and μ  causes L. This then is a case of simultaneous causation in which the asymmetry of causation breaks down and we have an interactive process. Kistler uses this example in order to criticize AIM theories of causality. He specifically mentions Woodward’s theory, but his argument applies more generally. According to Kistler (2013, p. 70): The asymmetry of cause and effect belongs to the conceptual core of causation: …  it is part of the content of the notion of causation that if x causes y, then y cannot at the same time cause x. Since AIM theories of causality lead to the conclusion that, in the case of the magnetic stirrer, x causes y and y causes x, it follows, according to Kistler, that such theories are not acceptable. Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:47:55.

Copyright © 2018. Routledge. All rights reserved.

General discussion of AIM theories of causality  43

My reply to this is that Kistler has stated the principle of the asymmetry of causation in too strong a form. It is correct that if x causes y, then usually y does not cause x, but it is also true that, in the case of simultaneous causation exceptions to the asymmetry of causation do sometimes occur. Although the actionrelated theory of causality cannot give a general criterion for showing when the asymmetry of causation holds in the case of simultaneous causation, it can explain whether this asymmetry holds or breaks down on a case by case basis. Thus, the agency theory, and its generalization the action-related theory of causation, provide a simple and fully satisfactory explanation of causal asymmetry. These theories explain not only the cases where causal asymmetry holds, but also the exceptional cases of simultaneous causation where it breaks down. No other theories of causality are so successful in this regard. Woodward’s theory, however, loses this attractive feature of the other AIM theories of causality because his definition of intervention severs the link with human action. An intervention IN on his account is defined in terms of causality and need not be a human action. It follows that we cannot obtain the time asymmetry of IN except by assuming that causation is asymmetric. On his theory, the asymmetry of causation has to be postulated, and cannot be derived. Specifically, in the argument given above, the derivation of (2.4) from (2.3) goes through in the agency and action-related theories, but fails in Woodward’s theory. Perhaps for this reason, the asymmetry of causation does not appear in the index of his 2003 book, or, as far as I have been able to ascertain, in the text of the book. That concludes my account of the action-related theory of causality in the case of deterministic causality. It still remains to extend this theory to indeterministic causality, and this will be carried out in Part III of the book. There are two reasons for treating the deterministic case separately, and before dealing with the indeterministic case. First of all, the deterministic case is simpler and, in particular, does not require a consideration of the problematic relationship between probability and causality. This means that the main features of the action-related theory can be more easily understood in the deterministic case. Secondly, the deterministic case is still important for medicine. The pioneers of what was then known as the germ theory of disease, namely Lister, Pasteur and Koch, all used a deterministic notion of causality. They were investigating what are now known as infectious diseases produced by bacteria. Their results are still largely accepted today, and so the deterministic notion of causality is still used in the context of bacterial diseases, even though now other areas of medicine very often use indeterministic causality. It is also worth noting that both Pasteur and Koch regarded causes as necessary as well as sufficient for their effects. The necessity of causes follows from Koch’s postulates which we will give in the next chapter (section 3.3), while Carter (2003, pp. 114–115) quotes Pasteur as saying in 1878: “two phenomena are in a relation of cause and effect if when one of the two exists the other follows”. So, if either Pasteur or Koch claim that A causes D where D is some disease, it follows that D can be avoided by ensuring that A does not occur, or that its usual effects are blocked. Vonka’s Thomist maxim: sublata causa, tollitur effectus (if the cause is removed, the effect is taken away) applies.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:47:55.

44  General discussion of AIM theories of causality

Copyright © 2018. Routledge. All rights reserved.

The aim of this book is to produce a philosophical account of causality in medicine. So, any theoretical views of causality which are developed, should be tested out by seeing whether they apply satisfactorily to important areas of medicine. For the notion of deterministic causality, the obvious choice of examples to consider are some of the results concerning bacterial diseases, which were discovered by Lister, Pasteur and Koch. In fact, I have already given (in section 1.1) an example of a discovery of Pasteur’s. In Chapter 3, I will consider some of the discoveries of Koch, focussing particularly on his work on tuberculosis and cholera. Koch not only researched into the causes of bacterial diseases, but formulated some criteria, which, if satisfied, were supposed to justify us in asserting that a particular bacterium is the cause of a particular disease. These criteria, known as Koch’s postulates, are still cited in contemporary textbooks of bacteriology. It will be an interesting task to assess them in the light of the action-related theory of causality.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:47:55.

3 AN EXAMPLE FROM MEDICINE Koch’s work on bacterial diseases and his postulates1

Copyright © 2018. Routledge. All rights reserved.

3.1  The background to Koch’s work Koch started his research in medicine about a decade after the first introduction of the germ theory of disease. Previous to this theory, diseases had been largely explained in terms of miasmas or chemical contagions. The idea behind the germ theory was instead that diseases were caused by micro-organisms which entered the body and multiplied there, bringing about the damage which the disease caused. Of course, the germ theory of disease was never applied to all diseases, but it was applied to the most serious diseases of the time. The germ theory of disease is now so generally accepted that it comes as rather a surprise that, apart from a few precursors who did not gain any general acceptance, it was only introduced around 1865. The two individuals mainly responsible for introducing the germ theory were Lister in Britain and Davaine in France. Both were strongly influenced by Pasteur’s work on fermentation and putrefaction. Pasteur had established that both fermentation and putrefaction were caused by micro-organisms, and Lister and Davaine thought of extending this approach to medicine. Lister was a surgeon, and therefore concerned with the main problem which affected surgery in the early 1860s, namely wound suppuration. Anaesthetics had been introduced for surgery in 1846, and this made possible longer and more complicated operations. However, the difficulty with all operations was that the wounds involved often became septic, and the sepsis would often spread killing the patient. Death rates following amputations were about 25% in the best hospitals, and could be as high as 40%, 50% or even 60% in less satisfactory hospitals. Clearly surgery could not make much progress until this problem was overcome. Inspired by reading Pasteur’s papers on fermentation and putrefaction of foodstuffs, Lister conjectured that wound suppuration was caused by pathogenic

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:48:33.

Copyright © 2018. Routledge. All rights reserved.

46  An example from medicine

micro-organisms entering the wound. If these germs were killed by a suitable application of antiseptics to the skin round the wound, the wound itself and the dressings applied, then sepsis would be avoided. Using carbolic acid as the antiseptic, Lister applied this approach, and achieved his first major success in 1865. At almost the same time in France, and also inspired by Pasteur’s work, Davaine was studying the disease anthrax. This was mainly a disease of animals – particularly cows and sheep, but it could affect humans as well, mainly those who came into contact with infected animals such as shepherds and butchers. Deaths of livestock from anthrax were very common in the 1860s, and the disease led to considerable economic losses to farmers. There was thus a strong motivation to seek for a way of curing or preventing the disease. Davaine examined the blood of animals infected by anthrax under a microscope and found that it contained rod-shaped micro-organisms which he named “bacteridia”. In 1863 he published a paper in which he postulated that these bacteridia caused anthrax. However, his theory was not well-received. It was strongly criticized by two other scientists ( Jaillard and Leplat) in 1865, and did not find general acceptance. Despite this set-back, anthrax proved to be a very suitable disease for the development of the germ theory. Since it was a disease of both humans and animals, it was easy to carry out experiments on animals, while still dealing with a human disease. Moreover, the bacterium involved, Davaine’s bacteridium, now called bacillus anthracis, is one of the biggest of the pathogenic bacteria in both length and thickness, and so is easy to study under the microscope. Robert Koch (1843–1910), was undoubtedly one of the greatest medical researchers of the late 19th and early 20th centuries. He was born on 11 December 1843 at Clausthal in the Upper Harz Mountains. His father was a mining engineer. In 1862, Koch went to the University of Gö t tingen, where he was able to study under some of Germany’s leading medical researchers and to carry out some research himself. He took his M.D. degree in 1866, and then went on to some further studies in Berlin under the famous Rudolf Virchow. One might have expected Koch to pursue a research career, but he wished to marry, which he did in 1867, and consequently to have an income to support his wife and family. So, he became a country doctor. From 1872 to 1880, he was based in the country town of Wollstein, which had about 3000 inhabitants. In 1873 he took up research in his spare time, buying the microscope and other equipment he needed. Remarkably, working on his own in this fashion, he reached results which brought him international fame. The area where Koch was a country GP was much affected by anthrax, and he started his research career by studying anthrax. Despite the criticisms of Davaine’s opponents, Koch seems to have accepted the germ theory of anthrax from the beginning. His early researches, published in his 1876 paper, provided much more evidence for this theory. In particular, he discovered that the anthrax bacillus formed spores, which are very resistant to damage, but which, in a living animal, can turn back into the normal bacillus and lead to disease. This discovery

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:48:33.

An example from medicine  47

shed light on many hitherto unexplained aspects of anthrax. Koch’s next paper of 1878 is entitled ‘Investigations of the Etiology of Wound Infections’, and so he is clearly carrying further the research of Lister. He was also influenced by the wound infections which he had witnessed while serving in the Franco-Prussian war (1870–1871). Despite its title, the paper also contains some new material on anthrax. We see from this, that Koch began his research career by investigating the two diseases which had seen the introduction of the germ theory of disease, namely anthrax and wound infections. His work provided a great deal of further evidence that these diseases were indeed caused by micro-organisms (bacteria), and also further information about the nature of the bacteria involved. This early work brought him fame, and he was appointed to a research position at the Imperial Health Office in 1880. Here he worked with two assistants: Loeffler and Gaff ky, who became well-known bacteriologists in their own right. Perhaps surprisingly, Koch, who had started as a lone researcher, proved to be a very successful team leader, and many of his subsequent research assistants, such as von Behring, Ehrlich, and Kitasato, made important discoveries.

3.2  Koch’s investigations of tuberculosis and cholera After his successful start with anthrax and wound infections, Koch turned his attention to the most dreaded disease of his day, tuberculosis. At the beginning of his 1882 paper, he says (p. 83):

Copyright © 2018. Routledge. All rights reserved.

Statistics show that one-seventh of all human beings die of tuberculosis, and that if one considers only the productive middle-age groups, tuberculosis carries away one-third and often more. However, the bacterium responsible for tuberculosis, now known as mycobacterium tuberculosis, was a great deal harder to study than bacillus anthracis. It was very much smaller, being about a quarter the length of bacillus anthracis, and considerably thinner. To make it visible, significant improvements in microscopy where needed. In addition, the bacillus had to be stained, and this was difficult because of its waxy surface. However, Koch found a method of staining using some of the artificial dyes which the new German dye industry was now producing. The tubercle bacillus is also difficult to culture since it grows slowly. However, Koch overcame all these difficulties, and his 1882 paper, containing the results of his research, convinced the medical community that mycobacterium tuberculosis really was the cause of tuberculosis. As we can see, Koch’s research so far had been in the framework of the new germ theory of disease, and had the objective of providing sufficient observational and experimental evidence to establish that a particular micro-organism was the cause of a particular disease. Naturally carrying out this work led Koch to reflect on what evidential criteria needed to be satisfied in order to show that

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:48:33.

Copyright © 2018. Routledge. All rights reserved.

48  An example from medicine

a micro-organism was the cause of a disease. His papers up to 1882 contain a number of passages on the necessary criteria, and from these passages have been drawn a set of criteria, known as Koch’s “postulates”. Koch’s postulates are still cited in contemporary textbooks of bacteriology. For example, Ronald Hare’s An Outline of Bacteriology and Immunity, gives a formulation of Koch’s postulates (2nd Edition, 1963, p. 2) . So Koch contributed not just to medical theory, but to the epistemology of causality in medicine. Like many brilliant scientists, Koch was not just a scientist, but a philosopher of science. Moreover, Koch’s postulates have been cited in quite recent debates about the causation of various diseases. A striking example is the discovery by two Australian researchers – Robin Warren and Barry Marshall that peptic ulcers are caused by a bacterium. I will now give a brief account of this development.2 The first surprise came in 1979 when Warren observed spiral bacteria in a biopsy specimen from a human stomach. Previously it had been thought that the stomach was too acid for any bacteria to exist there. In 1981 Marshall joined Warren to help him investigate the new bacterium. At first it was thought to belong to the Campylobacter genus, but further investigation showed this to be false and that the genus was named Helicobacter pylori. Warren and Marshall went on to suggest that their newly discovered stomach bacterium might be the cause of peptic ulcers. When this suggestion was first published in 1983, it was greeted by widespread scepticism. How then could its truth be established? Marshall’s answer was an appeal to Koch’s postulates. In 1985 Marshall et al published a paper with the title ‘Attempt to Fulfil Koch’s Postulates for Pyloric Campylobacter’. This attempt involved the drastic step of Marshall swallowing a sample of the new bacteria to see if this would produce a stomach disease in him. This dramatic example shows that Koch’s postulates were still of great influence more than a hundred years after they were first formulated. In the light of this, the story of how Koch himself used his postulates turns out to be a strange one. Koch’s evidence for the tubercle bacillus being the cause of tuberculosis satisfied all of Koch’s postulates, and this evidence did indeed convince the medical community that Koch had found the cause of tuberculosis. Fresh from this triumph, Koch now turned his attention to another dread disease – cholera. Cholera was endemic in India, but not in Europe. However, it did visit Europe from time to time during the 19th century, causing terrible epidemics. Koch went first to Egypt and then to India to investigate cholera. Naturally he supposed that cholera too would be caused by some micro-organism, and he found what he thought was the cause, namely a bacterium which he called the comma bacillus, because it was shaped like a comma. The comma bacillus is today known as the vibrio cholerae, and is indeed accepted as the cause of cholera. However, when Koch put forward his claim that the comma bacillus caused cholera in two papers (1884a and 1884b), he had to admit that the evidence presented did not satisfy all of Koch’s postulates. There was one postulate in particular which had been satisfied in the case of tuberculosis, but was not satisfied

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:48:33.

An example from medicine  49

in the case of cholera. Nonetheless, Koch argued that his evidence was sufficient to establish his causal claim. This time, however, not all the medical community accepted his hypothesis. Coleman in his 1987 paper: ‘Koch’s Comma Bacillus: The First Year’ writes (p. 337, Footnote 34): In England, France, and elsewhere, serious doubts came to be expressed regarding the identity of the vibrio, its relation (if any) to Asiatic cholera, and much else besides. Howard-Jones shows how little enthusiastic, in fact, was the initial response to the discovery of the comma bacillus. Coleman does say, however, that Koch had more support in his native Germany, but that, even here, the important Munich school headed by Pettenkofer rejected Koch’s claim that the comma bacillus was the cause of cholera. Pettenkofer had his own theory of what caused cholera, and he and his followers continued to hold this theory and reject Koch’s alternative theory for the next decade. Indeed, Pettenkofer may have influenced opinion in England as well, as he published a defence of his own theory and a criticism of Koch’s in the English journal The Lancet in 1884. Here he says (1884, p. 904):

Copyright © 2018. Routledge. All rights reserved.

The further one investigates the drinking-water theory the more and more improbable does it appear. Robert Koch, too, the famous bacteriologist, has hitherto failed to substantiate the drinking-water theory, and I feel convinced that the time is not far distant when he will own that he has gone in the wrong direction. Koch has succeeded in finding the comma bacillus in a water tank in a region where cholera was prevalent. I have the greatest respect for this important discovery, not as a solution of the cholera question, but only as a very promising field for pathological, not epidemiological inquiry. Later in this chapter, I will outline Pettenkofer’s alternative theory of the causation of cholera, and show how the dispute between him and Koch was resolved. Despite this set-back, Koch did eventually succeed in getting his view of the aetiology of cholera generally accepted, but this was not until he managed to collect some more evidence in favour of his position. This evidence was provided by the Hamburg cholera epidemic of 1892–3. This was the last major cholera epidemic to strike Europe – largely because of the acceptance of Koch’s account of the aetiology of cholera, and the consequent adoption of appropriate preventative measures in European cities. But now comes the twist to the story, because although this new evidence did establish Koch’s theory to the satisfaction of the medical community, Koch’s own postulates were still not satisfied, even by the augmented evidence. This suggests that Koch’s postulates were inadequate in some way, and in need of modification. I will argue that this was indeed the case. Before starting on this argument, however, it will be necessary to take a closer look at Koch’s postulates. Although the content of these postulates is indeed to

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:48:33.

50  An example from medicine

be found in Koch’s papers, Koch himself never gave an explicit and definitive statement of his postulates. As a result, several different versions of these postulates exist today. I will look into this question in the next section (3.3) where I will present a version of the postulates closely based on Koch’s own writings. Then in section 3.4, I will turn from history to philosophy and consider Koch’s postulates in the light of the action-related theory of causality. I will argue, on the basis of this philosophical view, that Koch’s postulates are inadequate, and need to be supplemented. I will accordingly formulate a modified version of Koch’s postulates. Then in section 3.5, I turn back to history and examine how Koch did manage to establish the comma bacillus as the cause of cholera, using his observations on the Hamburg cholera epidemic of 1892–3. I will show how Koch’s practice implicitly conformed to the modified Koch’s postulates of section 3.4, although Koch himself never explicitly suggested any change to his original postulates, despite their not being satisfied in the case of cholera.

3.3  Koch’s postulates As we have already remarked, Koch never gave an explicit and definitive formulation of his postulates, and, as a result, there are several different versions of these postulates. Brock in his admirable 1988 life of Koch remarks (p. 180):

Copyright © 2018. Routledge. All rights reserved.

Surprisingly, the enunciation of Koch’s postulates in their final “textbook” form occurred not in a paper by Koch, but in the paper by Loeffler on diphtheria, dated December 1883. As has already been mentioned, in the period 1882–4 Koch worked at the Imperial Health Office with two collaborators: Loeffler and Gaff ky. The three of them were engaged in the same research programme of trying to identify micro-organisms which were the causes of various diseases. They all used the same techniques and methods. Loeffler succeeded in identifying the bacterium which caused diphtheria, and Gaff ky the bacterium which caused typhoid. Very likely the postulates were the result of discussions between these three research workers. However, we can nonetheless find most of the postulates in Koch’s own papers, and it is interesting to analyse these passages from Koch, as they give an idea of how Koch’s thinking on this subject evolved. Of course, Koch’s views were influenced by earlier thinkers as well, but I will not consider his predecessors here. Two works which deal with this topic are Carter (1985) and Evans (1993, Chapter 2, pp. 13–39). Koch first mentions the problem of establishing, or, as he says, proving that a micro-organism is the cause of a disease, in his 1878 paper on wound infections. He says (p. 27): there are justified objections to the assumption that bacteria cause infected wound diseases. To prove this assumption it would be necessary to Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:48:33.

An example from medicine  51

demonstrate bacteria in every case of such a disease. Moreover, the number and distribution of the bacteria must be appropriate to explain completely the disease symptoms. In similar vein, he writes on p. 29: a proof would require that we find the parasitic micro-organisms in all cases of the disease, that their presence is in such numbers and distribution that all the symptoms of the disease can be explained …  We will take Koch’s postulates as being designed to establish that a particular micro-organism causes a particular disease. We can then obtain the first 2 postulates by paraphrasing the above passages as follows:

Copyright © 2018. Routledge. All rights reserved.

1. The micro-organisms must be shown to be present in all cases of the disease. 2. Their presence must be in such numbers and distribution that all the symptoms of the disease can be explained. At this stage Koch seems to have relied on just these two postulates, but developments between 1878 and 1882 led him to add another two postulates in his discussion of tuberculosis. In 1881 Koch made one of his great contributions to bacteriology by discovering a method of producing a pure culture of a bacterium. Up to that point, bacteria had normally been grown in laboratories in liquid media in test tubes. However, Koch had the idea of growing them on solid media instead. The solid media could be the surfaces of cut potatoes, or existing liquid nutrient media with added gelatin which caused them to solidify when poured onto glass plates. In a liquid medium, the bacteria of interest would mix with other contaminant bacteria, and it was hard to separate them. On a solid medium it was possible to find a colony of the bacteria of interest which was separate from the other colonies. By taking a sample from this colony and growing it on another plate, the culture would become purified, and one could repeat this procedure as often as one liked until a pure culture of the required bacterium was obtained. Koch demonstrated this new technique at the international medical conference in King’s College London in August 1881. Lister and Pasteur were present, and Pasteur is said to have taken Koch by the hand and exclaimed: “C’est un grand progrè s, Monsieur”. In his 1882 paper on tuberculosis, Koch added some new postulates which referred to the pure cultures which he was now capable of producing. The relevant passage runs as follows (p. 87): To prove that tuberculosis is caused by the invasion of bacilli, and that it is a parasitic disease primarily caused by the growth and multiplication of bacilli, it is necessary to isolate the bacilli from the body, to grow them in pure culture until they are freed from every disease product of the animal

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:48:33.

52  An example from medicine

organism, and, by introducing isolated bacilli into animals, to reproduce the same morbid condition …  Paraphrasing this passage as postulates 3 and 4, we get the following version of Koch’s postulates:

Copyright © 2018. Routledge. All rights reserved.

1. The micro-organisms must be shown to be present in all cases of the disease. 2. Their presence must be in such numbers and distribution that all the symptoms of the disease can be explained. 3. The micro-organisms must be isolated and grown in pure culture. 4. It must be possible to reproduce the disease by introducing this pure culture into animals. I have formulated the postulates in terms of micro-organisms, though Koch himself often speaks of bacteria or bacilli. The reason for this was that Koch’s postulates were later applied to viruses as well as bacteria. This raised many problems and good accounts of this development are to be found in Evans (1993), and Vonka (2000). Quite recently an appeal to Koch’s postulates was made in the debate about whether HIV causes AIDS (see Harden [1992], and Evans [1993, pp. 53–65]). This is another proof of the contemporary relevance of Koch’s postulates. It is of interest to compare the above version of Koch’s postulates with other versions in the literature. My version is more or less the same as that in Hare (1963, p. 2), or the version given by Loeffler in 1883, which is quoted in English translation in Brock (1988, p. 180). The only difference is that both Hare and Loeffler combine my postulates (1) and (2) into a single postulate, and so number my postulates (3) and (4) as (2) and (3). Evans (1993, p. 30) gives a version in which my postulates (1) and (2) are combined into a single postulate (1), and my postulates (3) and (4) are combined into a single postulate (3). Evans then adds another Postulate (2), which is different from any of my postulates. Carter (1987, p. xviii) gives a version of Koch’s postulates which is similar to Evans’ except that he omits my postulate (2) altogether rather than combining it with my postulate (1). In his 2003, p. 136, Carter gives another five postulate version. Here my postulate (2) is included as his postulate (3). However, his postulates (1) and (2) include more material than my postulate (1). I am not saying that one version of Koch’s postulates is to be preferred to another. Since Koch never gave an explicit and definitive version of his postulates, they have to be reconstructed from statements he makes in different places, and it is possible to make different reconstructions which all have some textual justification. Luckily, however, this complexity does not affect the main argument of this paper, since all the versions include somewhere a version of my postulate (4), and it is this postulate, which is going to be the focus of our attention. My general account of Koch’s postulates does, however, differ considerably from the account to be found in Gradmann (2014). So, I will now briefly compare the two positions.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:48:33.

Copyright © 2018. Routledge. All rights reserved.

An example from medicine  53

Gradmann writes (2014, p. 886): “The blueprint for Koch’s postulates appears to have originated in a paper written by Friedrich Loeffler … ”. He goes on to say (p. 886): “Not only did Koch not write any postulates … ”, and (p. 888): “Koch never conceived any postulates”. Moreover, he refers in his paper to the postulates as Loeffler’s postulates. It is true that Koch never used the term “postulates” and did not produce any explicitly numbered postulates. Still, the key question is surely whether Koch formulated the content of the postulates in his published papers. In fact, he clearly did, as the passages from him cited in this section demonstrate. Regarding Loeffler’s contribution, I do not wish to minimize this. He, and indeed Gaff ky, might well have contributed something to the passage which I cite from Koch’s 1882 paper. However, neither Loeffler and Gaff ky could have contributed to the passage I cite from Koch’s 1878 paper, as this was written while Koch was still working on his own and before Loeffler and Gaff ky had become his assistants. As the first two of Koch’s postulates, in the version I give, are based on the 1878 passage, it is clearly mistaken to attribute these postulates to Loeffler rather than Koch. It also seems to me at best misleading to say that Koch did not write any postulates or that Koch never conceived any postulates. Let us now return to the contemporary relevance of Koch’s postulates. The fact that they have been cited in recent debates on the causes of peptic ulcers and of AIDS shows that they have had an enduring importance for medical research. In the light of this, it is surprising that the postulates, which had worked perfectly for tuberculosis, failed for the very next disease which Koch himself investigated, namely cholera. I will now explain why this happened. The problem was that, although cholera is a violent and often fatal disease in humans, other animals appear to be immune so that it was not possible to satisfy postulate (4) by reproducing the disease through the introduction a pure culture of the comma bacillus into experimental animals. On his expedition to Egypt to investigate the cholera epidemic there, Koch took 50 mice from Berlin. Yet it was not possible to infect them by inoculations with cholera material (Brock, 1988, p. 150). Koch and his fellow researchers had no more luck when they tried to infect animals with cholera in India. Koch in his 1884a paper on cholera admits the failure with characteristic openness and honesty. He writes (1884a, pp. 160–1): One should show that cholera can be generated experimentally by comma bacilli. Indeed we have tried in every imaginable way to meet this condition. …  To become certain about the possibility of infecting animals with cholera, I inquired all over India whether similar diseases had been observed among animals. In Bengal, however, I was assured that this never happened. This province is especially thickly populated, and many animals live with the people. One would assume that in such a land, where cholera is everywhere and always present, animals would often be infected with cholera materials, and that such infections would be as effective as among humans. But no one ever observes animals with cholera.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:48:33.

54  An example from medicine

Therefore, I believe that all the animals available for experimentation and those that often come into contact with people are totally immune. True cholera processes cannot be artificially created in them. Therefore, we must dispense with this part of the proof. Koch argued that the rest of the evidence was sufficient, even without postulate (4), to establish that the comma bacillus was the cause of cholera. Yet, as we have seen, not everyone accepted this conclusion, and Koch’s opponents used Koch’s postulates to argue that Koch had not established his case. In his next paper (1884b), Koch was heartened by the partial success of an animal experiment in France. He writes (1884b, p. 176): during the last cholera epidemic in Marseille, Professors Rietsch and Nicati were able to generate a cholera-like condition in dogs and guinea pigs. They did this by binding the ductus choleochus of the animals and by injecting a certain quantity of pure comma bacilli culture into the duodenum. Supposedly, the experiment was also successful later without tying off the ductus choledoctus.

Copyright © 2018. Routledge. All rights reserved.

Obviously, such experiments were not very convincing. They involved a drastic intervention, and only produced a “cholera-like condition” rather than a disease, which was unmistakeably cholera. One further possibility for satisfying Koch’s postulate (4) remained. This was to produce cholera in humans by infecting them with the comma bacillus. Considering the very dangerous nature of cholera, Koch concludes his paper by advising against such an expedient and continuing to try animal experiments. As he says (1884b, p. 176): it would certainly be wiser to continue experiments with guinea pigs and other animals than to follow the recent suggestions that human volunteers consume pure comma bacilli cultures. Despite Koch’s warning, some human volunteers did resort to consuming pure comma bacilli cultures, as we shall see later on, and, as we mentioned earlier, Marshall also resorted to self-experimentation in the 1980s in order to satisfy Koch’s postulate (4). In fact, however, no one ever did succeed in satisfying Koch’s postulate (4) for the comma bacillus, and yet as a result of new evidence collected in the Hamburg cholera epidemic of 1892–3, Koch managed to convince the medical community that the comma bacillus was indeed the cause of cholera. This suggests that Koch’s postulates are not adequate, and that some modification of them is needed to take account of a kind of evidence, which they do not consider. In the next section (3.4), I will examine Koch’s postulates in the light of the actionrelated theory of causality developed in Chapters 1 and 2. This philosophical

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:48:33.

An example from medicine  55

analysis will lead to the conclusion that Koch’s postulates are inadequate, and I will also suggest how they should be modified. Then in section 3.5, I will return to history and show that Koch’s new arguments, arising out of his study of the Hamburg cholera epidemic of 1892–3, implicitly assumed the modified Koch’s postulates, which will be formulated in section 3.4, even though Koch himself never recognized this.

Copyright © 2018. Routledge. All rights reserved.

3.4 Modification of the postulates in the light of the action-related theory of causality Let us now examine Koch’s postulates in the light of the action-related theory of causality. The action-theory, like any AIM theory of causality, stresses the connection between causal laws and interventions. As was argued in section 1.2, this leads to what I called the Principle of Interventional Evidence, which states that a causal law cannot be taken as established unless it has been confirmed by some interventional evidence. This principle of interventional evidence does not apply universally since there are some areas of the social sciences where it is not possible to gather interventional evidence, but it does apply in the case of medicine. If we turn now to Koch’s postulates, we see that they do indeed satisfy the principle of interventional evidence. Koch’s postulate (4) requires that it must be possible to reproduce the disease by introducing a pure culture of the microorganism into animals. This is clearly an intervention and results in interventional evidence. So Koch’s postulates do gain support from an AIM theory of causality. However, the analysis in section 1.3, does also show that there is a lacuna in Koch’s postulates. His postulate (4) is a productive action. It is using the micro-organism to produce the disease. But it was argued that causal laws should also be related to avoidance actions, and indeed that avoidance actions are more important than productive actions for medicine. This suggests that Koch’s postulates should be extended by including a clause which refers to avoidance actions. Here is a way in which this could be done. Postulate (4) could be replaced by a two clause postulate, Postulate (4a and/or 4b). 4a would be exactly the same as the original Postulate (4), and so would refer to productive actions. Postulate (4b) might be formulated as follows: 4b. It must be shown that if the micro-organisms are prevented from multiplying in the patient’s body, then the patient will not have the disease. There are in fact three ways in which the micro-organism can be prevented from multiplying in the patient. The first (vaccination) is the production of a successful vaccine through killing or attenuating the micro-organism. This is a blocking action. The second (prevention) consists in manipulating the environment so that the micro-organism is prevented from entering the patient’s body. The third (antibiotics) is finding a substance which destroys or severely inhibits the microorganism in the patient’s body without harming the patient. The third way only

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:48:33.

Copyright © 2018. Routledge. All rights reserved.

56  An example from medicine

became possible in the 20th century, but the first two ways were available in the 19th century. It should be noted 3 that there is a difference between 4a and 4b. 4a refers to animals while 4b refers to humans. From a strictly epistemological point of view, both 4a and 4b should both refer to humans, since we are investigating a disease of humans. However, for ethical reasons, animals must obviously be substituted for humans in 4a. This shows that considering avoidance actions (as in 4b) is actually epistemologically better than considering productive actions (as in 4a). If we replace Postulate (4) by Postulate (4a and/or 4b), then the principle of interventional evidence is still satisfied. The interventional evidence can result from a productive action (4a) or from an avoidance action (4b) or from both. In fact, the evidence used to establish that the bacillus anthracis was the cause of anthrax satisfied 4a and 4b. Both Koch and Pasteur showed that inoculation of experimental animals with a pure (at least fairly pure) culture of the bacillus resulted in the animal getting the disease. So Postulate 4a was satisfied. However, the evidence, which completely convinced any remaining doubters, was the production by Pasteur and his colleagues of a successful vaccine against anthrax by the attenuation of cultures of bacillus anthracis. The famous and dramatically successful trial of the new vaccine took place in May and June 1881 at Pouillyle-Fort. In this trial 25 sheep were vaccinated, and a further 25 sheep were left unvaccinated as controls. Then all 50 sheep were given a fatal injection of bacillus anthracis. All 25 unvaccinated sheep died of anthrax, while 24 of the 25 sheep which had been vaccinated remained healthy. Only one of the vaccinated sheep was sickly and later died, but it turned out that this sheep was a pregnant ewe which died of the complications of the pregnancy rather than anthrax.4 Obviously, Koch knew of the striking evidence in favour of bacillus anthracis being the cause of anthrax provided by Pasteur’s successful vaccination. So why did he not include evidence provided by preventative measures in his postulates? There were certainly important differences between Koch and Pasteur, particularly as regards Pasteur’s views on the attenuation of pathological bacteria, but this does not explain the bitterness of the controversy that broke out between them. In this controversy Koch was undoubtedly the aggressor. Pasteur referred in favourable terms to Koch’s paper on anthrax of 1876, but in 1881 Koch wrote a paper attacking Pasteur’s work on anthrax. Some of Koch’s criticisms of Pasteur are undoubtedly correct, but the tone of the paper is exaggeratedly hostile. Koch goes as far as to say (1881, p. 65): “Only a few of Pasteur’s beliefs about anthrax are new, and they are false.” The timing of this attack was very badly judged, since, shortly after it appeared in print, Pasteur had his great triumph at Pouillyle-Fort. Pasteur, who did not know German, had not read Koch’s 1881 paper, when the two of them met at the London conference. This was perhaps fortunate for, as we have seen, Pasteur greeted Koch kindly and congratulated him on his new solid medium technique. In 1882, however, a French translation of Koch’s paper appeared which produced much bitter feeling in France. Pasteur replied to Koch, and there were further exchanges in the controversy.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:48:33.

An example from medicine  57

Pasteur was 20 years older than Koch, and generally thought of as the grand old man of the subject. Why then did Koch make such a bitter attack on him? Perhaps Koch wished to displace Pasteur as the leading researcher in the field, but, if he did have such a personal ambition, it was certainly intensified by the national rivalry, which then existed between France and Germany. The FrancoPrussian war had taken place in 1870–71, and had resulted in a catastrophic French defeat. This led to bitter feelings on the part of the French (including Pasteur) and a desire for revenge. On the German side, there was a wish to consolidate their superiority to France. Given this background, it is perhaps not surprising that Koch did not include the evidence provided by a successful vaccine among his postulates. When it came to cholera, however, Koch did appeal to the evidence provided by successful prevention as will be shown in the next section. Thus, although Koch never satisfied his own postulates for cholera, he did satisfy the modified postulates presented in this section.

Copyright © 2018. Routledge. All rights reserved.

3.5 Koch establishes that the comma bacillus is the cause of cholera The new evidence, which led to the general acceptance of Koch’s claim that the comma bacillus was the cause of cholera, came from the cholera epidemic in Hamburg in 1892.5 The first cases of cholera seem to have occurred about the middle of August. The epidemic had definitely started by 25 August. It began to decline in September and had largely disappeared by mid-October. During this short space of time, according to the official statistics, the disease killed 8,616 people, i.e. 13.4 per thousand of Hamburg’s population. About half of those who contracted the disease recovered, giving a figure of about 17,232 for those who suffered from cholera. However, the official figures may well have underestimated the numbers. Evans calculates that over 21,000 may have contracted the disease and over 10,500 died from it (see Evans, 1987, pp. 292–293, and 599–605). Robert Koch was told on 23 August that he was being sent to Hamburg as the chancellor’s official representative, and arrived there on 24 August. Koch inspected conditions in Hamburg with his usual thoroughness, and recommended measures to control the epidemic. In May of the next year, he published his conclusions in his 1893 paper: ‘Water-Filtration and Cholera’. In order to understand this classic paper of scientific medicine, it is important to realise the situation that prevailed in the medical community when Koch wrote it. Koch had published his theory that cholera was caused by the comma bacillus in 1884. It was accepted by some members of the medical community, but not by all. The main rival theory was that of Max von Pettenkofer, a leading figure among German doctors, and head of the influential Munich school. On his return from India in April 1884, Koch decided to visit Pettenkofer to see if he could convince him of the new bacteriological approach to cholera.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:48:33.

Copyright © 2018. Routledge. All rights reserved.

58  An example from medicine

Koch did see Pettenkofer on 29 April (Brock, 1988, p. 167), but Pettenkofer was clearly unconvinced since he published a statement of his own theory and a criticism of Koch in 1884. Pettenkofer continued to have many supporters as is illustrated by what happened when Koch, in 1891, resigned his position as Professor of Hygiene and Director of the Hygiene Institute in Berlin in order to take up a similar position at the new Institute of Infectious Disease in Berlin. Koch naturally hoped that he would be replaced by one of his own followers, but instead a follower of Pettenkofer’s was appointed as the new Director of the Hygiene Institute in Berlin. In his 1893 paper, therefore, Koch not only tried to confirm his own theory empirically, but to show that the evidence refuted Pettenkofer’s theory. Before examining Koch’s paper, therefore, we must have a brief look at Pettenkofer’s theory of cholera. Before the rise of the germ theory of disease, the two principal ways of explaining disease causation were contagion and miasma. Contagion was a mechanism by which someone suffering from a disease would transmit it to a person in close contact. Contagions were usually thought of as chemical poisons which were passed, by those suffering from the disease, to anyone near them. Miasmas were putrid atmospheres or bad airs, which transmitted the disease to anyone who breathed them. There was much evidence to support the miasma view, since, for example, malaria occurred in marshlands, and diseases of all kinds were more common in overcrowded slums, barracks, ships and workhouses where the atmosphere often was evil smelling. Moreover, the miasma theory did lead to valuable reforms in preventive medicine. It was held by Chadwick who was perhaps the principal advocate of the construction of sewers in Britain, and who also advocated improved drainage, cleaning and sanitary regulation of buildings. In hospitals it led to a belief in cleanliness, fresh air, avoidance of overcrowding, and so on. Florence Nightingale based her reforms on the miasma theory, and was never converted to the germ theory. Pettenkofer always accepted the contagion and miasma framework, and he continues to use it in his 1884 article on cholera, but by that time, the germ theory had made considerable advances, and Pettenkofer combines the old framework with some ideas from the germ theory. Thus he begins his 1884 article as follows (p. 769): Cholera is an infectious disease. By infectious diseases are meant those diseases which are caused by the reception from without of specific infective material into healthy bodies, which material acts like a poison. …  Infective material differs essentially from lifeless chemical poison in being composed of the smallest possible units of living matter which when taken into healthy bodies rapidly increase and multiply under certain conditions and by their life-growth disturb the health of the body. These germs of disease belong to the smallest units of life, to the schizomycetes, which lie on the borderland of the invisible, and which, according their form, are known as cocci, bacteria, bacilli, vibriones and spirilla.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:48:33.

An example from medicine  59

So far Pettenkofer looks like a supporter of the germ theory of disease, but now he introduces the old contagion/miasma framework as follows (1884, p. 769):

Copyright © 2018. Routledge. All rights reserved.

Infective material is derived partly from sick individuals, in which case the disease is termed “contagious” and partly from locality (earth), in which it has developed, in which case the resulting disease is termed “miasmatic.” Pettenkofer did not deny that microbes, of which the comma bacillus might be one, were involved in the causation of cholera, but he held that such microbes could not cause cholera directly. The microbes would not cause the disease if they were ingested, or if someone came in contact with them in the stools of a cholera patient. So, cholera was not contagious, or at least only very slightly contagious. Cholera could not be spread through soiled linen, contaminated drinking-water etc. If we call the relevant microbes “x”, then x alone will not cause cholera. x has to come in contact with another factor “y” in the soil which causes x to germinate, and produce the actual infective material of cholera “z” in the form of a miasma. This theory explained why cholera could move from place to place. The factor x, which was harmless in itself, could be transported by humans from one cholera area to another. If x were deposited in the new site through human excrement, then, if the soil conditions were appropriate, it would in conjunction with y produce a cholera epidemic. Pettenkofer studied local soil conditions, the movement of the water-table, conditions of temperature and rainfall, and so on, in the hope of learning the conditions which would cause x to germinate and produce an epidemic. Through such studies, he hoped that means would be found to prevent cholera epidemics. His theory was called the “soil theory (Bodentheorie) of cholera”. Although Pettenkofer made some concessions to the new germ theory of disease, he unfortunately retained what was to prove the principal weakness of the contagion/miasma framework. Within that framework, there was no plausible mechanism by which a disease could be transmitted through drinking-water. Drinking-water did not involve any close contact with a person suffering from a disease, and so could not be considered as a form of contagion. Nor did drinkingwater generate a harmful miasma. As we have seen, the miasmatists advocated the construction of sewage systems, and greater cleanliness, and both these factors did reduce the incidence of many diseases, but these measures were powerless against the diseases transmitted through drinking-water, which included both cholera and typhoid. Cholera and typhoid epidemics continued to affect towns, like Hamburg, which had installed excellent sewage systems. Pettenkofer was responsible for the installation of a new water supply for Munich. He arranged for spring-water to be transported from the mountains to the city, and to replace the previous supply of water from wells in the city. He did not think it necessary to filter the water in the new supply, and shortly after it came into operation, a massive typhoid epidemic hit the city (Evans, 1987, p. 241).

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:48:33.

60  An example from medicine

Despite this set-back, Pettenkofer did not alter his views, and in fact most of his 1884 paper on cholera is taken up with arguing that cholera cannot be transmitted by drinking-water. He argues against advocates of the drinking-water view, both the earlier John Snow and the contemporary Robert Koch. Here is one of the arguments, which he uses against Koch (1884, p. 904):

Copyright © 2018. Routledge. All rights reserved.

Koch was further strengthened in his views … from the fact that after Fort William in Calcutta was supplied with pure water no more cases of cholera occurred there, although it had formerly been ravaged by the disease. The gentlemen in Calcutta had not, however, told Koch the whole truth. For it was a fact that cholera had begun to decrease in Fort William since 1863, and yet the fresh water-supply was introduced as late as March 25th, 1873. Moreover, it was not true that the only improvement then effected was a change in the water-supply, for many other changes were carried out, the fortress being made a model of cleanliness. Alterations in the drainage of the soil were effected in and around the foundations of the building, which before this was nothing more than a morass during the rainy season; so that, inasmuch as the nature of the soil, as well as the drinking water, was changed, the case of Fort William affords an argument as much in favour of the localists as it does for the contagionists. Essentially Pettenkofer is arguing that while there was a correlation between the improvement of the water-supply at Fort William and the decline of cholera there, this was just a correlation and not causation. Other changes such as the improvement in cleanliness, and the drainage of the soil were made at the same time, and perhaps these changes, rather than the improvement in the watersupply, caused the decline in cholera. It is worth noting that, if Koch had been able to satisfy his 4th postulate for cholera, he could easily have refuted Pettenkofer’s theory. Suppose that experimental animals inoculated with a pure strain of the comma bacillus (x) had acquired cholera, this would have shown that there was no need for a factor in the soil (y) to germinate x, and produce the miasma z in order to cause cholera. However, Koch was not able to satisfy his 4th postulate in the case of cholera. Pettenkofer published his discussion of Fort William in The Lancet issue of 22 November 1884, and in the issue of 13 December 1884, The Lancet published a criticism of Pettenkofer’s treatment of the case of Fort William by A.C.C. De Renzy, who was Deputy Surgeon-General of the Bengal Army. De Renzy questions many of Pettenkofer’s factual claims. For example, Pettenkofer says that the new water-supply for Fort William was introduced only in 1873, but that incidence of cholera in the Fort had begun to decline as early as 1863. De Renzy replies (1884, p. 1043): Professor Pettenkofer … seems quite unaware of the fact that changes in the water-supply began even earlier than 1863, that they had been carried out to Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:48:33.

An example from medicine  61

very considerable completeness by 1865, and that in 1872 the existing state of things was attained when the Calcutta municipal water was laid on in pipes all over the fort. De Renzy also disputes Pettenkofer’s claim that drainage of the soil had been carried out round the foundations of the fort. He writes (1884, p. 1044):

Copyright © 2018. Routledge. All rights reserved.

Then as to the alterations in drainage, I do not know on what authority Pettenkofer writes. …  The fact is, there are insurmountable difficulties in the way of drainage owing to the lowness of the site, which is only one foot above the mean level of high water of the Hooghly during the rains. The site is surrounded by rice fields and vast swamps which it is utterly impossible to drain. Thus, drainage cannot have been the factor, which caused the decline of cholera at Fort William. Nor could it have been cleanliness, because there are many forts in India, which, according to De Renzy, are just as clean as Fort William, but which continue to suffer from cholera epidemics. De Renzy holds the view that (1884, p. 1043): “cholera is mainly communicated through the medium of water”, and he makes a good case in favour of this opinion and against that of Pettenkofer. However, having given many observations in favour of his opinion, he does raise the question of “how far these phenomena are related as cause and effect” (1884, p.1045). Thus, he does seem to have residual doubt as to whether he is dealing with a correlation or a genuine causal link. This interchange between Pettenkofer and De Renzy gives a flavour of the debate about the aetiology of cholera in the decade before the Hamburg epidemic. I will now turn to an analysis of Koch’s 1893 paper, which, in the opinion of the overwhelming majority of the medical community, decided the argument in favour of Koch’s theory of cholera. In the first paragraph of his paper, Koch states his views about the transmission of cholera (1893, p. 183) 6 : I have always maintained that in the light of our experiences to date direct infection from person to person is possible but from every appearance it occurs not very frequently, that on the contrary in real epidemics and mass outbreaks of cholera the principal role devolves on indirect infections from the many carriers of the cholera germ, and that among these carriers water is one of the most important. Koch here wants to stress that cholera is not transmitted exclusively by water, but that water transmission is one of the most important ways in which the disease spreads. This opinion agrees with that of De Renzy, but completely contradicts that of Pettenkofer.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:48:33.

62  An example from medicine

Koch now turns to the evidence from the Hamburg epidemic which he thinks strongly confirms his views and refutes those of Pettenkofer. This evidence concerns the different rates of cholera in Hamburg and Altona. Altona was a small town downstream along the Elbe from Hamburg. Originally it must have been an independent town, but by 1892 there was no open space between Altona and Hamburg, so that Altona had become a suburb of the big city. Altona had, however, a different water-supply from Hamburg. Hamburg took its water from the Elbe before the river entered the city, and its water was unfiltered. Altona took its water from the Elbe after it had flowed through Hamburg, but Altona’s water was filtered by slow sand filters. Presumably filtration had been thought necessary in Altona because, at the point where it took its water, the river water was polluted with the wastes of Hamburg, whereas the Elbe water above Hamburg was thought to be purer, and so not in need of filtration. Koch begins by pointing out that the difference in the incidence of cholera in Altona and Hamburg was very striking (1893, p. 184):

Copyright © 2018. Routledge. All rights reserved.

Cholera behaved most surprisingly at the boundary between Hamburg and Altona. On both sides of the boundary the character of the soil, buildings, sewers, population, in effect everything that is relevant here is completely the same, and yet cholera in Hamburg went only right up to the boundary of Altona and halted here. In one street, which marks the boundary for a long stretch, cholera occurred on the Hamburg side, while the Altona side remained free of the disease. Evans (1987, p. 291) gives some detailed statistics, which completely bear out what Koch is saying here. For example, in the area near the Elbe, the incidence of cholera on the Altona side of the boundary was 0–5 per thousand inhabitants, while on the Hamburg side it was 25–30 per thousand inhabitants. It is worth noting that Koch stresses that all the conditions apart from the watersupply were the same on both sides of the boundary. Koch must have been aware, from Pettenkofer’s earlier criticisms, how Pettenkofer was likely to argue against Koch’s claim that cholera was transmitted by the water-supply. Pettenkofer’s line, as in the Fort William example, would probably be that, although there was a correlation here between water-supply and cholera incidence, this was not causal in character, and that the real cause of the difference lay in other factors. Indeed, Pettenkofer would have had to argue that the cholera microbes had germinated on the Hamburg side of the boundary to produce a miasma, but had failed to do so on the Altona side. To counter such a claim, Koch stresses that the character of the soil, sewers etc. were the same on both sides of the boundary. He elaborates this argument in the next passage where he claims that the Altona/Hamburg case is like a laboratory experiment (1893, p. 184): Here we have therefore to do with a kind of experiment, which has been carried out with more than a hundred thousand people, but despite its Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:48:33.

An example from medicine  63

enormous dimensions satisfies all the conditions, which are imposed on an exact and completely decisive laboratory experiment. In two large population groups, all the factors are the same, except a single one which is different, namely the water-supply. The group, which was supplied with unfiltered water from the Elbe, was heavily affected by cholera, the group supplied with filtered water to a very small extent. This difference must be considered of yet greater weight, since the Hamburg water is taken from a place, where the Elbe is comparatively little polluted, but Altona must use water from the Elbe after it has received the complete liquid waste, including the faeces, from nearly 800,000 people. Under such conditions there is for the scientific thinker first of all no other explanation, except that the difference, which the two population groups show regarding cholera, is caused by the difference in water-supply, and that Altona was protected against cholera by the filtration of the water of the Elbe. Koch’s next point is that the difference between Hamburg and Altona is very easily explained on the assumption that cholera is a disease caused by bacteria. As he says (1893, p. 185):

Copyright © 2018. Routledge. All rights reserved.

For the bacteriologist nothing is easier than to give an explanation for the confinement of cholera to the area of Hamburg’s water-supply. He needs only to point out, that cholera bacteria reached Hamburg’s water-supply from the output of Hamburg’s sewers, or, which is much more probable, from the dejecta of those with cholera, who were to be found on the numerous small boats anchored in the Elbe above the place from which Hamburg’s water is taken and that, after this had happened, among the people who used the water, according to the degree of its pollution, more or less numerous cholera cases must have occurred. …  Altona takes water which initially is much worse than that of Hamburg, but through careful filtration is freed wholly or almost completely from cholera bacteria. So it is easy to explain the contrast between Hamburg and Altona, as regards cholera, if we assume a bacterial cause for the disease. Conversely, however, as Koch goes on to say, it is very hard, if not impossible, to explain this difference on other hypotheses, such as the claim that cholera is caused by a miasma. As Koch writes (1893, p. 185): Why anyone would want to derive the behaviour of the Hamburg-Altona cholera from cosmic-telluric, or from purely meteorological factors is a puzzle to me; since sky, sun, wind, rain and so on were distributed absolutely equally on both sides of the boundary. Despite the seeming impossibility of the task, Pettenkofer continued to defend his theory of cholera in the light of the Hamburg epidemic, in an article of 1892 Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:48:33.

64  An example from medicine

Copyright © 2018. Routledge. All rights reserved.

cited by Koch. Snow had earlier used an argument involving the water-supply by two different water companies, which is quite similar to Koch’s Hamburg/Altona example. Pettenkofer had criticized Snow’s argument in his 1884 paper (p. 863), and he uses the same line of reasoning in the Hamburg case. Pettenkofer’s argument was that water does not directly cause cholera, but that a water-supply may predispose towards cholera. If an area, such as Hamburg, has a polluted watersupply, the use of the water in the houses and the soil may result in the deposit of cholera germs in the soil which predisposes the area to cholera should these germs later give rise to a miasma. Koch criticizes this argument in no uncertain terms (1893, pp. 185–186): The localist conception thus claims, that water has had not an infecting, but a ‘distributing’ effect in the manner, that the unfiltered water brought dirty material into the houses, on to the streets, and into the soil and thereby so to speak provided appropriately fertile soil for the development of the cholera germ. Nevertheless the honourable Localist gentleman in his trouble has certainly not thought, how infinitely small the quantity of dirty material which could be deposited through the water used in the most unfavourable cases in the houses and in the soil, compared with the infinitely much greater masses of dirty material, which human households daily bring to their houses and which is deposited continually by humans and animals on streets and courtyards. He has further not considered, that the water from the Elbe which reaches the water-supply can get an addition of the dirty water of the city of Hamburg not regularly, but only entirely exceptionally especially when there is serious flooding. Moreover he appears to have wholly forgotten, and that is the weightiest objection which is made to him here, that Hamburg, one of the cities with the best kind of sewers, is provided with arrangements, from which we know, that they take the dirty waters out of the houses, from the courtyards and streets by the shortest routes out of the area of the town. What then would be the use of the sewers after all, if they should not be capable any day of removing and rendering harmless this small increase of organic substance, which unfiltered water carries with it. The soil theory could not on the whole give a more striking proof of its complete bankruptcy than this failed attempt at explanation. Having thus disposed of Pettenkofer’s soil theory of cholera, Koch now proceeds to give some more evidence in favour of his own theory. So far, he has shown that cholera is transmitted by drinking-water, and that this is easily explained on the bacteriological theory of the disease. However, to confirm this explanation, it was necessary to show that the slow sand filtration carried out on the Altona water really did remove bacteria. This is what Koch now does. Koch begins by making some technical remarks about sand filtration. If it is to remove bacteria, it must be slow, and the water must trickle through a sufficiently

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:48:33.

An example from medicine  65

thick layer of sand. Koch mentions a speed of 100 mm per hour, and a thickness of 30 cm (1893, p.187). He remarks that not all sand filtration systems satisfy these criteria, but that the water filtration at Altona, although one of the oldest in Germany, fortunately does satisfy them, and it is to this fact that the citizens of Altona owe their preservation from the worst ravages of cholera. In the water of the Elbe there were between a thousand and a hundred thousand germs per cubic centimetre. Since the summer of 1890, the filtered water at Altona had been tested for bacteriological content weekly. Koch gives the following results (1893, pp. 192–3):

Copyright © 2018. Routledge. All rights reserved.

During a period of two years up to the summer of 1892, the number of germs in a cubic centimetre of the filtered water remained, with the exception of a short period in January 1891, always under a hundred. Numbers under 20 were the norm …  This showed that the filtration system did indeed remove most of the bacteria from the water of the Elbe. Koch does mention a short period in January 1891 when the filtration system did not work so well. He goes on to explain this as caused by ice forming in the cold weather and interfering with the filtration process. The problem of ice did not affect cholera, which only flourishes in warm weather. The Hamburg epidemic of 1892 took place in a very hot summer. However, epidemics of the other water-borne disease (typhoid) could occur in winter. So, Koch discusses how the problem of ice forming in the filtration system could be overcome. So far Koch has not mentioned the comma bacillus, which he regarded as the cause of cholera. In a way it was not necessary for him to do so. If cholera is a bacterial disease, then the fulfilment of his first two postulates would show that the comma bacillus was the relevant bacterium. If Altona’s slow sand filtration system removed nearly all bacteria, it would remove the comma bacillus along with the others. However, he did also test for the presence of the comma bacillus, which he here refers to simply as the cholera bacterium, in the water systems of Hamburg. He states the results as follows (1893, p. 199): We succeeded in detecting the cholera bacteria in the water of the Elbe …  They were also detected in the water immediately before filtration. They were not found in the filtered water. This completes the evidence, drawn from the Hamburg cholera epidemic of 1892, which Koch presents in his 1893 paper. This new evidence led to the overwhelming majority of the medical community abandoning Pettenkofer’s soil theory of cholera and accepting Koch’s view that cholera was caused by the comma bacillus. Koch’s recommendation that all water-supply systems should use properly filtered water was accepted and implemented in Germany and the rest of Europe, and, as a result, there were no more European cholera epidemics.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:48:33.

66  An example from medicine

Yet though Koch thus succeeded in establishing the cause of cholera, he never succeeded in satisfying his 4th postulate. Instead he succeeded in satisfying what we have suggested as a postulate 4b; that is to say, he showed that if the comma bacillus was prevented from entering someone’s body by removing it from the supply of drinking water, then that person would not get cholera. It is a perfect instance of Vonka’s Thomist maxim: sublata causa, tollitur effectus (if the cause is removed the effect is taken away). In his 2005 book on Koch, Christoph Gradmann writes (2005, pp. 176–177):

Copyright © 2018. Routledge. All rights reserved.

after combating cholera in Hamburg in 1892, Koch increasingly turned toward epidemiological questions and on this basis developed preventive measures designed to control epidemics in entire populations. Despite this new focus of his interests, Koch never modified his postulates to take account of evidence from successful preventive interventions. Perhaps it would have been better if he had done so, as is shown by the reaction to Koch’s success of Pettenkofer and his followers. This will now be described. When Koch returned to Berlin in 1892, an expert commission on epidemics was convened, meeting for six days from 26 September to 1 October. Pettenkofer was naturally appointed to this commission, and he opposed Koch’s ideas. Yet Koch’s position was accepted on every point, and Pettenkofer overruled. Pettenkofer returned to Munich in early October, and asked Koch’s colleague, Georg Gaff ky, for a sample of the comma bacillus. This request alarmed Gaff ky, but he could hardly refuse to send the sample and he did so. Pettenkofer had decided to use Koch’s Postulate 4 to refute Koch. If he ingested the sample and did not get cholera, this would show that the comma bacillus was not after all the cause of cholera as Koch maintained. Accordingly, on 7 October, Pettenkofer drank the sample on an empty stomach and after taking an alkaline solution to neutralize the stomach acids which might destroy the bacillus. The result of this experiment was that Pettenkofer suffered from severe diarrhoea for a few days, and his stool contained many comma bacilli. However, by 16 October, he was back to normal. The experiment was repeated by Pettenkofer’s loyal disciple Emmerich on 17 October in front of a hundred witnesses. Emmerich suffered more than Pettenkofer, but he too managed to survive. These results were of course rather inconclusive. Pettenkofer maintained that he and Emmerich had not had cholera so that Koch’s theory was refuted. However, others thought that they had suffered from mild attacks of cholera which they were fortunate enough to survive. Gaff ky claimed that he guessed what Pettenkofer planned to do, and so sent him a weakened culture, hoping that it would have a mild effect. Pettenkofer’s self-experiment might be an indication that he had developed suicidal tendencies. The theories, which he had spent so many years developing and defending, had been rejected in favour of those of a rival. This would be a severe blow for any scientist, but the situation was worse in this case. As Pettenkofer himself wrote (1884, p. 769):

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:48:33.

An example from medicine  67

Copyright © 2018. Routledge. All rights reserved.

As is always the case in medicine, the conflict of views is important, ­inasmuch as the measures to be adopted in the healing and prevention of a disease depend on the theoretical conception of it. It was a consequence of Pettenkofer’s theory that there was no need to filter the water-supply, whereas on Koch’s theory slow sand filtration was essential to avoid the disease. If Koch was right, this meant that Pettenkofer’s views were responsible for several epidemics and the consequent loss of life. We have already mentioned that Pettenkofer’s new water-supply may have been responsible for a major typhoid epidemic in Munich. Dr Kraus, a convinced supporter of Pettenkofer, was appointed Chief Officer in Hamburg in 1871, and was still in post in 1892. Koch’s discoveries did not lead Kraus to change his mind about cholera (Evans, 1987, pp. 252, 254). Kraus views may well have had some influence on Hamburg’s failure to construct a filtered water-supply in the period 1873–1892, after the Hamburg cholera epidemic of 1873. If Koch’s views came to be generally accepted in the medical community, then Pettenkofer’s theories would be seen as not only misguided, but also partly responsible for several epidemics. This would have been very hard to accept for a man who had once been Germany’s leading Professor of Hygiene – highly respected and very influential. At any rate, Pettenkofer committed suicide by shooting himself in the head on 10 January 1901. Curiously enough Emmerich, who had followed Pettenkofer in the selfexperiment, remained, until his death in 1914, faithful to Pettenkofer’s soil theory of cholera. With some other loyal disciples, he continued to write papers and books maintaining that drinking-water plays no part in the causation of cholera epidemics. However, after 1893, this group was completely isolated in the medical profession (Evans, 1987, pp. 502–503). One of the advantages of replacing Koch’s original postulate 4, by postulates 4a and 4b is that it means that the dramatic self-experimentation of Pettenkofer and Emmerich, and later of Marshall, become no longer necessary. It does indeed follow from the action-related theory of causality (and other AIM theories of causality) that causality in medicine cannot be established without some interventional evidence. However, this interventional evidence does not have to take the form of giving an animal or human the disease (a productive action), as in Koch’s original postulate 4. It can instead take the form of preventing humans getting the disease, or curing them if they do have the disease. Such avoidance actions, as specified in the modified postulate 4b, are quite as satisfactory as productive actions in supplying the necessary interventional evidence.

Notes 1 Much of the material in this chapter is taken from my paper (Gillies, 2016b). 2 For a more detailed account with analysis, see Thagard (1999, pp. 39–97). 3 The points in this paragraph were kindly suggested to me in correspondence by Ladislav Kvasz.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:48:33.

68  An example from medicine

Copyright © 2018. Routledge. All rights reserved.

4 For a fuller account of the trial at Pouilly-le-Fort, see Debré  (1994, Chapter 14, pp. 378–413). 5 My account of the Hamburg cholera epidemic is mainly based on Richard Evans’ classic 1987 book: Death in Hamburg. I have also found the relevant section in Brock’s life of Koch (Brock, 1988, pp. 229–232) very helpful. 6 There is no English translation of this paper of Koch’s. So the quotations given are my own translations, which were kindly revised and corrected by Christian Wallmann.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-15 22:48:33.

PART II

Copyright © 2018. Routledge. All rights reserved.

Causality and mechanisms

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 19:04:38.

4

Copyright © 2018. Routledge. All rights reserved.

MECHANISTIC THEORIES OF CAUSALITY AND CAUSAL THEORIES OF MECHANISM

The basic idea of mechanistic theories of causality is that A causes B if there is a mechanism M linking A to B. One of the principal advocates of this view was Wesley Salmon, and, in the introduction, I illustrated his position by giving his striking example of a dog whistle. We observe a dog owner who from time to time blows a whistle. No sound comes from the whistle, and yet each time it is blown the owner’s dog comes.We accept that blowing the whistle is really the cause of the dog’s coming, because there is a mechanism linking the two. The dog whistle emits a sound which is too high for humans to hear, but which can be heard by the dog. Salmon developed several versions of the mechanistic theory of causality. He began, by trying to define causal processes in terms of the transmission of marks, an approach, which had originally been suggested by Reichenbach in 1928. There were, however, many criticisms of the method of mark transmission. So, Salmon decided to adopt instead an alternative approach, which Dowe had suggested in his 1992 paper. This involved the use of conserved quantities. In his 1994 paper, Salmon began by summarizing Dowe’s views, and then modified them to produce a new version of the mechanistic approach to causality. I will describe this Dowe–Salmon theory of causality in the next section.

4.1  The Dowe–Salmon theory of causality The general plan is to begin by defining causal interactions. These are defined without using any causal notions in order to avoid circularity. A definition is then given of causal processes. The two definitions in Dowe’s version of the theory are stated by Salmon as follows (1994, pp. 253–4): DEFINITION 1. A causal interaction is an intersection of world-lines which involves exchange of a conserved quantity.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:52:50.

72  Mechanistic theories of causality

DEFINITION 2a. A causal process is a world-line of an object which manifests a ­conserved quantity. Salmon remarks (1994, p. 254): In his elaboration of the foregoing definitions, Dowe mentions massenergy, linear momentum, angular momentum, and electric charge as examples of conserved quantities.

Copyright © 2018. Routledge. All rights reserved.

So, in terms of the dog whistle example, when the owner blows the whistle, his or her breath has a causal interaction with the whistle which transfers mass-energy and linear momentum to some sound waves. These sound waves are causal processes because they manifest these conserved quantities. Another causal interaction occurs at the dog’s ear where some of the mass-energy and linear momentum is transferred to the dog’s ear drum, and so on. Note that in Definition 1, causal interaction is defined in terms of the non-causal notions of world-line and conserved quantity, and in Definition 2, causal process is defined in terms of causal interaction without using any other causal notions.Thus, the two definitions give a non-circular definition of cause in terms of non-causal notions. Salmon next proposes modifying Dowe’s theory by changing conserved quantity to invariant quantity, and by introducing the concept of transmission. This is done by changing Definition 2a to Definition 2e, and introducing a Definition 3 as follows (1994, p. 257): DEFINITION 2e. A causal process is a world-line of an object that transmits a nonzero amount of an invariant quantity at each moment of its history (each spacetime point of its trajectory). DEFINITION 3. A process transmits an invariant (or conserved) quantity from A to B (A ≠  B) if it possesses this quantity at A and at B and at every stage of the process between A and B without any interactions in the half-open interval (A, B] that involve an exchange of that particular invariant (or conserved) quantity. Salmon comments on Definition 3 (1994, p. 257): This definition embodies the at-at theory of causal transmission …  which still seems to be fundamental to our understanding of physical causality. These then are the fundamental points in the mechanistic theory of Dowe and Salmon. I will proceed to criticize this theory in the next section.

4.2  Criticism of the Dowe–Salmon theory of causality My basic criticism of the Dowe–Salmon theory is that it lays too much emphasis on physics and so does not apply to causality in other areas such as medicine, which is

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:52:50.

Mechanistic theories of causality  73

Copyright © 2018. Routledge. All rights reserved.

the main focus of the present book. Most of Salmon’s examples are either everyday examples which can be analysed purely in terms of physics, or examples directly from physics. Thus, for example he discusses a bullet hitting a screen and causing a hole in it (1993, p. 17), a solidly hit baseball causing a window to shatter (1994, p. 253), and a gas molecule colliding with another molecule causing changes in the momentum of both molecules (1994, p. 258). For such examples, the definition of causality proposed by Salmon could indeed be criticized, but it is quite plausible. In examples from medicine, however, it is not at all plausible. In medical examples such as A causes D, where D is some disease, there is often a mechanism M linking A to D, but M is not a physical mechanism. Usually it will be some bodily process involving biochemical and physiological changes. These biochemical and physiological changes are certainly not directly describable in terms of the kind of physical quantities considered by Dowe and Salmon, such as massenergy, linear momentum, angular momentum, and electric charge. Against this point of view, it could perhaps be said that biochemistry and physiology are ultimately reducible to physics. Thus, chemistry and so biochemistry is reducible to physics. Physiology is reducible to biochemistry and so to physics. Salmon seems to presuppose such a reductionist position since he gives what purports to be a general definition of causal processes in terms of the conserved (or invariant) quantities of physics, such as mass-energy, linear momentum, angular momentum, and electric charge. The problem with the theory that everything can be reduced to physics is that it remains a philosophical speculation, which has not been established in any detail. A few simple results of chemistry have indeed been reduced to physics, but that really is as far as it goes at present. This is a far cry from reducing the whole of medicine to physics. I would argue that it is wrong to make an analysis of the notion of causality depend on such a highly speculative philosophical claim as reductionism.Thus, I prefer to allow the mechanisms linking cause to effect to be of different kinds, and not to presuppose that all these mechanisms can be reduced to one single kind of mechanism – a physical mechanism.

4.3  More general definitions of mechanism The criticism just given of the Dowe–Salmon mechanistic theory of causality seems to have been quite widely accepted and led to attempts to formulate more general definitions of mechanism, designed to apply not just to physics, but to all sciences. The most notable attempts in this direction are to be found in Glennan (1996, 2002, and 2017) and Machamer, Darden, and Craver (2000). I will now give an account of these investigations. Glennan’s (1996) aim is to develop a mechanistic (or, in his terminology, mechanical) theory of causation. He characterises such a theory as follows (1996, p. 64): Roughly put, a mechanical theory of causation suggests that two events are causally connected when and only when there is a mechanism connecting them.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:52:50.

74  Mechanistic theories of causality

Glennan stresses that for this task we need an analysis of mechanism, which applies not just to physics but to all the sciences (1996, p. 61): my analysis is in no way limited to mechanisms that are physical in nature. It is meant equally to apply to chemical, biological, psychological and other higher level mechanisms. I emphasize this point because the generality of the analysis is key if it is to provide a foundation for a theory of causation. Glennan in his (2002) makes an explicit criticism of Salmon on this point. Glennan writes (2002, p. S346): Salmon has recently (1994) characterized them (i.e. interactions – D.G.) in terms of exchange of conserved quantities. For instance, two colliding particles interact when they exchange a conserved quantity like momentum. …  To talk in terms of exchange of conserved quantities requires us to treat mechanisms at a level at which such talk is intelligible, namely at the level of physics. But different tokens of a single higher level event type (e.g., a phone call event type) may have nothing in common in terms of their micro-physical descriptions. Thus, if interactions are characterized in terms of exchange of conserved quantities, tokens of higher level interactions cannot be recognized as forming higher-level kinds. This is a good argument against the reductionism, which, as we saw earlier, is involved in the Dowe–Salmon approach to causality. Glennan’s 1996 definition of mechanism is the following (p. 52):

Copyright © 2018. Routledge. All rights reserved.

(M) A mechanism underlying a behavior is a complex system which produces that behavior by the interaction of a number of parts according to direct causal laws. This leads to his mechanical theory of causality, which is stated as follows (1996, p. 56): a relation between two events (other than fundamental physical events) is causal when and only when these events are connected in the appropriate way by a mechanism. Glennan goes on to discuss (1996, pp. 64–66) two objections to his theory.The first of these could be called the circularity objection. It runs as follows (1996, pp. 64–65): any analysis of causation that relies on mechanisms is circular, since any explication of the concept of mechanism requires the use of causal concepts. Indeed, Glennan’s definition of mechanism (M) explicitly refers to “direct causal laws”. So, to use (M) to define causal relations does seem to involve a circularity. Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:52:50.

Mechanistic theories of causality  75

The second objection is concerned with Glennan’s exclusion of the fundamental laws of physics from his definition of causality. Glennan himself regards this second objection as more serious than the first, but I take the opposite view. I will begin therefore by saying why I don’t think that the exclusion of the fundamental laws of physics from a definition of causality is a problem, and then go on to consider the circularity objection. Earlier in Part I of the book, we considered Russell’s arguments that causality is not needed in theoretical physics and should be eliminated therefrom. These arguments are convincing. The further conclusion which Russell drew, namely that causality should be eliminated from all advanced sciences was incorrect, since medicine is an advanced science in which the concept of causality plays an essential role. However, if Russell was wrong about medicine, which he never considered, he was right about theoretical physics; and so, excluding fundamental laws of physics from a definition of causality is not at all problematic but highly reasonable. Turning now to the circularity objection, this seems to me more serious. This is how Glennan tries to answer it (1996, p. 65):

Copyright © 2018. Routledge. All rights reserved.

This circularity is only apparent. In describing the mechanism that connects the two events I have explained how these events are causally connected. How the parts are connected is a different question. I can try to answer this second question by offering another account of the mechanisms which connect them. …  in giving account of how two events are causally connected, I refer to a mechanism which in turn refers to causal relations, but these latter causal relations are different (and more basic) relations than the one which I am seeking to explain. The idea is this. We account for A causes B in terms of a mechanism M1 which connects them. However, M1 is described in terms of causal laws, and to account for these we have to introduce further mechanisms M 2 , M 2’, M 2’’, …  . These new mechanisms are described in terms of further causal laws, which are accounted for by further mechanisms and so on. However, this process does not lead to an infinite regress, because we eventually reach the fundamental laws of physics, which are non-causal in character. A circularity is indeed avoided here, but only at the expense of introducing another kind of reduction to physics, which is not dissimilar to that involved in the Dowe–Salmon approach. At all events, this reduction is liable to the same objections which apply to the Dowe–Salmon reduction to physics. I stated some of these objections in section 4.2 above, and have just quoted a further good argument against reduction given by Glennan (2002, p. S346). These difficulties connected with reduction to physics seem to me to show that the circularity objection is a strong, perhaps fatal one. Let us now turn to Machamer, Darden and Craver (2000). Like Glennan, they think that while Salmon’s approach may be satisfactory for physics, it is

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:52:50.

将因果性从理论 物理领域排除是 合理的。

76  Mechanistic theories of causality

unsatisfactory for many other sciences, and particularly for biological sciences such as neurobiology and molecular biology. They say (2000, p. 7): Salmon identifies interactions …  more recently, in terms of exchanges of conserved quantities …  . Although we acknowledge the possibility that Salmon’s analysis may be all there is to certain fundamental types of interactions in physics, his analysis is silent as to the character of the productivity in the activities investigated by many other sciences. …  As our examples will show, much of what neurobiologists and molecular biologists do should be seen as an effort to understand these diverse kinds of production and the ways that they work. So Machamer, Darden and Craver try to give an account of mechanism which applies, not just to physics, but to other sciences as well. As they say (2000, p. 2): Our goal is to sketch a mechanistic approach for analysing neurobiology and molecular biology that is grounded in the details of scientific practice, an approach that may well apply to other scientific fields. Their definition of mechanism is the following (2000, p. 3): Mechanisms are entities and activities organized such that they are productive of regular changes from start or set-up to finish or termination conditions.

Copyright © 2018. Routledge. All rights reserved.

They particularly stress the importance of having activities as well as entities in a satisfactory account of mechanisms. The role of activities is described as follows (2000, p. 6): Activities are types of causes. Terms like “cause” and “interact” are abstract terms that need to be specified with a type of activity and are often so specified in typical scientific discourse. Anscombe …  1971 …  noted that the word “cause” itself is highly general and only becomes meaningful when filled out by other, more specific, causal verbs, e.g. scrape, push, dry, carry, eat, burn, knock over. An entity acts as a cause when it engages in a productive activity. While traditionally philosophers have tried to analyse science using the general and abstract term “cause”, Machamer, Darden and Craver suggest that we should instead, in each particular scientific case, try to specify a type of activity and use this activity in our analysis. This is analogous to Anscombe’s consideration of specific causal verbs such as scrape, push, etc., though Anscombe here gives examples from everyday life rather than science. Machamer, Darden and Craver then go on, in

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:52:50.

Mechanistic theories of causality  77

the rest of their paper, to illustrate this approach by analysing in detail a number of examples from the biological sciences, including (2000, section 6, pp. 18–21) the discovery of the mechanism of protein synthesis. Machamer, Darden and Craver also discuss other definitions of mechanism. In particular, they consider Glennan’s 1996 definition of mechanism as part of an “interactionist” tradition which divides mechanisms into “parts” and “interactions”, and comment on this approach that (2000, p. 4): the interactionist’s reliance on laws and interactions seems to us to leave out the productive nature of activities. Glennan accepted this criticism, and in his most recent treatment of mechanisms (Glennan, 2017), he uses a new definition of mechanism which includes activities. As we have seen, Machamer, Darden and Craver (MDC) in their 2000 paper focussed mainly on neurobiology and molecular biology, but they thought that their definition of mechanism (2000, p. 2) “may well apply to other scientific fields.” Darden suggests that it applies generally in biology. She writes (2013, p. 21): “The MDC characterization of mechanisms is …  a characterization to capture the way biologists use the term.” Glennan goes further than this by stating that he wants a notion of mechanism that will (2017, p. 18): “capture most of the wide range of things scientists have called mechanisms”. This notion he calls “minimal mechanism” and defines it as follows (2017, p.17): A mechanism for a phenomenon consists of entities (or parts) whose activities and interactions are organized so as to be responsible for the phenomenon.

Copyright © 2018. Routledge. All rights reserved.

Here we see the introduction of “activities”, though “interactions” are still retained. As Glennan says, his definition was influenced not only by Machamer, Darden and Craver, but also by Illari and Williamson, who gave the following definition (2012, p. 120): A mechanism for a phenomenon consists of entities and activities organized in such a way that they are responsible for the phenomenon. Despite changing his definition of mechanism, Glennan continues to support the mechanistic theory of causation. As he says (2017, p. 145): What is the relationship between mechanisms and causality? Put briefly, it is just that causes and effects must be connected by mechanisms. Once again, the question of circularity arises because the definition of a minimal mechanism involves causal concepts. Activities are causes, and the phrase “so as to be responsible for” could be rendered “so as to cause”. Glennan resolves the

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:52:50.

78  Mechanistic theories of causality

problem of circularity just as before. The causes in his definition of a particular mechanism are defined in terms of more basic mechanisms and so on down. He even considers the possibility of this process never ending, writing (2017, p. 185): Perhaps …  there is no end to the nesting and there are mechanisms “all the way down”. An infinite regress here seems, however, just as bad as a vicious circle. So, a better strategy would seem to be that of looking for a grounding of the process, but now another problem arises. The downward nesting appears to end in the quantum world, but the quantum world does not seem to contain the entities engaging in productive activities which are required by the definition of minimal mechanism. Such entities look much more like classical objects. Glennan resolves this problem in an ingenious fashion by appealing to what he calls “decoherence theory”. The quantum world requires coherent superpositions like those created by the double slit experiment. However, the constant interactions, which macroscopic systems have with the environment, cause such coherent systems to decohere, producing a classical system which can serve as a foundation for Glennan’s regress of mechanisms. As he says (2017, p. 192):

Copyright © 2018. Routledge. All rights reserved.

If there are not mechanisms all the way down, a mechanistic account of causal production will need some kind of bottom-out level of classical objects that can engage in productive interactions, and decoherence theory explains the conditions that will lead to the appearance of such a level. In that sense, decoherence gives an account of the transition from quantum systems to a “fundamental classical level”. Here Glennan has successfully specified a “bottom-out level of classical objects”, and thus produced a mechanistic theory of causation which avoids both a vicious circle and an infinite regress. However, this theory does still involve what amounts to a reduction to physics, and incurs the difficulties connected with reductionism which I have pointed out in this section and the preceding one (section 4.2). For this reason, I cannot accept a mechanistic theory of causation and prefer another approach. A mechanistic theory of causation is an attempt to define, or at least characterize, causality in terms of mechanisms, I suggest that this should be reversed, and 应该用因果性 that we should try to define mechanisms in terms of causality. To put it another 定义机制而不 way, instead of a mechanistic theory of causality, we should try to produce a 是相反。 causal theory of mechanisms. Causality itself is no longer defined or characterized in terms of mechanism, but rather in terms of human actions, using the action related theory of causality developed in Part I. In the next section (4.4) I will expound a causal theory of mechanisms. However, this attempt differs from the approaches in Machamer, Darden and Craver (2000), Illari and Williamson (2012), and Glennan (2017) in that I am

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:52:50.

Mechanistic theories of causality  79

not attempting to give a theory of mechanisms which applies generally across the sciences. My aim is the limited one of developing an account of mechanisms in medicine. Now it might be thought that this approach rejects all the features of mechanistic theories of causality, but this is not the case. It preserves many of the valuable points of mechanistic theories of causality, as I will explain in section 4.5. The causal theory of mechanisms differs from the Machamer, Darden and Craver approach in that it uses the general concept of cause throughout, while they prefer to focus on activities which are more specific causal verbs. This difference is rather less than it might appear at first, as I will explain in section 4.6. There I will show that my causal account of mechanisms does contain activities, though perhaps in a more hidden form. I will even show that Anscombe’s account of scraping, pushing, etc. can be fitted quite smoothly into my approach.

Copyright © 2018. Routledge. All rights reserved.

4.4  A causal theory of mechanisms in medicine The general plan is this. The notion of cause is taken as characterised by the action-related theory of causality given in Part I, and not by any kind of mechanistic (or mechanical) theory of causality. Having given a foundation to the notion of causality in this way, mechanisms are defined using the concept of causality. I will not try to define mechanisms in general, but only those mechanisms which characteristically appear in medicine. Medicine deals with hypotheses such as (i) A causes D, where D is a disease, or (ii) B prevents D, or (iii) C cures D. In this chapter, I will, for ease of exposition, focus on (i), but the results obtained can easily be extend to (ii) and (iii) since “prevents” and “cures” are both causal notions. When dealing with a causal hypothesis in medicine of the form A causes D, where D is some disease, I will, for convenience, write A causes D as A →  D. When handling hypotheses like this, it is often very useful, if not essential, for reasons which will be explored later, to devise a linking mechanism M, such that A→ M→ D, and to test whether the existence of M is confirmed or disconfirmed by evidence. Let us make the distinction between basic mechanisms, and mechanisms in general. I will start by considering basic mechanisms, and then introduce mechanisms in general. A basic mechanism M is defined to be a sequence of causes C1→ C2→ C3→ … → Cn which describe some biochemical/physiological processes going on in the body. I will next illustrate this by an example. My example is based on the treatment of anthrax which is given by Koch in his 1876, 1878, and 1881 papers. Anthrax was discussed earlier in section 3.1. It is a disease which affects both animals and humans. However, it will be convenient to consider anthrax in sheep. It had been observed by farmers that if sheep grazed in particular fields which they called “anthrax fields”, then many would contract anthrax and die of it. There was thus an empirical causal law of the form: Sheep grazing in anthrax fields →  some of them to die with symptoms of anthrax.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:52:50.

80  Mechanistic theories of causality

Koch’s first great discovery regarding anthrax was that the bacillus can form spores which are very resistant to destruction, but which can turn back into bacilli when they enter an animal’s blood stream. He therefore postulated that what farmers called “anthrax fields” were in fact fields which contained many anthrax spores. He further postulated that the spores had got there from “the bloody refuse of sick and dying animals” which had earlier contracted anthrax (1881, p. 59). Thus, he replaced the farmers’ empirical causal law just given by the law: Sheep grazing in fields where there are many anthrax spores →  some of them to die with symptoms of anthrax. Koch then connected the cause and effect in this law with the following linking mechanism (M4.1) 1. Anthrax spores enter the blood stream of some of the sheep through abrasions on their body or via the intestine →  2. Spores in the blood stream to turn back into anthrax bacilli →  3. These anthrax bacilli to multiply rapidly →  4. Anthrax bacilli to enter the capillaries →  5. Almost all the capillaries in the lungs, liver, kidneys, spleen, intestines, and stomach to be filled with enormous numbers of anthrax bacilli.

Copyright © 2018. Routledge. All rights reserved.

In the papers just cited, Koch gives details of numerous experiments which he carried out, and whose results supported each step of the above mechanism. A basic mechanism is a simple linear sequence C1→ C2→ C3→ … → Cn. A mechanism in general is formed by connecting basic mechanisms together to obtain a causal network of the kind which is illustrated in the Introduction, Figure 0.1. (For the reader’s convenience, I will give this figure again in this chapter.) While such complicated branching network of causes is indeed necessary in some cases to describe a mechanism in medicine, the simple linear sequence is

FIGURE 0.1 

Diagrammatic representation of a causal network.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:52:50.

Mechanistic theories of causality  81

Copyright © 2018. Routledge. All rights reserved.

surprisingly effective, and all the examples, which I will give in the rest of the book, except for one, have this form. There are two further points I would like to make about this approach to mechanisms in medicine. It applies to physical illnesses, but is more problematic for mental illnesses. It is possible that mental disturbances could all be explained in terms of material processes going on in the brain. However, despite the advances of neuroscience, we are very far from being able to give satisfactory explanations of this sort in all cases. The mind-body problem is still an enigma, and to explain psychiatric illnesses, recourse is often made to psychological principles, or to sociological considerations. In this book, however, I am focussing on physical illnesses and will leave aside the tricky problems of psychiatry. Thus, my account of mechanisms is limited not just to medicine, but more narrowly to physical illnesses. Secondly, it should be stressed that when a causal sequence such as C1→ C2→ C3→  …  → Cn is given, it is always possible to make the mechanism more detailed by inserting further causal sequences between any two successive causes in the sequence. For instance, we could insert a sequence such as C’1→ C’2→ C’3 between say C2 and C3. This indeed often happens in practice by going from one level to a lower level – for example, from the level of bacilli to the level of their DNA or of receptors on their surface. It is usually a matter requiring skilled judgement to decide how detailed a mechanism should be. One needs a mechanism, which is sufficiently detailed for the purposes in hand, without being overloaded with unnecessary complications. The general point here was well expressed by Venn, when he suggested in 1866 that the phrase “chain of causation” could usefully be replaced with “rope of causation” (quoted from Salmon, 1998, p. 224). Really causal processes are continuous like a rope, and whenever we reduce them to a chain, more causes can usually be inserted between any two links in the chain. I now want to show that this causal theory of mechanisms retains some of the good features of mechanistic theories of causality. This will be done in the next section (4.5).

4.5 The usefulness of postulating mechanisms for the confirmation of causal hypotheses and for discovering cures Glennan, referring to his mechanical theory of causation, writes (1996, p. 64): The chief virtue of the theory is that it makes the connection-conjunction problem a scientific one. If one can formulate and confirm a theory that postulates a mechanism connecting two events, then one has produced evidence that these events are causally connected. I intend to argue that the causal theory of mechanisms in medicine just sketched preserves this “chief virtue”. Let us consider a causal hypothesis in medicine of

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:52:50.

82  Mechanistic theories of causality

the form A →  D, where D is some disease. If such a hypothesis is to be used in medicine, then it must be established by evidence, that is to say that it must be confirmed by empirical evidence to a degree which makes it acceptable for employment in practical applications. Now I will argue that to produce this degree of confirmation it is nearly always necessary to postulate a mechanism M linking A to D (A →  M →  D), where the existence of M is either plausible given background knowledge, or supported by new evidence, or, preferably, both. Without postulating such a mechanism, it is rarely, if ever, possible to confirm the hypothesis A →  D sufficiently for its use to be acceptable in practical applications. This is the thesis for which I will argue in detail in Chapters 5 to 10. There is moreover a second reason why it is usually essential to consider linking mechanisms when handling causality in medicine. This is that such linking mechanisms are often needed in order to discover cures or preventative strategies for diseases. I will argue for this second thesis in Chapter 11. So, to sum up, although I am rejecting a mechanistic theory of causation in medicine in favour of an action-related theory, I am not thereby denying that mechanisms play a crucial role in handling causality in medicine. On the contrary, the postulating and testing of mechanisms is usually essential both in order to confirm or disconfirm causal claims in medicine, and in order to discover new cures or preventative strategies in medicine. The causal theory of mechanisms in medicine proposed in section 4.4 differs from that of Machamer, Darden and Craver in that it uses the concept of cause, denoted by → , but contains no mention of activities. This looks like a big difference, but, as I will argue in the next section 4.6, the difference here is rather less than it might at first seem.

Copyright © 2018. Routledge. All rights reserved.

4.6  Causes, activities, and Anscombe In the approach of Machamer, Darden and Craver (henceforth MDC), mechanisms are defined using the term “activities”, but the definition contains no mention of the general concept of cause. Conversely in the causal theory of mechanisms proposed in section 4.4, mechanisms in medicine are defined in terms of the general concept of cause, but there is no mention of the more specific concept of activity. We must now compare the two approaches, and a useful way to begin is to consider some observations made by Darden in 2013. Darden argues that causal claims are impoverished compared to claims about the mechanisms operating. As she says (2013, p. 20): I will argue that within the mechanistic sciences, such as molecular biology and molecular medicine, the claim “C causes E” is impoverished compared to the claim that “this mechanism produces this phenomenon.” Knowledge of a mechanism in the biological sciences is usually more useful for explanation, prediction, and control than merely being able to label something as a cause.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:52:50.

Mechanistic theories of causality  83

Later in the paper, she goes on to illustrate this general claim by an example taken from the field of medicine. She writes (2013, p. 26):

Copyright © 2018. Routledge. All rights reserved.

One might say “A mutation in the CFTR gene causes the disease cystic fibrosis.” But this is an impoverished claim, compared to a description of the myriad mechanisms involved in the etiology of the disease.” This last statement of Darden’s seems to me quite correct. Let us consider the general case of A causes D (A →  D), where D is some disease. Suppose now a mechanism M is discovered linking A to D. Then the statement (A →  D) and M will in general be much more informative and yield more important consequences for medicine than the statement (A →  D) on its own. If the mechanism M is analysed in the MDC fashion, then it will be a statement about activities rather than causes, and so the extra important information will not involve the general concept of cause. Suppose, however, we adopt the causal theory of mechanisms in medicine proposed in section 4.4, then the linking mechanism will take the form A →  M →  D, and, if M is a basic mechanism, it will have the form C1→ C2→ C3→  …  → Cn. So, the new statement with the extra mechanistic information will involve many more instances of the general concept of cause than the original statement A →  D did. This is illustrated by the anthrax example given in section 4.4. The original causal law was Sheep grazing in anthrax fields →  some of them to die with symptoms of anthrax This involved one use of → . Once the mechanism developed by Koch was inserted, however, the number of uses of →  increased to six. Naturally the mechanistic additions greatly improved knowledge about anthrax. On the basis of the original causal law, the only step which could be taken to stop sheep getting anthrax was to prevent them from grazing in those fields which had been identified empirically as anthrax fields. Once a single causal link had been replaced by a chain of six causal links, it was possible to consider preventing anthrax by intervening on any one of the six links. Consider the link between 2 and 3 namely: 1. Spores in the blood stream to turn back into anthrax bacilli →  2. These anthrax bacilli to multiply rapidly. Pasteur succeeded in blocking the usual causal consequence of 2 by developing a vaccine which used attenuated anthrax bacilli. This primed the immune systems of the animals who received it to enable them to destroy any invading anthrax bacilli quickly, and so to prevent such bacilli multiplying rapidly. Such a method of prevention would have been quite impossible, even unthinkable, on the basis of the original causal law. Darden’s example of cystic fibrosis can be analysed along similar lines, as I will now show.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:52:50.

84  Mechanistic theories of causality

Darden gives details of several mechanisms which are involved in producing cystic fibrosis. For simplicity, I will discuss just one of these mechanisms, which, according to Darden (2013, p. 28): “occurs in about 90% of patients with cystic fibrosis in the USA.” She describes it as follows (2013, p. 28): Three bases are deleted in the CFTR gene. During protein synthesis, this deletion results in one missing amino acid: phenylalanine at position 508 (of the 1,480 amino acids). Although missing only one amino acid, such Delta F 508 mutant proteins do not fold properly. The misfolded proteins do not implant into the cell membrane to properly transport chloride ions in and out of the cell. This can easily be translated into my sequence of causes approach along the following lines (M4.2) 1. 2. 3. 4. 5.

Three bases are deleted in CFTR gene →  Protein synthesized to miss one amino acid →  Protein to fold wrongly →  Misfolded protein not to implant into the cell membrane →  Cell membrane not to properly transport chloride ions in and out of the cell.

Darden goes on to say (2013, p. 29):

Copyright © 2018. Routledge. All rights reserved.

Although some of the protein degrades, some of the misfolded protein remains in the cells. Therapy can be directed to finding drugs that aid in rescuing the undegraded misfolded protein so that it refolds and inserts into the cell membrane and functions (albeit at a reduced level) to transport chloride ions. Darden is quite correct that the single causal law “A mutation in the CFTR gene causes the disease cystic fibrosis” would, on its own, never have suggested developing the therapy just described. However, if the single causal law is made more detailed by inserting the causal sequence M4.2 (1 →  2 →  3 →  4 →  5), then there is ample justification for developing the therapy. In fact, the therapy tries to prevent the step from 4 to 5 by getting the misfolded protein to refold in the correct manner. The analysis just given of Darden’s cystic fibrosis example surely shows that the difference between the MDC approach and the causal theory of mechanisms of section 4.4 is less than it might appear at first sight. The mechanism M4.2 does contain both entities and activities. The entity (a CFTR gene missing three bases) has the activity of synthesizing an entity: a protein which misses an amino acid. This has the activity of folding wrongly, and so on. The difference is that the causal theory approach retains the general notion of cause (→ ) which marks each link in the chain, and the entities and activities are confined to the nodes in the network.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:52:50.

Copyright © 2018. Routledge. All rights reserved.

Mechanistic theories of causality  85

Despite this large overlap with the MDC approach, I will next argue that the causal theory of mechanisms of section 4.4 is to be preferred to the MDC approach when analysing mechanisms in medicine. There are in fact a number of reasons why the causal theory approach should be judged superior. First of all, it is closer to the way that medical researchers analyse mechanisms. What I have called a basic mechanism is more or less the same as what medical researchers call a pathway. Secondly it is useful to have the causal arrows (→ ) because they indicate the points at which it might be possible to intervene in order to avoid an undesirable consequence, or to produce a desirable one. This is in accordance with the action-related approach to causality discussed in Part I of the book. Thirdly the causal theory of mechanisms in medicine avoids the difficulty for the MDC approach produced by causation by absence. In fact, causation by absence is a very common feature of mechanisms in medicine. If a gene is missing, this may mean that a protein is not produced. The absence of this protein may in turn mean that a hormone is not produced, and the absence of this hormone may in turn cause a disease. This is a problem for the MDC approach since, if an absence causes something, there are no entities whose activities produce the result. On the other hand, there is no problem for an absence causing something on the action-related approach to causality. Suppose the causal law is that the absence of hormone H produces disease D. We can base actions and interventions on such a law just as on any other causal law. For example, to cure D we have only to supply hormone H, perhaps by means of a pill. This is a typical avoidance action based on a causal law A causes B, where, in order to remove B, we have to ensure that A does not occur (sublata causa, tollitur effectus). From the point of view of the action-related theory of causality, ensuring that an absence does not occur, is no different from ensuring that an undesirable presence does not occur. I next want to consider the passage from Anscombe cited by MDC. Anscombe’s point seems to me correct, but I will argue that it supports the causal approach to mechanisms in medicine (when combined with the action-related theory of causality) better than it supports the MDC approach. The discussion of this will lead to a further, and perhaps the most cogent, reason for preferring the causal approach to mechanisms to the MDC approach. Anscombe says (1971, pp. 92–3): The truthful …  answer to the question: How did we come by our primary knowledge of causality? is that in learning to speak we learned the linguistic representation and application of a host of causal concepts. Very many of them were represented by transitive and other verbs of action used in reporting what is observed. …  The word ‘cause’ itself is highly general. How does someone show that he has the concept cause? We may wish to say: only by having such a word in his vocabulary. If so, then the manifest possession of the concept presupposes the mastery of much else in language. I mean: the word ‘cause’ can be added to a language in which are already represented many causal concepts. A small selection: scrape,

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:52:50.

86  Mechanistic theories of causality

push, wet, carry, eat, burn, knock over, keep off, squash, make (e.g., noises, paper boats), hurt. But if we care to imagine languages in which no special causal concepts are represented, then no description of the use of a word in such languages will be able to present itself as meaning cause. It is often useful for understanding philosophical texts, to place them in their historical context. The above passage is taken from Anscombe’s inaugural lecture as professor of philosophy at Cambridge in 1971.1 From the end of the Second World War until the early 1970s, the dominant school of philosophy in the UK was ordinary language philosophy. This was partly based on the later philosophy of Wittgenstein, particularly his Philosophical Investigations (1953). Anscombe had been a student of Wittgenstein and was the principal translator of his works into English. The other branch of ordinary language philosophy was based in Oxford, and the two major figures there were J.L. Austin and Gilbert Ryle. Anscombe was an Oxford don from 1946 to 1970 before returning to Cambridge as professor, and she was influenced by the Oxford approach to ordinary language philosophy as well as by Wittgenstein. In 1956, J.L. Austin gave what became a classic statement and defence of ordinary language philosophy in his presidential address to the Aristotelian Society on 29 October. Austin begins his address by arguing that “excuses” are a good topic for philosophers to study. This particular example leads him on to some general observations about ordinary language philosophy. He says (1956, p. 7):

Copyright © 2018. Routledge. All rights reserved.

the study of excuses may throw light on ethics. But there are also reasons why it is an attractive subject methodologically, at least if we are to proceed from “ordinary language”, that is, by examining what we should say when, and so why and what we should mean by it. Perhaps this method, at least as one philosophical method, scarcely requires justification at present – too evidently, there is gold in them thar hills: …  I will, however, justify it very briefly. A key part of Austin’s justification of ordinary language philosophy is the following (1956, p. 8): our common stock of words embodies all the distinctions men have found worth drawing, and the connexions they have found worth making, in the lifetimes of many generations: these surely are likely to be more numerous, more sound, since they have stood up to the long test of the survival of the fittest, and more subtle, at least in all ordinary and reasonably practical matters, than any that you or I are likely to think up in our armchairs of an afternoon – the most favoured alternative method. Austin goes on to advocate what he calls (1956, p. 9) “field work in philosophy” which consists in studying how we use a whole range of words when Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:52:50.

Mechanistic theories of causality  87

discussing a topic in everyday life. He illustrates this idea by considering aesthetics (1956, p. 9): How much it is to be wished that similar field work will soon be undertaken in, say, aesthetics; if only we could forget for a while about the beautiful and get down instead to the dainty and the dumpy.

Copyright © 2018. Routledge. All rights reserved.

Austin here suggests that philosophers studying aesthetics have focussed too much on considering “the beautiful” and progress would be made by considering more specific terms such as “the dainty”, “the dumpy” and so on. It is obvious that Anscombe has been influenced by this passage from Austin when she moves from the general word “cause” to more specific words such as “scrape”, “push”, “wet”, “carry”, and so on. This is not to say that Anscombe is simply a follower of Austin. In order to understand her position, it is necessary to consider not just ordinary language philosophy but also some of its critics. In the 1950s, ordinary language philosophy was undoubtedly the dominant school of philosophy in the UK, but still there were quite a number of philosophers in the UK at that time who were strongly critical of that approach to philosophy. The most famous of these critics was Bertrand Russell. Russell was born on 18 May 1872, and so was 84 years old when Austin gave his presidential address to the Aristotelian Society. But Russell, despite his age, was still intellectually formidable, and the next year (1957) he published a critique of ordinary language philosophy in Mind. The article was a reply to some criticisms of Russell’s theory of descriptions which had been made by an Oxford philosophy don, Peter Strawson. The specifics of this controversy are interesting, but do not concern us here. Russell regarded Strawson as a member of the school of ordinary language philosophy, and, as part of his reply to Strawson gave the following criticism of this approach to philosophy (1957, pp. 387–8): Everybody admits that physics and chemistry and medicine each require a language which is not that of everyday life. …  Let us take, in illustration, one of the commonest words in everyday speech: namely, the word “day”. …  Astronomers …  have three sorts of day: the true solar day: the mean solar day; and the sidereal day. These have different uses …  the sidereal day is relevant if you are trying to estimate the influence of the tides in retarding the earth’s rotation. …  If astronomers were subject to the prohibition of precision which some recent philosophers apparently favour, the whole science of astronomy would be impossible. For technical purposes, technical languages differing from those of daily life are indispensable. Russell clearly intends his criticism to lead to a total rejection of ordinary language philosophy, but, if we compare what he says with what Austin says, a compromise seems possible. Austin defends ordinary language philosophy as Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:52:50.

Copyright © 2018. Routledge. All rights reserved.

88  Mechanistic theories of causality

“one philosophical method”, and specifies that it is to be used in “ordinary and r­ easonably practical matters”, while Russell argues that the approach is inadequate for dealing with “physics and chemistry and medicine”. These views are compatible. Ordinary language philosophy might be appropriate when dealing with problems of everyday life, but not when dealing with problems of the sciences. Bearing this in mind, let us look again at the passage from Anscombe quoted above. Anscombe does not say, as a strict follower of Austin might, that we should consider special causal concepts such as scrape, push, wet, carry, etc instead of considering the general concept “cause”. Her point is rather that the introduction of the general concept of cause can only occur in a language which already contains the specific causal concepts. Now this point is correct and very valuable in defending the action-related theory of causality against the objection that it involves a vicious circularity. The problem, as was discussed in section 2.4, is that this theory of causality characterizes causality in terms of human actions, but human actions involve some notion of causality. So there seems to be a circularity here. It is worth noting here that all of Anscombe’s special causal concepts (scrape, push, wet, carry, … ) can be taken as referring to human actions, though this is not the only interpretation in several cases. If we limit the interpretation to human actions, Anscombe’s point is that we cannot introduce the general concept “cause” unless we have mastered a considerable vocabulary dealing with human actions. To put it another way, the understanding of the general concept of “cause” as it is used in medicine in cases like “infection by a papilloma virus causes cervical cancer” presupposes the earlier understanding of concepts describing human actions. To use a Wittgensteinian terminology, which might well have appealed to Anscombe, we can say that this earlier understanding comes from participating in language games which involve words for human actions such as “scrape”, “push”, etc. Given that this is so, it is legitimate to characterize the concept of cause, as used in medicine, in terms of human actions, and thus the action-related theory of causality does not involve a vicious circularity. So Anscombe’s point provides a good defence of the action-related theory of causality against the objection of circularity. Russell argues that the sciences require technical languages differing from those of daily life. Such technical languages usually involve some mathematics, and mathematics has a feature which differentiates it sharply from ordinary language. Ordinary language characteristically has a considerable range of words describing different objects, activities, and qualities. Mathematical systems, by contrast, are usually based on just a few key concepts in terms of which any other concepts used are defined. These are concepts such as natural number, set, probability. Now medicine is not very strongly mathematized, but there is one situation in which mathematics is used. This is in analysing diseases which are produced by several indeterministic causes. This is a subject which will be dealt with in detail in Part III of the book. However, we can say that the basic technique which has evolved to deal with such cases is that of causal networks, such as the network illustrated in Figure 0.1. In these networks the arrow connecting

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:52:50.

Mechanistic theories of causality  89

Copyright © 2018. Routledge. All rights reserved.

two nodes A and C say, means that A causes C (or more strictly that A exerts a causal influence on C). These networks usually have associated probability distributions. The nodes {A, B, C, D, E} are regarded as random variables which have a joint probability distribution. In many cases statistics enable estimates to be made of the probabilities involved. Mathematical techniques have been developed which enable probability calculations to be made in such causal probability networks, and because of their connection with causality such probabilities can be used to guide actions regarding the cure and prevention of illnesses. Details of these mathematical procedures will be given in Part III of the book, but the point I want to make now is that, as is characteristic of mathematics, the validity of such techniques depends on the arrows in the network having a uniform interpretation. Suppose, for example, that the arrow joining A to C in Figure 0.1 meant that the activity of A produces C, while the arrow C to E meant that the activity of C produces E, and so on. One would need to take account of all these differences in trying to calculate the effect of A, B, and C on E. This would complicate the calculations, if indeed they could be made at all. It is much easier from the mathematical point of view to give the arrows a uniform interpretation throughout. This is my last reason for preferring a causal theory of mechanisms to the MDC approach, and it seems to me the strongest reason. I can now sum up as follows. Basic mechanisms in medicine are defined as finite linear sequences of causes (C1→ C2→ C3→  …  → Cn), which describe biochemical/physiological processes in the body. This definition corresponds closely to the term “pathway” often used by medical researchers. Such basic mechanisms can be fitted together to produce more complicated mechanisms which are represented by networks. Entities and activities are often involved, but these are confined to the nodes of the networks, while the arrows of the networks have a uniform causal interpretation in order to facilitate mathematical calculations. That concludes my analysis of mechanisms in medicine and in the next chapter I will consider the nature of the evidence which can confirm or disconfirm the existence of such mechanisms.

Note 1 I attended this lecture since I was a research fellow of King’s College Cambridge at the time.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:52:50.

5 TYPES OF EVIDENCE (i) Evidence of mechanism

Copyright © 2018. Routledge. All rights reserved.

5.1 Confirmation and disconfirmation of causal hypotheses in medicine In this chapter and the next we will consider the types of evidence which can be used to confirm, or disconfirm, a causal hypothesis in medicine. Very often the hope is to ‘establish’ a causal hypothesis, that is to say to confirm it sufficiently strongly for it to be used in medical practice. As we have seen, Koch’s postulates were designed to give evidential criteria, which, if satisfied, would establish causal hypotheses of the form: such and such a microbe causes disease D. I will generalize this task in two ways. First of all, I will consider not just the case where the causal agent is a microbe, but any causal hypothesis of the form A causes D, where D is a disease. In addition, there are in medicine causal hypotheses of the form A cures D, and A prevents D. Clearly ‘cures’ and ‘prevents’ are words, which stand for a causal link. For the sake of simplicity, I will initially deal just with the case of A causes D, written A → D, since the cases of A cures D, and A prevents D can be treated in a very similar fashion. Secondly, although the case of establishing a causal hypothesis in medicine is very important, and attracts a lot of attention, the more Popperian case of evidence disconfirming a hypothesis, and perhaps leading to the hypothesis being rejected by the medical community, should not be forgotten. Strong disconfirmations of causal hypotheses leading to their rejection are just as important as strong confirmations of causal hypotheses leading to their acceptance. Pettenkofer wrote (1884, p. 769): As is always the case in medicine, the conflict of views is important, inasmuch as the measures to be adopted in the healing and prevention of a disease depend on the theoretical conception of it.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:15:20.

Copyright © 2018. Routledge. All rights reserved.

Types of evidence  91

This is very true, though it is somewhat ironical that Pettenkofer should have made such a point. As we saw in Chapter 3, Pettenkofer’s theoretical conception of cholera and typhoid made him draw the conclusion that these diseases could not be transmitted by drinking water, and that therefore the filtration of a town’s water supply was not necessary. Thus, Pettenkofer’s theoretical conception did not lead him to adopt the correct measures for the prevention of cholera and typhoid, and indeed he and his followers may have been partly responsible for some cholera and typhoid epidemics. It was thus very important for the progress of medicine that Pettenkofer’s theories should be disconfirmed and rejected. Koch indeed produced evidence which not only confirmed his own theory of cholera and led to its general acceptance, but also disconfirmed Pettenkofer’s alternative theory of cholera and led to its rejection. Often, of course, confirmation of one theory is closely connected with the disconfirmation of a rival theory. For this reason, in what follows, I will not focus exclusively on confirmation, but, from time to time, mention disconfirmation as well. The main thesis for which I will try to argue, in this chapter and in Chapters 6 to 11, needs, however, to be stated in terms of confirmation rather than disconfirmation. It can be put as follows. Suppose we are trying to establish a causal law of the form A causes D (A → D), where D is some physical disease. My thesis is that, in order to do so, it is nearly always necessary to postulate a mechanism M linking A to D (A → M → D), where the existence of M is either plausible given background knowledge, or supported by new evidence, or, preferably, both. Without postulating such a mechanism, it is rarely, if ever, possible to confirm the hypothesis A → D sufficiently for its use to be acceptable in practical applications. For simplicity, I will limit myself to what were earlier (section 4.4) described as ‘basic mechanisms’. So M can be taken to be a sequence of causes C1→C2→C3→ … →Cn which describe some biochemical/physiological processes going on in the body. Basic mechanisms are adequate for most cases, but occasionally more complicated mechanisms may be need and my account can be adjusted accordingly. It should also be stressed that by “evidence” I mean empirical evidence, that is to say the results of observation and experiment.

5.2  Two kinds of evidence Our model A → M → D leads to the distinction between two kinds of evidence. The first kind of evidence is statistical evidence about human populations, assuming that one is dealing with human disease. In Chapter 4, we considered the disease anthrax in sheep. If one were focussing on the diseases of sheep, then the statistical evidence would be about sheep populations. However, apart from this example, we will in this book limit statistical evidence to human populations. Statistical evidence can be obtained by observational studies, such as the cohort studies carried out in epidemiology.1 It can also be obtained from interventional studies, such as controlled trials, particularly the popular randomized controlled trials. The second kind of evidence is evidence of mechanism. This is evidence which either confirms or

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:15:20.

92  Types of evidence

disconfirms the correctness of the postulated mechanism M. Since M is concerned with biochemical/physiological processes going on in the body, the evidence for M will not usually be concerned with whole human beings, and so differs from what we have called statistical evidence. Evidence of mechanism characteristically comes from laboratory studies of cells, including the molecules of which they are composed, tissues, or animal models for the disease. Another source of evidence of mechanism is the pathological anatomy revealed by autopsies of those who have died from the disease. In section 1.2, a distinction was made between interventional and observational evidence. Interventional evidence is obtained by making an intervention and recording its results. Observational evidence arises from simply observing without making any interventions. The interventional/observational distinction cuts across the statistical/mechanistic distinction. Statistical evidence can be obtained by observational studies of human populations, or by clinical trials which involve interventions. Evidence of mechanism usually comes from laboratory experiments which, like all experiments, are interventional in character. However, evidence of mechanism also comes from autopsies which are purely observational. We are thus proposing a two by two classification of evidence. However, there are interesting connections between the two different binary distinctions, and we will explore some of these connections in later chapters. Abstract distinctions are of little interest unless they can be shown to be useful in handling important examples in medicine. To illustrate the distinction between statistical evidence and evidence of mechanism, I will give some examples from the study of coronary heart disease. In the next section, I will make some introductory remarks about this illness, and how it came to be classified.

Copyright © 2018. Routledge. All rights reserved.

5.3  Coronary heart disease (CHD) When Koch wrote his paper on tuberculosis in 1882, he gave statistics showing that it was one of the major killer diseases (section 3.2). By the 1950s, however, the situation had changed considerably. Infectious diseases were no longer such a threat, at least in the developed world, because of preventive measures, such as sewers, water filtration and chlorination, and antiseptic precautions. In addition, the development of antibiotics had made most infectious diseases, including tuberculosis, curable. The major threat to health in the developed world had become coronary heart disease. Table 5.1 gives an analysis of the causes of death in 1959 among the white population in the United States. It shows that 40% of this population died of heart disease, and 30% of just one form of heart disease, namely coronary heart disease. So coronary heart disease was the number one killer disease in the white population of the United States at that time. Males were more liable to the disease than females. Data on the black population of the United States is not given in Table 5.1, but we can summarize the situation by saying that they were healthier than the white population as far as heart disease was concerned

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:15:20.

Types of evidence  93 TABLE 5.1  Deaths in the white population of the United States in 1959

All Causes All Heart Disease Coronary Heart Disease Other Myocardial Degeneration Hypertensive Heart Disease Rheumatic Heart Disease Cerebro-vascular Disease

Males

Females

834,651 345,557 277,674 22,075 23,208 8,000 79,795

626,189 239,571 166,140 25,588 30,729 8,741 87,866

Source: Keys and Keys, 1963, p. 13.

in 1959: 29% of the black population of the United States died of heart disease in 1959 as against 40% of the white population, and 16% of the black population died of coronary heart disease as against 30% of the white population. Keys and Keys comment on the situation as follows (1963, p. 13):

Copyright © 2018. Routledge. All rights reserved.

Finally we come to coronary heart disease … this one disease accounts for some three-fourths of all heart deaths and is killing Americans at a rate of more than half a million a year. Heart attacks kill nearly twice as many people as die from all cancers and other tumors, the second leading cause of death. Moreover, it is not true, as sometimes imagined, that coronary heart disease is mainly a problem for aged people. Among U.S. white males in 1959, coronary heart disease was blamed for 30 per cent of all deaths under the age of 70. Currently, each year heart attacks are killing about 100,000 Americans under the age of 60. But what exactly is coronary heart disease, and how does it differ from other forms of heart disease? The identification of coronary heart disease as a specific disease entity was the result of an important development in the 19th century – the pathological anatomy movement. The idea behind this movement was simple enough. After a patient had died of some disease, a careful autopsy would be carried out to determine what internal damage had occurred, and an attempt made to see how this internal damage related to the patient’s symptoms. A natural result of this approach was to define diseases less in terms of their symptoms, and more in terms of internal pathological changes. One quite early discovery was that patients with a variety of different diseases had arteries in whose walls plaques had developed which restricted the flow of the blood. The German-born French surgeon and pathologist Jean Lobstein in 1829 introduced the term ‘arteriosclerosis’, or hardening of the arteries, for this phenomenon.2 Nowadays the term ‘atherosclerosis’ (from the Greek words: athera = gruel, and sclerosis = hardening) is usually employed. Atherosclerotic plaques restrict the blood flow in an artery, and, when strongly developed, can lead to thrombosis, that is the formation of a blood clot which may close the artery completely. If this happens in an artery supplying blood to the brain the result is a stroke.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:15:20.

Copyright © 2018. Routledge. All rights reserved.

94  Types of evidence

If the arteries supplying the legs and feet are affected by atherosclerosis, the result is peripheral vascular disease. A typical symptom when blood flow is restricted without being cut off altogether is intermittent claudication, which consists of pains in the legs when walking which disappear after a short rest. The arteries are no longer able to supply sufficient oxygenated blood to fuel the muscles needed in walking. In more extreme cases, perhaps after a thrombosis, gangrene can set in. Coronary heart disease is produced by atherosclerosis of the coronary arteries which supply the heart with blood. If these arteries become partially blocked, and blood flow to the heart is restricted, the result is angina pectoris. This consists of chest pains which arise from exertion, but disappear with rest. It is similar to intermittent claudication, but affects the heart rather than the legs. A thrombosis in one of the arteries supplying the heart leads to a myocardial infarction (or heart attack). If the blood clot closes off the blood supply from the artery, a portion of the cardiac muscle (myocardium) dies. In 25 to 35% of cases, a first heart attack is fatal. The 19th century pathological anatomists identified atherosclerosis, and were able to relate atherosclerotic plaques to the symptoms of various diseases. They also started making some conjectures as to how and why atherosclerotic plaques formed. Carl von Rokitansky (1804–1878) was a leading Austrian pathologist, who performed over 30,000 autopsies. He suggested that atherosclerotic plaques were formed by the deposition of blood elements on the lining of the artery. However, this theory was challenged by von Rokitansky’s German rival: Rudolph Virchow (1821–1902). Virchow pointed out that the plaques were not situated on the inner surface of the artery wall (the endothelium), but just inside the endothelium in the so-called intima of the arterial wall. A further important discovery was made in 1910 when a German chemist (Windaus) showed that atherosclerotic plaques were composed of calcified connective tissue and cholesterol. This completes my account of the background to the researches into coronary heart disease, which took place from 1910 onwards. If we followed the logic of the pathological anatomy movement, we should really consider researches into atherosclerosis rather than coronary heart disease, for, as we have seen coronary heart disease is only one of the many manifestations of atherosclerosis. Perhaps because of its great importance, however, there tends to a focus on coronary heart disease (or CHD), and our account will go along with this. Research into coronary heart disease (CHD) since 1910 is a good area for providing examples which are relevant to our investigation into the different types of evidence for a causal hypothesis in medicine. It will be possible to give examples of evidence of mechanism, observational evidence from an epidemiological cohort study, and interventional evidence from a controlled trial. In each case I will try to point out not just the strengths, but also the weaknesses of the evidence. This is designed to argue that, to produce strong confirmation, we need to combine different types of evidence. Suppose for example that the weaknesses of evidence A are the strengths of evidence B, and vice versa. Then, by putting A and B together we can have much stronger evidence than either separately could provide.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:15:20.

Types of evidence  95

Research into CHD since 1910 is also a good area for assessing the strengths and weaknesses of types of evidence, since the main conclusion which was eventually reached was at first hotly disputed, and, as part of this dispute, the cogency of different types of evidence was debated. When I speak of the ‘main conclusion’, I have in mind the two hypotheses: H1: High blood cholesterol levels cause coronary heart disease, and H 2 : Eating a diet high in saturated fat causes high blood cholesterol levels. To show that these are now part of the paradigm of contemporary scientific medicine, I will quote from Dr Mike Laker’s admirable 2012 book Understanding Cholesterol. This book was published by Family Doctor Publications in association with the British Medical Association, and so has a guarantee that it represents contemporary mainstream medical thinking. In it Laker writes (2012, p. 14): “Do high blood cholesterol levels cause CHD? The short answer is yes …”, and (2012, p. 85): “most cholesterol in your body is made in the liver from dietary saturated fats.” Daniel Steinberg’s excellent 2007 book stresses the debates that took place about hypotheses H1 and H 2. This is indicated by the title of the book: The Cholesterol Wars. The Skeptics vs. the Preponderance of Evidence. Steinberg covers the examples which I will consider below: Anitschkow in this chapter, and Ancel Keys in Chapter 6; and I will make use of his work in my own accounts of these cases. Steinberg also supports the idea that it is desirable to combine different types of evidence. He speaks of (2007, p. 197):

Copyright © 2018. Routledge. All rights reserved.

the need to synthesize evidence of several different kinds – epidemiologic evidence, experimental observations in animals, genetic evidence, clinical observations, and clinical trial data … A major message from the history … is … cumulative evidence and evidence of different kinds must be taken into account in evaluating postulated causal relations. My own views on combining evidence of different types were formed in discussion with my fellow researchers in various projects, namely: Brendan Clarke, Phyllis Illari, Federica Russo, and Jon Williamson. Our consensus position is expounded in two papers: Clarke et al. 2013 and 2014. I will refer to these two papers from time to time in what follows. After these preliminaries, we are now in a position to consider the first of our examples. This concerns the evidence of mechanism which the Russian scientist Nikolai Anitschkow obtained from his study of experimental atherosclerosis in rabbits.

5.4 An example of evidence of mechanism: Anitschkow’s study of experimental atherosclerosis in rabbits Nikolai Anitschkow (1885–1964) was a Russian scientist. He studied at the Military Medical Academy in Russia, and also abroad in Germany at Freiburg. Although of aristocratic descent, and never a member of the communist party, he

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:15:20.

96  Types of evidence

had a successful career in the Soviet Union. He was Professor in the Department of Pathological Physiology in the Military Medical Academy (1920–39), and became President of the Academy of Medical Sciences (1946–53). His work on experimental atherosclerosis was the development of the work of another researcher at the Military Medical Academy, Alexander Ignatowski. Ignatowski studied the effect on herbivorous rabbits of eating food not natural to them such as meat, milk and eggs. Ignatowski may here have been following the lead of Claude Bernard, who, in his well-known 1865 work, described experiments in which he showed that rabbits could become carnivorous. Bernard says (1865, p. 153):

Copyright © 2018. Routledge. All rights reserved.

A carnivorous rabbit had to be experimentally produced by feeding it with meat … . So I had rabbits fed on cold boiled beef (which they eat very nicely when they are given nothing else). … as long as the animal diet was continued, the rabbits kept their clear and acid urine. This type of urine is characteristic of carnivores, but, when the rabbits returned to eating grass, their urine returned to normal – the turbid and alkaline urine, characteristic of herbivores. Ignatowski was interested in studying the effects of eating animal protein on rabbits. He was influenced by Metschnikow’s theory that too much protein in the diet was harmful and accelerated aging. Ignatowski found that the rabbits fed with animal protein developed atherosclerotic plaques. As such ‘hardening of the arteries’ was then taken as a sign of aging, he concluded that Metschnikow’s theory had been confirmed. Ignatowski published these results in his 1909. Anitschkow, however, working in collaboration with Chalatow, was able to show in 1913 that atherosclerosis could be produced in rabbits by feeding them pure cholesterol dissolved in vegetable oil. This diet did not contain any animal protein. Having in this way discovered a technique for producing atherosclerosis experimentally in animals, Anitschkow devoted the next two decades to a careful study of this experimental atherosclerosis. He did not always feed his rabbits with pure cholesterol, sometimes using egg yolks, the most common dietary source of cholesterol. He killed the rabbits after time intervals of varying lengths, and was able to observe how far atherosclerosis had developed, and to compare the experimental rabbits with controls, who had been fed their normal herbivorous diet. He gave a survey of his main results from these investigations in his 1933 paper, written in English for an important international collection on arteriosclerosis. This paper in effect gives Anitschkow’s evidence of the mechanism by which atherosclerosis develops. I will now briefly summarize this evidence, so that we can try to assess its strengths and weaknesses. Anitschkow does not explicitly formulate a mechanism, but we can easily reconstruct the mechanism he advocates by paraphrasing passages from his 1933 paper, as we did earlier (section 4.4) when giving Koch’s mechanism linking

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:15:20.

Types of evidence  97

sheep grazing in ‘anthrax fields’ with some of these sheep developing anthrax. As in that case, I will use ‘→’ to mean ‘causes’. Anitschkow’s mechanism for experimental atherosclerosis in rabbits is then the following (M5.1): 1. Feeding of experimental rabbits with cholesterol-rich diet → 2. Cholesterol content of blood to increase several times → 3. Infiltration of some parts of the arterial wall by lipoids (mainly cholesterol esters) → 4. Lipoids to be absorbed by macrophages, which become foam cells → 5. Full development of atherosclerotic plaques.

Copyright © 2018. Routledge. All rights reserved.

As regards the time needed for this causal process to unfold, Anitschkow states that, in the case of rabbits fed daily with 0.5 gm of cholesterol dissolved in vegetable oil, it takes about a month to reach stage 3, and about three to four months to reach stage 5 (Anitschkow, 1933, pp. 286–7). Anitschkow presents detailed experimental evidence for each step in the causal chain. One of the most significant links in this chain is that between 2 and 3. Lipoids in the form of fatty streaks can be observed under the inner surface of the artery (the endothelium) in the intima of the arterial wall. But how did they get there? Anitschkow’s claim is that they came from the blood circulating in the interior of the artery (its lumen) by penetrating the artery wall. This is how he defends this view (1933, p. 297): The lipoidal deposits certainly cannot possibly be regarded as products of degeneration of the arterial wall itself. The whole morphologic picture of the lipoidal accumulations contradicts that idea. For the normal structure of the arterial wall remains unimpaired after the lipoidal substances have made their appearances. The latter are found as deposits within this unimpaired, normal structure of the wall, and more especially in its ground substance. The hypothesis that such large quantities of lipoid could be derived from the decomposition of the mucoid ground substance of the arterial wall is quite untenable from the chemical point of view. Thus the result of direct microscopical examination is quite clear – we have to deal not with lipoid degeneration, but with lipoid impregnation of the arterial wall. The experimental conditions under which these lipoid deposits are formed in the arteries also point in the same direction. For these deposits are observed exclusively in experimental animals that have been fed for a considerable time with a mixture of cholesterin [Anitschkow uses this term for what we now call cholesterol – D.G.] and oil. Anischkow’s last point is that the appearance of lipoids in the artery walls occurs only in the experimental rabbits which have been fed a cholesterol-rich diet causing the cholesterol content of their blood to increase several times (the transition from stage 1 to stage 2). In normal rabbits which have a much lower cholesterol content in

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:15:20.

98  Types of evidence

their blood, no lipoid deposits occur in the artery walls. This indeed is evidence that the lipoid deposits in the artery walls come from the lipoids circulating in the blood. Anitschkow gives one further piece of experimental evidence for the same conclusion. If a colloidal stain (trypan blue) is introduced into the blood, it can be seen to enter the arterial wall in approximately the same places where the lipoid deposits are found. This shows that the inner lining of an artery (its endothelium) is at least partly porous. All this evidence for the causal step from stage 2 to stage 3 was most important. It will be remembered that Virchow criticized von Rokitansky’s theory that atherosclerotic plaques were formed by deposits from the blood on the grounds that these plaques formed not on the surface of the arterial wall but in its intima. Anitschkow, however, revived von Rokitansky’s theory in a developed form. Anitschkow writes (1933, p. 306):

Copyright © 2018. Routledge. All rights reserved.

the experimental investigation of cholesterin atherosclerosis has thrown much light on the pathogenesis of human atherosclerosis. Above all it has made possible definitely to demonstrate that the initial stages of the process are in the nature of a lipoid infiltration or imbibition of the intima, and that the lipoids enter the arterial wall from the lumen. This has greatly strengthened the “infiltration theory” of atherosclerosis. The remaining stages of Anitschkow’s causal mechanism for atherosclerosis need little comment. The macrophages mentioned in stage 4 are the white blood cells which normally engulf and dispose of invaders from outside such as bacteria. The macrophages try to dispose of the lipoids in the arterial intima, but with little success. Instead of digesting these lipoids successfully, the macrophages become swollen with lipoids and turn into what Anitschkow calls in his 1933 (p. 303) “xanthoma cells”. I have changed this to the usual contemporary term ‘foam cells’. Anitschkow was the first, in 1912, to describe foam cells. That concludes my account of Anitschkow’s causal mechanism for atherosclerosis, and the experimental evidence he presents for it. We must next consider the strengths and weaknesses of this evidence of mechanism. I think it can be agreed that Anitschkow presents very strong experimental evidence for his causal account of the formation of atherosclerosis in rabbits fed with a cholesterol-rich diet. However, he wants to argue that this provides evidence for the human pathogenesis of atherosclerosis. Now this holds only if we regard his experimental rabbits as providing a good model for the human case, and this could be questioned. This leads us at once to the key weakness of much evidence of mechanism. Very often this evidence consists of laboratory experiments carried out with animals. Now the use of animals enables detailed experiments to be carried out which obviously would not be possible on humans. So very strong evidence can be built up for the causal processes occurring in the animals. However, these will only apply to the real case of interest (the human case) if the animal provides a good model for what happens in humans, and this may not be the case.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:15:20.

Types of evidence  99

Copyright © 2018. Routledge. All rights reserved.

Unfortunately, in Anitschkow’s case, there are many reasons for supposing that his experimental atherosclerosis in rabbits is not a good model for human 在这个例子中， atherosclerosis. To begin with, the rabbit is a herbivore, and the experimental 兔子不是人类的 atherosclerosis is only produced by feeding the rabbit with a cholesterol-rich diet 一个好模型。 made of foods such as egg yolks which are a completely unnatural food for rabbits. Humans, on the other hand, are omnivores, and it is quite natural for us to eat meat and eggs. Thus, the rabbit case and the human case look very different. Next comes the time factor. The experimental atherosclerosis in rabbits is created in three or four months, whereas the atherosclerotic plaques found in humans take decades to form. Is a rapid experimental process really the same as such a slow natural one? Now admittedly in his famous 1882 paper on tuberculosis, Koch reports a considerable number of animal experiments and these were accepted as being relevant to human tuberculosis. However, there is a difference between Koch’s case and that of Anitschkow. Koch used a great variety of animals and was always successful. He mentions among others: guinea pigs, rabbits, cats, mice, rats, hedgehogs, hamsters, pigeons, and frogs. By contrast Anitschkow was unable to produce experimental atherosclerosis in some animals. He says (1933, p. 301): “notwithstanding repeated attempts, it has not proved possible as yet to produce arterial lipoidosis by feeding dogs and cats with cholesterin”. These objections must have been made to Anitschkow himself, because in his 1933 paper he argues against them, and in favour of the view that his proposed causal mechanism for atherosclerosis produced experimentally in rabbits holds also for human atherosclerosis. I will now give a brief account of Anitschkow’s arguments on this point. First of all, regarding the question of why experimental atherosclerosis cannot be produced in dogs and cats by feeding them with cholesterol, Anitschkow has this to say (1933, p. 301): This is probably to be accounted for by the circumstance that carnivorous animals are normally adapted to the ingestion of large quantities of cholesterin in their food, so that the cholesterin is promptly excreted from the body … . The extremely abundant lipoid infiltration of the bile-passages and of the wall of the gall-bladder in such animals would seem to indicate that this excretion takes place together with that of the bile. A second objection was that experimental atherosclerosis is formed in a very short time in rabbits, while the development of atherosclerotic plaques takes much longer in humans. Anitschkow responded to this objection by carrying out new slower experiments in rabbits (1933, p. 307): comparatively acute atherosclerotic changes are produced in rabbits by feeding them with cholesterin for 3 to 4 months, whereas the analogous changes of human atherosclerosis need a much longer time for their Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:15:20.

100  Types of evidence

development. But if we feed the rabbits during a considerably longer period, say two and one-half years, with much smaller quantities of some food that is not too rich in cholesterin, such as milk, we can produce typical atherosclerotic changes in their arteries, while the lipoid changes in the internal organs and the hypercholesterinemia do not reach a very high degree (Anitschkow, 1932). In other words, the total picture of the changes produced in the rabbit under the experimental conditions just described resembles that of human atherosclerosis much more closely than that observed during experiments of a shorter duration, in which the rabbit is literally swamped with cholesterin.

Copyright © 2018. Routledge. All rights reserved.

However, Anitschkow’s main response to the objection that his animal experiments are irrelevant to human atherosclerosis is to point out that the development and character of atherosclerosis in his experimental rabbits is very similar in detail to the development and character of human atherosclerosis. To begin with, the initial lipoidal deposits occur in exactly the same places in the arteries of humans and experimental rabbits. Typical locations in both cases are (Anitschkow, 1933, p. 302): “the places where the large arteries branch off from the aortic arch.” Moreover, the lipoid deposits have the same shapes in the experimental rabbits which they have in the corresponding positions in the human case. Examples are (Anitschkow, 1933, p. 302): “fan-shaped lipoidal streaks in the ascending aorta … long lipoidal streaks in the left coronary artery of the heart.” If the development of the atherosclerosis is examined under the microscope, it goes through exactly the same stages in the experimental rabbits as it does in humans. The lipoidal deposits are absorbed by macrophages which turn into foam cells. These are then transformed by the development of fibrous connective tissue into atherosclerotic plaques. Finally, from the chemical point of view, the two cases are just the same. The lipoid deposits consist largely of cholesterol and cholesterol esters, and the plaques contain calcium deposits. Anitschkow concludes (1933, p. 305): And thus we see that between the atherosclerotic changes in human arteries on the one hand and the analogous experimental changes observed in cholesterinized rabbits on the other there is a remarkable similarity; not only from the morphologic and morphogenetic but also from the chemical point of view. We are therefore fully justified in describing the two processes as analogous in all essential respects. This is a convincing argument, and we can draw the conclusion that the experimental evidence, which Anitschkow provides for the mechanism of formation of atherosclerotic plaques in rabbits, is also evidence that the same mechanism applies to the formation of such plaques in humans. Anitschkow obtained his information about the development and character of atherosclerosis in humans from the results of autopsies, which provided

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:15:20.

Types of evidence  101

Copyright © 2018. Routledge. All rights reserved.

observational, but not experimental, evidence of pathological anatomy. He combined this with the experimental, interventional, evidence provided by his cholesterol-fed rabbits. Now, as usual, controlled experiments can provide much more information than can be obtained by observation alone. Consider, for example, the experiment of injecting a dye (trypan blue) into a rabbit’s blood stream, killing the rabbit a little later, and then checking that the dye had penetrated the arterial walls in exactly the points where lipoidal deposits characteristically appear. This experiment provides strong evidence that such lipoidal deposits come from the blood stream, and are able to penetrate the arterial wall into the intima. However, such an experiment could obviously not be carried out on humans. So, work on experimental animals produces strong evidence for the mechanisms going on in those animals. The weakness of this approach is that we cannot be sure that the same processes go on in the human case. Conversely, observational evidence from autopsies is weaker than experimental evidence on living creatures, but we can be sure that what is observed applies in the human case. If we put the two types of evidence together, however, and manage to show that the observational evidence on humans completely agrees with the animal case, we have so to speak cancelled out the weaknesses of the two types of evidence, and produced strong evidence for the mechanism which operates in human beings. This is an example of an evidential principle which I will call strength through combining. If we combine two types of evidence, A and B say, such that the strengths of A compensate for the weaknesses of B, and vice versa, and if the two types of evidence support the same hypothesis, then we produce much stronger evidence for that hypothesis than could be provided by either type of evidence on its own. I will give further examples of strength through combining as we go along. That concludes my account of an example of evidence of mechanism in the case of coronary heart disease. In the next chapter I will give examples of a quite different type of evidence, namely statistical evidence in human populations.

Notes 1 For a good recent account of epidemiology and the philosophical problems it generates, see Broadbent (2013). 2 The account of 19th-century pathological anatomy which follows is largely based on Abela (2003) and Mayerl et al. (2006).

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:15:20.

通过合并统计证 据和机制证据来 抵消两种证据的 弱点。

6 TYPES OF EVIDENCE (ii) Statistical evidence in human populations

Statistical evidence in human populations can be divided into two types, namely (a) observational statistical evidence, generally obtained by epidemiological studies, and (b) interventional statistical evidence, generally obtained by clinical trials. I will begin with an example of (a). In fact, my example is one of the most famous in the history of epidemiology, namely Ancel Keys’ study of coronary heart disease in seven countries. Before I come on to this example, however, it will be useful to say something about Ancel Keys and his work, and this I will do in the next section.

Copyright © 2018. Routledge. All rights reserved.

6.1  Ancel Keys and the dangers of saturated fat Ancel Keys, who made a most important contribution to the study of coronary heart disease, was also a rather fascinating character. He was born in Colorado Springs on 26 January 1904. The family moved to San Francisco just before the earthquake of 1906. Having fortunately survived, they moved to Berkeley, and in due course, Ancel Keys became a student of the University of California at Berkeley. He began studying chemistry, but took time off to work as an oiler on a ship sailing to China. On his return, he switched majors and graduated in economics and political science in 1925. These events show two features of his character which remained for the rest of his life, namely his love of travel and the breadth of his intellectual interests. In fact, after an undergraduate degree in economics and political science, he took an M.S. in zoology (1928), followed by a PhD in oceanography and biology from UC Berkeley. He then obtained a post-doc fellowship to carry out research in Denmark, where he worked for two years on the physiology of fishes, writing several papers on the subject. His next move was to Cambridge in the UK where, in 1936, he took a second PhD – this time in physiology. His research in this area included an expedition to the Andes to study the effects of high altitude on the body.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:56:22.

Types of evidence  103

When war broke out, Ancel Keys received a call from the Quartermaster Corps of the US Defense Department for advice on food for paratroops. He says (1990, p. 288) that he does not know why he was consulted, but that it may have been because he survived for six days on condensed food at the top of a mountain in the Andes. In response to the call, Ancel Keys devised a ration, which came to be known as the K Ration. This made him well known, but, when he wrote some reminiscences in 1990, he does not appear very enthusiastic about this part of his work. He writes (Keys, 1990, p. 288):

Copyright © 2018. Routledge. All rights reserved.

The only nutritional criterion for a paratroop ration was calories. The K Ration, produced in vast quantity by the Wrigley Company in Chicago, had a piece of hard sausage, dry biscuits, a block of chocolate, a stick of chewing gum, matches, and a couple of cigarettes, all in a waterproof package to fit in a uniform pocket. … I … know that many thousands were issued to troops where better food could have been provided … Some soldiers lived on it for a month in islands in the Pacific. “K Ration” became a synonym for awful food. Certainly, the K Ration very much contradicts the ideas on nutrition, which Ancel Keys was to develop later in his life. During the war many people suffered from lack of food, and even starvation. Keys decided to study the effects of starvation, and the best way to help people to recover from a period of starvation. His plan was to set a baseline by feeding a controlled diet for three months, then introduce a starvation diet for six months and study its physical and mental effects on those who volunteered to take part, and finally to examine the effects of three different ways of refeeding over a period of three months. The volunteers were recruited from conscientious objectors, and in all 32 of them completed the study. When the war ended, Keys sent his findings to international relief agencies dealing with victims of starvation. He wrote up his research as a two volume 1385-page book: Biology of Human Starvation, which is still one of the most important works in the field. Before this was published, however, the mercurial Ancel Keys had already acquired a new research interest – coronary heart disease (CHD). As we have seen, coronary heart disease had become by this time (the late 1940s) the major killer disease in America. So, it is not surprising that an ambitious researcher such as Ancel Keys should turn his attention to it. His first effort was empirical in character, and, it must be said, rather floundering. Keys was in the University of Minnesota, Minneapolis. So he selected a cohort of middleaged business men (aged 45–55) from the Twin Cities with a view to following them for some years. The idea was to keep track of a number of different characteristics and see if significant correlations could be found with any of them among the sub-group who contracted CHD. However, as Keys himself says (1990, p. 289): “Among 60 characteristics measured, the diet had little attention.” In fact, this study produced few interesting results.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:56:22.

104  Types of evidence

Keys, however, was also pursuing some more theoretical lines of enquiry. In a paper with collaborators published in 1950, Keys does not directly refer to Anitschkow, but as he says (1950, p. 79): “when certain herbivorous animals like the rabbit are fed enormous amounts of cholesterol the blood level rises and atherosclerosis develops”, he must have been reading about Anitschkow’s work. Very likely his source was Anitschkow’s 1933 paper, written in English in a well-known international collection on arteriosclerosis. Keys’ reaction to Anitschkow’s work is, however, very negative and critical. He goes on, immediately after the passage just quoted to say (1950, p. 79): “but the situation may well be very different in man when subsisting on more natural diets.” In fact, Keys and his collaborators carried out some experiments which showed that the situation in humans was different from that in rabbits. Their result was (Keys et al., 1950, p. 79):

Copyright © 2018. Routledge. All rights reserved.

the serum cholesterol level of “normal” men as represented here is not significantly related to differences in the habitual cholesterol intake over a range of something like 250–800 mg per day. So, feeding humans large amounts of cholesterol is not likely to raise their blood cholesterol level very much. This important finding has been confirmed by subsequent research. Dietary cholesterol intake up to about 300 mg per day does raise blood cholesterol, but thereafter a saturation effect occurs, and further cholesterol dietary intake has little effect on blood cholesterol levels. The main dietary source of cholesterol is egg yolks, but, for those whose intake is already 300 mg, it has been shown that adding a further two eggs per day to the diet does not increase blood cholesterol levels. (See Steinberg, 2007, p. 47.) But are there other dietary changes which have a bigger effect on blood cholesterol changes? In their 1950 work (p. 80), Keys et al. note that three weeks on a rice-fruit diet did reduce the mean blood cholesterol level for four patients from 232.5 mg/100 ml to 151.8 mg/100 ml. However, they do not explore the question further. Perhaps Keys was not very interested at this stage in what raised or lowered blood cholesterol levels, because he says (1950, p. 79): “it is by no means certain that moderate differences in the serum cholesterol level in man have any influence on development of atherosclerosis.” Some crucial developments in the next few years changed Keys attitude to the importance of blood cholesterol levels for atherosclerosis. In 1951, Keys was Chairman of the Joint FAO-WHO United Nations Committee – no doubt because of his work on starvation. This committee had a meeting for ten days in Rome that year, and at the meeting Keys met Gino Bergami from the University of Naples. According to Keys’ 1983 memoir (p. 2): Gino Bergami … told me about the typical low-fat diet of Naples, insisted that coronary heart disease was not a big problem in that city, and invited me to Naples to see for myself. Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:56:22.

Types of evidence  105

Naturally Keys, a great lover of travel, could not resist this offer, and in January 1952 he set off for Naples with his wife Margaret. Margaret Keys was a medical technician who assisted her husband by taking blood samples and measuring cholesterol levels. She collaborated with him in his research in other ways as well. Their findings were very much in agreement with what Gino Bergami had told him (Keys, 1983, pp. 2–3):

Copyright © 2018. Routledge. All rights reserved.

No tests of probability were needed to show a vast difference between these workers in Naples and their counterparts in Minnesota. The average serum cholesterol was about 170 mg/dl … . The dietary survey was crude but there was no mistaking the general picture. A little lean meat once or twice a week was the rule, butter was almost unknown, milk was never a beverage except in coffee or for infants, ‘colazione’ on the job often meant half a loaf of French bread crammed with boiled lettuce or spinach. Pasta was eaten every day, usually also with bread (no spread), and something like a fourth of the calories in the diet were provided by olive oil and wine. There was no evidence of nutritional deficiency … As to coronary heart disease … the disease was rare in the hospitals. … During the time we worked in Naples we learned of no acute myocardial infarction patients in the 60-bed medicine service in the University Hospital next door … The only qualification which Keys makes, is that these observations applied to the working class. Wealthy Neapolitans suffered more from coronary heart disease than the workers, but they also ate more meat and rich food. From Naples, Ancel and Margaret Keys went on to Madrid, where they found much the same situation. Ancel Keys’ new focus on fat in the diet was also suggested by some of his reading. He mentions (1983, p. 1), in particular, a 1950 paper by Malmros which deals with atherosclerosis in the Scandinavian countries during the war. The surprising result was that shortages and rationing produced a reduction in atherosclerosis. Speaking of deaths from arteriosclerosis in Sweden, Malmros says (1950, pp. 140–1): The mortality curve shows a rising tendency up to 1941, but it falls steeply between 1942 and 1943 in order after that to climb to its former level again. … the rationing of foodstuffs … during the years 1942–3 led to a reduction in the consumption of butter, eggs and meat. This is illustrated by Figure 6.1, taken from Malmros (1950, p. 142). Figure 6.1 shows that the same dip in the death rate from arteriosclerosis occurs in Finland and Norway as well as Sweden, but it does not occur in the USA where the curve rises steadily in the same period. Of course, food stuffs were not restricted in the USA during the war.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:56:22.

106  Types of evidence

rate from arteriosclerosis (atherosclerosis) in Finland, Norway, Sweden, and USA 1935–47.

Copyright © 2018. Routledge. All rights reserved.

FIGURE 6.1  Death

Studies such as that of Malmros linked Keys’ previous work on starvation to his new research interest in coronary heart disease. Like his investigations in Naples and Madrid, they showed the value of examining the effects of different diets in different countries. With his characteristic love of travel, Ancel Keys extended his research in the next few years to Finland and Japan. His visits to Finland led him to question some commonly held ideas about the aetiology of coronary heart disease. The high levels of this disease among American businessmen were often attributed to their sedentary life, which led them to become fat and physically unfit, and to the stresses of their life in the high-pressured atmosphere of American cities. If this were so, then a lean, physically fit population in a calm rural environment should have a very low level of coronary heart disease. Finland showed that this was not the case. Speaking of Finland, Keys says (1957, p. 1917): a large proportion of the population is strictly rural and lives widely scattered in the quiet countryside. There is little housing congestion, motor traffic is light, hard physical work is the rule rather than the exception, and

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:56:22.

Types of evidence  107

probably no other country can boast so high an average level of physical fitness. According to some ideas, Finland should be peculiarly free from coronary heart disease, but according to vital statistics it competes with the United States in the great frequency of the disease. In the later seven countries study, the cohort with the highest level of coronary heart disease was the one from East Finland. Keys tells us what he learnt about the local diet from his visit to the area (1983, p. 7):

Copyright © 2018. Routledge. All rights reserved.

we went into the woods to have sauna with some lumberjacks. Two of them confessed to being slowed up by angina pectoris but more interesting was a glimpse into local eating habits. A favorite after-sauna snack was a slab of full-fat cheese the size of a slice of bread on which was smeared a thick layer of ‘that good Finnish butter’. A later detailed dietary survey by Maija Pekkarinen and the late Paavo Roine found that butter, milk, and cheese accounted for 40% of the average total dietary calories of the middle-aged men in that area … . We were not surprised when Minneapolis reported serum cholesterol values over 300 mg/dl in more than 15% of the samples we sent from East Finland. Japan was in terms of the prevalence of coronary heart disease, the other extreme from Finland. Keys states (1957, p. 1916) that the death rate in Japan from coronary heart disease in 1953–54 for men aged 50–54 was less than a tenth that for white American men in the same age group. However, the groups, which Keys and his collaborators studied in rural southern Japan, had a lifestyle very similar to that of East Finland. They were farmers or fisherman, used to carrying out hard physical labour, and, as a result, lean and physically fit. The great difference between them and the Finns was the character of the diet, which in the case of Japan was very low in fat. These observations led Keys to the conclusion that a high level of fat in the diet was a crucial factor in the aetiology of coronary heart disease. But what kind of fat was relevant? Keys and his associates carried out some clinical trials on this question, reporting the results in Keys et al. (1957). Dietary fats were divided into saturated, monounsaturated, and polyunsaturated. The experimental results showed that saturated fats raised blood cholesterol levels, monounsaturated fats left blood cholesterol levels largely unchanged, while polyunsaturated fats reduced blood cholesterol levels, though by much less than they were raised by the corresponding amount of saturated fats. The main source of saturated fats is animal fats in meat, butter, cheese, cream, etc. By 1957 Keys had developed causal hypotheses regarding CHD, which in our notation can be stated as follows (CH 6.1): 1. Having a diet high in saturated fats → 2. Blood cholesterol level to become high → 3. Coronary heart disease. Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:56:22.

108  Types of evidence

Despite his PhD in Physiology, Keys never expresses much interest in the mechanism which connects 2 and 3. However, he could, if he had wished, have adopted the mechanism connecting 2 and 5 in Anitschkow’s scheme given above (M5.1, section 5.4). Keys did disagree with Anitschkow, but only in the step from 1 to 2 in M5.1, and so could have accepted the rest of Anitschkow’s mechanism. Keys’ difference from Anitschkow is however of great importance. Keys stressed the importance of a diet high in saturated fats, while Anitschkow had used diets high in cholesterol. This emphasis on saturated fat is a strikingly original innovation, and one which is accepted today by the medical mainstream. For example, Laker writes (2012, pp. 84–85): Although the amount of cholesterol in your diet is important, it is probably less so than the amount of saturated fat that you eat, because most cholesterol in your body is made in the liver from dietary saturated fats. Cholesterol is also absorbed rather poorly from the gut with over half remaining behind. Laker quantifies this claim earlier by saying (2012, p. 27): “Typically, around seven times more cholesterol is made in your body from dietary saturated fats (animal and dairy fat) than is absorbed as cholesterol from your food.” Laker concludes (2012, pp. 85–86):

Copyright © 2018. Routledge. All rights reserved.

The main objective of changing your diet is to reduce your saturated fat intake. … foods rich in cholesterol with a modest amount of saturated fats are allowed in moderation, including eggs and shellfish. It is OK to eat two eggs and one portion of shellfish per week. Previous to Keys, most nutritionists would have considered meat and milk to be very healthy foods. It was a revolutionary step to propose that such foods, because they were high in saturated fats, might be a danger to health. This suggestion was very unwelcome to many people both at the time and even today. Keys must have known that his theory would meet with strong opposition, and it was in order to provide strong empirical confirmation of the theory that he organized his seven countries study, which will be considered in the next section. It is worth noting that this study, unlike Keys’ earlier study of business men in the Twin Cities, was designed to test out a theory which he had explicitly formulated before the study began.

6.2 An example of observational statistical evidence: The seven countries study The seven countries study is an example of what epidemiologists call a “cohort study” or “prospective study”. Suppose such a study is undertaken to investigate

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:56:22.

Types of evidence  109

a particular disease D say. First of all, a cohort is selected, and the individuals are examined at the beginning of the study, and thereafter at regular intervals, to find what values they have of a number of chosen characteristics such as blood pressure. In due course, some of the individuals may develop disease D, and it can then be checked whether interesting correlations can be found between occurrence of D and the measured values of the chosen characteristics. A cohort study is purely observational in character, since there is no intervention on the part of the researchers, and the result of the investigation consists of some statistical data. In the case of the seven countries study, 16 different cohorts were chosen – at least one in each of the seven countries. The investigations were started between 1958 and 1964, with the plan of examining the participants at the beginning and then every five years. By 1969, the five-year results were in, and in 1970, the results were published in Ancel Keys (ed.). I have given Ancel Keys as the editor, since the individual chapters were generally written by the local organizers, sometimes with a contribution from Ancel Keys and one or more of his collaborators at the University of Minnesota, Minneapolis. For example, Chapter 11, pp. 113–122 of the report deals with ‘Five Years of Follow-up of Railroad Men in Italy’, and is authored by Henry L. Taylor and Ancel Keys from Minneapolis, and Alessandro Menotti, Vittorio Puddu, and Mario Monti from Rome. The choice of the seven countries seems to have been a mixture of chance and planning. Ancel Keys says that he started by including all those in which he had already worked in his preliminary studies. However, some of these were excluded for various reasons. For example (1970, p. 3):

Copyright © 2018. Routledge. All rights reserved.

Spain was eliminated because of lack of funds and interested local personnel; anyway it seemed unlikely that work in Spain would produce anything much different from what would be found by the same kind of work in Italy. Greece was added because Keys had friends there, and he thought that (1970, p. 3) “the reputed large use of olive oil in the diet should be interesting.” The Netherlands was added because the government offered personnel and financial help, and also because it looked to be an interesting case, since (1970, p. 3): “Mortality ascribed to CHD was very low, although the Dutch were pictured as growing fat on a diet high in saturated fats.” The final choice of the seven countries was: Finland, Greece, Italy, Japan, the Netherlands, USA, and Yugoslavia. The selection of the cohorts within each country is also of interest. Keys considered men aged between 40 and 59, as being the group most at risk of CHD. In the USA, there was one large cohort of 2,571 men who were employees of the American railroad industry. In the other countries, the usual procedure was to fix on some location or locations within the country and choose all the men of the requisite age in that area. An exception was Zutphen in the Netherlands where a selection of all the men in the town would have produced a cohort too big to be followed with the resources available. Accordingly, a random sample

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:56:22.

Copyright © 2018. Routledge. All rights reserved.

110  Types of evidence

was taken of four out of nine of Zutphen men aged between 40 and 59. The selection of the cohorts in Italy gives an idea of the general approach. In a feasibility study carried out in 1957 before the start of the main programme, Keys had selected the village of Nicotera in Calabria in the toe of Italy. However, in the main study he decided that the traditional Mediterranean cuisine of Nicotera was very similar to that of Crete. So, he included Crete, and selected instead of Nicotera, two villages in the centre of Italy near Bologna and Modena, where the diet contained more meat and dairy products. In addition to the cohorts from these two villages (Crevalcore and Montegiorgio), a cohort of employees of the Italian railroads was selected to match that of the USA. We can see that the selection was made to produce variety and useful comparisons. Leaving aside the very large American cohort, the participants in the 15 remaining cohorts numbered 10,199, giving an average of 680 per cohort. The smallest of these cohorts was 504, and the largest 993. The next problem was what characteristics of the participants to measure. Naturally the blood cholesterol level was measured, but so also was height, weight, blood pressure, and fatness (measured by the combined thickness of two skinfolds). A note was made of smoking habits, and also of level of physical activity, which might be high for those engaged in heavy manual work, and low for those in sedentary occupations. Finally, of course, the diet was recorded. Samples from the cohorts were taken, and the foods they ate in a week were carefully recorded and weighed. This was repeated at different seasons of the year. The foods consumed were analysed to find their constituents in terms of fats, carbohydrates, proteins, etc. This analysis was carried out in different ways, and it was checked that the different estimates agreed. (See Ancel Keys (ed.), 1970, pp. 162–9 for details.) The study had to record how many of the participants had coronary heart disease and how many died from it. Here Keys was very concerned to standardize the diagnostic criteria used, and the way in which the cause of death was determined. He was afraid that different medical practices and different ways of preparing death certificates in the different countries might distort the findings. Emphasis was placed on hard diagnostic criteria such as ECGs. It is difficult not to be impressed by the close attention to detail with which every aspect of the project was implemented. Equally impressive is the overall design. In fact, there is very considerable individual variation in the effect of diet on blood cholesterol levels. In an extreme case, those who suffer from a genetic condition known as familial hypercholesterolemia have very high blood cholesterol levels whatever diet they eat. This problem was overcome by comparing, not individuals, but cohorts of at least 500 members who all ate very similar diets. Average values for such cohorts gave a good indication of the general effect of diet on blood cholesterol level. Another very good feature of the design was that there was a great deal of variation in the diets of cohorts in different countries, and even, to some extent, between cohorts in the same country. This variation was a good way of detecting the key causal factors in

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:56:22.

Types of evidence  111

coronary heart disease. As Keys himself says, in a passage which contains an implicit criticism of his study of business men in the Twin Cities, (1970, p.1):

Copyright © 2018. Routledge. All rights reserved.

limiting studies to urban white Americans must be a serious handicap if it is hoped that relationships between factors in the mode of life and susceptibility to CHD will be discovered. Variety – and not a high degree of homogeneity in culture and habits – must be sought. So it is necessary to look at other populations. The seven countries study is thus an excellent example of the view of causality proposed by Russo in her 2009 book. This is based on what she calls “the rationale of variation”, writing (p. 4): “the rationale of variation conveys the idea that, in causal models, to test causal relations means to test variations among the variables of inter- 从variation中推 est.” Later, she contrasts this with a more Humean approach, saying (p. 109): “Hume 论因果关系。 inferred causation from regularity, whereas my claim is that we infer causation from variation because variation conceptually and empirically comes before regularity.” That completes my description of the seven countries study. Let us now see what results it brought to light after five years. After all the complexities of setting up the seven countries study, the results obtained turn out to be quite simple. Keys was working with a causal hypothesis stated above (CH6.1). The causal links postulated in this hypothesis were between stages 1 and 2, between stages 2 and 3, and consequently between stages 1 and 3. If these causal links do indeed hold, then there should be strong positive correlations between percentage of saturated fats in the diet (stage 1) and serum cholesterol level (stage 2), between serum cholesterol level (stage 2) and incidence of coronary heart disease (stage 3), and between percentage of saturated fats in the diet (stage 1) and incidence of coronary heart disease (stage 3). Thus, the main results of the five-year study boiled down to three correlations, which we will consider in turn. Note that the correlations were not worked out for individuals, but for cohorts. For each cohort, the average percentage of calories which came from saturated fats, and the median serum cholesterol level were calculated from the data. The first correlation was between the percentage of calories in the diet from saturated fats, and the median serum cholesterol level. Its value was 0.89, and the relationship is illustrated in Figure 6.2. Here the broken circles indicate that (Keys, 1970, p. 171): “the fatty acid values are estimated from the dietary records rather than determined by direct chemical analysis as in the other cohorts.” As can be seen, linear regression provides a good fit. The second correlation was between the median serum cholesterol level, and the incidence of coronary heart disease (CHD). This incidence was measured in two different ways. The first was the percentage of deaths from CHD, and myocardial infarctions (heart attacks) established by “hard criteria” such as ECG results. The second was the percentage who were diagnosed as having CHD. In both cases, consideration was limited to men CHD-free at entry, and the results were age-standardized. The results for both measures are given in Figure 6.3.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:56:22.

112  Types of evidence

FIGURE 6.2 

Copyright © 2018. Routledge. All rights reserved.

Source: Keys, 1970, Figure XVII. 2, p. 170.

Once again, the linear regression model gives a good fit, and the correlations are 0.76 and 0.81. The third correlation was between the percentage of calories in the diet from saturated fats, and the incidence of coronary heart disease. Again the two methods were used to measure CHD incidence. The result of the first method, i.e. deaths from CHD and myocardial infarctions is given in Figure 6.4. The result of the second method, i.e. percentage who were diagnosed as having CHD is shown in Figure 6.5. In both cases, the linear regression model gives a good fit, and the correlation is 0.84. This evidence undoubtedly confirms Keys’ causal hypothesis. For each causal link postulated in that hypothesis, the data yields a good linear regression with a strong positive correlation. The correlations are all between 0.76 and 0.89. This is the strength of the evidence. But what is its weakness? Its weakness can be summed up in the well-known maxim that correlation is not causation. It is perfectly possible to have a very strong correlation between two variables which are not causally connected at all. The classic example is that of barometers and rain. Suppose I have a particular barometer. I might well establish by careful observation that the barometer’s reading falling to a low level is strongly correlated with rain occurring. Obviously, however, the barometer’s reading is not a cause of rain. Rain is caused by a fall in atmospheric pressure, and such a fall also causes the barometer’s reading to fall. The barometer’s reading and the rain have a common cause, and this is why they are correlated, even though the barometer’s reading has no causal influence on the rain.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:56:22.

Copyright © 2018. Routledge. All rights reserved.

Types of evidence  113

FIGURE 6.3 

Source: Keys, 1970, Figures XVII. 3 and 4, p. 172.

Generalizing from this example, suppose that two variables A and B are strongly correlated. This could indicate that there is a causal link between them, but this is not necessarily the case. The principal cause of B might be some other variable C, and A might have no causal influence on B, or, at best, might have a very weak causal influence. This could happen, as in the barometer/rain example, if C is a common cause of both A and B, though there could be other explanations of the correlation of A and B. In this situation C is known as a confounder. It confounds the claim that the correlation between A and B is causal in character. So, the weakness of the kind of observational statistical evidence which Keys presents in his seven countries study is that it is liable to confounding. I will consider in the next section how this problem of possible confounders might be dealt with.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:56:22.

FIGURE 6.4 

Copyright © 2018. Routledge. All rights reserved.

Source: Keys, 1970, Figure XVIII. 6, p. 174.

FIGURE 6.5 

Source: Keys, 1970, Figure XVII. 8, p. 176.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:56:22.

Types of evidence  115

Copyright © 2018. Routledge. All rights reserved.

6.3 The problem of confounders, and the disconfirmation of causal hypotheses It is difficult to illustrate the problem of confounding using the data from all the seven countries; but the nature of the problem can be shown conveniently by confining ourselves to the results from just two countries, namely Japan and the USA. There were two cohorts from Japan. One was from the farming village of Tanushimaru and had 509 members. The other was from the fishing village of Ushibuka and had 504 members. Both cohorts were located in Kyushu in Southern Japan. There was only one large cohort from the USA consisting of 2,571 railroad men. The various statistics for the five-year period were as follows. The percentage of calories from saturated fats was 3% for Japan and 17–18% for USA. The USA figure was thus between 5.7 and six times that of Japan. (Keys, 1970, p. 168). The median serum cholesterol levels were 170 mg/dl in Tanushimaru, the farming village, and 140 mg/dl in Ushibuka, the fishing village, giving an average of 155 mg/dl. (Keys, 1970, p. 111) The median serum cholesterol level of the USA cohort was 237 mg/dl. (Keys, 1970, p. 54). So, the USA figure was 82 mg/dl higher than that of Japan. To measure the incidence of coronary heart disease in the two countries, I will simply use the death rates from CHD in the five years (see Keys, 1970, p. 7). Among the Japanese cohorts there were five deaths from CHD, giving a death rate per 1,000 of 4.9. In the USA cohort, there were 62 deaths from CHD, giving a death rate per 1,000 of 24. The U.S.A death rate from CHD was thus 4.9 times that of Japan. These figures certainly suggest a causal link between dietary saturated fat and coronary heart disease, but could there be an alternative explanation? The USA railroad men were all white (Caucasian). Could the Japanese have a group of genes which protected them from CHD, and which was lacking among Caucasians? The lower death rates from CHD among the Japanese might be nothing to do with their diet, and be the result rather of their genes. To put it another way, there might be a genetic confounder of the hypothetical causal link between dietary saturated fat and CHD. In fact, it turns out to be quite easy to test out the hypothesis of a genetic confounder against data, and this was done by Keys and his colleagues. If there is a genetic confounder, then groups of Japanese who start eating a typical American diet will have an incidence of CHD much lower than that of white Americans, because they will be protected by their anti-CHD genes. If on the other hand, there is no genetic confounder, then groups of Japanese who eat a typical American diet will have an incidence of CHD much the same as that of white Americans. All that needs to be done is to compare data on incidence of CHD among Japanese living in Japan and eating the traditional Japanese diet with that among Japanese who have emigrated to the USA and started eating a typical American diet. Keys et al. (1958) carried out an investigation of this sort by comparing Japanese in Japan with Japanese in Hawaii and in California. The overall finding was that death rate from CHD among the Japanese in California

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:56:22.

116  Types of evidence

was more or less the same as that of white Californians, and much higher than that of Japanese in Japan. The death rate from CHD for Japanese in Hawaii was intermediate between the two values. The observations indicated that as Japanese, who had emigrated, became increasingly Americanized and adopted the American diet, so their death rate from CHD rose to become more or less the same as that of white Americans. There was no indication that the Japanese had any special genetic protection against CHD. These findings were confirmed by a later investigation along similar lines (Robertson et al., 1977). In effect the hypothesis that there might be a genetic confounder was strongly disconfirmed. This example suggests a general strategy for dealing with the problem of confounding1. Suppose there is a strong correlation between A and B, and we think that this might indicate that A causes B. We should then try to think of as many possible confounders C as we can, and in each case test whether the confounder C causes B. If the hypothesis that C causes B is disconfirmed for each confounder C considered, then we can conclude that the link between A and B may indeed be causal in character. What makes this strategy viable is that it is much easier to disconfirm or even refute a claim that A causes B than it is to confirm that A causes B. The situation here is the opposite of what Popper claims in a famous passage where he writes (1963, p. 36):

Copyright © 2018. Routledge. All rights reserved.

It is easy to obtain confirmations … for nearly every theory – if we look for confirmations. In the case of causal theories of the form A causes B, it is in fact easier to find disconfirmations than confirmations. From “A causes B” it can often be deduced that, in specified circumstances, A should be strongly correlated with B. If frequency data is available, we can easily check whether A and B are correlated, and, if it turns out, that they are not correlated, or correlated to only a very small extent, then we will have obtained a disconfirmation of A causes B. Suppose, however, we observe that A is strongly correlated with B, this on its own does not confirm that A causes B, because of the maxim that correlation is not the same as causation. In order to confirm that A causes B, we need to provide some evidence in addition to that of correlation. In order to disconfirm that A causes B, the evidence of correlation is often sufficient. Hence in the case of causal hypotheses disconfirmations are easier to obtain than confirmations. We can illustrate this point further by mentioning two further disconfirmations which Keys et al. make on the basis of their data from the seven countries study. The first of these hypotheses is that the percentage of calories in the diet from proteins is a causal factor in coronary heart disease. As we mentioned in section 5.4, Alexander Ignatowski, Anitschow’s colleague and predecessor, had held that consumption of protein was harmful and caused hardening of the arteries. Similar views were held by others as well. This is probably why Keys et al. (1970) examine what their data show on this question. Their results are given in Figure 6.6.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:56:22.

Types of evidence  117

FIGURE 6.6 

Copyright © 2018. Routledge. All rights reserved.

Source: Keys, 1970, Figure XVII. 10, p. 178. The letters have the same meaning as in Figures 6.2 to 6.5, except that, in Figure 6.6, H seems to be a misprint for N = Zutphen.

This shows the age-standardized CHD incidence, measured by all diagnoses of CHD among men CHD-free at entry, against the percentage of calories from proteins in their diet. As can be seen the correlation is very low (0.14), and there is no sign of a linear relation between the variables. A contrast is provided by the result when CHD incidence is plotted against the percentage of calories from saturated fats in their diet, which was shown in Figure 6.5. Here the correlation is 0.84 and there is a clear sign of a relationship between the variables. These results strongly disconfirm the claim that the percentage of calories from protein in the diet is a causal factor for coronary heart disease. The second hypothesis was that a sedentary life, a lack of physical fitness, and the stress of existence in the big city were causal factors for coronary heart disease among American business men. This was a popular theory in Keys’ time, and has by no means disappeared today. However, as we saw, in section 6.1, Keys produced evidence, which strongly disconfirmed this hypothesis, from his study of the East Finns. The cohort from East Finland engaged in heavy manual work, and were lean and physically fit as a result. They also lived in the tranquil countryside. Yet their rate of coronary heart disease was the highest of all the cohorts in the seven countries study. This seems to me an important result which is not widely known or considered. Many believe that strenuous exercise resulting in a lean and muscular physique is the road to health. The problem is that those who follow this road focus too much on the external body and do not devote enough, if any, thought to what is going on inside the body, and in particular in the arterial system. For long term health, it is just as important, if not more important, to keep the arterial system fit. As the case of the East Finns shows, it is perfectly possible to have a very fit external body, and a degenerating arterial system.Those who seek to remain healthy in the longer

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:56:22.

118  Types of evidence

term should think as much, if not more, about keeping their arterial system in good shape, as about keeping their external body in good shape.

Copyright © 2018. Routledge. All rights reserved.

6.4 Some further general points regarding the seven countries study Before leaving the fascinating seven countries study, I would like to make a few concluding points about it. It has already been remarked that most of the cohorts chosen were simply all the men of the specified age (40 to 59) in the locality selected. This meant not only that all the cohorts were large in size, each containing more than 500 men, but also that there was no possibility of a biased sample arising from the way in which the cohort was selected from the eligible members of the population under consideration. Another important consideration is the time factor. I have discussed the results after five years, which were written up in Keys (1970). However, this was by no means the end of the seven countries study. On the contrary, it continued with regular investigation of the situation of the cohorts every five years. The results after ten years were written up in detail in Keys (1980). A couple of examples will show that these results were quite similar to those obtained after five years. After five years the correlation between the median cholesterol level and the percentage of deaths from CHD and non-fatal myocardial infarctions was 0.76 (see Figure 6.3). After ten years the correlation between the median cholesterol level and the deaths from CHD per 10,000 men was 0.8 (Keys, 1980, p. 122). After five years the correlation between the percentage of calories in the diet from saturated fats and the percentage of deaths from CHD and non-fatal myocardial infarctions was 0.84 (see Figure 6.4). After ten years the correlation between percentage of diet calories from saturated fats and the deaths from CHD per 10,000 men was 0.84 (Keys, 1980, p. 252). Some results after 25 years are to be found in Menotti et al. (1999). The seven countries study has even continued 50 years after it started in 1958. The later results have not significantly altered the picture presented by the results after five years. Still the continuation of the study for such a long time is undoubtedly a source of strength, and shows that the results obtained are robust in character. The main result of the seven countries study is that the amount of saturated fat in the diet is a principal cause, perhaps the principal cause, of heart disease. However, Keys never maintained that high saturated fat intake, and the consequent high blood cholesterol level, were the only causes of coronary heart disease. Quite to the contrary he explicitly states, on the basis of some of the data, that there must be other causal factors for CHD. In this connection Keys was particularly impressed by the results from the Cretan cohort. He writes regarding the concentration of cholesterol in the blood serum (1970, p. 99): The general level of that variable is clearly far more favorable for the Greeks than for the Americans and even more strikingly better than for the Finns. Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:56:22.

Types of evidence  119

Copyright © 2018. Routledge. All rights reserved.

So that the variable might seem to be adequate to explain the general gradient in CHD incidence from the Greeks to the Americans and the Finns. Yet why should the Greeks have less CHD than their rural counterparts in six other cohorts, the men in Croatia, Serbia, and Italy? The general level of cholesterol in the blood serum in those groups is as low as, and in some of the groups lower than, that in Crete and Corfu. Moreover, if serum cholesterol is of such crucial importance, why should the incidence of CHD be so much lower – and with statistical significance too – in the Cretans than in the men of Corfu? Note, too, that blood pressure tends to be, if anything, lower in Corfu than in Crete. The conclusion seems to be forced that the risk factors so far identified … by no means provide full explanation of differences between some populations in the incidence of CHD. The Cretan cohort is indeed something of an anomaly in the seven countries study. Despite having both an intake of saturated fat and blood cholesterol level much higher than the Japanese cohorts, the Cretan cohort had a lower incidence of CHD than either Japanese cohort. Its incidence of CHD was also much lower than that of the Italian cohorts from Crevalcore and Montegiorgio, who had comparable intakes of saturated fat and comparable blood cholesterol levels. The Cretan anomaly held not just after five years, but also in subsequent phases of the seven countries study. Keys writes (1983, p. 10): “when the 10-year follow-up of the Cretans examined in 1960 was completed; they had the lowest death rate of the 16 cohorts. Follow-up of the Cretans in 1979 again confirmed their longevity.” Menotti et al. (1999, p. 509) gives some figures for age standardized death rates from CHD per 1,000 for the 25-year follow-up. Once again, the Cretan cohort on 25 was the lowest. The two Japanese cohorts, both of which had lower intake of saturated fat and lower blood cholesterol levels than the Cretan cohort, had slightly higher death rates from CHD. Tanushimaru was 30 and Ushibuka was 36. Two of the Italian cohorts which had saturated fat and blood cholesterol levels comparable to the Cretan cohort had much higher death rates from CHD. Crevalcore was 93 and Montegiorgio 60. These results show the value of continuing an investigation over many years. The Cretan anomaly was not just a statistical fluctuation as might have been supposed after five years, but was a robust result which continued over 25 years. That concludes my account of the seven countries study as an example of observational statistical evidence. In the next section, I will consider the last type of evidence in my classification, namely interventional statistical evidence.

6.5 Examples of interventional statistical evidence: Some clinical trials 干预统计证据的来源：临床试验 Interventional statistical evidence comes from clinical trials. These are carried out using a group of volunteers, who may be patients, an intervention is made on some members of the group, and its effects are studied. Most clinical trials are

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:56:22.

120  Types of evidence

randomized controlled trials (RCTs). Here, typically a control group is chosen from the group of volunteers by random selection. The remaining volunteers could be called the “experimental group”. Some intervention is made on the experimental group, but this intervention is not made on the control group. The outcomes for the experimental group are then compared to the outcomes for the control group. Despite the ubiquity of the RCT, there are other designs for clinical trials, and my first example will be one of these other designs, known as a cross-over study. In this design, the effect of an intervention on a person during one-time period is compared to the effect of non-intervention on the same person in a different time period. Keys and his colleagues carried out a cross-over clinical trial to study the effects of the consumption of various types of fat on blood cholesterol levels. This trial is described in Keys et al. (1957, p. 960), and I will now describe the procedure and results on the basis of this paper. The trial took place at Hastings State Hospital in Minnesota. The participants were 66 men between 32 and 56, except for one man who was 62. According to Keys et al., 1957, p. 960:

Copyright © 2018. Routledge. All rights reserved.

The men at Hastings were “stabilised” schizophrenics judged, on the basis of extensive examinations and tests, to be physically and metabolically normal. These men did not volunteer to take part in the trials. Instead (Keys et al., 1957, p. 960): “Permission for each subject to participate in these trials was obtained in writing from his nearest relative.” In his reminiscences, Keys observes (1990, p. 289): “It should be mentioned that our … experiments would now be illegal; current laws forbid such involuntary experiments.” The men being in a mental hospital could presumably be carefully monitored to make sure they stuck to the diet, which was assigned to them. The diets all had fixed adequate amounts of calories, proteins, and vitamins, but varied in the amounts and types of fat which they contained. According to Keys et al. (1957, p. 959): Most of the dietary comparisons were made from both forward and backward dietary changes – i.e., changing from diet X to diet Y and vice versa. Standard periods on each diet were from two to nine weeks, usually four weeks. In all the experiments the experimental diets were preceded by at least four weeks’ standardisation on fixed “normal” diets. The general result obtained was as follows (Keys et al., 1957, p. 962): A highly significant change in the serum-cholesterol level takes place within a week after a change in diet fat; by the end of the second week a relative plateau is reached, and no further significant change can be observed within the next month or two.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:56:22.

Types of evidence  121

Keys et al. illustrate this by an example given in Table V, p. 962 of their 1957 work. This shows the result of the change of 14 of the participants from the house diet to a low-fat diet. At the start the mean serum-cholesterol level of the 14 was 220 mg/dl. After two weeks this had reduced on average by 37 mg/ dl. These averages, however, conceal very considerable individual variations. Serum-cholesterol levels on the house diet varied from 291 mg/dl (the highest) to 166 mg/dl (the lowest). The biggest reduction in serum-cholesterol level was that of number 7, who went down 75 mg/dl from 238 mg/dl to 163 mg/dl. The serum-cholesterol level was reduced for all the participants, except number 12, who started at the low level of 173 mg/dl, which then increased slightly to 175 mg/dl. In the trials, Keys at al. divided fats into S = saturated fats (mainly animal fats such as meat, butter, cheese, cream), M = monounsaturated fats, such as olive oil, and P = polyunsaturated fats such as corn oil. Their main finding was that saturated fats increase blood cholesterol levels, polyunsaturated fats decrease blood cholesterol levels, by about half the amount, while monounsaturated fats leave blood cholesterol largely unchanged. Keys et al even used statistical analysis of their data to derive an equation relating change in blood cholesterol level Δ Chol to ΔS and ΔP. The equation (Keys et al., 1957, p. 963) is:

Copyright © 2018. Routledge. All rights reserved.



D Chol = 2.74 DS - 1.31 DP (6.1)

Keys et al. (1957) have therefore presented evidence from the results of clinical trials about how blood cholesterol levels change as the quantity and type of dietary fat changes. But how convincing is this evidence? What are its strengths and weaknesses? These are the questions to which we must turn next. The evidence obviously has considerable strengths. The clinical setting of the trials enabled the relevant factors to be carefully controlled and measured. The cross-over design eliminates a number of possible confounders. However, these clinical trials, like most clinical trials do have weaknesses as well as strengths. I will draw attention to two such weaknesses, namely (i) time limitations, and (ii) sample limitations. 缺点：时间限制和样本限制 Practical considerations normally limit the time during which the clinical trial must be carried out. In this case, the limitation is obvious. The experimental diets were eaten for an average of about four weeks, and not longer than nine weeks. The effect of the diet in this time period was noted, but how much does that tell us about the effect of a similar diet for a period of years? We have learnt about short-term effects, but not about the long-term effects which are surely much more relevant. Let me turn to the second general weakness of clinical trials, including the Hastings trial. This is concerned with sample limitations. Typically, a clinical trial will be used to test out some causal law in medicine, and, if the trial results in confirmation of the law, it will be used in the treatment of a population of patients. This procedure is only satisfactory, however, if the sample of participants

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:56:22.

122  Types of evidence

in the trial is a reasonable representation of the population of patients to whom the causal law is later applied. Yet often the trial sample is not a good representation of the patient population. This is what I call a sample limitation. Clarke and Russo (2017) give a very striking example of this kind of limitation, based on the results reported in Fortin et al. (2006). Fortin et al. considered a database of 980 primary care patients in Canada, and a particular condition, hypertension (high blood pressure). They found that a very large percentage of patients in their database who had hypertension also had some other condition (a comorbidity). Thus, for example, 54.5% of these patients had hyperlipidemia (high blood cholesterol level), 40.1% had heart disease, 34.7% urinary tract or kidney disease, and 23.6% diabetes (Fortin et al., 2006, p. 105). Many of the patients with hypertension had several of these conditions. Now, as Clarke and Russo say (2017, p. 304):

Copyright © 2018. Routledge. All rights reserved.

the presence or absence of other diseases is likely to change the effectiveness of an intervention. For example, hypertension in those with hyperlipidemia carries a much greater risk of adverse cardiovascular outcomes than without. … the aim of prescribing antihypertensive treatments is not to reduce blood pressure per se, but to reduce the incidence of these adverse cardiovascular events … responses to a particular treatment may be very different in individuals with isolated hypertension compared with individuals with hypertension secondary to renal disease, largely because the mechanisms by which these two kinds of hypertension come about are different. In the light of all this, how much guidance for treatment is given by the results of clinical trials for antihypertensives? Fortin et al. (2006, pp. 105–6) examine five clinical trials chosen at random. For illustration purposes, I will consider only one of these. In this trial, possible participants were excluded who suffered from heart disease, renal insufficiency, poorly controlled hyperlipidemia or diabetes mellitus, or diabetes requiring insulin. Thus, this trial was not relevant to a large number of patients in the database. However, even for those in the database who did not contravene these exclusion criteria, no less than 89% suffered some form of comorbidity. There is no information on how many (if any) patients with these kinds of comorbidities were included in the clinical trial. In short, the relevance of the results of this clinical trial to the treatment of patients in the database is doubtful. The problem clearly is that the sample of participants who took part in the clinical trial is not a good representative of the population of primary care patients requiring treatment. The clinical trial of Keys et al. (1957) had a cross-over design, but the vast majority of clinical trials nowadays are randomized controlled trials (RCTs). It is time to say something about this standard design of clinical trial. RCTs, like other clinical trials, have the two weaknesses just mentioned, namely (i) time limitations, and (ii) sample limitations; and it is important to bear this in mind. However, they do have considerable strengths as well. Randomization was introduced by

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:56:22.

Types of evidence  123

Ronald Fisher in the 1920s in the context of agricultural trials, and it was then extended to medicine in the late 1940s and early 1950s by Austin Bradford Hill. Randomization has a number of advantages, which I will now describe. Teira argues that randomization helps to make RCTs impartial. As he says (2013, p. 413):

Copyright © 2018. Routledge. All rights reserved.

an experiment will be impartial if it incorporates methodological devices preventing the experimenter from manipulating the results at will. … If, for instance, a researcher wants to increase the chances of an experimental drug to succeed in a trial … allocating the therapy to the healthier group of patients in the trial will do it. As a debiasing method, randomization protects the experimental outcome from manipulation in the allocation of treatments. Teira cites some authors who have claimed that the alleged impartiality of RCTs was a mere rhetorical device, but he disputes this view. Using the example of early British clinical trials, including the streptomycin trial, he argues, on the basis of his account of impartiality, that the impartiality of early British RCTs was real, and played a role in getting RCTs accepted as regulatory yardsticks. I will consider these early British trials of streptomycin and other anti-­t uberculosis agents in more detail in sections 9.2, 9.3, and 9.4. Another great advantage of randomization is that it provides a second strategy for dealing with the problem of confounding. Suppose that A is strongly correlated to B, and we think A may cause B, but are doubtful whether there may instead be a confounder C which causes B. The first strategy for dealing with this doubt was to think of all the possible confounders C that we can, and then test them out against the data to see if we can refute the hypothesis that C causes B. If we can indeed refute all such hypotheses, this provides good evidence that A causes B. However, it could now be objected that this is all very well, but that there may still be a confounder we have not thought of. Randomization provides a technique for dealing with unknown confounders. Suppose there is a confounder C, which, unknown to us, causes both A and B, and that A does not really have any causal influence on B. In our trial, let us divide the participants into the control group, and the intervention group randomly by tossing a fair coin. Then if a particular participant has C, he or she is just as likely to be in the control group as the intervention group. Since C is a cause of B, and A is not a cause of B, then the difference between the intervention group and the control group will disappear. If there continues to be a difference between the control group and the intervention group, we can conclude that the intervention variable really is a cause and that confounders have been dealt with. Of course, this argument is not entirely conclusive, since random choice by tossing a fair coin only produces an equal distribution in the limit. In a particular trial, an unbalanced distribution is quite possible. Still randomization must be regarded as an effective tool, and its use constitutes the second strategy for dealing with confounding.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:56:22.

124  Types of evidence

We can illustrate these points by considering a randomized controlled trial carried out to investigate the relationship between the dietary intake of saturated versus unsaturated fat and the development of atherosclerosis. This clinical trial was carried out by Seymour Dayton and his colleagues. A preliminary report on the trial and its results was given in Dayton et al. (1968), and a more extensive report in Dayton et al. (1969). The trial was partly inspired by some of Keys’ work. At the beginning of Dayton et al (1969), the authors cite Keys et al. (1958), and some other works in the area, and go on to say (p. 1): As information along these lines developed, many investigators proposed that habitual intake of abundant saturated fat of animal origin was a causative factor in atherosclerosis and coronary heart disease. This was, however, only hypothesis; epidemiological relationships do not prove causality. Indeed, as we know, correlation is not the same as causation because of the problem of confounding. Hence the desirability of carrying out a randomized controlled trial, since randomization is a strategy for dealing with confounding. The nature of the trial is described by Dayton et al. (1968, p. 1060) as follows:

Copyright © 2018. Routledge. All rights reserved.

Male veterans, aged 54–88 years (mean 65.5 years), living in the domiciliary unit of the Los Angeles Veterans Administration Center, were asked to volunteer for this study. A total of 1095 men volunteered, but exclusions and dropouts during the preliminary phases reduced this number so that 846 men were ultimately randomised. Of these, 422 received the conventional control diet, and 424 the experimental diet. The control diet was a standard American diet containing 40% fat calories, mostly of animal origin. The experimental diet was designed to be as similar to the standard diet as possible, except that about two thirds of the animal fat calories were replaced by vegetable oils, such as corn, cottonseed, safflower, and soybean. In effect, the total fat content of the diet was about the same, but two thirds of the saturated fat had been replaced by unsaturated fat. Two other small changes were that the mean iodine value was increased from 53.5 in the control to 102.4 in the experimental diet; while the cholesterol content was reduced from 653 mg/day in the control to 365 mg/day in the experimental diet. The latter change was brought about by having less eggs – the usual source of dietary cholesterol. In addition to being randomized, the trial was double-blind. Indeed, Dayton et al. (1969, p. 55) say: “The study described herein is, as far as we know, the first application of the double-blind technique to a nutritional problem.” The double blinding for the participants was achieved in a rather ingenious fashion, described as follows (Dayton et al, 1969, p. 4): We planned … to establish double-blind conditions by providing a control diet which, like the experimental diet, appeared to be a modification of the Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:56:22.

Types of evidence  125

Copyright © 2018. Routledge. All rights reserved.

regular institutional fare. Therefore, both groups of participants were told at the outset that their diets would differ from the regular diet but would resemble ordinary institutional food. Opportunities for critical comparison of the diets were minimized by serving them in separate dining rooms. The participants were examined at regular intervals, and the values of various variables recorded. However, the physicians carrying out these examinations were also “blinded” by not being told which diet the participant was on. Recruitment started in the summer of 1959, and the trial was concluded with a final re-examination of a participant eight years after he had been recruited. No further recordings for the purposes of the trial were made after 15 January 1968. The serum cholesterol level of the two groups was about the same at the beginning (234 mg/dl for the control, and 233 mg/dl for the experimental). However, as a result of the change of diet, the serum cholesterol level of the experimental group was lower than of the control group throughout the trial by an average of about 12.7%. A feature of this trial was that it recorded all atherosclerotic events and not just coronary heart disease. The events recorded were those established by hard criteria, and, in addition to myocardial infarction, and sudden death due to coronary heart disease, they included definite cerebral infarction (i.e. stroke), ruptured aneurysm, and even amputation necessitated by atherosclerosis of arteries in the leg. This choice of outcomes is more logical than focussing exclusively on coronary heart disease, since the other outcomes have the same underlying cause, namely atherosclerosis. The results were as follows (Dayton et al., 1969, pp. 26–27). In the control group 70 men died of an atherosclerotic event, as opposed to 48 in the experimental group. The difference was significant at the 5% level. In addition, there were some nonfatal atherosclerotic events established by hard criteria. 96 men in the control group were affected by a fatal or nonfatal atherosclerotic event, as opposed to 66 in the experimental group. The difference was significant at the 1% level. These results clearly give empirical confirmation to the causal hypotheses (CH 6.2): 1. High dietary intake of saturated fat → 2. Raised level of serum cholesterol → 3. Coronary heart disease and other atherosclerotic events. I will next try to analyse the strengths and weaknesses of this interventional statistical evidence as compared to the observational statistical evidence provided for similar causal hypotheses by the seven countries study. The seven countries study was observational in character, and so strictly yielded only correlation not causation. The study of the Los Angeles veterans,

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:56:22.

Copyright © 2018. Routledge. All rights reserved.

126  Types of evidence

by contrast, was a randomized, and even double blind, clinical trial, and so had a built-in remedy against confounding. This is the great strength of the Los Angeles veterans trial. In favour of the seven countries study, however, it should be added that Keys et al. did consider at least one possible confounder (racial/ genetic factors), and showed that it was refuted by further observational evidence. Let us now turn to the weaknesses of the evidence from the Los Angeles veterans trial. Dayton et al. are quite self-critical and have a section (1969, p. 56) discussing weaknesses of the trial. One rather specific point they mention is that, since the veterans did not take all their meals at the institution, there was not very strict adherence to the diet. In fact, Dayton et al. estimated the adherence rate as approximately 50% overall. In addition, their RCT, like any other clinical trial, has the two weaknesses mentioned earlier, namely (i) time limitations, and (ii) sample limitations. The veterans trial continued for eight years, which is a very long time for a clinical trial. Even so, it is much shorter than the seven countries study which, as we mentioned earlier, has continued for over 50 years. There is another sense in which the seven countries study has a time advantage. The veterans, who started the experimental diet low in saturated fat, had up to that point in their lives eaten the standard American diet. Thus, they had had a diet relatively high in saturated fat for at least 54 years, as against a maximum of eight years on the diet low in saturated fat. By contrast, the men in the seven countries study had eaten more or less the same diet all their lives. Thus, the seven countries study does give a better account of the long term effects of diets high or low in saturated fats. Turning now to sample limitations, the number involved in the veterans study was 846, which was much smaller than the 12,770 who took part in the seven countries study. In addition, the sample in the veterans study is not very representative of the target population to which the results might be applied. This target population is adult males in a country prone to atherosclerosis and coronary heart disease. The hope would be to lower the rate of these diseases by persuading the target population to change its dietary habits. However, our knowledge of the development of atherosclerosis suggests that such a dietary intervention should take place early, say in an individual’s 20s. By contrast, the veterans’ study had a sample of elderly men between 54 and 88. To judge the efficacy of a dietary intervention in the target population, it would have been better to have a sample of much younger men. Indeed, it is rather surprising that dietary intervention in a sample of such elderly men had such a striking effect. Moreover, Dayton et al. do say (1969, p. 60): Stratification of the data by age demonstrated that most of the prophylactic affect occurred in the younger half of the study population, less than 65.5 years old at the start of the study. Taking the evidence of the veterans study together with that of the seven countries study gives a striking instance of the principle of strength through

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:56:22.

Types of evidence  127

combining. The weakness of the seven countries study was concerned with the problem of confounding, while the randomization used in the veterans study is one of the standard ways of dealing with the problem. The weaknesses of the veterans trial were those characteristic of clinical trials in general, namely time limitations and sample limitations. The sample was relatively small, and the participants were too elderly to be representative of the target population. However, both these weaknesses were overcome in the seven countries study. The results of the two studies empirically confirmed broadly the same causal hypotheses. So, taken together, they give very strong evidence for these hypotheses. So far, however, I have not considered the most important example of the evidential principle of strength through combining, namely the combining of evidence of mechanism with statistical evidence from human populations. I will deal with this in the next chapter.

Note

Copyright © 2018. Routledge. All rights reserved.

1 This is not, however, the only strategy for dealing with the problem of confounding. A second strategy will be described in section 6.5, and a third in section 7.2.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:56:22.

7 COMBINING STATISTICAL EVIDENCE WITH EVIDENCE OF MECHANISM

7.1 Combining the results of Anitschkow, Dayton et al., and Keys et al.

Copyright © 2018. Routledge. All rights reserved.

We must next examine how the statistical evidence, produced by Keys et al. in the observational seven countries study, and by Dayton et al. in the Los Angeles veterans randomized controlled trial, can be combined with the evidence of mechanism produced by Anitschkow.To do so, we have first to merge the mechanism produced by Anitschkow, given earlier as M5.1, with causal hypotheses proposed by Keys (CH6.1) and Dayton et al. (CH6.2). To carry out this merger, we have to correct Anitschkow’s mechanism in the light of Ancel Keys’ criticism that cholesterol levels in humans are raised mainly by dietary saturated fat, while dietary cholesterol plays a minor part. This means in effect omitting stage 1 in M5.1. If we do this and make a few minor adjustments to CH6.1 and CH6.2, we get the following (CH7.1): 1. High levels of saturated fat in the diet →  2. Cholesterol content of blood to become raised →  3. Infiltration of some parts of the arterial wall by lipids (mainly cholesterol esters) →  4. Lipids to be absorbed by macrophages, which become foam cells →  5. Full development of atherosclerotic plaques →  6. Coronary heart disease and other atherosclerotic events. Here the causal steps 1 →  2 →  6 are empirically confirmed by both the observational evidence of the seven countries study, and by the Los Angeles veterans randomized controlled trial. The causal steps 2 →  3 →  4 →  5 can be thought of as the mechanism linking 1 to 6. This mechanism is supported by the observational evidence from autopsies, and by the interventional evidence of Anitschkow’s work on experimentally induced atherosclerosis in rabbits.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:56:34.

Combining statistical evidence  129 TABLE 7.1  Two by two classification of types of evidence in medicine

Copyright © 2018. Routledge. All rights reserved.

Statistical Evidence Evidence of Mechanism

Observational

Interventional

Epidemiological surveys Autopsies

Clinical trials Laboratory experiments on animals, tissues, cells, etc.

As we remarked earlier, the seven countries study produced very high correlations regarding the steps 1 →  2 →  6, where stage 6 was in this case limited to coronary heart disease. The weakness of this evidence was that correlation is not causation, and there is always the problem of confounding. However, in CH7.1, stages 1 and 6 are now linked by a mechanism 2 →  3 →  4 →  5, and this in turn is confirmed by experimental evidence which effectively eliminates the possibility of confounding since it involves interventions. Thus, combining Keys et al.’s observational evidence with the experimental evidence of Anitschkow greatly strengthens the force of either type of evidence taken on its own. This is an example of a third strategy for dealing with confounding. Suppose we have found a strong correlation between A and B, and suspect that they may be causally linked. If we can then find a mechanism M linking A and B, and also produce experimental evidence which confirms M, then by combining the statistical evidence of the correlation of A and B, and the experimental evidence for the mechanism M, we have produced strong evidence for the causal hypothesis A →  M →  B. So, our third strategy for dealing with confounding consists in postulating a linking mechanism and finding evidence that it really exists. In our analysis of evidence, we have introduced four different types of evidence which can be conveniently displayed in the two by two table which is shown in Table 7.1. It is worth noting that CH7.1 is supported by all four types of evidence. Earlier (section 1.2), I formulated the following principle of interventional evidence. A causal law cannot be taken as established unless it has been confirmed by some interventional evidence. This principle is strongly supported by the action-related theory of causality given in Part I of this book. However, it may not apply in all cases of causality, since there are some areas of the social sciences where it is not possible to gather interventional evidence. In medicine, however, interventional evidence, as we see from Table 7.1, is supplied by clinical trials, and by laboratory experiments on animals, cells, etc. Thus, as the examples we have given have shown, there is no difficulty about collecting interventional evidence in medicine, and the principle of interventional evidence should be taken as holding for all cases of causality in medicine.

7.2  Strength through combining In this section, I will discuss various points connected with the evidential principle of strength through combining. Let me first state this principle in a case where Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:56:34.

130  Combining statistical evidence

we are considering a causal hypothesis H say which is confirmed by two different types of evidence, α  and β  say. Each type of evidence will naturally have strengths and weaknesses. Suppose, however, that the weaknesses of evidence of type α  are compensated for by the strengths of evidence of type β , and vice versa. Then the combination of evidence of type α  with evidence of type β  will give an overall confirmation to the hypothesis H much greater than would be obtained by a comparable amount of evidence just of type α  or a comparable amount of evidence just of type β . Naturally this principle of strength through combining can be generalized to more than two different types of evidence. The evidential principle of strength through combining gives rise to a methodological principle, namely the following. When we are trying to assess some causal claim against evidence, we should look for evidence of different types rather than concentrating exclusively on one type of evidence. Illari seems to have been the first to formulate the principle of strength through combining, though she does not use this name (2011). Speaking of the value of combining two different types of evidence she says (2011, p. 146): “each addresses the major weakness of the other evidence” (italics in original). In her analysis of evidence of mechanism, Illari points out a problem with mechanisms. We might establish that a mechanism exists, but because of the complexity of the human organism, there might be another mechanism which cancels the effect of the first mechanism. Illari illustrates this with the example of exercise and weight loss taken from Steel (2008, p. 68). She writes (2011, p. 146):

Copyright © 2018. Routledge. All rights reserved.

Increased exercise leads to more calories being burned, via well-known mechanisms. But it also leads to increased appetite. Until we investigate further, we do not know the overall effect of exercise on fatness or thinness, and we cannot claim either that exercise makes you fat or makes you thin. The operation of one mechanism might mask or hide the operation of the other. This example exposes a weakness of evidence of mechanism. Let us suppose that we have strong evidence that mechanism M1 exists. M1 may indeed exist, but perhaps a masking mechanism M 2 also exists, so that the effects of M1 are cancelled out. We need to eliminate this possibility, and this can be most easily done by using non-mechanistic evidence. Using the example of masking mechanisms, Illari formulates a principle similar to what has here been called the evidential principle of strength through combining. She updated her 2011 formulation for the 2014 joint paper. So, I will quote her views (from Clarke et al., 2014, p. 351): What this means is that evidence of mechanisms and evidence of correlation do not act independently, each suggesting separately that A does cause B. They integrate in a special way. … 

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:56:34.

Combining statistical evidence  131

1. Evidence of a correlation relation between A and B: Its problems are that of confounding and non-causal correlations. Its advantage is that it can reveal masking, and can help assess the net effect of a complex mechanism. 2. Evidence of a mechanism linking A and B: Its problem is masking, and being too complex to assess a net effect. Its advantage is that it can reveal confounding and non-causal  correlations. Evidence of a linking mechanism helps to show that the overall relationship between A and B is genuinely causal. But evidence of correlation helps to determine the net effect of a mechanism, and to show that it is not masked by further unknown mechanisms. Together, evidence of these two different things is very much stronger than evidence of one alone.

Copyright © 2018. Routledge. All rights reserved.

The example given in section 7.1, where A = High levels of saturated fat in the diet, and B = Coronary heart disease is an excellent illustration of this. In the course of discussions leading to the joint paper (Clarke et al., 2014), I suggested that there was a good analogy for this principle of combining evidence, namely reinforced concrete. Here are the details (p. 351): We can describe this situation by an analogy to reinforced concrete, which is formed by placing steel grids into concrete. Now most concrete mixes have high resistance to compressive stresses, but any appreciable tension (e.g., due to bending) will break the microscopic rigid lattice, resulting in cracking and separation of the concrete. Steel, however, has high strength in tension. So, if steel is placed in concrete to produce reinforced concrete, we get a composite material where the concrete resists the compression and the steel resists the tension. The combination of two different materials produces a material that is much stronger than either of its components. In the same way …  it is the combination of two different types of evidence which produces much stronger overall confirmation than would either type of evidence on its own. So far, I have emphasized the use of different types of evidence to strengthen the confirmation of a causal hypothesis, but this use can also be valuable in disconfirming or even refuting causal hypotheses. If a causal hypothesis is in fact false, its falsity may be brought to light more easily by one type of evidence than by another. So, in order to achieve a rigorous empirical testing of a causal hypothesis, it is always better to collect, and make use of, different types of evidence. This leads in turn to a criticism of the view that randomized controlled trials (RCTs) constitute “the gold standard” of evidence.1 The idea behind this slogan is that the evidence supplied by RCTs is much superior to other forms of evidence.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:56:34.

132  Combining statistical evidence

We should accept causal hypotheses, which are confirmed by RCTs, but should be less inclined to accept causal hypotheses which are only confirmed by “inferior” forms of evidence, such as epidemiological surveys. Now my criticism of this posi- 没有单一的证 tion is not based on denying that RCTs have great strengths. On the contrary, 据可以是“黄金 I have stressed that randomization, and careful experimental controls, have very 标准”。合并证 great value in overcoming the problem of confounding. But RCTs have weak- 据更好。 nesses as well as strengths. In particular, they suffer from time limitations, and sample limitations. Moreover, these weaknesses can be overcome by another type of evidence, such as that provided by epidemiological surveys. My claim is that there is no gold standard in evidence. We should not try to obtain strong empirical confirmation of a causal hypothesis by using just one type of evidence, which is claimed to be superior to all other types. Instead we should seek to provide empirical confirmation of the causal hypothesis from different types of evidence, which are such that the weaknesses of one type of evidence are compensated for by the strengths of another type. In fact, in the next chapter (8), I will give an example of a causal law which was established without using RCTs, while in the chapter after (9), I will give an example in which a RCT, taken on its own, gave the wrong result, but gave the correct result when combined with some evidence of mechanism. These examples will be considered in connection with another evidential principle, the RussoWilliamson thesis. This thesis, as I will show, is related to, but stronger than, the principle of strength through combining. As a consequence of its strength, the Russo-Williamson thesis is controversial, and has led to a lot of debate. However, exactly for this reason, its consideration is helpful in studying how evidence confirms or disconfirms causal hypotheses in medicine. Consequently, the next three chapters will be devoted to the pros and cons of the Russo-Williamson thesis.

Copyright © 2018. Routledge. All rights reserved.

Note 1 This view is held by many supporters of evidence-based medicine (EBM), as I will show in section 9.1, which discusses EBM.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:56:34.

8 THE RUSSO–WILLIAMSON THESIS (i) Effects of smoking on health

8.1  Statement of the Russo–Williamson thesis The Russo–Williamson thesis was first proposed by Russo and Williamson in their 2007 paper, where they write (pp. 158–9):

Copyright © 2018. Routledge. All rights reserved.

the health sciences infer causal relations from mixed evidence: on the one hand, mechanisms and theoretical knowledge, and, on the other, statistics and probabilities. …  To establish causal claims, scientists need the mutual support of mechanisms and dependencies. …  The idea is that probabilistic evidence needs to be accounted for by an underlying mechanism before the causal claim is established …  The general sense of this thesis is quite clear, but, like most philosophical theses, it can be formulated in a number of slightly different ways. Russo and Williamson speak in one place of “statistics and probabilities” and in another of “dependencies”. Here I will develop the thesis along the lines given earlier in my 2011 work. I will confine myself to “statistical evidence” understood as referring to human populations when we are dealing with human diseases. Similarly, I will use the phrase “evidence of mechanism”. The concepts of “statistical evidence” and “evidence of mechanism” as I will use them have been explained in Chapter 5, section 5.2, and illustrated in Chapter 7, Table 7.1. In terms of these concepts, the Russo–Williamson thesis can be stated as follows: A causal hypothesis in medicine can be established only by using both statistical evidence and evidence of mechanism. It is clear that this is a very strong thesis, and, in particular, a thesis which is stronger than the principle of strength through combining which was given earlier. According to the principle of strength through combining, if statistical evidence and evidence of mechanism both confirm a causal hypothesis, and are such that

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:57:04.

134  The Russo–Williamson thesis

the weaknesses of the statistical evidence are cancelled out by the strengths of the evidence of mechanism and vice versa, then combining the two types of evidence produces a great increase in the degree of empirical confirmation of the hypothesis in question. This is in agreement with what the Russo–Williamson thesis claims, but the principle of strength through combining does not rule out the possibility of a causal hypothesis becoming established through statistical evidence alone, if such evidence is very strong, or through evidence of mechanism alone.The Russo– Williamson thesis does rule out this possibility, and so it is a very strong claim. Perhaps it is even a bit too strong, and in fact, as we go along, I will introduce some qualifications to the version of the thesis just given. One such qualification will be given in section 8.3 of this chapter, and a further two in Chapter 10. However, these qualifications are relatively minor, and I will argue that the Russo–Williamson thesis is broadly correct, and a very useful evidential principle for medicine. Before developing the thesis further, however, it will be helpful to examine how it relates to other important historical views of the roles of statistical evidence and evidence of mechanism in medicine. This I will do in the next section.

8.2 Some different views concerning the roles of statistical evidence and evidence of mechanism in medicine

Copyright © 2018. Routledge. All rights reserved.

Let me begin with the views of Claude Bernard, expressed in his famous and influential essay of 1865: ‘An Introduction to the Study of Experimental Medicine’. Bernard’s aim was to turn medicine into a science by basing it on a scientific physiology. This in turn would be based on animal experiments of the kind which Bernard himself carried out. Thus, for Bernard, the key type of evidence in medicine was evidence of physiological mechanisms established by animal experiments. While stressing evidence of mechanism in this way, Bernard regarded statistical evidence as of no value whatever. He states this view very strongly and clearly as follows (1865, p. 138): never have statistics taught anything, and never can they teach anything about the nature of phenomena. I shall further apply what I have just said to all the statistics compiled with the object of learning the efficacy of certain remedies in curing diseases. Aside from our inability to enumerate the sick who recover of themselves in spite of a remedy, statistics teach absolutely nothing about the mode of action of medicine nor the mechanics of cure in those in whom the remedy may have taken effect. Bernard held statistics in low esteem because he believed that absolute determinism was the basis of science. As he says (1865, p. 136): Bernard不看好统计学因为他认为决定论是 科学的基础。

I acknowledge my inability to understand why results taken from statistics are called laws; for in my opinion scientific law can be based only on certainty, on absolute determinism, not on probability. Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:57:04.

The Russo–Williamson thesis  135

As his overall project was to turn medicine into a science, this meant eschewing the use of statistics. As he says (1865, p. 139): if based on statistics medicine can never be anything but a conjectural science; only by basing itself on experimental determinism can it become a true science, i.e., a sure science. I think of this idea as the pivot of experimental medicine …  It should be remembered that the use of statistics in medicine was a relatively recent innovation when Bernard was writing. Although there had been some suggestions in Britain in the 18th and early 19th centuries that statistics should be used in medicine, as far as France is concerned, Pierre-Charles-Alexandre Louis’s research on the efficacy of blood-letting is probably the first significant use of statistics in medicine. Louis published a paper on this subject in 1828 and a revised and extended version as a book in 1835.11 This was only three decades before Bernard’s essay. Clearly statistics was a new technique in medicine in 1865, and one which Bernard rejected in favour of determinism and animal experiments. It is interesting to look again at Koch’s postulates in the light of Bernard’s Koch一直使用决 定论的因果关 views of 1865. The postulates were formulated in the years 1878 and 1882, 13 系。 and 17 years after the appearance of Bernard’s influential essay. Moreover, as we have seen, Koch always employed deterministic causality. The question can then be raised of whether Koch’s postulates satisfy the Russo–Williamson thesis. Koch’s first two postulates in the version we gave in section 3.3 are as follows:

Copyright © 2018. Routledge. All rights reserved.

1. The micro-organisms must be shown to be present in all cases of the disease. 2. Their presence must be in such numbers and distribution that all the symptoms of the disease can be explained. Here postulate 1 gives statistical evidence. All patients suffering from the disease must be shown to have the micro-organisms, presumably by taking a blood sample and examining its contents. This gives statistical evidence in a human population. Postulate 2 gives evidence of mechanism, presumably obtained by autopsies, and animal experiments, where available. Thus, Koch’s postulates do seem to satisfy the Russo–Williamson thesis. However, a point should be noted. The statistical evidence here is supposed to produce a result of 100%, to be in effect completely deterministic in accordance with the requirements of Claude Bernard. Now, when he compared the cholera situation of Hamburg with that of Altona in 1892, Koch did use more genuinely statistical evidence. The difference between the two places was very striking, but fell short of being 100%. Koch relied on this statistical evidence to establish his case, and so, in his practice, recognized the value of statistical evidence. However, he never modified his postulates to allow them to include statistical evidence of this kind. Perhaps this was partly because statistical evidence was still at that time held to be a bit suspect, as a result of Claude Bernard’s attack.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:57:04.

Copyright © 2018. Routledge. All rights reserved.

136  The Russo–Williamson thesis

By the time of the rise of the evidence-based movement (EBM) in the 1990s, the pendulum had swung to the other extreme. I will discuss the EBM position in more detail in the next chapter, but, in general terms we can say that EBM advocates almost exclusive use of statistical evidence, and allows evidence of mechanism only a very subordinate place. The Russo–Williamson thesis (2007) represents a middle way between the extremes of Claude Bernard on the one hand, and EBM on the other. Its rehabilitation of evidence of mechanism against the strictures of EBM was influenced by the rise of the new mechanism in the philosophy of science community in the preceding decade, which we described in Chapter 4. Of course, there were other influential approaches, which embodied a more pluralistic attitude than that of Bernard on the one hand, and EBM on the other. One such is that of Sir Austin Bradford Hill, which he put forward in his 1965 paper, one hundred years after Bernard’s famous essay. This paper is analysed by Russo and Williamson (2007, pp. 161–2). Hill sets out nine viewpoints, which may support the claim that an association is causal in character. His idea is not that any of these viewpoints are necessary for establishing causality, but that all of them can yield useful evidence as to whether an association is causal. Now Hill’s nine viewpoints include both statistical evidence and evidence of mechanism. In connection with his first viewpoint, he says (1965, pp. 295–6): “prospective inquiries into smoking have shown that the death rate from cancer of the lung in cigarette smokers is nine to ten times the rate in non-smokers and the rate in heavy cigarette smokers is twenty to thirty times as great.” This is clearly statistical evidence. In connection with his seventh viewpoint, he speaks of (1965, p. 298): “the histopathological evidence from the bronchial epithelium of smokers and the isolation from cigarette smoke of factors carcinogenic for the skin of laboratory animals.” This is clearly evidence of mechanism. Moreover, his eighth viewpoint is Experiment, and laboratory experiments often provide evidence of mechanism. The Russo–Williamson thesis has been placed in a spectrum of views concerning the roles of statistical evidence and evidence of mechanism in medicine. I will now go on to consider the development of the thesis in the light of an example from the history of medicine, namely the investigations of the influence of smoking on health. In the next section, I will start this example by considering the question of smoking and lung cancer.

8.3  Smoking and lung cancer In 1950, Austin Bradford Hill began to suspect that smoking might be a cause of lung cancer, and in 1951 he started an epidemiological cohort study to test this hypothesis. He and Doll wrote at the end of October that year to all the doctors on the British Medical Register who were believed to be resident in the United Kingdom to ask them if they would participate in a survey concerning smoking. 34,440 agreed to take part and they were then followed for the next 40 years.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:57:04.

The Russo–Williamson thesis  137

Their smoking habits were monitored from time to time, and when they died the cause of death was noted. Reports on the results were published occasionally as the study progressed. Here I will quote some of the results to be found in Doll and Peto (1976). Peto had by this time replaced Hill in handling the survey. The 1976 paper deals with the mortality rates of the male doctors over the 20 years from 1 November 1951 to 31 October 1971. During that time, 10,072 of those who had originally agreed to participate in the survey had died, and 441 of these had died of lung cancer. Doll and Peto calculated the annual death rate for lung cancer per 100,000 men standardized for age. The results in various categories were as follows (1976, p. 1527): Non-smokers Smokers 1–14 gms tobacco per day 14–24 gms tobacco per day 25 gms tobacco per day or more

 10 104  52 106 224

Copyright © 2018. Routledge. All rights reserved.

(A cigarette is roughly equivalent to 1 gm of tobacco)

These results do indeed show a striking correlation between smoking and lung cancer. Smokers are on average more than 10 times more likely to die of lung cancer than non-smokers, and this figure rises to more than 22 times for heavy smokers who consume 25 gms or more of tobacco per day. The results are highly significant statistically. Doll and Peto said explicitly (p. 1535) that the excess mortality from cancer of the lung in cigarette smokers is caused by cigarette smoking, but were they justified in doing so? They have demonstrated a correlation, indeed a very striking one, between smoking and lung cancer. However, we know that correlation is not causation, and that not all correlations can be taken as establishing causal links. This point is emphasized by some research results given by Freudenheim et al. (2005), which reports on a series of studies of alcoholics carried out between 1974 and 1981. These showed that a high consumption of alcohol was accompanied by high levels of morbidity and mortality from lung cancer. So, we have a strong correlation between heavy drinking and lung cancer. However, few would accept that there is a significant causal link here. Instead the correlation would be attributed to the confounding factor of smoking. A large percentage of alcoholics are also heavy smokers, and most researchers would judge that it was the smoking of alcoholics rather than their drinking which increased the risk of lung cancer. Admittedly Freudenheim et al. (2005) think that heavy drinking might slightly increase the risk of lung cancer, but they add (p. 657): “Residual confounding by smoking may explain part of the observed relation.” I will here assume that the correlation between heavy smoking and lung cancer does show a causal relation, while the correlation between heavy drinking and lung cancer does not indicate any causal relation, or, at most, shows a very weak causal relation.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:57:04.

138  The Russo–Williamson thesis

What more is needed then to establish that a correlation is causal in character? I earlier sketched a possible answer based on the action-related theory of causality, which was developed in Part I. This theory relates causality to intervention, and hence shows the need for some interventional evidence in order to establish causality. A similar point of view is put forward by Pearl (2000). Pearl argues for this position by giving an interesting criticism of Hume’s constant conjunction view of causality, which resembles a criticism of Hume given by Collingwood in 1938, and which we quoted in section 1.2. Pearl quotes the following passage in which Hume formulates his constant conjunction theory (1738, A Treatise of Human Nature, Part III, Section VI, p. 90): We remember to have had frequent instances of the existence of one species of objects; and also remember, that the individuals of another species of objects have always attended them, and have existed in a regular order of contiguity and succession with regard to them. Thus we remember to have seen that species of object we call flame, and to have felt that species of object we call heat. We likewise call to mind their constant conjunction in all past instances. Without any further ceremony, we call the one cause, and the other effect, and infer the existence of the one from that of the other.

Copyright © 2018. Routledge. All rights reserved.

Pearl then comments as follows (2000, pp. 336–7): It is hard to believe that Hume was not aware of the difficulties inherent in his proposed recipe. He knew quite well that the rooster crow stands in constant conjunction to the sunrise, yet it does not cause the sun to rise. …  Today these difficulties fall under the rubric of spurious correlations, namely “correlations that do not imply causation.” …  What difference does it make if I told you that a certain connection is or is not causal? Continuing our example, what difference does it make if I told you that the rooster does cause the sun to rise? …  The obvious answer is that knowing “what causes what” makes a big difference to how we act. If the rooster’s crow causes the sun to rise, we could make the night shorter by waking up our rooster early and making him crow …  Pearl’s criticism of Hume seems to me to be correct. A Humean constant conjunction of A and B is just a correlation whose value is 1. If such a constant conjunction is established, we cannot, as Hume claims, “without any further ceremony …  call the one cause, and the other effect.” But what more is needed? If we take Hume’s own example of flame causing heat, our evidence is not just observational. We also know how to intervene and to produce heat by making a fire. As Pearl points out, a similar intervention with the rooster would

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:57:04.

Copyright © 2018. Routledge. All rights reserved.

The Russo–Williamson thesis  139

produce evidence that his crow does not cause the sun to rise. In general, we need i­nterventional as well as observational evidence to establish causation. Let us now return to our example of smoking and lung cancer. The Hill, Doll, Peto cohort study of doctors produced a very striking correlation between smoking and lung cancer, but it was a purely observational study, and correlations from such studies need not be causal in character. The above line of reasoning suggests that we need some interventional evidence in addition to the statistical evidence of correlation to establish causation, but what form should this interventional evidence take? We have here to deal with the problem of possible confounding, and earlier I formulated three strategies for dealing with this problem. The first strategy (section 6.3) was to think of as many confounders as possible, and then try to find evidence to disconfirm each one. The problem with this strategy is that it need not involve any interventional evidence, and therefore is less satisfactory than the other two strategies which do involve interventional evidence. Let us therefore consider strategies 2 and 3 in turn, and see how they apply to this case. Strategy 2 (section 6.5) is essentially to use a randomized controlled trial. The problem with the smoking causes lung cancer example is that we cannot set up an RCT to test the causal claim directly. To do so we would have to proceed in something like the following fashion. First, we take a sample of 30,000 humans chosen at random. We then select at random 15,000 who are forced to become smokers. The remaining 15,000 are forced to become non-smokers. We then follow up these individuals over a period of say 20 years and see what the differential rates of lung cancer are in the two groups. However, as Pearl remarks (2000, p. 353), controlled experiments of this sort “are impossible (and now also illegal) to conduct.” The example of smoking causing lung cancer is, in this respect, typical of very many causal claims in medicine: namely those of the form: X causes D, where D is a disease and X is a putative cause of D. In testing such claims it is obviously impossible for ethical, practical, and usually legal reasons, to set up an RCT in which X is implemented for a group of humans under controlled conditions in order to see how often D results. If then RCTs are ruled out in many medical situations, such as our example of smoking and lung cancer, it is obvious that we must seek some other form of interventional evidence. This brings us to our third strategy for dealing with the problem of confounding (section 7.1). This involved the consideration of mechanisms. Suppose we have established a correlation between A and B by means of statistical evidence. To establish that A causes B, we have to show that there is a mechanism linking A to B. For simplicity, I will consider this to be a basic mechanism of the form:

A ® C1 ® C2 ® ¼ ® Cn ® B (8.1)

where as usual “→ ” means “causes”. However, there seems to be an obvious flaw in this approach. Our original problem was to establish the existence of one causal link, namely A causes B. Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:57:04.

Copyright © 2018. Routledge. All rights reserved.

140  The Russo–Williamson thesis

We have now replaced this with the problem of establishing n+2 causal links as given in formula (1). Seemingly then we have replaced an easier problem by a harder one. Here, however, our earlier principle of the need for interventional evidence comes to our rescue. According to this principle, we need to make interventions in order to establish causality. Now the connection between A and B may have been established by observational statistical evidence, and it may be difficult, if not impossible, to intervene directly with an RCT to test the character of the link between A and B. However, in each of the sub-links of the mechanism given in formula (8.1), it may be possible to intervene and so establish that we have a genuine causal link. Indeed, this is likely to be the case, because in the case of medicine, the mechanisms will usually be of a physiological and biochemical character, and so can be studied by experimental methods in the laboratory, using animals, tissues, cells, etc. Thus, by postulating and testing out a mechanism, we may be able to establish a causal link which cannot easily be established directly using an RCT. An analogy should explain this situation. Suppose we wish to connect two posts with a piece of string. Unfortunately, we do not have a piece of string long enough to do the job. However, we do have a pile of short lengths of string. We can then tie these short pieces of string together in a sequence which will eventually yield a length of string long enough to solve our original problem of connecting the two posts. We have now reached the point where we can conveniently introduce the first qualification which needs to be made to our specific formulation of the Russo– Williamson thesis. This is the distinction between a plausible mechanism, and a confirmed mechanism. A plausible mechanism is one, which is confirmed by our general background knowledge but not necessarily by particular investigations and experiments designed to test it out. This distinction can be illustrated by the example of the difference between the cases of smoking and drinking in relation to lung cancer. Cigarette smoke contains a large number of chemicals of various types and, in the case of those who smoke regularly, some of these are likely to find their way into the lungs. It seems likely that some of these chemicals will damage the tissue of the lungs, and some indeed might be carcinogens which initiate a cancerous tumour. It was part of the background knowledge of 1976 that there existed chemical carcinogens capable of initiating cancerous tumours. Indeed this had been established by experimental evidence. Consequently, it was known in 1976 that there was a plausible mechanism linking smoking to lung cancer. By contrast there does not seem to be an obvious mechanism connecting the consumption of alcohol to lung cancer. Of course, the body is very complicated, and such a mechanism might turn out to exist after all. However, background knowledge does not make the existence of such a mechanism very plausible. This suggests that we may be more prepared to go from correlation to causation if there exists a plausible mechanism linking the two factors. In contrast to smoking and lung cancer in 1976, a mechanism may, in some other cases, be uncovered by a lengthy series of experiments, and its correctness confirmed by these results. Here I will speak of a confirmed mechanism. Of course, there is really a continuum here, which depends on the degree to which

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:57:04.

The Russo–Williamson thesis  141

the mechanism has been confirmed. However, the rough distinction between plausible and confirmed mechanisms is helpful for clarifying the situation. I will understand confirmed mechanisms to be also plausible, but not vice versa. Corresponding to this distinction, we can have two forms of the Russo– Williamson thesis – a strong form and a weak form. According to the strong form, a causal link between A and B can only be said to be established if it has been shown that there is a confirmed mechanism linking A and B. For the weak form, it may suffice just to show that there is a plausible mechanism linking A and B. I prefer the weak form of the thesis since it fits better with the classic example of smoking and lung cancer. I would argue that Doll and Peto were quite justified in claiming in 1976 to have established that smoking causes lung cancer, even though the mechanism linking the two was, at that stage, only plausible rather than confirmed. Actually, I have simplified this historical example somewhat to illustrate my distinction between plausible and confirmed mechanisms, for, already by 1976, there was some specific evidence for a link between smoking and lung cancer. In fact, Austin Bradford Hill speaks of (1965, p. 298), “the histopathological evidence from the bronchial epithelium of smokers and the isolation from cigarette smoke of factors carcinogenic for the skin of laboratory animals.” Yet even if this additional evidence of mechanism had not existed, it still seems to me that the claim that smoking causes lung cancer would have been established by the existence of a plausible mechanism, that is to say by the existence of a mechanism which was plausible in the light of background knowledge. Since this background knowledge is itself confirmed empirically, this gives us a sort of evidence of mechanism.This point can be further clarified by comparing the case of “smoking causes lung cancer” with the case of “smoking causes heart disease”. This I will do in the next section.

Copyright © 2018. Routledge. All rights reserved.

8.4 Smoking and heart disease: Is there a linking mechanism? Let us now return to the Doll and Peto paper of 1976. Their survey of mortality among male doctors in the UK did not confine itself to lung cancer, but investigated all the recorded causes of death. In fact, the most common cause of death among the doctors was what they called “ischaemic heart disease”. This is just an alternative terminology for what we have called “coronary heart disease”. It accounted for 3,191 deaths as compared to 441 from lung cancer. Now death from ischaemic heart disease was also positively correlated with smoking. The annual death rate from ischaemic heart disease per 100,000 men, standardized for age, in various categories was as follows (Doll and Peto, 1976, p. 1527): Non-smokers Smokers 1–14 gms tobacco per day 14–24 gms tobacco per day 25 gms tobacco per day or more

413 565 (+37%) 501 (+21%) 598 (+45%) 677 (+64%)

(A cigarette is roughly equivalent to 1 gm of tobacco) Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:57:04.

142  The Russo–Williamson thesis

These figures are less striking than those relating smoking to lung cancer, but they still show a positive correlation, which is highly significant statistically. Moreover, Doll and Peto say (1976, p. 1534): “Our data for ischaemic heart disease are similar to those that have been reported in many other studies throughout the world.” They cite in this context six studies carried out between 1965 and 1975. Doll and Peto are, however, cautious about drawing causal conclusions from this correlation. Ischaemic heart disease is one of the conditions, which they list in their table III, and they say (1976, p. 1528): Half the conditions in table III were positively related to smoking, some very strongly so, …  . To say that these conditions were related to smoking does not necessarily imply that smoking caused …  them. The relation may have been secondary in that smoking was associated with some other factor, such as alcohol consumption or a feature of the personality, that caused the disease. Alternatively, smoking habits may have been modified by the disease or the relation may have been an artefact due to misdiagnosis … 

Copyright © 2018. Routledge. All rights reserved.

As we have seen, Doll and Peto (1976) did definitely regard the link between smoking and lung cancer as causal in nature. However, as regards ischaemic heart disease, they make the weaker claim that (p. 1535) the excess mortality from ischaemic heart disease in cigarette smokers is probably wholly or partly attributable to smoking. This caution is very much in agreement with the Russo– Williamson thesis, because the background knowledge of heart disease in 1976 did not support any plausible mechanism linking smoking to heart disease. In fact, the background knowledge regarding coronary (or ischaemic) heart disease in 1976 can be largely summed up in what we described in section 7.1 as causal hypotheses CH7.1. These were as follows: 1. High levels of saturated fat in the diet →  2. Cholesterol content of blood to become raised →  3. Infiltration of some parts of the arterial wall by lipids (mainly cholesterol esters) →  4. Lipids to be absorbed by macrophages, which become foam cells →  5. Full development of atherosclerotic plaques →  6. Coronary heart disease and other atherosclerotic events. So, in 1976 it was known that a high concentration of cholesterol in the blood favoured the development of atherosclerosis, via the steps listed above as 3 to 6. But why should smoking accelerate this process? The background knowledge of the time suggested no plausible mechanism linking smoking and atherosclerosis. However, research into heart disease between 1976 and the late 1990s greatly increased medical knowledge of how and why atherosclerotic plaques form, and this new knowledge does support a plausible mechanism by which smoking may

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:57:04.

The Russo–Williamson thesis  143

increase the rate of formation of such plaques. I will give a brief account of some of the research into atherosclerosis between 1979 and the late 1990s in the next two sections.

8.5  Research into atherosclerosis 1979–89

Copyright © 2018. Routledge. All rights reserved.

To follow the course of this research, two technical terms must first be mastered, namely macrophages and monocytes. Macrophages (literally “large eaters”) are well-known to the general public. They are white cells which attack and destroy microbes, such as bacteria, which invade the body. Macrophages have receptors or binding sites by means of which they attach themselves to their prey. This prey is then engulfed and destroyed. At least the macrophages attempt to destroy their prey. They are not always successful. Monocytes are circulating predecessors of macrophages. If additional macrophages are needed at any point, monocytes arrive and turn into macrophages at the location of the action in the tissue. Let us now consider the problem of how LDL carried round by the blood stream can be converted into atherosclerotic plaques. A crucial step in elucidating the process was taken by Goldstein et al. in their paper published in 1979. In this paper they showed that if LDL is modified by being acetylated, it gets taken up in large quantities by macrophages using a specific binding site or receptor, which later became known as the “scavenger receptor”. The macrophages which take up the modified LDL become, in their attempts to destroy it, bloated with lipid and resemble the foam cells to be found in the fatty streaks which are the first stage in the formation of atherosclerotic plaques. This was a most suggestive observation, but of course acetylation of LDL in the body seemed an unlikely occurrence. As Goldstein et al. say (p. 337): Although in vivo acetylation of plasma LDL seems unlikely at this point, some chemical or physical alteration of LDL occurring in plasma or interstitial fluid may make it susceptible to recognition by the macrophage binding site. Naturally this remark prompted the search for the appropriate in vivo alteration of LDL, and it was duly found in the next few years (Henriksen et al. [1981], Steinbrecher et al. [1984]). The alteration was the oxidation of LDL. This can occur quite naturally in the body. It is induced by incubation of LDL with the endothelial cells which line the arteries. It can also be induced by incubation with smooth muscle cells, or with macrophages ( Jü rgens et al. [1987]). Moreover, oxidized LDL, just like acetylated LDL, is taken up by the scavenger receptor on macrophages resulting in the formation of foam cells bloated with lipid. Oxidized LDL has two further properties of great interest ( Jü rgens et al. [1987]). It inhibits the movement of macrophages, while attracting monocytes. Putting all these elements together, we arrive at a mechanism by which fatty streaks can form within the artery walls. This is described in Steinberg et al. (1989),

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:57:04.

Copyright © 2018. Routledge. All rights reserved.

144  The Russo–Williamson thesis

a paper with the significant title: ‘Beyond Cholesterol. Modifications of Low-Density Lipoprotein That Increase its Atherogenicity’. The key point is that LDL, even in large quantities, causes no problems as long as it remains in its natural state. However, if it is oxidized, then trouble starts. Oxidized LDL is attacked by macrophages which bind it with their scavenger receptor, and then attempt to destroy it.The result is the formation of foam cells. If some of these are in the wall of an artery, then, because oxidized LDL inhibits the movement of macrophages, they remain there. Moreover oxidized LDL attracts monocytes, which turn into macrophages and generate more foam cells. Since macrophages also oxidize LDL, a self-reinforcing process can start, resulting in the formation of fatty streaks in the artery walls. It may seem a little odd that macrophages should try to dispose of oxidized LDL since the results of this attempt are somewhat unfortunate. However, it should be remembered that oxidized LDL can be very damaging to the cells of the body, so that its disposal, even at some cost, may be on balance justified. It would obviously, however, be better for LDL not to be oxidized, and in fact there are many devices to protect LDL from oxidation. LDL carries round with it in the blood stream a whole package of anti-oxidants which protect it against oxidation. The principal component of this package is vitamin E, but the package also contains (Esterbauer et al.,1989, p. 256) beta-carotene, lycopene, and retinylstearate. Moreover, in the plasma of normal blood there are large quantities of two powerful anti-oxidants – vitamins C and E. These devices have obviously evolved to prevent the oxidation of LDL and the harmful consequences of this oxidation. So, by the end of the 1980s it was established that the oxidation of LDL was an important step in the process which led to atherosclerotic plaques. Now this seems at once to explain why smoking accelerates atherosclerosis. Cigarette smoke gives rise to what are known as reactive oxygen species (or ROS) (Lehr et al., 1994, p. 7691). These include superoxide and hydrogen peroxide. Their effect would be to increase the tendency towards oxidation in the body, to introduce what is called: “oxidative stress”. The existence of such oxidative stress in smokers was strongly confirmed in a study by Morrow et al. (1995). They introduced a new and superior index of the amount of lipid peroxidation occurring in vivo. This was the level of F2 -isoprostanes. Sure enough, this level proved to be much higher in smokers than non-smokers. Morrow et al. conclude (1995, pp. 1201–2): Our finding that the production of F2 -isoprostanes is higher in smokers than in nonsmokers provides compelling evidence that smoking causes oxidative modification of biologic components in humans. This conclusion is greatly strengthened by the finding that levels of F2 -isoprostanes in the smokers fell significantly after two weeks of abstinence from smoking. These results provide a basis for hypotheses that link oxidative damage of critical biomolecules to the pathogenesis of diseases caused by smoking.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:57:04.

The Russo–Williamson thesis  145

At this point it may well look as if we have not just a plausible mechanism, but a confirmed mechanism linking smoking to heart disease. It is the following (M8.1): 1. 2. 3. 4. 5. 6. 7. 8.

Smoking →  Reactive oxygen species (ROS) in the blood →  Oxidation of LDL →  Attack on oxidized LDL by macrophages →  Formation of foam cells →  Attraction of monocytes →  Further macrophage attacks and formation of foam cells →  Formation of atherosclerotic plaques.

Yet there is a difficulty which shows that this mechanism, however convincing it may appear at first sight, cannot be correct. The difficulty concerns the place at which the oxidation of the LDL takes place. Steinberg et al. argue that LDL is not oxidized in the blood stream but within the artery wall. This is what they say (1989, p. 919):

Copyright © 2018. Routledge. All rights reserved.

For several reasons, it seems that the oxidation of LDL probably occurs not in the circulation but within the artery wall. First, even if oxidized LDL were generated in the plasma, it would be swept up within minutes by the liver. Second, oxidation is inhibited by plasma and so probably requires the favorable conditions of a sequestered microenvironoment. As we pointed out, normal blood plasma contains large quantities of vitamins C and E, and these strongly inhibit the oxidation of LDL. If, however, the blood contains a great deal of LDL, some LDL may diffuse into the artery wall. Here it is no longer protected by the anti-oxidants of the blood plasma, and so, in the presence of oxidizing agents such as macrophages, its own package of protective anti-oxidants may be exhausted, and it may become oxidized. Now smoking introduces additional oxidizing agents into the blood stream (the ROS), but these do not penetrate into the artery wall, and so should have no effect on the formation of the atherosclerotic plaques as so far described. Therefore, the mechanism (if any) by which smoking accelerated the formation of atherosclerotic plaques remained a mystery. However, as I will describe in the next section, further research in the 1990s that shed new light on the question.

8.6  Research into atherosclerosis in the 1990s In the 1990s, investigations began into another aspect of the process of formation of atherosclerotic plaques. If monocytes adhere to the endothelial cells lining the arteries, they can work their way into the artery wall, turn into macrophages and accelerate any on-going process of plaque formation. Now cigarette smoke consists of 92% gaseous components and 8% particulate constituents. These particulate

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:57:04.

146  The Russo–Williamson thesis

constituents are known as cigarette smoke condensate. In 1994, Kalra et al. ­discovered that cigarette smoke condensate increases significantly the tendency of monocytes to adhere to the endothelial cells. As they say (Kalra et al., 1994, p, 160): It thus appears that cigarette smoke particulate constituents potentiate adherence of monocytes to the vascular endothelial cell lining. This presumably is followed by trans-migration of adhered monocytes into the subendothelium space to form foam cells and subsequent atherosclerotic lesion formation. Interestingly Kalra et al. elucidate the mechanism which brings about the increased adherence of the monocytes. They describe it as follows (1994, p. 155): the recruitment of monocytes to the endothelial surface could occur as a result of change in the adhesive properties of the endothelial surface. The results presented here show that cigarette smoke condensate (CSC), the particulate phase of cigarette smoke, causes an increase in the expression of CD11b on monocytes and ICAM-1, ELAM-1, and VCAM-1 adhesion molecules on endothelial cells with a concomitant increase (70–90%) in the basal adherence of monocytes to cultured endothelial cells.

Copyright © 2018. Routledge. All rights reserved.

Daniel Steinberg, who played a key part in the elucidation of the mechanisms described in the previous section, made a useful comment on the situation in his 1995 paper. Here he pointed out that the studies carried out so far had been mainly concerned with the formation of fatty streaks in the artery walls, but that these are only the earliest atherosclerotic lesions. He writes (1995, p. 37): How long does it take for a new fatty streak to become a clinically threatening lesion? We cannot be sure, but we know that by age 25 some 20–30% of the aorta is covered by fatty streak lesions and yet myocardial infarction seldom occurs before age 50. If lesion progression is more or less linear as a function of time, we might conclude that it takes 20 years or more for a new fatty streak to become the nidus for coronary thrombosis. This suggests that more attention should be devoted to the development of atherosclerotic plaques than to their initial stages. One factor to which attention is drawn by Poston and Johnson-Tidey concerns the rate at which monocytes adhere to the endothelium of the artery. They write (1996, p. 75): at 20 or 37oC, human peripheral blood monocytes …  showed selective binding to atherosclerotic plaques, compared with non-atherosclerotic arterial intima …  Adhesion occurred to the endothelium of the plaque area …  The binding to non-atherosclerotic artery endothelium was much less, and the difference was highly significant … 

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:57:04.

The Russo–Williamson thesis  147

They go on to suggest that a positive feedback mechanism may develop in which atherosclerotic plaques as they develop attract more and more monocytes which cause them to develop further. As they point out (p. 73) there could be a link here to the question of the effects of smoking on atherosclerosis in the light of some earlier work of Lehr et al. 1994, to which we now turn. Lehr et al. investigated the effects of cigarette smoke on hamsters. They showed that cigarette smoke increased the rate of adhesion of leukocytes to arterial endothelium, and that this rate of adhesion was significantly reduced by vitamin C, but not by vitamin E. They argue that the increase in the rate of adhesion was due to the reactive oxygen species (ROS) produced by cigarette smoke (CS). From this they draw the following conclusion (Lehr et al., 1994, p. 7691):

Copyright © 2018. Routledge. All rights reserved.

The fact that the water-soluble vitamin C, but not the lipid-soluble antioxidants vitamin E and probucol (which contribute little to serum antioxidant activity), afforded protection from CS-induced changes indicates that CS-induced leukocyte adhesion and aggregate formation with platelets involves isolated, direct attacks of aqueous-phase ROS, rather than the sequelae of membrane lipid peroxidation. Like dietary vitamin C, a single intravenous injection of vitamin C just 5 min prior to CS exposure resulted in a similar protection from CS-induced leukocyte/platelet/endothelium interaction, suggesting that vitamin C does not need to be incorporated into cells in order to be effective, but that it merely needs to be circulating in the bloodstream in order to neutralize aqueous phase ROS …  We saw earlier that the oxidation of LDL which is relevant to the formation of atherosclerotic plaques was thought in the 1980s to occur within the artery walls rather than in the blood stream and so was unlikely to be produced by smoking. However, the results of Lehr et al. indicated that there was another oxidation process which did occur in the blood stream and which affected atherosclerosis. This was the process which resulted in an increased rate of activation, aggregation, and adhesion of leukocytes to the endothelium of the artery. The process was known to involve oxidation because it was inhibited by anti-oxidants, and it was known to occur in the bloodstream because it is inhibited by vitamin C but not vitamin E. Lehr et al. remark (1994, p. 7692): Corroborative evidence can be derived from epidemiological surveys which consistently demonstrate a significant consumption of vitamin C, but not of vitamin E, in the plasma of smokers. Here then, at last, was a mechanism linking smoking to atherosclerosis. Smoking produced oxidative stress. This increased the adhesion of leukocytes to the endothelium of the artery, which in turn accelerated the formation of atherosclerotic plaques. This mechanism was certainly plausible, and indeed could be regarded as

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:57:04.

148  The Russo–Williamson thesis

at least partially confirmed. So, it was sufficient for the Russo–Williamson thesis in the form in which I have formulated and defended it.

Copyright © 2018. Routledge. All rights reserved.

8.7 Implications of our medical examples for the Russo–Williamson thesis Let us now consider our various medical examples in relation to the Russo– Williamson thesis (RWT) as formulated earlier. The first example is the claim that smoking causes lung cancer. This was taken as established by Doll and Peto in 1976 on the basis of strong observational statistical evidence, which accords with our version of the RWT because there was a plausible mechanism linking smoking and lung cancer. By the early 1980s, a number of observational studies had shown that there was also a strong correlation between drinking heavily and lung cancer. However, this was not taken as showing a causal connection, but rather as being explained by confounding factors such as smoking. This again agrees with our version of the RWT since there was, and is, no plausible mechanism linking heavy drinking to lung cancer. The third example concerned smoking and heart disease. We saw that Doll and Peto established a strong correlation between the two in 1976, and yet were hesitant about inferring a causal connection. They thought that causality was probable but had not been established. This again agrees with our version of the RWT because no plausible mechanism linking smoking and heart disease had at that time been shown to exist. However, I went on to show that between 1979 and the late 1990s research into atherosclerosis did bring to light a mechanism linking smoking to an acceleration of the rate of formation of atherosclerotic plaques. This mechanism was at least plausible and perhaps even confirmed. So, it justified the acceptance that smoking causes heart disease, and indeed it is now generally accepted by the medical community that smoking is a causal factor in atherosclerosis. However, an account of the research which led to this result, shows that the path to a plausible mechanism here was a winding one. Some earlier results suggested a mechanism which for quite subtle reasons could not be correct, and the linking mechanism which now looks plausible has a rather different character. This is a nice illustration of the difficulties of establishing plausible mechanisms through research in the medical field. In all three of the examples considered in this chapter, the statistical evidence has been observational in character, because of the impossibility of running a randomized controlled trial when investigating the health effects of smoking. This feature of the smoking case applies to the investigation of many cases in medicine. Suppose we are examining a causal hypothesis of the form A causes D, where D is some quite serious disease. It would obviously be unethical and quite probably illegal to carry out a randomized controlled trial in which for half the participants selected at random, A is implemented, whereas for the controls A is not implemented. In such cases we can only conduct observational surveys. Now observational statistics are never sufficient to establish causality, because we need to have interventional evidence to establish causality in medicine. Since RCTs are ruled

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:57:04.

The Russo–Williamson thesis  149

out, this means that we need evidence of mechanism, which is interventional in character, to establish causality. This gives a derivation of the Russo–Williamson thesis in this case. It also gives the examples, promised in section 7.2, of causal laws in medicine which were established without the use of randomized control trials in human populations. In fact, the two causal laws: (i) smoking causes lung cancer, and (ii) smoking causes heart disease, are now generally accepted by the medical community. This was on the basis of observational statistical evidence, and evidence of mechanism, without the use of randomized control trials in human populations. In some cases of causal hypotheses in medicine, it is, however, possible to carry out randomized controlled trials on human populations. Might the statistical evidence from such RCTs be sufficient to establish the causal hypothesis? This would contradict the Russo–Williamson thesis, which claims that we always need some evidence of mechanism in addition to statistical evidence in order to establish a causal hypothesis in medicine. In the next chapter, I will consider whether the Russo–Williamson thesis no longer applies when we can use randomized controlled trials.

Note

Copyright © 2018. Routledge. All rights reserved.

1 More details about early uses of statistics in medicine are given in section 10.5.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:57:04.

9 THE RUSSO–WILLIAMSON THESIS

Copyright © 2018. Routledge. All rights reserved.

(ii) The evaluation of streptomycin and thalidomide

Consider a causal hypothesis of the form: A causes D, where D is a disease. It has already been argued that it is difficult to assess hypotheses of this type using randomized controlled trials. Suppose we have some volunteers to take part in such a trial, and divide them into two groups by tossing a fair coin. The cause A is imposed on those selected, while the control group is shielded from A. We then examine the results of this intervention. It might just about be possible to carry out such a trial for a very mild illness, such as the common cold. However, for a serious disease, it would obviously be unacceptable morally and perhaps legally, as was illustrated previously (section 8.3) by the example of smoking and lung cancer. However, causal hypotheses of the form: A causes D are not the only causal hypotheses to be found in theoretical medicine. A cures D or A prevents D are also causal hypotheses in medicine, because “cures” and “prevents” are causal expressions. Indeed “A cures D” can be roughly rendered as “A causes D to disappear”, and “A prevents D” as “A causes not-D”. Now randomized controlled trials can be used for this type of causal hypothesis in medicine, and indeed they 药物测试。 are characteristically used to test drugs to see if they are genuine cures. So, can a causal hypothesis of the form “A cures D” be established solely on the basis of the results of randomized controlled trials? Let us next examine the views of evidence-based medicine (EBM) on this question.

9.1  Evidence-based medicine My account of evidence-based medicine is based on Jeremy Howick’s excellent book on the subject (Howick, 2011a). Howick broadly supports the evidence-based medicine position and I will consider his objections to the Russo–Williamson thesis in the next chapter. According to Howick, the proponents of EBM arrange

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:57:14.

The Russo–Williamson thesis  151

evidence for medical claims in a hierarchy (see Howick, 2011a, p. 5). At the top are randomized control trials, then come other types of statistical evidence concerning human groups. However, evidence about mechanisms is considered to be much less important. As Howick puts it (2011a, p. 23):

Copyright © 2018. Routledge. All rights reserved.

when it comes to appraising evidence, randomized trials remain at the pinnacle of a hierarchy …  , while expertise and mechanistic reasoning are omitted altogether …  or are at the very bottom …  . As an instance of EBM proponents, who omit evidence about mechanisms altogether, Howick cites Guyatt, Oxman, Vist et al. (2008), and of those who place it at the very bottom of the hierarchy, he cites Phillips, Ball, Sackett et al. (2001). The latter is a publication of the Oxford Centre for Evidence-Based Medicine. There is a problem here because neither Howick nor the other proponents of EBM use the phrase “evidence of mechanism”, but instead the phrase “mechanistic reasoning” (see quote from Howick, 2011a, p. 23 given above). Now “mechanistic reasoning”, if the words are used in their ordinary sense, means “reasoning about mechanisms”, and so need not be a form of evidence at all. For example, I can reason that if such and such a mechanism (M say) is really operative, then such and such results will follow. However, this “mechanistic reasoning” does not give any evidence as to whether M is operative or not. I am using evidence to mean empirical evidence, that is to say the results of observation and experiment, while reasoning might be ordinary deductive reasoning which does not involve either observation or experiment. Yet EBM proponents include “mechanistic reasoning” in a hierarchy of evidence, which seems to imply that it is a form of evidence. I will resolve this difficulty by assuming that when the EBM proponents use the phrase “mechanistic reasoning” they mean the same as what is here called “evidence of mechanism”. With this interpretation, it turns out that there are two EBM positions. Hard EBM is the view that the evidence used to assess medical claims should be entirely statistical in character, and not include any evidence of mechanism. This point of view is to be found in Guyatt, Oxman, Vist et al. (2008). This expounds GRADE, which the authors claim represents an emerging consensus on rating quality of evidence. In fact, they write (2008, pp. 925–6): The GRADE system is used widely: the World Health Organization, the American College of Physicians, the American Thoracic Society, UpToDate (an electronic resource widely used in North America, www. uptodate.com), and the Cochrane Collaboration are among the more than 25 organisations that have adopted GRADE. In their exposition of GRADE, Guyatt, Oxman,Vist et al. (2008) mention observational studies and randomized trials. They say (2008, p. 926): “observational studies (for example, cohort and case-control studies) start with a ‘low-quality’ rating.” Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:57:14.

152  The Russo–Williamson thesis

However, they add that in some circumstances this can be upgraded. By contrast (2008, p. 995): “In the GRADE approach to quality of evidence, randomized trials without important limitations, constitute high quality evidence.” They also say (2008, p. 925): Expert reports of their clinical experience should be explicitly labelled as very low quality evidence, along with case reports and other uncontrolled clinical observations. These are the only types of evidence considered in the GRADE system. Evidence of mechanism is not mentioned at all. The position is the exact opposite of Claude Bernard’s. Not all followers of EBM are quite so hard, however. There is also soft EBM, an exposition of which is to be found in Sackett, Rosenberg, Gray et al. (1996). Soft EBM does allow the inclusion of some evidence of mechanism. For example, Sackett, Rosenberg, Gray et al. (1996) do say (p. 72): And sometimes the evidence we need will come from the basic sciences such as genetics and immunology. This does seem to be a concession to evidence of mechanism, but the authors then go on to stress that for them, and they think, most people, randomized trials have become the gold standard (1996, p. 72):

Copyright © 2018. Routledge. All rights reserved.

Because the randomised trial, and especially the systematic review of several randomised trials, is so much more likely to inform us and so much less likely to mislead us, it has become the ‘gold standard’ for judging whether a treatment does more good than harm. Either version of EBM must surely lead to the conclusion that causal claims in medicine can be established using only statistical evidence, and without taking account of evidence of mechanism. In particular, the EBM position is that causal claims such as “drug M, taken in such and such quantities, cures disease D without harming the patient” could be established by randomized controlled trials (RCTs) without using any evidence of mechanism. The EBM position is very different from the Russo–Williamson thesis. This holds that statistical evidence and evidence of mechanism should be treated on a par. They complement each other, and a causal hypothesis in medicine can only be established by using both types of evidence. I next want to argue in favour of the Russo–Williamson thesis, and against the EBM position by giving a case in which exclusive reliance on RCTs and neglect of evidence of mechanism would have led researchers to the wrong conclusion regarding the efficacy of a proposed medicine.The case is a famous one. It concerns

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:57:14.

The Russo–Williamson thesis  153

the three trials of streptomycin and other anti-tuberculosis chemical agents, which were carried out by the British Medical Research Council (MRC) in the period 1947–51.1 These trials are of considerable importance in the history of medicine, because they were among the first RCTs used in medicine, and they were one of the strong influences, which led to the increasing use of RCTs to test the efficacy of proposed medicines. One might therefore have thought that these trials would support the EBM position, and this view is put forward by Solomon in her 2015 book, where she writes (pp. 12–13): Evidence-based medicine …  has core successes that serve as exemplars in the manner of a traditional Kuhnian paradigm: A. Bradford Hill’s 1948 trial of streptomycin for tuberculosis functions much like the case of the inclined plane for Galileo’s mechanics. When we examine the streptomycin trials in the next three sections, it will become obvious that they were very far from being a core success for evidencebased medicine.

Copyright © 2018. Routledge. All rights reserved.

9.2  The trial of streptomycin against bed-rest Streptomycin was discovered in America in 1944 by Schatz, Bugie, and Waksman. It was shown that it strongly inhibited tubercle bacilli in vitro, and that it was also successful in vivo in treating experimental tubercular infections in guinea pigs. The new antibiotic even produced some quite spectacular cures of patients suffering from tuberculosis. In fact, so promising did streptomycin appear that it might have seemed immoral to conduct a randomized controlled trial of the new drug, since those who were unlucky enough in the random allocation to be assigned the then standard treatment (prolonged bed-rest) might thereby have been deprived of an excellent hope of cure. Possibly for this reason, no controlled trial of streptomycin in pulmonary tuberculosis was undertaken in 1946 in the USA. In England, however, Austin Bradford Hill, some of whose work was considered in sections 8.2 and 8.3, was a firm believer in the desirability of randomized controlled trials and wanted to carry out one with streptomycin. Curiously it was Britain’s poverty in the immediate aftermath of the Second World War, which gave him his opportunity. As he wrote in his reminiscences of 1990 (p. 78): It was just after the Second World War when we were trying to make the trial, in 1946, and Britain literally had no currency. We had exhausted all our supply of dollars in the war and our Treasury was adamant that we could have only a very small amount of streptomycin. This, I think, turned the scales: …  . I could argue that in this situation it would not be immoral to make a trial – it would be immoral not to make a trial since the opportunity would never rise again … 

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:57:14.

154  The Russo–Williamson thesis

So Bradford Hill managed to persuade the Medical Research Council (or MRC) to carry out an RCT, and the first patients for it were recruited in January 1947. The report on trial published in the British Medical Journal on 30 October 1948 (MRC, 1948) contains an account of both the procedure and the results. The procedure was fairly straightforward. The first requirement was to make the patients and their disease as uniform as possible. The type of case to be considered was therefore defined quite precisely as follows (MRC, 1948, p. 770): acute progressive bilateral pulmonary tuberculosis of presumably recent origin, bacteriologically proved, unsuitable for collapse therapy, age group 15 to 25 (later extended to 30)

Copyright © 2018. Routledge. All rights reserved.

Between January 1947 and September 1947, 109 patients had been accepted. Two of these died in the preliminary observation week, and the remaining 107 were assigned randomly to either the control group C or the streptomycin group S. There were 52 in C, and 55 in S. The control group C received the standard treatment of the time, which was six months of bed-rest. The S group received, in addition to bed-rest, a dose of 2g of streptomycin per day, given in four injections at six-hourly intervals. The streptomycin was continued for four months, but the patients were observed for a further two months. So the trial was brought to a close for each patient after six months. The improvement or deterioration of the patients in the six months of treatment was assessed by X-rays, “the radiological picture …  being in our opinion the most important single factor to consider” (MRC, 1948, p. 771). The results obtained are shown in the following table (MRC, 1948, p. 771). Radiological Assessment

Streptomycin Group Control Group

Considerable improvement Moderate or slight Improvement No material change Moderate or slight Deterioration Considerable deterioration Deaths Total

28 10

  51%   18%

 4 13

 8%  25%

 2  5

 4%  9%

 3 12

 6%  23%

 6  4 55

  11%  7% 100%

 6 14 52

 11%  27% 100%

These figures show that the S group did very considerably better than the C group. 51% of the S group showed considerable improvement as against only 8% of the C group. 7% of the S group died as against 27% of the C group. These differences are highly significant statistically. In the light of such good results from the RCT, one might have expected that the MRC would have declared that treatment with streptomycin had been shown to work, and was to be recommended. Instead of giving such an endorsement

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:57:14.

The Russo–Williamson thesis  155

of streptomycin therapy, however, the MRC conclude on a very cautious note, saying (1948, p. 780): This planned group investigation has demonstrated both the benefit and the limitations of streptomycin therapy in pulmonary tuberculosis. This caution proved to be amply justified. The same patients were investigated after five years, and it was then found that 58% of the S group had died as against 67% of the C group. The difference here is not statistically significant. These figures are taken from Florey (1961, p. 133) where she comments: “it was obvious that the encouraging promise at an earlier time had not been fulfilled.” It should be noted that the patients were all between 15 and 30 years old at the start of the trial. So it is unlikely that any of them would have died from causes other than tuberculosis in the succeeding five years. What seems to have happened in the S group is that, after the encouraging initial improvement, many relapsed. This example is an instance of a general problem with RCTs, which we earRCT受到时间 lier (section 6.5) described as time limitations. Such trials have to come to an 限制 end after some time period t. Suppose the RCT shows that the treatment has produced a marked improvement by t, can we then be sure that this will not be followed by a relapse later on? How can this problem be overcome? Well, those who conducted the streptomycin trial did seem to overcome the problem. They did foresee that the long-term results might not be so good as was suggested by the short-term improvements; and, for this reason sounded a note of caution. How did they manage this? The answer is that they took account of evidence about the mechanism of the treatment. I will give some details about this in the next section.

Copyright © 2018. Routledge. All rights reserved.

9.3  The investigation of the treatment mechanism Already by 1947 many researchers in the area had become aware that there might be a problem with streptomycin therapy (see Florey, 1961, pp. 136–7). While some antibiotics such as penicillin could dispose of the pathogenic bacteria, which they targeted, in a week or two, streptomycin took many weeks, even months, to deal with a patient’s tubercle bacilli. Now Darwinian evolution as applied to bacteriology strongly suggested that, in such a time period, strains of the tubercle bacillus might develop which would be resistant to streptomycin. Such resistant strains posed a very considerable threat to streptomycin therapy. They might well increase in numbers producing a relapse, and, in this new condition, a fresh treatment with streptomycin would obviously be useless. Because of an awareness of this difficulty, those, who carried out the streptomycin RCT, at the same time carried out an investigation into the mechanism of the treatment. In the paper, the results of this investigation are reported in a separate section entitled ‘Bacteriology’ (MRC, 1948, pp. 778–780). The resistance of a strain of tubercle bacilli to streptomycin was measured by the resistance ratio (R.R.),

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:57:14.

156  The Russo–Williamson thesis

which is the ratio of the minimum concentration of streptomycin to which the tubercle bacilli of the patient are sensitive to the corresponding figure for the standard strain H37Rv. At the beginning of the streptomycin treatment no bacilli from the patients were found to be resistant. However, by the end of the trial (MRC, 1948, p. 781), Strains showing streptomycin resistance over 10 times that of the original strain or of the standard H37Rv were isolated from 35 of 41 patients [85% – D.G.] for whom data are available: in 13 of the 35 cases [37% – D.G.] the strains had a resistance over 2,000 times that of the control organism. The times at which resistance developed were also of interest, and are given in MRC, 1948, p. 779 as follows: Of 35 cases showing resistance over 10 times H37, this resistance emerged in five cases in the first month, in 21 in the second month, four in the third, four in the fourth, and one as late as the fifth month. So by the end of the second month, 63% of the cases in the S group, which were examined, had developed resistance to streptomycin. We will describe the reaction of the researchers to this result in the next section.

9.4  The trials of streptomycin and PAS

Copyright © 2018. Routledge. All rights reserved.

The thinking of researchers in the MRC group in the light of the evidence about the mechanism of streptomycin treatment is clearly stated in MRC (1949, p. 1521). A major disadvantage in the use of streptomycin is that the period of effective therapy is limited in many patients by the emergence of streptomycinresistant strains of tubercle bacilli after five or more weeks of treatment. It has been thought by many workers that the addition of another tuberculostatic agent might be sufficient to suppress the resistant strains, which in the initial phases are present in very small numbers. In fact Jorgen Lehmann, a Danish doctor working in Sweden, had announced in 1946, the existence of a tuberculostatic agent, namely para-amino-salicylic acid or PAS. PAS is a variant on aspirin (acetyl-salicylic acid), and this is no accident. Lehmann had got the idea of looking for PAS when his friend Bernheim told him in 1940 the result of an experiment in which adding aspirin to a culture of tuberculous bacilli increased their consumption of oxygen. As soon as the first streptomycin trial was over, the MRC researchers started a second trial. This was conducted along similar lines to the first trial, except that the patients were divided into three groups. The S group received streptomycin only, though only 1g a day, given by one injection at 8am. The P group received

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:57:14.

The Russo–Williamson thesis  157

20g daily of PAS by mouth in four doses of 5g at 8am, noon, 4pm and 8pm. The SP group received both streptomycin and PAS in doses as for the other two groups. These treatments were continued for three months, and the patients were observed for a further three months. Patients improved most in the SP group, but the most striking results concerned the difference in streptomycin resistance between the S group and the SP group. As it is stated in MRC (1950, p. 1081): At the end of the six months 89% of the SP patients producing positive cultures had completely sensitive strains, and only 21% of the S patients. Moreover the resistant strains took a longer time to appear in the SP group, and some subsequently disappeared (MRC, 1950, p. 1084): Streptomycin resistance (as judged by a “resistance ratio” greater than 8) was detected in 33 of the 49 S cases [67% - D.G.] from which results were available from the first month onwards, whereas streptomycin-resistant strains occurred in only 5 of the 48 SP cases [10% - D.G.] (in 4 of these 5 cases resistance emerged late in treatment or after its cessation; in 2 of them only a single resistant culture was obtained, subsequent cultures being sensitive). The association of P.A.S. with streptomycin had therefore a most marked effect in delaying or preventing the emergence of streptomycinresistant variants of the infecting organisms.

Copyright © 2018. Routledge. All rights reserved.

By December 1949, the results concerning streptomycin resistance were already so striking that an interim communication was made (MRC, 1949). The authors conclude their report on the full trial by saying (MRC, 1950, p. 1085): Combination of P.A.S. with streptomycin not only renders effective administration of streptomycin possible for longer periods than previously, but probably permits also of repeated effective courses. The only problem with PAS was that it produced gastric disorders in the patients, such as nausea, diarrhoea, and vomiting. This is hardly surprising since PAS is a variant on aspirin. The third MRC trial in this series, reported in 1952, was designed to see if smaller doses of PAS (10 g per day or 5 g per day) were as effective as 20 g per day in preventing streptomycin resistance. If so, the gastric discomfort of the patients could be avoided. Unfortunately, the lower doses of PAS, though they did prevent streptomycin resistance to some extent, were not nearly as effective as the strong 20 g per day dose of PAS. This concludes my account of the MRC RCTs of streptomycin and of streptomycin and PAS. In the next section, I will consider the impact of this example on the debate about EBM and the Russo–Williamson thesis (RWT).

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:57:14.

158  The Russo–Williamson thesis

Copyright © 2018. Routledge. All rights reserved.

9.5  The streptomycin trials in relation to EBM and RWT There seems no doubt that the results just presented very much undermine the EBM position. Proponents of EBM, as we have seen, argue that evidence about mechanisms has little or no value, while the evidence of RCTs is at the top of the evidence hierarchy. If, however, the researchers of the MRC in the period 1947 to 1951 had followed the EBM guidelines, then they would have reached a completely erroneous conclusion. The first streptomycin trial showed that the patients given streptomycin improved dramatically over six months when compared to the controls. If this had been accepted, and all evidence, concerning the mechanism by which the streptomycin therapy worked, had been ignored, then the conclusion would inevitably have been reached that streptomycin on its own was an excellent therapy for tuberculosis. However, the reality was that streptomycin on its own was, over a longer period, no better than bed-rest. Fortunately, however, the MRC researchers of that period did not follow EBM guidelines. They were almost certainly following the pluralist approach to evidence, which Austin Bradford Hill elaborated in 1965 (see section 8.2). Bradford Hill was not of course opposed to RCTs. He was the first to introduce such randomized trials into medicine and supported them strongly. However, in accordance with his pluralist approach to evidence, he regarded the results of the RCT in terms of patient improvement as only part of a picture, which should be supplemented by as many other considerations as possible. More specifically he and his researchers took both the evidence from the RCT and the evidence about the mechanism of streptomycin therapy into account, and, when both were weighed they reached the correct conclusion that there were problems with using streptomycin on its own as a treatment for tuberculosis.This is a very good instance of the Russo–Williamson thesis, which states that a causal claim in medicine should in general only be accepted if it is supported both by statistical evidence, and by evidence about mechanisms. Considering both types of evidence not only enabled a serious error to be avoided, but it also suggested a way out of the problem, which had been brought to light.This was the conjecture that the combination of streptomycin with PAS might prevent the emergence of strains of tubercle bacilli resistant to streptomycin. This conjecture was tested out, and proved to be correct.Together with further research and development, it led to satisfactory treatments for tuberculosis. If EBM guidelines had been followed, and evidence about mechanisms had been ignored, it is hard to see how the conjecture about combined therapies would ever have emerged. In short, an adherence to EBM guidelines would have led to a serious error, and would have held up the development of satisfactory treatments for tuberculosis. In the light of all this, EBM’s disparagement of evidence about mechanisms seems to be radically mistaken.2

9.6 Generalizing from the streptomycin case, and the example of thalidomide The streptomycin case goes strongly against the EBM position, but is it a rather peculiar special case? Or can we draw more general conclusions from it? In

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:57:14.

Copyright © 2018. Routledge. All rights reserved.

The Russo–Williamson thesis  159

fact, for practical reasons, every randomized controlled trial lasts for a limited period of time. This is what we earlier (section 6.5) called the time limitations of clinical trials. Such time limitations mean that, given a positive result, we can always ask whether it is a genuine cure, or a temporary improvement which could be followed by a relapse. To put the point another way, we can never be sure whether the trial has lasted long enough to give reliable results. To resolve this doubt, we need, as in the streptomycin case, to consider evidence of mechanism. Thus the streptomycin case is generalizable, and this supports the Russo– Williamson thesis that, in addition to RCTs, we need evidence of mechanism to establish causality in medicine. This conclusion is further supported by another consideration. To establish that a drug is a satisfactory cure, we need to show not only that it is effective in removing the disease, but also that it is safe, and does not have any unacceptably harmful and unpleasant side effects. I will now argue that considerations of establishing safety also support the Russo–Williamson thesis. To do so, it is worth examining the most famous case of a drug having disastrous side effects, namely thalidomide, and seeing what safeguards could be put in place to avoid a similar tragedy. Thalidomide3 was originally proposed as a treatment for anxiety and insomnia, which were quite widespread in the 1950s – perhaps not surprisingly. The Second World War had ceased a few years before in 1945, and many had had, during the War, the kind of horrifying and traumatic experiences which leave lasting effects. The War had ended with the explosions of two atomic bombs, and had been succeeded by the Cold War, which, in the nuclear age, seemed to threaten the continued existence of mankind. Moreover, the Cold War was not entirely cold, and the Korean War raged from 1950 to 1953. Given this general background, it is not surprising that the use of sedatives was very common. In Britain it was estimated that a million people took sedatives, and in the USA that one person in seven took sedatives (Brynner and Stephens, 2001, p. 4). These sedatives were barbiturates, and an overdose could be fatal. There were many tragic accidents in which children took several of their parents’ sleeping pills, thinking them to be sweets, and died as a result. The pharmaceutical industry was looking for a sedative, which could be taken in large quantities without ill effects. In 1954, a German pharmaceutical company, Chemie Grü nenthal, came up with what they thought was the answer, a drug which they named thalidomide. Thalidomide was indeed a powerful sedative and effective sleeping pill for humans. Moreover it appeared in animal trials to be completely non-toxic. In such trials, animals are fed a chemical to determine the dosage at which half of the tested animals die; this is called LD50. In the case of thalidomide, they could not find a dose large enough to kill rats. This was most unusual. Moreover, the drug appeared to be non-toxic for the further animals tested, namely mice, guinea pigs, rabbits, cats, and dogs (Brynner and Stephens, 2001, p. 9). Grü nenthal concluded that thalidomide was much safer than the barbiturates currently used as sedatives. Indeed, thalidomide was believed to be so safe that it was released as an over the counter drug in Germany on 1 October 1957.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:57:14.

Copyright © 2018. Routledge. All rights reserved.

160  The Russo–Williamson thesis

The advertising campaign for thalidomide stressed that thalidomide was much safer than the other sedatives currently available. Even a determined suicide could not take enough of it to die, and tragic accidents with children would no longer be possible (Brynner and Stephens, 2001, pp. 14–15). The theory was that thalidomide was an effective sedative and sleeping pill, which could be used safely by anyone. There was some, though not enough, empirical support for this theory when thalidomide was released, but it was soon to be refuted in a tragic fashion. Thalidomide, as we now know, produces horrifying birth defects in babies when it is taken by pregnant women. When thalidomide’s ability to cause birth defects came to light, it was withdrawn in Germany in November 1961, and, though there were unfortunate delays, in the rest of the world by the end of 1962. During the four to five years when it was on the market, thalidomide produced between 8,000 and 12,000 deformed babies of whom about 5,000 survived beyond childhood (Brynner and Stephens, 2001, p. 37). The first question which the thalidomide case raises is whether the disaster could have been avoided by testing the drug more severely before putting it on the market. Nowadays a drug would never be marketed unless it had been subjected to clinical trials, which are usually randomized controlled trials. Surprisingly no clinical trials were conducted on thalidomide before its release, since such trials were not required at that time. Unfortunately, however, RCTs conducted on thalidomide would never have revealed its potential to cause birth defects when taken by pregnant women. The reason is simple. As a precautionary measure, pregnant women are never allowed to take part in clinical trials of new drugs. This is a very reasonable requirement to make, but it does mean that we can never be sure on the basis of RCTs that a drug will not have a harmful effect on the foetus if taken by pregnant women. This is an example of the second main weakness of clinical trials which was discussed in section 6.5, namely limitations of sample. Any clinical trial is carried out with a limited sample of participants, and, if this sample is not representative of the population as a whole to whom the drug might be prescribed, then the results of the trial could be misleading as far as that population is concerned. Granted then that RCTs could not have brought to light thalidomide’s potential to cause birth defects if taken by pregnant women, what other ways might there be to detect such a potential harmful effect of a new drug? One obvious idea would be to conduct careful animal trials. As we have seen, some animal trials were conducted on thalidomide, but not very many. However, it is not clear that more animal trials would have brought the problems of thalidomide to light. The difficulty is that thalidomide in small doses does not cause birth defects in any animals other than primates. After the disaster, birth defects produced by thalidomide were shown in rabbits, but only with doses 150 times greater than the therapeutic dose. Moreover, it was difficult to show this effect. Dr Helen Taussig, a leading American researcher, said that she was unable to obtain abnormalities in baby rabbits using thalidomide, because the massive doses needed produced abortions (Brynner and Stephens, 2001, p. 13).

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:57:14.

The Russo–Williamson thesis  161

More consideration of the evidence of mechanism would, however, have helped to avoid the disaster. Brynner and Stephens (2001, pp. 12–13) write:

Copyright © 2018. Routledge. All rights reserved.

it had been known since 1955 that any substance with a molecular weight of less than 1,000 could cross the placenta and enter the fetal blood. The molecular weight of thalidomide is 258. …  it had been demonstrated in 1948 that the dye known as trypan blue could cause birth defects in rat embryos, whereas the mother rats exhibited no symptoms. Evidence of this kind should have raised doubts in the minds of the regulators whose job it was to approve thalidomide. Indeed, perhaps it did in some cases. Although thalidomide was widely approved, it was not approved in the USA. In fact a major disaster with thalidomide was narrowly avoided in the USA. A leading American pharmaceutical company (Richardson-Merrell) was preparing to launch thalidomide in March 1961. Ten million tablets were already manufactured, when they submitted an application to the American regulatory authority (the FDA) on 8 September 1960. It was expected that the application would go through with no problems, but, unexpectedly, difficulties were raised by the FDA officer assigned to the case (Dr Frances Kelsey). Kelsey delayed granting approval until 29 November 1961 when Grü nenthal withdrew the drug in West Germany. After that, it was obvious that thalidomide would not be approved in the USA. In August of the next year, Kennedy presented Dr Kelsey with the President’s Award for Distinguished Federal Civilian Service (Brynner and Stephens, 2001, p. 55). Obviously it was richly deserved. Now the interesting fact here is that, before joining the FDA, Dr Kelsey had carried out research with her husband on malaria. Among other things they had examined the effects of quinine on pregnant rabbits, and discovered that quinine is toxic to the foetus, but not to the mother (Brynner and Stephens, 2001, p. 45). It is clear from this that the terrible danger of thalidomide could not have been revealed by RCTs. It could only have been brought to light, before the drug was released, by a careful consideration of evidence of mechanism. First it would have had to be realized that thalidomide’s molecular weight meant that it could cross the placenta and enter the foetus. Second, account would have needed to be taken of the fact that there were drugs and chemicals, which, in animal experiments, had been shown to harm the foetus while leaving the mother unaffected. Putting together the evidence for each of these two mechanisms would have shown the potential danger of thalidomide. All this makes clear that, when trying to evaluate the safety of new drugs, account needs to be taken not just of the results of RCTs, but also of some quite sophisticated evidence of bodily mechanisms. This result gives strong support to the Russo–Williamson thesis. I have now presented in this chapter and the previous one, my arguments in favour of the Russo–Williamson Thesis. Yet this thesis remains controversial, and in the next chapter (10) I will give an account of some of the objections, which have been raised to the thesis.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:57:14.

162  The Russo–Williamson thesis

Notes

Copyright © 2018. Routledge. All rights reserved.

1 MRC (1948, 1949, 1950, 1952) contain the reports on these trials published in the British Medical Journal. There is an overview of the three trials in in Daniels and Bradford Hill (1952), while Bradford Hill (1990) gives some interesting reminiscences. 2 This is perhaps putting the point in an overly critical and negative fashion. To be more constructive and positive, one could say that the evidence criteria proposed by EBM should be supplemented to include evidence of mechanism. This is the point of view of the EBM+ consortium, of which I am a member together with Brendan Clarke, Phyllis Illari, Federica Russo, Jon Williamson and others. More details about EBM+ can be found on ebmplus.org. 3 My account of thalidomide is largely based on Brynner and Stephens’ 2001 book: Dark Remedy. This is an excellent book, which I strongly recommend. It was written by a historian with a M.A. in Philosophy (Rock Brynner), and a Professor of Anatomy and Embryology who had spent 25 years researching into thalidomide (Trent Stephens). The book contains a highly informative account of all aspects of the thalidomide case – scientific, methodological, sociological, political, and historical.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 13:57:14.

10 OBJECTIONS TO THE RUSSO– WILLIAMSON THESIS

Copyright © 2018. Routledge. All rights reserved.

10.1  Causal pluralism The first objection we shall consider is based on causal pluralism. The Russo–Williamson thesis (RWT) presupposes that the same causal hypothesis in medicine (A causes B say) can be supported by two different types of evidence, namely statistical evidence and evidence of mechanism. The objection based on causal pluralism is that there are really two different types of cause involved here.The statistical evidence supports (or undermines) a causal claim in the statistical sense (A statistically causes B), while the evidence of mechanism supports (or undermines) a different causal claim (A mechanistically causes B). I have already described a position of this sort, due to Campaner and Galavotti (2007), in the introduction, section 0.3. Another advocate of this type of causal pluralism is Reiss (2009). Reiss attributes to Russo and Williamson, a position which he calls evidential pluralism, and goes on to describe the conditions for this to be acceptable as follows (2009, p. 28): Evidential pluralism could not work if every evidential method defined its own concept because when moving from method to method we would in fact change the hypothesis to be tested. If (say) “X causes Y (as supported by probabilistic evidence)” means something different from “X causes Y (as supported by mechanistic evidence),” evidential pluralism does not get off the ground because instead of having one hypothesis that is being supported by two sources of evidence, we in fact have two separately supported hypotheses. In other words, we cannot be operationalists about the concept of cause. Rather, we require an independent concept of cause that, nevertheless, bears some systematic relationship with different evidential methods.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:18:38.

164  Objections to the Russo–Williamson thesis

However, Reiss goes on to argue that the combination of an independent concept of cause (causal monism) and evidential pluralism simply does not work in many cases. Russo and Williamson, however, argue that “pluralism about the nature of causality faces a crucial problem – it cannot account for the homogeneity of causal language” (2007, p. 166). They seem to be right here. Let us take an example such as “smoking causes heart disease”. As we saw in Chapter 8, this is supported by both statistical evidence in human populations, and by evidence of mechanism. Yet nowhere in the extensive literature on the topic do scientists distinguish between two senses of cause, namely cause (in the statistical sense) and cause (in the mechanistic sense). They simply use the word “cause” without any attempt to distinguish different senses of the term. Moreover, it is not clear how one might distinguish two senses of cause, and what significance such a distinction would have. Altogether, this sort of causal pluralism seems very unrealistic.

Copyright © 2018. Routledge. All rights reserved.

10.2  McArdle disease Earlier (in section 8.1) I formulated the RWT as follows: a causal hypothesis in medicine can be established only by using both statistical evidence and evidence of mechanism. Leaving aside causal pluralism, there are two kinds of counterexample which could be produced to this thesis. The first kind of counter-example (a mechanistic counter-example) would consist of a causal hypothesis in medicine which is established using only evidence of mechanism and no statistical evidence. The second kind (a statistical counter-example) would consist of a causal hypothesis in medicine which is established using only statistical evidence and no evidence of mechanism. In this section I will consider a mechanistic counterexample, and in the next three sections some statistical counter-examples. My general conclusion will be that there is some validity in these counter-examples. However, they should not lead us to reject the RWT completely, but rather to introduce some modifications and qualifications in our formulation of the thesis. The mechanistic counter-example is McArdle disease. This counter-example was introduced by Clarke in 2011, where he discusses how it is related to the Russo–Williamson thesis. Later in 2017, Clarke considers McArdle disease again, this time in connection with the problems of discovery in medicine. My own account is largely based on Clarke (2011 and 2017). McArdle disease is an unusual disease, but Clarke has done a considerable service by introducing it into philosophy of medicine, since this strange disease has, as we shall see, many interesting philosophical implications. I will first give a brief discussion of McArdle disease, and then consider the evidence which led to the acceptance of the existence and cause of this disease. The disease is produced by a malfunctioning in the skeletal muscles (a metabolic myopathy). The normal functioning of skeletal muscles is shown, in a simplified form, in Figure 10.1. Muscles naturally require energy to work, and much of this is obtained from glucose. Some glucose circulates in the blood, but this is soon exhausted by exercising

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:18:38.

科学家没有 区分两种意 义上的原 因：这能表 明什么呢？

Objections to the Russo–Williamson thesis  165

Copyright © 2018. Routledge. All rights reserved.

muscles. The muscles then have to utilize their stores of glycogen. This is broken down into glucose, and this glucose is converted into energy in order to fuel the muscle. Glycogen stores are therefore depleted when the muscle is working, but in quiescent periods they are replenished. For this system to function, glycogen has to be broken down into glucose during periods of exercise, and glycogen has to be built up from glucose during periods of rest. The breaking down of glycogen into glucose, and the building up of glycogen from glucose are both accomplished by the use of enzymes, as illustrated in Figure 10.1. The first enzyme needed to break down glycogen is myophosphorylase, that is to say phosphorylase in the muscle. McArdle disease is caused by a lack of this phosphorylase. The reason for this lack is thought to be genetic and specifically a “single, completely recessive, rare, autosomal gene” (Pearson et al., 1961, p. 516).

FIGURE 10.1  Simplified

mechanism of interconversion of glucose and glycogen in skeletal muscle.

Source: Clarke, Brendan, 2011, p. 96. Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:18:38.

Copyright © 2018. Routledge. All rights reserved.

166  Objections to the Russo–Williamson thesis

If an individual lacks the phosphorylase enzyme in his or her muscles, the muscles will still perform perfectly normally when the individual is at rest, since the muscle can derive energy from the glucose in the blood stream, as shown in Figure 10.1. As soon as the individual starts to carry out more energetic exercise, however, the glucose from the blood will be soon exhausted. In these circumstances, a normal person will derive more glucose by breaking down their stores of glycogen. A person who cannot do this, however, will soon be overcome with exhaustion accompanied by pains and cramps in the muscles. The first account of the disease was published by McArdle in 1951. McArdle had carried out extensive investigations at Guy’s Hospital in London of a patient George W., a 30-year-old man, who had suffered all his life from fatigue, and muscle stiffness and pain on exertion (Clarke, 2017, pp. 287–8). It might have been suspected that this condition was psycho-somatic in origin, since a physical examination showed that George W. was normal in most respects. However, McArdle noticed some subtle biochemical abnormalities which suggested that the condition had a physical basis. One of these abnormalities was that George W.’s level of lactate, a breakdown product of glucose, was significantly lower during exercise than that of a normal person. These and other anomalies led McArdle to the conclusion given in the title of his paper, namely that this was a “myopathy due to a defect in muscle glycogen breakdown.” So far, McArdle’s conclusion is still accepted, but he made a mistake regarding the exact causal mechanism involved. He speculated that the problem was a defect in glyceraldehyde phosphate dehydrogenase (GPD) rather than an absence of phosphorylase. The next important step in the investigation of the disease was triggered by the discovery of another patient with symptoms similar to those of George W. This was D.G., a 19-year-old youth of Scotch-Irish descent, who was investigated by Mommaerts and his group at Los Angeles in the late 1950s. Mommaerts et al. (1959) went further than McArdle by taking a muscle sample by biopsy from D.G.’s right thigh. They found (1959, pp. 792–3) that this sample contained much more glycogen than a normal human muscle. This showed that the usual breakdown of glycogen to glucose was not occurring in D.G.’s case. Moreover (1959, p. 793): “the pathological muscle contained no detectable amounts of either phosphorylase a or b.” The failure in glycogen breakdown was due to the absence of this crucial enzyme. Further Mommaerts et al. discovered (1959, p. 795): “that other enzymes associated with phosphorylase are present in normal amounts.” They concluded therefore that this was: “a metabolic disorder which is referable to the absence of a single enzyme.” Let us now return from this historical example to the philosophical issue. We have here an example of a disease whose cause was discovered and came to be generally accepted purely on the basis of evidence of mechanism. No statistical evidence was collected and brought to bear on the question of McArdle disease. Moreover, it would have been impossible to obtain any relevant statistical evidence for the simple reason that McArdle disease is so rare. Clarke gives some current estimates of the frequency of the disease in footnote 28 on p. 95 of his 2011 work.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:18:38.

Objections to the Russo–Williamson thesis  167

A case of McArdle disease occurs in between one in a million and one in eight hundred thousand live births, and only 55 individuals with McArdle disease are currently known in the UK. This might just be enough to obtain some statistical evidence, but at the time when McArdle disease was identified, and its cause came to be generally accepted, no more than three patients with the disease were known. Here then we have acceptance of causality solely on the basis of evidence of mechanism without any statistical evidence from human populations; and so the Russo– Williamson thesis seems to break down. Moreover the same is likely to apply to any very rare disease. Reliable statistics often cannot be collected for such diseases, and so any knowledge about them must depend on evidence of mechanism. This argument does not, however, lead Clarke to reject the RWT, but rather to suggest a modification of the thesis. He writes (2011, p. 99):

Copyright © 2018. Routledge. All rights reserved.

We could take McArdle disease as … an exception to the RWT. This is by no means out of the question. But I think a more conceptually satisfactory alternative is to make a slight modification to the RWT to save this, and similar cases. This is to modify …  by substituting ‘evidence of differencemaking’ for ‘probabilities’. So, applied to our formulation, the RWT now becomes: a causal hypothesis in medicine can be established only by using both statistical evidence (or difference-making evidence in the case of very rare diseases), and evidence of mechanism. But what is this “difference-making evidence”? This is the next question to which must be examined. My own suggestion is that we should take difference-making evidence, in cases like McArdle disease, to consist in making an intervention on the basis of the postulated mechanism which successfully ameliorates the patient’s symptoms. Now, in addition to the obviously mechanistic evidence obtained from a biopsy, Mommaerts et al. did make just such an intervention, which I will now describe. To test D.G.’s muscle weakness, Mommaerts et al. used walking on a treadmill inclined at ten degrees at the rate of four miles per hour. A normal subject can walk for at least an hour at this pace. D.G., who incidentally co-operated most willingly and helpfully in all the experiments, produced a very different outcome. As Pearson et al. say (1961, pp. 504–5): The fatigue end point was very distinct in the patient. It usually came on suddenly and was manifested by very marked weakness of the thigh and pelvic muscles, leading to near collapse on many occasions. …  At 4 miles per hour …  there was a reproducible end point with complete fatigue always by the sixth minute, with an average of 4.2 minutes and a range of three to six minutes. On the basis of their model of the disease, Mommaerts et al. decided that if they could infuse glucose in D.G.’s blood stream, this would prevent the early onset of Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:18:38.

168  Objections to the Russo–Williamson thesis

fatigue. They therefore fitted D.G. with an intravenous infusion of 10% glucose. They then carried out eight tests on the treadmill while glucose was infused at a constant rate after a 20-minute priming infusion. The results were as follows (Pearson et al., 1961, p. 506): Marked improvement in ability to exercise was apparent in all tests …  , with a range of from thirty-seven minutes to more than sixty minutes (a fifteen-fold improvement). The exercise period was arbitrarily discontinued at sixty minutes on three occasions when it was apparent that the subject was still walking easily. So, this intervention eliminated D.G.’s customary fatigue, and brought his exercise performance almost up to the normal level. We see then that Mommaerts et al., in their investigation of McArdle disease, used not only the evidence of mechanism obtained from the biopsy, but also the evidence of an intervention, based on their model of the disease, which greatly ameliorated the patient’s condition. It is this second type of evidence which I propose to call: “difference-making evidence”, and the suggested modification of the RWT is that this type of evidence can substitute the more usual statistical evidence from human populations in the case of very rare diseases, where there are not enough patients to obtain statistical results. That concludes my discussion of mechanistic counter-examples to the RWT in which a causal hypothesis in medicine apparently comes to be accepted solely on the basis of evidence of mechanism without using statistical evidence from human populations. I now turn, in the next three sections, to the other kind of counter-example (a statistical counter-example), in which a causal hypothesis in medicine is apparently accepted solely on the basis of statistical evidence in human populations in a situation where there is no plausible mechanism linking cause to effect.

Copyright © 2018. Routledge. All rights reserved.

10.3  The Semmelweis case Both Broadbent (2011) and Howick (2011a and b) give statistical counterexamples to the Russo–Williamson thesis. Broadbent mentions Semmelweis and Snow, writing (2011, p.57): The opponents of Snow and Semmelweis are generally considered to have been unreasonable to doubt the extremely convincing evidence for a causal link; this is difficult to explain if warrant for causal inference requires the identification of a mechanism. Of course, in these cases, a mechanism was eventually discovered. But …  the discovery came after the causal hypothesis was well-established. Howick also mentions Semmelweis explicitly. He writes (2011b, p. 930) of the dangers with requiring mechanistic reasoning alongside comparative clinical studies to establish causation. Countless mothers and babies would Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:18:38.

Objections to the Russo–Williamson thesis  169

have been saved had Semmelweis’s intervention been adopted after the results from his comparative clinical study became available. He goes on to say (2011b, p. 930):

Copyright © 2018. Routledge. All rights reserved.

Moreover, there are many counterexamples where medical interventions have been accepted on the basis of evidence from comparative clinical studies alone. Howick goes on to give a list of three counter-examples, and refers also to his book (2011a), where on pp. 131–2, there is a list of eight counter-examples which includes Snow’s work on cholera. In fact, the Snow case is in all essential respects the same as the Semmelweis case. So, I will now deal in detail with Semmelweis. Then in section 10.5 I will deal more briefly with Howick’s other suggested counter-examples. I will begin by giving a brief account of Semmelweis’ work on puerperal fever. This is based on my longer treatment in Gillies (2005b), which also contains references to some of the considerable literature on Semmelweis. Ignaz Semmelweis (1818–1865) carried out his research between 1 July 1846 and 20 March 1849. During most of this time he was assistant at the first maternity clinic in Vienna. The General Hospital of Vienna contained a Lying-In or Maternity Hospital as part. This was enormous by contemporary standards. By the 1850s and 1860s it catered for about 7,000 to 8,000 patients a year compared with around 200 to 300 in the General Lying-In Hospital in London. When Semmelweis began his research, one of the worst risks of childbirth was contracting a terrible disease, now fortunately eliminated, called puerperal fever because it normally struck during the puerperium, or approximately six weeks after childbirth when the womb returns to its normal shape. A woman who contracted puerperal fever would become feverish, then delirious with agonizing pains, and would often die within a few days. The Vienna Maternity Hospital was divided into two clinics from 1833. Patients were admitted to the two clinics on alternate days thereby producing, unintentionally, a system of random allocation. Between 1833 and 1840, medical students, doctors and midwives attended both clinics, but, thereafter, although doctors went to both clinics, the first clinic only was used for the instruction of medical students who were all male in those days, and the second clinic was reserved for the instruction of midwives. When Semmelweis began working as an assistant in 1846, the mortality statistics showed a strange phenomenon. Between 1833 and 1840, the death rates in the two clinics had been comparable, but, in the period 1841–46, the death rate in the first clinic was 9.92% and in the second clinic 3.88%. The first figure is more than 2.5 times the second – a difference which is certainly statistically significant. The quoted figures actually underestimate the difference since some severe cases of puerperal fever were removed from the first clinic to the general hospital where they died – thereby disappearing from the first clinic’s mortality statistics. This rarely happened in

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:18:38.

Copyright © 2018. Routledge. All rights reserved.

170  Objections to the Russo–Williamson thesis

the second clinic. Semmelweis was puzzled and set himself the task of finding the cause of the higher death rate in the first clinic. The first hypothesis considered by Semmelweis was that the higher death rate in the first clinic was due to “atmospheric-cosmic-terrestrial” factors. This sounds strange but is just a way of referring to the miasma theory of disease, which was standard at the time. According to this theory, disease was due to a putrid atmosphere or miasma. Pettenkofer used miasmas as part of his account of cholera of 1884, and we gave an account of the miasma approach in the course of expounding this theory of Pettenkofer in section 3.5. However, Semmelweis pointed out that miasma theory could not explain the different mortality rates in the first and second clinics. These were under the same roof and had an ante-room in common. So, they must be exposed to the same “atmospheric-cosmic-terrestial” influences. Yet the death rates in the two clinics were very different. The next hypothesis was that overcrowding was the key factor, but this too was easily refuted since the second clinic was always more crowded than the first, which, not surprisingly had acquired an evil reputation among the patients, almost all of whom tried to avoid it. In this sort of way Semmelweis eliminated quite a number of curious hypotheses. One concerned the appearance of a priest to give the last sacrament to a dying woman. The arrangement of the rooms meant that the priest, arrayed in his robes, and with an attendant before him ringing a bell had to pass through five wards of the first clinic before reaching the sickroom where the woman lay dying. The priest had, however, direct access to the sickroom in the case of the second clinic. The hypothesis then was that the terrifying psychological effect of the priest’s appearance debilitated patients in the first clinic, and made them more liable to puerperal fever. Semmelweis, however, persuaded the priest to come by a less direct route, without bells and without passing through the other clinic rooms. This change was introduced, but had no effect on the difference in mortality rates between the two clinics. After trying out these and other hypotheses unsuccessfully, Semmelweis was in a depressed state in the winter of 1846–7. However, a tragic event early in 1847 led him to formulate a new hypothesis. On 20 March 1847, Semmelweis heard with sorrow of the death of Professor Kolletschka. In the course of a post-mortem examination, Professor Kolletschka had received a wound on his finger from the knife of one of the students helping to carry out the autopsy. As a result, Kolletschka died not long afterwards of a disease very similar to puerperal fever. Semmelweis reasoned that Kolletschka’s death had been owing to cadaverous matter entering his bloodstream. Could the same cause explain the higher death rate of patients in the first clinic? In fact, professors, assistants and students often went directly from dissecting corpses to examining patients in the first clinic. It is true that they washed their hands with soap and water, but perhaps some cadaverous particles still adhered to their hands. Indeed, this seemed probable since their hands often retained a cadaverous odour after washing. The doctors and medical students might then infect some of the patients in the first clinic with these cadaverous particles, thereby giving

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:18:38.

Copyright © 2018. Routledge. All rights reserved.

Objections to the Russo–Williamson thesis  171

them puerperal fever. This would explain why the death rate was lower in the second clinic, since the student ­m idwives did not carry out post-mortems. In order to test this hypothesis, from sometime in May 1847 Semmelweis began to stipulate that everyone wash their hands in disinfectant before making examinations. At first, he used chlorina liquida, but, as this was rather expensive, chlorinated lime was substituted. The result was dramatic. In 1848 the mortality rate in the first clinic fell to 1.27%, while that in the second clinic was 1.30%. In fact, in the period 1848 to 1858, during which Semmelweis’ antiseptic precautions continued to be applied, there was no statistically significant difference in the death rates of the first and second clinics. Semmelweis developed his theory by claiming that it was not just cadaverous particles, but any kind of decaying animal-organic matter which caused puerperal fever. Such matter was normally conveyed to the patients by the hands of doctors, and, if it entered the blood stream through a lesion, would produce the disease, which Semmelweis considered to be an infection of the blood. Puerperal fever could therefore be prevented by the doctors washing their hands in disinfectant before examinations, and Semmelweis demonstrated statistically that this was the case. Anyone, nowadays, who reads Semmelweis’ own account of his investigations (Semmelweis, 1861), and studies the statistics which he uses to support his case, is likely to think that he has produced very strong empirical confirmation for his causal hypothesis. But now comes the surprise. The great majority of Semmelweis’ contemporaries did not accept his hypothesis, or adopt his antiseptic precautions. Why was this? In an earlier work, I have considered the various theories which have been put forward to explain Semmelweis’ failure with his contemporaries (Gillies, 2005b). I opted for a Kuhnian approach. The idea is that in the 1840s, medicine was dominated by what could be called the miasma-contagion paradigm. According to this, diseases were caused either by miasmas or by contagion, or by some combination of the two. This remained the dominant paradigm until it was replaced by the germ theory of disease. The germ theory was first introduced in the early 1860s and became generally accepted by the late 1880s. As we saw in section 3.5, Pettenkofer’s theory of cholera of 1884 was to some degree a compromise, since it involved both the miasma-contagion framework, and some ideas from the germ theory of disease. However, by the 1890s, the old miasma-contagion paradigm had effectively been eliminated. In the 1840s, however, the miasmacontagion paradigm was accepted by nearly all doctors; but Semmelweis’ theory of puerperal fever completely contradicted this paradigm. As we have seen, the very first hypothesis which Semmelweis considers is that puerperal fever is caused by a miasma, and he rejects this hypothesis. Semmelweis also states quite explicitly that, in his view, puerperal fever is not caused by a contagion. As he says (1861, p. 117): Childbed fever is not a contagious disease. A contagious disease is one that produces the contagion by which the disease is spread. This contagion brings about only the same disease in other individuals. Smallpox is

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:18:38.

172  Objections to the Russo–Williamson thesis

a contagious disease because smallpox generates the contagion that causes smallpox in others. Smallpox causes only smallpox and no other disease. …  Childbed fever is different. This fever can be caused in healthy patients through other diseases. In the first clinic it was caused by a discharging medullary carcinoma of the uterus, by exhalations from a carious knee, and by cadaverous particles from heterogeneous corpses. …  However, childbed fever cannot be transmitted to a healthy maternity patient unless decaying animal-organic matter is conveyed.

Copyright © 2018. Routledge. All rights reserved.

So, Semmelweis’ views as to the causation of puerperal fever contradicted the two dominant theories of disease held at the time – the miasma theory and the contagion theory. It is not so surprising, therefore, given a Kuhnian model of science, that Semmelweis’ theory was not accepted. Between the 1840s and the 1890s, however, a revolution in medicine occurred in which the miasma-contagion paradigm was replaced by the germ theory of disease. According to the germ theory, puerperal fever was caused by a bacterial infection. Now Semmelweis’ cadaverous particles and other decaying animalorganic matter are good sources of pathogenic bacteria, and so, within this new paradigm, Semmelweis’ theory becomes very plausible. That concludes my brief historical account of Semmelweis’ researches, and of the problems they raise. Let us now turn to considering how this example relates to the Russo– Williamson thesis. As we have seen, Broadbent (2011) and Howick (2011a and b) regard the Semmelweis case as a counter-example to the Russo–Williamson thesis (RWT). Surprisingly, however, Russo and Williamson in their 2007 mention the Semmelweis case and argue that it supports their thesis. They say (2007, p. 162): The history of medicine presents many cases in which causal claims made solely on the basis of statistics have been rejected until backed by mechanistic or theoretical knowledge. …  Semmelweis’ claim about cadaverical contamination and puerperal fever was accepted only after the germ theory of disease was developed. I agree with Russo and Williamson that the historical sequence is indeed in accordance with the RWT. However, I don’t think, for that reason, that the RWT was implicitly accepted by the medical community at the time. Semmelweis was avant-garde in his use of statistics, as well as in most other things. As we saw in section 8.2, Claude Bernard, a leading medical researcher, stated in 1865 that statistical evidence was of no value, and even Koch in the 1880s and 1890s seemed to prefer evidence of mechanism to statistical evidence. This is perhaps another reason for Semmelweis’ failure. Many of the medical community at the time (the late 1840s) had a low opinion of the value of the kind of statistical evidence on which Semmelweis relied. Instead of statistical evidence being treated on a par with evidence of mechanism (as in the RWT), evidence of mechanism was

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:18:38.

Copyright © 2018. Routledge. All rights reserved.

Objections to the Russo–Williamson thesis  173

definitely prioritized at the time. Semmelweis did indeed propose a mechanism to explain how puerperal fever occurred. Cadaverous particles, or other decaying animal-organic matter, were conveyed on the hands of doctors to the patients. These particles entered the patient’s blood stream through lesions associated with childbirth, and hence produced the disease. However, as we have seen, this mechanism was highly implausible given the dominant paradigm of medicine at the time. As the community relied mainly on evidence of mechanism, this explains why they rejected Semmelweis’ theory. But why then do Broadbent (2011) and Howick (2011a and b) regard the Semmelweis’ case as a counter-example to the RWT? They do not regard the RWT as contradicting the history, but they argue that the RWT fails as a normative, as opposed to historical, thesis. Yes, the medical community did reject Semmelweis’ theory, but they should not have done so. They should have given more weight to statistical evidence and accepted Semmelweis’ theory on the basis of that evidence. If they had done so, many lives would have been saved and medical science would have developed more rapidly. Thus, Broadbent and Howick recommend that, to avoid disasters like the Semmelweis episode in future, medical researchers should reject the RWT, and be prepared to accept causal hypotheses in medicine solely on the basis of statistical evidence. The problem with this recommendation, however, is that, while it deals satisfactorily with the Semmelweis case, it gets into difficulties with the streptomycin case which we considered in Chapter 9. There is a certain parallel between the Semmelweis and streptomycin cases. In both cases the statistical evidence strongly supported the causal claim that the suggested intervention (antiseptic handwashing in one case, taking streptomycin in the other) would be effective in preventing or curing the disease. The RCT showed that the stretomycin group did significantly better than the control group. Semmelweis’ antiseptic handwashing reduced mortality significantly in the first clinic. However, in both cases one of the dominant paradigms in medicine cast doubt on whether the suggested mechanism would be effective. The miasma-contagion paradigm made Semmelweis’ mechanism look very implausible. The Darwinian evolution paradigm made it probable, given the length of time needed for streptomycin to kill tubercle bacilli, that some of these bacilli would become resistant and produce a relapse in the patients over a longer period than the RCT. Now it turned out that the doubts based on the miasma-contagion paradigm were unjustified, while those based on the Darwinian evolution paradigm were entirely justified. Why was this? Obviously, the explanation, from the modern point of view, is that the miasma-contagion paradigm was incorrect, while the Darwinian evolution paradigm was correct. Now we know that any dominant paradigm in any branch of science might be incorrect in some respects. Kuhnian revolutions, which overthrow well-established paradigms and replace them by new ones, do occur regularly, though not very frequently, in science. Does this mean that it is wrong to rely on the established paradigm when applying a particular branch of science? Surely not. The established paradigm gives the results which are best

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:18:38.

174  Objections to the Russo–Williamson thesis

confirmed empirically at the moment, even though they may be overturned in the future. This line of argument is reassuring in that it suggests that the Semmelweis debacle is likely to be a rare case. A similar episode would only occur in the run-up to a scientific revolution, and scientific revolutions are infrequent. Normally reliance on the established paradigm of the field is perfectly justified, and likely to produce good results. Nonetheless, however many excuses we can find for the medical community of the 1840s and 1850s, it remains true that they made a mistake as regards Semmelweis’ theory, and this naturally raises the question of whether similar mistakes can be avoided in the future. I think that a second modification of the RWT would help to avoid something similar to the Semmelweis case occurring again. It will help the formulation of this modification if we consider another related example – that of alternative medicine. I will do this in the next section.

Copyright © 2018. Routledge. All rights reserved.

10.4  Alternative medicine Alternative medicine is by no means a unified field. It contains many different approaches, such as aromatherapy, homeopathy, traditional Chinese medicine, traditional Indian medicine (Ayurveda), and so on. Here I will concentrate on just one example: acupuncture. This is based on the traditional Chinese theory of qi (a kind of energy). Qi (pronounced “chee”) flows through a system of channels in the body, known in English translation as “meridians”. These meridians intersect and form a web-like structure in the body which connects the principal organs. Sometimes the flow of qi becomes unbalanced, causing pain and illness. In acupuncture fine needles are inserted in carefully defined points of the body in order to correct disruptions in the harmony of the qi. These insertion points are located at important points on the web of meridians. Unfortunately, qi theory completely contradicts modern scientific medicine. For example, the position of the meridians on which acupuncture is based, does not agree at all closely with the paths of nerves as accepted by standard anatomy and physiology. Moreover, there is no empirical evidence which confirms qi theory, so that this theory has to be regarded as speculative and metaphysical. Despite its unscientific basis, however, it could be that treatments using acupuncture do help to improve the patient’s condition. As a result, attempts have been made to test the effectiveness of acupuncture. A good account of these tests and their results is to be found in Kaptchuk (2002). Kaptchuk writes (2002, p. 376): Since the early 1970s, when acupuncture began to capture the popular imagination of the West, almost 500 randomized, controlled trials (RCTs) have evaluated its efficacy. Kaptchuk gives some of the results of these trials (pp. 378–9). It seems that RCTs definitely show that acupuncture is effective for adult postoperative and Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:18:38.

Objections to the Russo–Williamson thesis  175

chemotherapy nausea, and for acute dental pain. Some studies indicate that acupuncture gives relief of pain on other diverse conditions, but the evidence here is a bit inconclusive. The situation here is very similar to that of the Semmelweis case. To begin with, the interventions in both cases are very unlikely to cause any harm to the patient, provided acupuncture is performed with carefully sterilized needles. In both cases, there is statistical evidence in favour of the proposed interventions. There are some RCTs, which show that acupuncture is effective for particular conditions. However, in neither case were the mechanisms proposed to explain why the interventions worked at all plausible relative to the dominant paradigms of the time. The traditional mechanisms used to explain the effectiveness of acupuncture, involving as they do the theory of qi and meridians, completely contradict contemporary paradigms of scientific medicine. Such mechanisms must therefore be regarded, if we assume the truth of such paradigms, as highly implausible. What is to be done in this situation? What in practice was done was to try to explain the effectiveness of acupuncture in terms not of qi and meridians, but using the standard theories of contemporary scientific medicine. Kaptchuk states that extensive research has been carried out into the pain-killing effects of acupuncture, and that this has shown that they may be initiated (2002, p. 379):

Copyright © 2018. Routledge. All rights reserved.

by stimulation, in the muscles, of high-threshold, small-diameter nerves. These nerves are able to send messages to the spinal cord and then activate the spinal cord, brain stem (periaqueductal gray area), and hypothalamic (arcuate) neurons, which, in turn trigger endogenous opioid mechanisms. These responses include changes in plasma or cortico-spinal fluid levels of endogenous opioids (for example endorphins and enkephalins). This extensive research included animal experiments in which the effects of acupuncture in one rabbit were transmitted to another rabbit by cerebrospinal fluid transfusions. Moreover, it was shown that the pain-killing effects acupuncture could be reversed by administrating an endorphin antagonist in a dose-dependent manner. In effect, the claim, based on this research, is that the pricks of the acupuncture needles stimulate the body to produce its own natural pain-killers (endorphins and enkephalins), and that this account for the pain-killing effects of acupuncture. All this provides strong evidence of the mechanism of the pain-killing effects of acupuncture of a kind which is completely compatible with the paradigms of contemporary scientific medicine. In fact, Kaptchuk argues that this evidence of mechanism is more important than the results of the RCTs in rendering acupuncture acceptable to the contemporary medical community. As he says (2002, p. 379): Numerous surveys show that, of all the complementary medical systems, acupuncture enjoys the most credibility in the medical community …  . Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:18:38.

176  Objections to the Russo–Williamson thesis

Copyright © 2018. Routledge. All rights reserved.

The RCT research is probably not the main basis for this positive ­opinion. A more likely reason is the existence of a substantial body of data showing that acupuncture in the laboratory has measurable and replicable physiologic effects that can begin to offer plausible mechanisms for the presumed actions. I have now given three historical episodes bearing on the Russo–Williamson thesis (RWT). These are the streptomycin case (Chapter 9), the Semmelweis case (section 10.3) and the case of acupuncture (section 10.4). I will now give my second suggested modification of the RWT designed to take account of all three cases. It runs as follows: Suppose we have strong statistical evidence in favour of the efficacy of a treatment, but that the only proposed treatment mechanisms are very implausible given accepted background knowledge. Suppose further that there is good evidence that the treatment will not harm the patient. Then, despite the RWT, it is justified to adopt the treatment provisionally, provided that research is carried out into possible mechanisms of the treatment and that the treatment is modified if necessary in the light of this research. This modification covers all three cases. In the streptomycin case, Darwinian evolution applied to bacteria, suggested that some of the tubercle bacilli would develop resistance to streptomycin given the length of time that streptomycin needed to kill the bacteria. Research was carried out into the treatment mechanism, which revealed that some tubercle bacteria were indeed developing resistance to streptomycin. This suggested modifying the treatment by combining streptomycin with PAS, and this new treatment proved effective. In the Semmelweis case, if this modification had been followed, then his antiseptic hand-washing would have been adopted in hospitals and many lives saved. At the same time, research into his suggested treatment mechanism would have hastened the development of the germ theory of disease. Thirdly, in the acupuncture case, the modification was followed, and this resulted in the discovery of a mechanism for the pain-killing effects of acupuncture which was compatible with contemporary scientific theory, and led to the acceptance of acupuncture for some conditions by the medical community. Thus, the suggested modification of the RWT seems satisfactory, but, there turns out to be a difficulty about implementing it in the Semmelweis case – always the most problematic case. The point to note is that the research in the cases of streptomycin and acupuncture was quite compatible with the dominant paradigms accepted by the medical community. However, the research required for the Semmelweis case would have involved criticizing the dominant disease paradigm of the time – the miasma-contagion paradigm. Now it is always difficult to gain funding for research which contradicts the dominant paradigm. This was true in Semmelweis’ day, and is even more true today. In fact, the most popular research funding systems of the present almost guarantee that any research which contradicts a paradigm dominant in the research community, will not be

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:18:38.

Objections to the Russo–Williamson thesis  177

funded. Usually research funding is allocated by a competition between rival projects in which the outcome is judged by peer review. However, if a project contradicts the dominant paradigm, then most peer reviewers, who accept this paradigm, will judge that it is unacceptable, and the project will not be funded. Almost all Semmelweis’ contemporaries were opposed to his ideas, and would have assessed any research project he proposed negatively. So Semmelweis would not have got any research funding. This poses a serious problem about research funding. This is obviously not the place to tackle it in detail, but the interested reader can explore the issue further in my 2008 book and a later 2014 paper, the Semmelweis case is discussed in relation to the research funding problem in the former (Gillies, 2008, pp. 21–5).

10.5  Other statistical counter-examples

Copyright © 2018. Routledge. All rights reserved.

I will now consider, as promised earlier, Howick’s list of counter-examples to the Russo–Williamson thesis (RWT). Howick begins his critical discussion of the RWT in 2011a, p. 130. On p. 131, he mentions the Semmelweis case, and then goes on (2011a, pp. 131–2) to give his full list of statistical counter-examples to the RWT as follows: there are many … examples where treatments were widely accepted before any semblance of a mechanism was established. To name a few, Percival Pott’s hypothesis that soot caused scrotum cancer (1775) was accepted years before benzpyrene was identified (1933). Edward Jenner introduced smallpox vaccines (1798) decades before anyone really understood how they worked. John Snow helped eliminate cholera with cleaner water (1849) years before the Vibrio cholerae was identified (1893), and Carlos Finlay reduced the rates of yellow fever by killing mosquitoes (1881) decades before flavivirus was identified (1927). In the last century, general anaesthesia, aspirin, and the steroids were widely used for decades before their mechanisms were understood. In this century, deep brain stimulation has been used to suppress tremors in patients with advanced Parkinson’s disease, and also to cure other motor function disorders such as dystonia or Tourette’s syndrome, yet researchers have not been able to identify its mechanism of action with any certainty. Earlier I discussed the Semmelweis case in some detail, but Howick’s list here is so long that there is not space enough to discuss each example in the same detail. That would require a book in itself! What I will do therefore is begin by making two general points, and then say a few things about some of the examples in the list. This will give a reasonable idea of how I would deal with each of Howick’s alleged counter-examples. The first of my two general points is that, in giving historical counter-examples to the RWT, one must beware of being anachronistic. A statistical counter-example

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:18:38.

178  Objections to the Russo–Williamson thesis

to the RWT, as I have formulated it, consists of a case in which a causal hypothesis in medicine was accepted purely on the basis of statistical evidence without any evidence of mechanism. Now, if one chooses examples rather far back in the history of medicine, then one can find hypotheses which were indeed accepted, but not on the basis of statistical evidence, because statistical evidence was either not collected at all at the time, or, if some statistical evidence was collected, it did not carry much weight with the contemporary medical community. What led the medical community to accept a causal hypothesis in the pre-statistical era? I will illustrate this by giving an example from Robert Burton’s The Anatomy of Melancholy. This was first published in 1621, with subsequent editions in 1624, 1628, 1632, 1638, 1651–2, 1660, and 1676. It deals with “melancholy” which is approximately the same as what we now call “depression”, though the 17th-century term covered a somewhat broader range of conditions. It is a very learned work in which Burton cites almost all the leading contemporary medical authorities, and his work became a standard text on the condition for most of the 17th century. Burton discusses the causes of melancholy and also cures for the condition, including pharmaceutical cures. Among the latter he lists Melissa balm, about which he says (Burton, 1621, p. 289):

Copyright © 2018. Routledge. All rights reserved.

Melissa balm hath an admirable virtue to alter melancholy, be it steeped in our ordinary drink, extracted, or otherwise taken. Cardan, lib.8. much admires this herb. It heats and dries, saith Heurnius, in the second degree, with a wonderful virtue comforts the heart, and purgeth all melancholy vapours from the spirits, Matthiol. in lib. 3. cap. 10. In Dioscoridem. The justification for the effectiveness of Melissa balm in treating melancholy consist of an appeal to three medical authorities: Cardan, Heurnius, and Matthiol. Presumably they had used Melissa balm in their practice and thought that it had worked. So we might call this anecdotal, but not statistical evidence. There is also a hint here of evidence of mechanism, but this is hardly the same as evidence of mechanism as we know it today. The background theory of the time was that of the four humours which existed in the body, namely blood, yellow bile, black bile, and phlegm. These humours had properties described by the two Aristotelian dichotomies: hot/cold, and dry/wet. Thus blood was hot and wet, yellow bile was hot and dry, black bile was cold and dry, and phlegm was cold and wet. Illnesses were cause by an imbalance of the humours. Melancholy, in particular, was the result of black bile. One way of curing melancholy was to convert black bile (cold and dry) into yellow bile (hot and dry), and this could be done by a herb which “heats and dries” as Melissa balm was supposed to do. The theory of the four humours had been accepted as the basis of European medicine since ancient Greek times. It is a traditional theory like the ancient Chinese theory of qi and meridians, and like the traditional Chinese theory is metaphysical and speculative in character, and was never systematically tested and confirmed empirically. If a mechanism is plausible in the light of the theory of the four humours, or of the theory of qi and the meridians, we cannot say

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:18:38.

Objections to the Russo–Williamson thesis  179

it is supported by any empirical evidence. It is very different if a mechanism is plausible in the light of modern theories of anatomy and physiology. These are very well corroborated empirically, and so give genuine empirical support to any mechanism which accords well with them. It is clear from all this that to try to apply the RWT to cures which were accepted by leading experts in the 17th century would be simply anachronistic. The 17th century medical community simply did not have the concepts of statistical evidence and evidence of mechanism which we use today, and which are involved in the RWT. When did the use of statistical evidence begin in medicine? Ulrich Trö h ler has shed light on this question in his valuable 2000 book: “To Improve the Evidence of Medicine.” The 18th Century British Origins of a Critical Approach. Trö h ler shows that there was a group of British doctors in the 18th century who collected medical statistics, and advocated the use of such statistics in assessing claims in medicine. A typical member of this group was John Millar, a Scot who was associated with the Westminster Dispensary in London. Millar wanted to collect records of the success and failure of cures at the dispensary, and indeed appointed a clerk entrusted with this task “so that nothing might depend on the testimony of physicians” (Trö h ler, 2000, p. 32). Millar wrote in a book published in 1777 (quoted from Trö h ler, p. 32): detached cases, however numerous and well attested are insufficient to support general conclusions; but by the recording every case in a public and extensive practice, and comparing the success of various methods of cure with the unassisted efforts of nature some useful information may be obtained; … 

Copyright © 2018. Routledge. All rights reserved.

This is a most impressive defence of medical statistics in theory and in practice, but, as Trö h ler goes on to say, Millar did not succeed in getting his views generally accepted in the medical community. Millar was attacked by Donald Monro and Sir John Pringle, and (Trö h ler, 2000, p. 35): In 1778, opposition also arose against the whole system of record keeping at the dispensary and Millar lost his clerk …  Millar’s experience seems to have been that of most of these early British pioneers of the use of statistics in medicine. They had some success in the army and navy, but, as Trö h ler remarks (2000, p. 107): If … I was to take examples from among famous London consultants, their emphasis would be on the superior value of direct clinical observation and their own subjective experience derived therefrom. The advocates of the use of medical statistics were mainly Scots, Quakers or Unitarians, in effect dissenters, and they remained marginal. Their views had Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:18:38.

180  Objections to the Russo–Williamson thesis

little impact on the Royal College of Physicians, whose fellows were restricted to graduates of Oxford and Cambridge alone (Trö h ler, 2000, p. 119). On the practical side, there continued to be “opposition to open publication of the results of hospital treatment from around 1780 until the 1820s” (Trö h ler, 2000, p. 117). The next milestone in the development of medical statistics was PierreCharles-Alexandre Louis’s statistical investigation of the effects of bloodletting. Louis published a paper on this in 1828, and expanded the paper into a book published in 1835. This was translated into English in 1836. It has sometimes been claimed that Louis’ scientific approach led to the demise of bloodletting, but this myth is exposed by Carter (2012) in his treatment of the subject. As Carter says (2012, p. 25):

Copyright © 2018. Routledge. All rights reserved.

Louis …  found that patients who were treated early on, recovered, on average, three days sooner than those not bled until later. This result, far from discrediting bloodletting, actually supports its use – at least if blood is taken in a timely manner. Carter also (2012, p. 25) quotes Louis as saying that bloodletting “should not be neglected in inflammations that are severe and are seated in an important organ.” Louis’ use of statistical evidence was taken seriously by some members of the medical profession in the 1830s; but, as was shown in section 8.2, even as late as 1865, serious medical researchers such as Claude Bernard, rejected statistical evidence as of no value. So the general use and acceptance of statistical evidence by the medical community cannot be dated much earlier than the middle of the 19th century, and possibly should not be considered as established till the 1890s. As the RWT has been here formulated here in terms of statistical evidence, it is clear that it cannot be applied until after the middle of the 19th century, and any claimed counter-example should be after that date. The middle of the 19th century can be taken as the approximate starting point of scientific medicine. Against this view, it could be objected that Vesalius (1514–64) published his famous work of anatomy in 1543 and William Harvey (1578–1657) published his theory of the circulation of the blood in 1628. These were surely scientific works, and appeared long before the middle of the 19th century. I agree of course that the works of Vesalius and Harvey were scientific, but they were works of anatomy and physiology, not of medicine. Theoretical medicine considers hypotheses about the causes of disease, and about cures of and ways of preventing diseases. Such investigations require a knowledge of anatomy and physiology, but these in themselves are branches of natural science and not medicine. It was a long time before the advances of Vesalius and Harvey in anatomy and physiology led to advances in medicine. Robert Burton (1576–1639) was a contemporary of Harvey’s, but, as we have seen, his account of herbal cures for melancholy can hardly be considered scientific. My second general point is that the requirement of evidence of mechanism in the RWT is not equivalent to the requirement of evidence for the exact

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:18:38.

Objections to the Russo–Williamson thesis  181

mechanism which is accepted today, still less of evidence for the complete ­mechanism involved. In section 4.4, I defined basic mechanisms in medicine as a sequence of causes C1→ C2→ C3→ … → Cn which describe some biochemical/ physiological processes going on in the body, and this is the definition of basic mechanism which has been used in all the subsequent examples. I also pointed out there that between any two of the causes Ci and Ci+1 say, it is always possible to insert a third intermediate cause C’ to give Ci →  C’ →  Ci+1. So no mechanism is ever complete, and it is always possible to develop it in more detail. In particular the evidence of mechanism when a causal hypothesis is accepted may be only evidence for a sketch of a mechanism which is elaborated in much more detail today. In the light of these two general points, let us consider briefly Howick’s list of counter-examples. The first two, dated 1775 and 1798, can be ruled out by our historical argument that genuine counter-examples to the RWT must occur after the middle of the 19th century, since it does not make much sense to apply the RWT before then. The case of Snow is also easily dealt with since it is so similar to that of Semmelweis. Snow had perhaps a bit more success than Semmelweis, but his recommendations were not accepted by the medical community as a whole until Koch’s paper of 1893. As late as 1884, Pettenkofer criticized not only Koch but also Snow. The reason for the delay in the acceptance of Snow’s work was that it contradicted the dominant miasma-contagion paradigm (see section 3.5). I have already discussed the problems which this poses in connection with Semmelweis. Howick writes (2011a, p. 132):

Copyright © 2018. Routledge. All rights reserved.

Carlos Finlay reduced the rates of yellow fever by killing mosquitoes (1881) decades before flavivirus was identified (1927). The mistake here is the second of my general points. It is to require evidence of the exact mechanism held to be correct today at the time when the causal hypothesis was first accepted. In 1881 the germ theory of disease was widely held in the medical community. Carlos Finlay had already accepted the theory by the time he graduated from Jefferson Medical College in Philadelphia in 1855, where he had been taught by two advocates of the germ theory. Given the germ theory, there was certainly a plausible mechanism for the transmission of a disease by mosquito bites. Such bites could cause the microbe which produced the disease to enter the patient’s blood stream. Thus there was support for the disease mechanism, even though the exact microbe causing the disease (the flavivirus) was only identified later. The situation was the same during Pasteur’s investigation of rabies. Pasteur was convinced, on the basis of the germ theory of disease, that rabies was caused by a microbe, but the microbe in question, being a very small virus, was only isolated later. General anaesthesia is easily dealt with because research into it belongs more properly to physiology than to medicine. Becoming anaesthetized is not an illness

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:18:38.

Copyright © 2018. Routledge. All rights reserved.

182  Objections to the Russo–Williamson thesis

or a disease, but a physiological reaction of a perfectly healthy human to certain substances. Of course, anaesthesia is used in the treatment of diseases, but then so are X-rays, and the study of X-rays belongs to physics not medicine. Similarly, the study of general anaesthesia belongs to physiology not medicine. As we saw earlier in comparing Harvey’s work on the circulation of the blood with Burton’s on the cures of melancholy, physiology became scientific long before medicine. The two are distinct. Now the RWT, as we have formulated it, applies to medicine, and indeed only to physical rather than psychological medicine. It does not apply to physiology. Aspirin was developed from a traditional herbal medicine – the bark and leaves of the willow tree. These chewed, or given as a powder or in an infusion, were considered to be a remedy for pain, fever and inflammation. The use of these herbs seems to date back to the ancient Babylonians and Egyptians, and certainly this herbal medicine is described by Hippocrates in the 5th century B.C. Willow bark and leaves appear in the pharmacopoeias of the European Middle Ages.The reasons for the acceptance of this herbal medicine, however, obviously had little to do with systematic testing and collection of statistics. They were probably like those which Robert Burton gave when he recommended Melissa balm as a cure for melancholy. It should be remembered, moreover, that while the traditional pharmacopoeias did contain a few remedies which are still recognized as effective, most of the herbal medicines listed were either useless or positively harmful. In the 19th century, developments took a more scientific turn when the active ingredients of willow bark and leaves were extracted and purified. These were salicin and the more powerful salicylic acid. These were also found in another herbal remedy – meadowsweet. The next step was to synthesize these compounds as they were difficult and expensive to extract from plants. Successful chemical synthesis of salicin, salicylic acid, and sodium salicylate was achieved by the 1860s. These medicines, however, had severe and unpleasant side effects including gastric irritation and sometimes stomach bleeding. The problem was to find a related medicine which did not have these side effects. It is here that consideration of mechanisms enters the picture. The chemist Felix Hoffmann had the idea that by adding a chemical group to salicylic acid, he could alter its shape and hence reduce its gastric irritation while retaining its ability to deal with pain, fever, and inflammation. He is said to have motivated in this research because his father was suffering the side effects of taking sodium salicylate for rheumatism. In 1897 Hoffmann found that by adding an acetyl group to produce acetylsalicylic acid, he could achieve his goal. Aspirin is just the common name for acetylsalicylic acid. So a consideration of mechanisms and evidence of the correctness of these mechanisms played a crucial role in the development of aspirin. This has perhaps been overlooked because it is forgotten that establishing a cure in medicine involves not just showing that the cure works, but also that it is safe and does not have harmful side effects. Evidence of mechanism is very important in establishing the safety of drugs as we saw in the analysis of thalidomide in section 9.6.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:18:38.

Objections to the Russo–Williamson thesis  183

Clarke and Russo discuss the case of aspirin in (2018). They make the point that (p. 320) “the analgesic, antipyretic, and anti-inflammatory effects are not the only effects of aspirin of interest to medical practice. …  regular low-dose aspirin improves cardiovascular outcomes.” Indeed, this second use of aspirin is now more common than the first. But how did this second use of aspirin come about? The answer given by Clarke and Russo was that it was based on the discovery that low-dose aspirin inhibited platelet function, and hence bloodclotting. Clarke and Russo give an account of how the mechanism, which produced aspirin’s antiplatelet action was established in (2018, pp. 320–1). The evidence of this mechanism is the basis for aspirin’s use in cardiovascular cases. The last item which I will consider on Howick’s list concerns a treatment for Parkinson’s disease. Howick writes (2011a, p. 132): In this century, deep brain stimulation has been used to suppress tremors in patients with advanced Parkinson’s disease … yet researchers have not been able to identify its mechanism of action with any certainty.

Copyright © 2018. Routledge. All rights reserved.

This example is analysed in some detail by Campaner, and it is worth quoting what she says about it (2012, pp. 74–5): During an operation, surgeons found by accident that partial destruction of a brain area suppressed tremors in a patient with Parkinson’s disease. …  Deliberate destruction of parts of the brain, though, is a very risky and irreversible operation. In the mid-Nineties researchers reported that there was a better option, namely electrical stimulation. …  The technique applied is known as DBS, and is targeted mainly at the subthalamic nucleus or the globus pallidus. The DBS system consists of three components: the electrode, the extension, and the neurostimulator. …  The tip of the electrode is positioned within the targeted brain area. …  Once the system is in place, electrical impulses are sent from the neurostimulator up along the extension wire and the electrode into the target area in the brain. The high-frequency stimulation produced by the electrodes causes a functional block of the anatomic structure: impulses interfere with and reduce the hyperactivity that causes Parkinson’s disease symptoms. DBS does not damage healthy brain tissue by destroying nerve cells, but simply blocks electrical signals from targeted areas in the brain. Reading this passage, it is clear that deep brain stimulation (DBS) is based on brain mechanisms which are supported by evidence. So this treatment is confirmed by evidence of mechanism and is hardly a statistical counter-example to the Russo– Williamson thesis (RWT). Howick may of course be right that our knowledge of the mechanism of action of DBS is far from complete, but, as I pointed out in the second of my general points earlier in this section, the RWT does not require evidence of the complete mechanism involved, which is, anyway, impossible to obtain.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:18:38.

184  Objections to the Russo–Williamson thesis

Copyright © 2018. Routledge. All rights reserved.

So I can sum up my discussion of Howick’s statistical counter-examples as follows.The Semmelweis case is his strongest counter-example, and this does require a modification of the RWT along the lines given in Section 10.4 of this chapter. Of his long list of further counter-examples, the Snow case is essentially the same as the Semmelweis case, but the further examples turn out on closer examination not to be genuine statistical counter-examples to the RWT. They do not therefore require any further modifications of the Russo–Williamson thesis.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:18:38.

11 DISCOVERING CURES IN MEDICINE AND SEEKING FOR DEEPER EXPLANATIONS

So far, we have discussed the usefulness of mechanisms in trying to establish causal claims in medicine, that is to say in trying to confirm empirically an already formulated causal hypothesis to a level where it is legitimate to use it in practice. However, mechanisms can be useful in other ways. This chapter will deal with two further uses. The first (section 11.1) is the usefulness of mechanisms in trying to discover cures of diseases or ways of preventing them, or at least ameliorating the symptoms. The second (section 11.2) is in seeking for deeper explanations of diseases.

Copyright © 2018. Routledge. All rights reserved.

11.1 The importance of mechanisms for discovering cures in medicine It is easy to see why the study of mechanisms is useful for discovering cures of diseases, as well as of ways of preventing diseases or at least ameliorating the symptoms. Suppose we have a disease D, which is caused by A. In our usual notation, this is written A →  D, where “→ ” stands for “causes”. Now on the basis of the causal hypothesis A →  D, we can seek a cure of D by stopping the causal transition from A to D, which can be done in turn by either trying to prevent A occurring or by blocking its usual effects. Now suppose we manage to establish that there is a linking mechanism M, such that A→ M→ D. As usual, I will take M to be a basic mechanism, that is a sequence of causes C1→ C2→ C3→  …  → Cn which describe some biochemical/physiological processes going on in the body. So we now have A→ C1→ C2→ C3→  …  → Cn→ D. Whereas previously our only research strategy was to work on stopping the causal transition from A to D, we now can work on trying to stop any of the n+1 transitions A to C1, C1 to C2, … , Cn-1 to Cn, Cn to D. If we can stop one or more of these causal transitions, we will have found a way of curing or preventing D. Naturally with a much-increased

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:18:52.

186  Discovering cures in medicine

number of research strategies to work on, our chance of succeeding becomes much greater. This is illustrated by an example which we gave in section 4.4. This was based on the causal hypothesis: Sheep grazing in anthrax fields →  some of them to die with symptoms of anthrax. As a result of the research of Koch and others, the cause and effect in this law were connected with the following linking mechanism (M4.1):

Copyright © 2018. Routledge. All rights reserved.

1. Anthrax spores enter the blood stream of some of the sheep through abrasions on their body or via the intestine →  2. Spores in the blood stream to turn back into anthrax bacilli →  3. These anthrax bacilli to multiply rapidly →  4. Anthrax bacilli to enter the capillaries →  5. Almost all the capillaries in the lungs, liver, kidneys, spleen, intestines, and stomach to be filled with enormous numbers of anthrax bacilli. On the basis of this kind of mechanism, Pasteur developed a successful vaccine against anthrax by attenuating anthrax bacilli. In effect, Pasteur blocked the step from 2 to 3. His vaccine had primed the immune systems of the sheep, so that when the first anthrax bacilli appeared, they were destroyed before they could multiply rapidly. Obviously, Pasteur was only able to discover a way of preventing anthrax by vaccination, because a linking mechanism, like that shown in steps 1 to 5 above, had been developed. In section 4.6, the example, due to Darden (2013), was considered of mutation in the CFTR gene causing cystic fibrosis. Here again, a mechanism linking “mutation in the CFTR gene” to “cystic fibrosis” was discovered (M4.2), and on the basis of this mechanism treatments for cystic fibrosis were developed which would have been unthinkable without knowledge of this mechanism. Another excellent example of the use of mechanisms to discover cures is given by Campaner in her 2012 work, Chapter 6, Section 2.1, pp. 86–91. This example is concerned with the development of treatments for AIDS (Acquired Immune Deficiency Syndrome). AIDS first appeared in 1981, and in 1984 and 1986 two hitherto unknown retroviruses HIV-1 and HIV-2 were isolated, and are now generally accepted as the cause of the disease. Somehow these viruses destroyed the immune system, and (Campaner, 2012, p. 87): During the Eighties the mechanism by which HIV destroys the immune system was gradually elucidated …  The virus’ favoured target is the T4 lymphocyte, damage to which in the immune system is responsible for the destruction of immune defences encountered in patients with AIDS. A simplified diagram of how an HIV virus infects a T4 lymphocyte, and uses it to multiply, is shown in Figure 11.1. Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:18:52.

Discovering cures in medicine  187

simplified diagram of how an HIV virus infects a T4 lymphocyte and uses it to multiply.

Copyright © 2018. Routledge. All rights reserved.

FIGURE 11.1 A

Source: Campaner, 2012, p. 88.

We can write the mechanism involved in our usual style as follows (where “virus” refers to an HIV virus, and “cell” to a T4 lymphocyte): Free virus approaches cell →  Virus to bind to cell at two receptor sites →  Virus to penetrate cell, and empty its contents into the cell →  Single strands of viral RNA to be converted into double-stranded DNA by the reverse transcriptase enzyme →  5. Viral DNA to be combined with the cell’s own DNA by the integrase enzyme →  6. The viral DNA to be read and long chains of proteins to be made when the cell divides →  1. 2. 3. 4.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:18:52.

188  Discovering cures in medicine

7. Sets of viral protein chains to come together →  8. Immature virus to push out of the cell, taking some cell membrane with it →  9. Immature virus to break free of the infected cell →  10. Protein chains in the new viral particle to be cut by the protease enzyme into individual proteins that combine to make a working virus. A great deal of the research into finding cures for aids has focused on trying to stop one or more of these causal transitions. For example, an attempt is made to stop the transition from step 3 to step 4 as follows (Campaner, 2012, p. 90): the activity of the viral RNA is stopped at the point in which it is leaving the virus and entering the human cell. This is pursued by making the viral RNA combine with a complementary synthetic sequence (“chimera”) However, all the drugs so far developed are not completely effective. As Campaner says (2013, p. 91): All the therapeutic steps …  only halt disease development temporarily: the virus action is significantly slowed down and controlled, but its presence is not completely eliminated. In this situation, the strategy, which was employed to develop cures for tuberculosis (section 9.4), was repeated. As Campaner says (2012, p. 91):

Copyright © 2018. Routledge. All rights reserved.

Drugs are prescribed in the form of a cocktail, so that several (usually three or four) concomitant pharmacological strategies are implemented. The key idea is that an intervention at level (2), for example, may fail, but then one at level (4) or (5) may succeed, and again, if not, the virus could be killed at level (7) or (8). Clearly the development of these treatments for AIDS could not have occurred until after the mechanism of the disease had been worked out.

11.2  Seeking deeper explanations in medicine and physics Let us turn now to a more theoretical issue. I will argue that mechanisms help in the development of deeper explanations in medicine. The question of deeper explanations can conveniently be tackled by comparing the situation in theoretical medicine with that in theoretical physics. This is connected with Russell’s views on cause, which were discussed in section 1.1. It will be recalled that Russell thought the notion of cause to be useful only in everyday life, and in the beginnings of science. As soon as a science became advanced, it would eliminate the use of the notion of cause. Russell illustrated his claim by considering theoretical physics, and argued that causal laws do not appear in this branch

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:18:52.

Copyright © 2018. Routledge. All rights reserved.

Discovering cures in medicine  189

of science. Instead theoretical physics uses mathematical laws of various kinds, which are not causal in character. Russell then postulated that other sciences, as they became more advanced, would abandon the notion of cause in the way that theoretical physics had done. My conclusion was that Russell was partly right and partly wrong. It is true, as he says, that theoretical physics has abandoned the notion of cause, but false to claim that all sciences, as they advance will follow the same path. On the contrary, medicine has become a very advanced science, but continues to use causal laws. Indeed, such laws are central to medicine. I now want to elaborate this contrast between theoretical physics and medicine by arguing that the two disciplines seek deeper explanations in different ways, between which there are nevertheless some parallels. In physics deeper explanations are obtained by proceeding to more general, but still non-causal, laws from which the earlier laws can be obtained. Let us take a classic case. At the beginning of the 17th century Kepler and Galileo formulated some important, but quite empirical laws, which were closely based on observation. Kepler produced the laws of planetary motion, and Galileo the law of falling bodies. Towards the end of the 17th century, Newton produced a much more general theory consisting of his three laws of motion and his law of gravity. This theory is normally regarded as deeper than the laws of Kepler and Galileo, and provides a deeper explanation of their laws. In the 20th century Einstein produced a new theory of gravitation – the General Theory of Relativity, which provided a deeper explanation of Newton’s theory. In medicine too, there is a search for deeper explanations, but it proceeds in a different way. Usually the first step (analogous to the discovery of Kepler’s and Galileo’s laws) is the formulation of an empirical causal law, such as A→ B, where “→ ” means “causes”. Then a mechanism M is found which links A and B, giving A→ M→ B. Such a mechanism is analogous to Newton’s theory and provides a deeper explanation of the causal law A→ B. M in its turn can be replaced by a more detailed and accurate mechanism M’, which provides a deeper explanation of the mechanism M, in the way in which Einstein’s theory provided a deeper explanation of Newton’s. Let us next compare these two ways of producing deeper explanations in more detail, and with some further examples. Suppose then that we have a conjectured causal hypothesis (H) that A causes B, and a linking mechanism (M say) is discovered connecting A to B. My thesis is that M is related to H in a way which is quite analogous to the way in which Newton’s theory (N) is related to Galileo’s law of falling bodies (G). G follows logically from N (taken here to include any necessary initial conditions), and in the same way, in the causal case, H follows logically from M. Deducing G from N, however, adds considerably to the confirmation of G. Before this deduction, G was supported only by the direct evidence afforded by objects dropped from the leaning tower of Pisa, observations of pendula, and of balls rolling down inclined planes, etc. N, however, was supported by a whole range of additional evidence concerning the motion of the planets and comets,

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:18:52.

Copyright © 2018. Routledge. All rights reserved.

190  Discovering cures in medicine

the tides, etc. By establishing a link between such phenomena and falling bodies, N channelled some of this extra support down to G and so provided additional support for G. In exactly the same way, the evidence supporting the existence of the various causal connections in the linking mechanism (M) becomes channelled into supporting the causal hypothesis H. This is the evidence of mechanism which adds to the empirical support of the causal law. Thus, in the smoking causes heart disease example, all sorts of experimental evidence showing that the oxidation of LDLs produces plaques in the arterial walls comes to support the original causal hypothesis, which, without the discovery of the linking mechanism, it would never have done. Deducing G from N also refines and modifies G itself. Thus, N predicts that the rate of acceleration of bodies falling in different parts of the Earth might be different depending on the shape of the Earth, that the rate of acceleration of bodies falling at the bottom of deep mines might be different from their rate of acceleration on the surface, and so on. These refinements of G could of course be checked experimentally. In the same way deducing H from M may well refine and modify H. The anthrax example, discussed in section 4.4 and in section 11.1 of this chapter, is a good illustration here. The original causal law (H) was: Sheep grazing in anthrax fields →  some of them to die with symptoms of anthrax. Here “anthrax fields” were identified empirically by farmers as those fields which had this unfortunate effect on their flocks. However, once Koch and other researchers had discovered the linking mechanism (M4.1) given in section 11.1, this empirical law was modified to become: Sheep grazing in fields where there are many anthrax spores →  some of them to die with symptoms of anthrax. This refinement of H by M, like the refinement of G by N, could of course be checked experimentally. The refinement of H by M could be supported by showing that the fields, identified empirically by farmers as “anthrax fields”, did in fact contain many anthrax spores. In his 1957b article ‘The Aim of Science’, Popper proposes a criterion for showing that one theory is deeper than another. He writes (p. 202): Here …  I am …  interested …  in the problem of depth. And regarding this problem, we can indeed learn something from our example. Newton’s theory unifies Galileo’s and Kepler’s. But …  it corrects them while explaining them. I suggest that whenever in the empirical sciences a new theory of a higher level of universality successfully explains some older theory by correcting it, then this is a sure sign that the new theory has penetrated deeper than the older ones. Newton’s theory did indeed correct Kepler’s. For example, one of Kepler’s laws was that planets move in ellipses with the sun at one focus. In Newton’s theory,

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:18:52.

Discovering cures in medicine  191

Copyright © 2018. Routledge. All rights reserved.

this would be true only if there was just one planet orbiting the sun. As there are several planets, however, the gravitational attractions of the other planets on the planet under study produce small deviations from the elliptical orbit. Indeed, the extent and position of these deviations can be calculated using Newtonian theory. Thus, Kepler’s law that planets move in ellipses is not strictly true, and Newton’s theory corrected it. But did Newton’s theory correct Galileo’s law? This is a bit more doubtful. Newton’s theory does not so much tell us that some of what Galileo said was wrong as give us more information about the acceleration of falling bodies, indicating, for example, how this may vary from one point on the Earth’s surface to another. This is why I have preferred to speak of Newton’s theory refining or modifying Galileo’s rather than correcting it. So, I would alter Popper’s criterion for the depth of a theory slightly by saying that, if a theory T1 corrects, refines, or modifies another theory T2, while explaining it, then T1 is thereby shown to be a deeper theory than T2. This criterion then shows, in many cases, that a linking mechanism M is deeper than the original causal hypothesis H. In the case of the transition from Kepler and Galileo to Newton, it was the production of a more general theory which enabled the scientific community to penetrate deeper. This is often so in theoretical physics. In the case of the transition from a causal hypothesis H to a linking mechanism M, it is the discovery of a specific mechanism rather than the discovery of a more general theory that enables the scientific community to penetrate deeper. This shows a respect in which the causal hypotheses of medicine differ from the non-causal hypotheses of theoretical physics. Yet, although they have used different methods to do so, both medicine and theoretical physics have succeeded in delving deeper into the secrets of nature.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:18:52.

Copyright © 2018. Routledge. All rights reserved. Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 15:18:52.

PART III

Copyright © 2018. Routledge. All rights reserved.

Causality and probability

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Copyright © 2018. Routledge. All rights reserved. Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

12

Copyright © 2018. Routledge. All rights reserved.

INDETERMINISTIC CAUSALITY

A causal claim of the form A causes B is deterministic, only if, ceteris paribus, whenever A occurs, it is followed by B. Otherwise, the claim is indeterministic. In Part I of this book, I developed an action-related theory of causality, but only for deterministic causality. In this part of the book (Part III) my aim is to show how this action-related theory can be extended from the deterministic to the indeterministic case. There were at least two reasons for treating deterministic causality before indeterministic causality. First of all, deterministic causality was the notion of causality used by the pioneers of scientific medicine such as Koch and Pasteur when they developed the germ theory of disease in the 19th century. Moreover, the type of causality used in medicine remained almost exclusively deterministic in the first half of the 20th century. Deterministic causality has continued to be used often in medicine since 1950, and is still used in some cases today. For example, an account of McArdle disease, which was discovered in the 1950s and 1960s, was given in section 10.2. This disease is caused by a single, completely recessive, rare, autosomal gene, and the causation here is deterministic. Anyone who has this gene suffers from McArdle disease. There is thus a considerable body of medical results which use only deterministic causality, and so a separate analysis of this notion of causality is definitely worthwhile. Secondly, indeterministic causality is a much more complicated notion than deterministic causality, and raises quite a number of problems which are far from being completely resolved. Galavotti, who refers to indeterministic causality as probabilistic causality, give a good account of these problems where she writes (2010, p. 140): The first problem that arises as soon as causality is taken as a probable rather than constant conjunction is that of identifying causal as opposed to

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

196  Indeterministic causality

spurious relations, without getting muddled with problems of the Simpson’s paradox kind. … Moreover, the virtuous circle linking causality, explanation and prediction within classical determinism (of the Laplacean kind) breaks down in the case of probabilistic causality. Galavotti is quite correct to draw attention to “problems of the Simpson’s paradox kind”, and we will encounter some of these problems in Chapter 14, where Simpson’s paradox will be stated and discussed. One might indeed wish to limit causality in medicine to deterministic causality, but this is hardly an option. Since the early 1950s, indeterministic causality has come to be used more and more in medicine, and nowadays its use is probably more common than that of deterministic causality. Thus, despite all the problems involved, an analysis of indeterministic causality is clearly essential. In Part II of the book, a number of examples of indeterministic causality were discussed. Here I will mention two. In section 8.3, the example

Copyright © 2018. Routledge. All rights reserved.



Smoking causes lung cancer (12.1)

was considered. As mentioned in the introduction, section 0.1, this was one of the first uses of indeterministic causality in medicine. 12.1 is now generally accepted, and the use of “causes” in it is clearly indeterministic. Not all smokers get lung cancer. In fact, only about 5% of smokers develop this disease. On the other hand, the probability of a smoker getting lung cancer, written Prob (lung cancer | smoker) is much higher than the probability of a non-smoker getting lung cancer, written Prob (lung cancer | non-smoker). In section 8.3, some data from an epidemiological study are given, and, if the frequencies are a good estimate of the probabilities, we have Prob (lung cancer | smoker) is more than ten times Prob (lung cancer | non-smoker). If we define a heavy smoker as someone who smokes 25 gms or more per day, then Prob (lung cancer | heavy smoker) is more than 22 times Prob (lung cancer | non-smoker). In section 6.1, the example

Having a diet high in saturated fat causes coronary heart disease (12.2)

was considered. This is similar to the smoking causes lung cancer example. 12.2 is clearly an example of indeterministic causality, since some people have a diet very high in saturated fat all their lives, and never get coronary heart disease (CHD). On the other hand, the probability of getting CHD for those who consume large amounts of saturated fat as part of their diet is much higher than for those who consume very little saturated fat. This was illustrated in section 6.3 by the following example drawn from Ancel Keys’ Seven Countries Study. The percentage of calories from saturated fats was 3% in the diet of the Japanese cohort, and 17–18% in the diet of the cohort from the USA. The USA figure was thus between 5.7 and six times that Japan. If we now consider death rates

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Indeterministic causality 

197

from coronary heart disease five years from the beginning of the study, and take the frequencies observed as good estimates of the probabilities, we get that Prob (Death from CHD | American Diet) is 4.9 times Prob (Death from CHD | Japanese Diet). These two examples suggest that we can deal with indeterministic causality by relating it to probability. If A causes B in the deterministic sense, then if A occurs, then, ceteris paribus, B will follow. If A causes B in the indeterministic sense, then it is no longer the case that if A occurs, B will follow, but perhaps, if A occurs, the probability of B occurring will increase. An approach along these lines seems to fit the cases of “smoking causes lung cancer” and “a diet high in saturated fat causes CHD”. Unfortunately, however, there are other cases in which this approach does not work, and these cases are indeed, as Galavotti states, connected with Simpson’s paradox. We must now enter this difficult terrain, but, before doing so, it is worth pointing out that we do not always have to consider probability when we are dealing with indeterministic causes. In some cases, the rather simpler methods developed in Part I can be carried over to the indeterministic case without any appeal to probability. One such case is the recent discovery that cervical cancer is caused by a preceding infection by a human papilloma virus. This case will be considered in the next section.

Copyright © 2018. Routledge. All rights reserved.

12.1  The causation of cervical cancer Cervical cancer is the second most common cancer among women on a global scale. There are about half a million new cases per year, and about 60% of those who get the disease die from it. For these reasons the discovery of the cause of cervical cancer is one of the most significant advances in medicine in the last few decades, particularly as knowledge of the cause has led to a way of preventing the disease. Clarke (2011, pp. 53–75, and 2012) gives a full account of the discovery of the cause of cervical cancer, and I will here give a brief summary of the main steps, which is based mainly on Clarke’s work.1 The first step consisted in epidemiological investigations of cervical cancer which took place in the late 1950s and early 1960s. These showed that cervical cancer (Vonka and Hamš i ková , 2007, p. 132) “had the character of an infectious disease and that the causative agent was transmitted by sexual intercourse.” In the second half of the 1960s, the first case of a human cancer caused by a virus was proposed and came to be increasingly accepted by the medical community. This was the causation of Burkitt’s lymphoma by the Epstein-Barr virus (see Clarke, 2011, pp. 25–52, and 2012, pp. 181–184 for details). Given this situation, it was naturally to look for a virus which was transmitted by sexual intercourse, and which was correlated with cervical cancer. Exactly such a virus was discovered independently by two research groups in the USA in the late 1960s. This was the herpes simplex virus type 2 (HSV2) which was transmitted sexually. Patients with cervical cancer had antibodies to this virus much more frequently than normal women. Moreover, the Epstein-Barr virus,

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

198  Indeterministic causality

which was known to cause cancer, was a herpes virus and so of the same type as HSV2. For these reasons, HSV2 was generally accepted as the cause of cervical cancer from about 1970 to about 1985. Yet there were still some reasons for doubt. In the case “Epstein-Barr virus (EBV) →  Burkitt’s lymphoma”, large parts of the genome of EBV could be found in the tumour. However, the identification of HSV2-specific macromolecules in cervical cancer tumours did not seem possible in every case. The correlation of cervical cancer patients and HSV2 infections was obtained from case control studies of women who had already been diagnosed as having cervical cancer. It was thus possible that these infections were the result of the disease rather than its cause. Because of these problems, Vonka and his group decided to carry out a prospective study of cervical cancer (CaCer). As Vonka and Hamš i ková  say (2007, p. 132): In order to supply, as it seemed at that time, the last missing piece of evidence that HSV2 was indeed the decisive etiological factor in the development of cervical neoplasias, including CaCer, an extensive prospective study was undertaken in Prague: it was carried out by our team in cooperation with Jiř í  Kaň ka and his co-workers …  . Over 10,000 initially normal women were followed up for six years in the late 1970s and early 1980s.

Copyright © 2018. Routledge. All rights reserved.

The advantage of the prospective study was that the time order of HSV2 infections and the onset of cervical cancer, if either or both occurred, could be checked. It was expected that the development of cervical cancer would be significantly higher in those women who had been infected with HSV2 than in those who had not. But when the results came in and were published in the two papers: Vonka, Kaň ka et al. (1984), it turned out that this was not the case. As Vonka and Hamš i ková  write (2007, p. 132): When, upon completion of the work in the field, we tested a large number of selected sera, the result was quite surprising: women who had developed cervical neoplasia (i.e, precancerous lesions and in a few instances microinvasive CaCer) in the course of the study had anti-HSV2 antibodies as frequently as those who had not …  . The implication was as clear: HSV2 was not the causative agent of CaCer. This is a striking instance of the Popperian principle which I stated in section 5.1, namely that “Strong disconfirmations of causal hypotheses leading to their rejection are just as important as strong confirmation of causal hypotheses leading to their acceptance.” HSV2 was strongly correlated with cervical cancer, but this turned out to be a correlation which was not causal in character. This realization opened the way to a serious consideration of alternatives which had been neglected up to that point.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Indeterministic causality 

199

One alternative to the view that HSV2 caused cervical cancer was the ­ ypothesis that this cancer was caused by a human papilloma virus (HPV). This h hypothesis had been advocated since the early 1970s by the German virologist Harald zur Hausen. However, up to 1984, his views had not been taken seriously by the medical community. Now, however, they were investigated thoroughly and, as the evidence in their favour mounted up, came to be generally accepted. Harald zur Hausen was awarded the Nobel Prize in 2008. Let us consider, therefore, the now established hypothesis:

Infection by a human papilloma virus causes cervical cancer (12.3)

This is clearly a case of indeterministic causality, because not every infection by the appropriate kind of HPV results in cervical cancer. As Vonka and Hamš i ková  say (2007, p. 133):

Copyright © 2018. Routledge. All rights reserved.

However, it is evident that HPVs alone do not induce the disease. Thus, these viruses are a necessary but not sufficient condition. The evidence for this is that only a very small minority of those infected develop CaCer. For CaCer to develop, the participation of certain other factors is necessary, the most important of which seem to be the carcinogens present in tobacco smoke and a variety of infectious agents that produce inflammatory lesions in the cervix. Since an HPV infection is not a sufficient condition for cervical cancer to develop, we are dealing with indeterministic causality. However, since an HPV is a necessary condition, we can easily use the causal law to prevent cervical cancer. Suppose in general that we have an indeterministic causal law “A causes B”, but that A is a necessary condition for B to occur. This means that B will only occur if A occurs. So if we can prevent A occurring or block its usual effects, then B will not occur. This is just an instance of what was earlier (section 1.3) called an avoidance action, and which is expressed by Vonka’s Thomist maxim: sublata causa, tollitur effectus (if the cause is removed, the effect is taken away). In section 1.3, avoidance actions were expounded in the context of the action-related theory of causality applied in the deterministic case, but it can be seen that, in this instance where A is a necessary condition for B, the earlier discussion carries over to the indeterministic case. In the case of cervical cancer, the disease was prevented by introducing a vaccine against HPV. This vaccine prevents the individual getting an HPV infection, and, since this is a necessary condition for developing cervical cancer, it also prevents cervical cancer. I have simplified the situation somewhat in my account. Human papilloma viruses are a complicated group of viruses, and only some of them have the property of causing cervical cancer. The most important of these are HPV16 and HPV18 which together are responsible for about 70% of cervical cancer cases.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

200  Indeterministic causality

GlaxoSmithKline’s vaccine, cervarix, was designed against HPV16 and HPV18, and so is protective only against about 70% of cervical cancer cases, or possibly 80% if there is cross immunity with other HPV types. However, as Vonka and Hamš i ková  point out (2007, p. 137), it should be possible to develop further vaccines, including some additional pathological HPV types, which will have a broader effect, and perhaps reduce CaCer incidence by 90% or more. Altogether this has been a remarkable story of the discovery of a way of preventing a widespread and very nasty form of cancer. From the point of view of analysing causality, this example shows that provided that, in the causal law “A causes B”, A is a necessary condition for B, we can handle indeterministic causality in the same way as deterministic causality. Specifically, there is no need to introduce probability. However, it is only in some cases of indeterministic causality that A is a necessary condition for B. Consider, for example, 12.1: Smoking causes lung cancer. Here smoking is not a necessary condition for lung cancer to develop. In fact, lung cancer can have quite different causes, such as, for example, exposure to asbestos. In cases of indeterministic causality of the form “A causes B”, where A is not a necessary condition for B, we do have to introduce probability and also explore how probability is related to causality. I will begin this task in the next section.

Copyright © 2018. Routledge. All rights reserved.

12.2  Probabilistic causality In the introduction, section 0.4, I gave a brief outline of probabilistic theories of causality. It was shown there that this approach to causality really begins seriously with Good (1959). In this paper, Good suggests that causality can be defined in terms of probability, where probability is given the propensity interpretation which Popper had introduced (1957a). Such a definition would have been a definition of indeterministic causality, but Good’s idea, which was later endorsed by Popper, was that the deterministic case could be recovered by setting probability equal to one. In the same paper, Good introduced a very plausible principle which I called “the principle that causes raise the probability of their effects”. Suppose A causes B, then according to this principle, the probability of B, given that A occurs [P(B | A)] is greater than the probability of B, given that A does not occur [P(B | ~A)], where ~A stands for not-A. In symbols: P(B | A) >  P(B | ~A). So, for example, P(lung cancer | smoker) >  P(lung cancer | non-smoker), given that smoking causes lung cancer. This principle looks as if it could be made the basis of a definition of causality in terms of probability, namely the following:

A causes B if and only if P ( B | A ) > P ( B |~ A ) (12.3)

Good’s programme for developing probabilistic causality was indeed a very promising one when he proposed it in 1959. It is not surprising therefore that, not only Good himself, but many other leading philosophers of science have Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Copyright © 2018. Routledge. All rights reserved.

Indeterministic causality 

201

worked on it. Unfortunately, the development of probabilistic causality brought to light many unexpected difficulties, and this has lead some researchers, notably Pearl, to reject the programme altogether. Despite these problems, I will try in this part of the book (Part III) to develop a version of this approach which is quite similar, in many respects, to that in Good’s original 1959 paper. The ambitious aim of defining causality in terms of probability will be given up, but, nonetheless, causality will be linked to probability, using a circumscribed form of the principle that causes raise the probability of their effects. Moreover, the probabilities involved will be interpreted as propensities. As a first step in developing this approach, I will now briefly sketch some of the further work that was done on Good’s probabilistic causality programme, and describe the difficulties which emerged. Good continued his work on the theory of probabilistic causality in two further papers (1961 and 1962). The development of this approach was then taken up by other researchers. It should be observed that not all these researchers agreed with Good in interpreting probability as propensity. Suppes made an important contribution in 1970; and the approach was developed further by, among others, Cartwright (1979), Salmon (1980), Eells (1991), and, more recently, Twardy and Korb (2004), and Galavotti (2010). Russo (2009) gives an excellent overview of the debates in this area. The original hope of those working on the programme was to define causality in terms of probability, but this hope was definitely abandoned by researchers in probabilistic causality from Cartwright (1979) onwards (see Twardy and Korb, 2004, p. 241). There were a number of reasons for this, but one factor, which we mentioned in the introduction, section 0.4, was the discovery by Paul Humphreys of Humphreys’ paradox, which was first published, with an acknowledgement of the author, by Salmon in 1979. To illustrate this paradox, let us take the example of mud and rain. The probability of mud given rain [P(mud | rain)] and the probability of rain given mud [P(rain | mud)] are both well-defined, and in a particular set-up can be evaluated using statistical frequencies. By contrast rain causes mud, but mud does not cause rain. Causality is asymmetrical, while probability is symmetrical. This casts doubt on whether we can define causality in terms of probability. Suppose, for example, that we tried to define causality using 12.3 above. If we set A = rain and B = mud, the righthand side of 12.3 becomes

P ( mud | rain ) > P ( mud | not-rain ) (12.4)

12.4 is true and so we can conclude correctly from 12.3 that rain causes mud. If, however, we now set A = mud and B = rain, the righthand side of 12.3 becomes

P ( rain | mud ) > P ( rain | not-mud ) (12.5)

12.5 is also true, and so it follows from 12.3 that mud causes rain. But it is not true that mud causes rain, and this shows that our attempted definition of Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

202  Indeterministic causality

causality (12.3) is not a correct one. Other proposed definitions of causality in terms of probability could be criticized in a similar fashion. It seems better then to regard probability and causality as distinct, though connected, concepts. But how do we connect causality and probability? A simple suggestion would be to weaken 12.3 by turning it from an “if and only if ” statement (a definition) to an “only if ” statement. I will call this weaker version of 12.3 the Causality Probability Connection Principle (CPCP or CP2). It can be stated as follows:

Copyright © 2018. Routledge. All rights reserved.

    If A causes B, then P ( B | A ) > P ( B |~ A )

(CPCP)

CP2 enables us to derive probability statements from causal statements. If these probability statements are interpreted objectively as propensities, they can be tested out against statistical data. If they are confirmed, this will confirm the causal statement from which they were derived. If they are disconfirmed, this will disconfirm the causal statement. Moreover, the probability statements obtained can be used as a basis for action. All this would be fine, if CPCP were generally valid, but, unfortunately, counterexamples to CPCP were discovered. Good accounts of these are to be found in Cartwright (1979) and Salmon (1980). Some of these counter-examples involve single-case causality, but, as I stated at the beginning of the introduction, this book deals exclusively with generic causality. It is this concern with generic causality which permits the interpretation of probabilities as propensities. It follows that the only relevant counter-examples are those which involve generic causality. The most famous such counter-example is due to Hesslow (1976). As the other suggested counterexamples involving generic causality can all be dealt with in the same way, I will confine discussion to this counter-example of Hesslow. It can now be stated as follows. Suppose that we have a reference class of young women, all with male partners, and for whom the only contraceptive method available is the pill. Suppose further that both pregnancy and taking the contraceptive pill cause thrombosis, but that the probability of getting thrombosis is higher for those who are pregnant than for those who take the pill. So, stopping taking the pill in this population makes pregnancy very likely, and that in turn gives a higher probability of thrombosis. It therefore follows that stopping taking the pill increases the probability of thrombosis, or in symbols

P ( thrombosis | pill ) < P ( thrombosis | not-pill ) (12.6)

However, CPCP in this case states that: If pill causes thrombosis, then P(thrombosis | pill) > P(thrombosis | not-pill) We are assuming that taking the pill causes thrombosis, so it follows from CPCP that:

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Indeterministic causality 



203

P ( thrombosis | pill ) > P ( thrombosis | not-pill ) (12.7)

12.7 is the consequence of assuming CPCP, but the true situation is 12.6 which contradicts 12.7. It follows that CPCP does not apply in this case. Hesslow’s counter-example and similar cases are a considerable blow to the probabilistic causality programme. In this programme, CPCP was supposed to provide the general connecting link between causality and probability. Yet it clearly does not apply in some cases. What should be done? There are two possible responses to this situation. The first is to try to find some way of circumscribing the use of CPCP which excludes cases like the Hesslow counter-example, and ensures that CPCP is always valid when it is applied. This would rescue the probabilistic causality programme. The second is to regard the counter-examples as refuting the probabilistic causality approach, and try to find some other method of connecting causality to probability. In this book, I adopt the first approach, and try to develop a modified version of the probabilistic causality approach which overcomes the difficulties. Pearl, however, adopts the second approach. He is a harsh critic of probabilistic causality, and suggests a different way of connecting causes and probabilities which depends on his structural theory of causation. Pearl’s alternative approach will be stated and criticized in Chapter 15. This disagreement with Pearl does not of course mean a rejection of all his significant contributions to causality. Pearl is well-known for his development of the important new theories of causal networks and Bayesian networks. These have become indispensable tools for handling indeterministic causality. I will therefore outline some of the basic ideas of causal networks in Chapter 13, and use this framework in the remaining chapters dealing with the link between causality and probability (Chapters 14, 15, and 16).

Copyright © 2018. Routledge. All rights reserved.

Note 1 Both Clarke and I have been greatly helped by conversations with Vladimir Vonka, whom we met in Prague on several occasions. As we shall see, Vonka’s own research played a crucial part in the scientific development which led to the discovery of the viral cause of cervical cancer. Vonka’s own account of the discovery and its consequences, written with Eva Hamš i ková , is to be found in Vonka and Hamš i ková  (2007), and I quote from this paper on several occasions. The history of this discovery is also interesting in relation to the question of research organization, and it is discussed from this point of view in Gillies (2014).

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

13

Copyright © 2018. Routledge. All rights reserved.

CAUSAL NETWORKS

Causal networks and Bayesian networks were devised in the 1980s. The ­principal figure in this development was Pearl who introduced and developed the concept of the Bayesian network in a series of papers: Pearl (1982; 1985a; 1985b; 1986), Kim and Pearl (1983), and a book by Pearl (1988). An important extension of the theory was carried out by Lauritzen and Spiegelhalter (1988). Now the theory of causal networks is mathematically rather complicated. In this book I give the precise mathematical terminology being used in Appendix 2, and some new theorems concerned with causal networks in Appendix 3 (Sudbury’s Theorems). However, causal networks have a value also for those who are less mathematically inclined, because they introduce a new diagrammatic way of representing causal relations, and this is very helpful, especially in dealing with indeterministic causes. In the main body of the text I will try to introduce this diagrammatic representation without too many mathematical complications, and then use it when discussing the philosophical problems, which arise. These are mainly concerned with extending the action-related theory of causality from the deterministic to the indeterministic case, and with interpreting the probabilities which are involved.

13.1  Conjunctive and interactive forks The first simple example of a causal network in the modern sense was introduced by Reichenbach in his 1956. Reichenbach was investigating a problem which we met in section 6.2. This is the problem of events which are correlated, but between which there are no causal links. The classic example is that of barometers and rain. A barometer’s reading falling to a low level might be strongly correlated with rain occurring. Obviously, however, the barometer’s reading is not a cause of rain. Rain is caused by a fall in atmospheric pressure, and such a

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Causal networks 

205

fall also causes the barometer’s reading to fall. The barometer’s reading and the rain have a common cause, and this is why they are correlated, even though the barometer’s reading has no causal influence on the rain. Reichenbach discusses the problem of the common cause using an example of geysers, as well as that of barometers. He writes (1956, p. 158):

Copyright © 2018. Routledge. All rights reserved.

Suppose two geysers which are not far apart spout irregularly, but throw up their columns of water always at the same time. The existence of a subterranean connection of the two geysers with a common reservoir of hot water is then practically certain. The fact that measuring instruments such as barometers always show the same indication if they are not too far apart, is a consequence of the existence of a common cause – here, the air pressure. Further illustrations would be easy to find. Reichenbach pursues his investigation of common causes by introducing (1956, p. 159) what he calls a conjunctive fork. Most significantly he also gives a diagram of a conjunctive fork, which is shown in Figure 13.1. Here C is the common cause of two effects A and B. A, B and C are said to be the nodes of the causal network. The causal link between C and one of its effects (e.g. A) is shown by an arrow running from C to A. One might be tempted to read this arrow as “C causes A”, but this is not strictly correct. Consider the example: C = Drinking a lot of red wine, A = Liver Disease, and B = Heart Disease. With this assignment of values to variables, Figure 13.1 is a genuine causal network. Here C causes A in the usual sense, but C prevents B. However, prevention is also a causal link. In general then, the arrows in causal networks should be understood as stating that there is a causal link between the two nodes they connect, but this link could be either causation in the positive sense or prevention. A diagram such as Figure 13.1 represents causal connections in a very clear fashion, but it has another advantage. We can easily add probabilities to such a diagram. The nodes {A, B, C} can be regarded as random variables with a joint probability distribution. Reichenbach takes exactly this step, and so can be seen

FIGURE 13.1 

Conjunctive or Interactive Fork.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

206  Causal networks

as the first to use the causal network approach in connecting causality with probability. More specifically, Reichenbach is interested in the case where A and B are correlated, and this is explained by their having a common cause C. To understand what is involved here, we must introduce the probabilistic notion of independence. Suppose A and B are two events with probabilities P(A) and P(B). We can also consider the event A&B, which is the event that A and B both occur. Then A and B are said to be independent if

P ( A & B) = P ( A ) P ( B) (13.1)

By the probability calculus, provided P(A) and P(B) are both non-zero, 13.1 is equivalent to either

P ( A | B ) = P ( A ) (13.2)

or

P ( B | A ) = P ( B ) (13.3)

Copyright © 2018. Routledge. All rights reserved.

13.2 says that the probability of A given B is the same as the probability of A, or, in other words that the occurrence of B does not raise or lower the probability of A. Similarly, 13.3 says that the occurrence of A does not raise or lower the probability of B. This explains why it is reasonable to take 13.1 as saying that A and B are independent events. Now in Figure 13.1, A and B are correlated, and Reichenbach want to explain this correlation by the fact that they have a common cause C. If the existence of this common cause is to give a complete explanation of the correlation between A and B, then conditional on C, A and B must become independent. In symbols, this is written

P ( A & B | C) = P ( A | C) P ( B | C) (13.4)

13.4 is of course just the same as 13.1 except that all the probabilities are conditional on C. Reichenbach defines a conjunctive fork to be a fork as shown in Figure 13.1 for which the conditional independence condition (13.4) is satisfied. In a conjunctive fork, C is said to screen off A from B, since A is independent of B given C. The conditional independence condition (13.4) is also known as a Markov condition, after the Russian mathematician Andrei Andreyevich Markov (1856–1922) who was one of the first to study conditional independence. Reichenbach’s conjunctive forks are a simple example of the Bayesian networks, which were developed, mainly by Pearl, in the 1980s. Bayesian networks can have a much more complicated network structure, such as that illustrated in the introduction, Figure 0.1. For such a network to be a Bayesian network, each node must satisfy a generalized

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Causal networks 

207

Copyright © 2018. Routledge. All rights reserved.

version of the conditional independence or Markov condition which applies to conjunctive forks. There are some slight variations in the use of the expression “Bayesian network” and “Markov condition”, and exact definitions of these terms as used in the present book are given in Appendix 2. Reichenbach introduced the Markov condition into the definition of his conjunctive forks in order to ensure that the common cause C gave a complete explanation of the correlation between A and B. So, in a sense, the Markov condition establishes a connection between causality and probability, since correlation is a purely probabilistic concept. This connection is rather different from the probability raising connection (CPCP), which we stated in section 12.2. This is the connection which was introduced by Good in 1959, and which has been adopted by the probabilistic causality approach. As we shall see in Chapter 15, Pearl rejects the probabilistic causality approach, and tries to link causality and probability using the Markov condition in a way, which generalizes Reichenbach’s approach. The problem with Pearl’s approach, however, is that the Markov condition is not always satisfied. Reichenbach went on to formulate what he called the principle of the common cause (1956, p. 157f.). This states that, if A and B are correlated, then either A causes B, or B causes A, or A and B have a common cause C which screens off one variable from the other. The principle of common cause has been much discussed by philosophers of science, and several counter-examples to it have been discovered. Williamson (2005, Section 4.2, pp. 51–7) gives a systematic account of these counter-examples. I will mention more of these later on in section 15.1, but here will confine myself to a significant contribution by Wesley Salmon, who did his PhD with Reichenbach, and went on to develop Reichenbach’s ideas on causality. Salmon found it necessary to introduce, in addition to Reichenbach’s conjunctive fork, a second kind of causal fork, which he called an interactive fork. As he says (1978, p. 134): It thus appears that there are two kinds of causal forks: (1) Reichenbach’s conjunctive fork, in which the common cause screens off the one effect from the other, … , and (2) interactive forks, exemplified by the Compton scattering of a photon and an electron. Interactive forks can be illustrated by the same diagram as conjunctive forks (Figure 13.1). The difference is that in a conjunctive fork the common cause C screens off A from B, but this is not the case in interactive forks. Salmon shows the need for interactive forks by his interesting example of Compton scattering. In a Compton scattering experiment, an energetic photon collides with an electron which can be regarded as more or less stationary. The collision is represented by the node C, where the variable C has the energy E as its value. As the result of the collision we get an electron with energy E1 represented by node A, and a photon with energy E2 represented by node B. Now because of the conservation of

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

208  Causal networks

energy, we have E = E1 + E2, and so A and B are highly correlated given C. C is a common cause of A and B, but it does not screen off A from B. So, this is a causal fork, which is not a conjunctive fork. It is an interactive fork.

Copyright © 2018. Routledge. All rights reserved.

13.2  Multi-causal forks As their examples show, neither Reichenbach nor Salmon were primarily concerned with medicine. It turns out that in medicine, another type of causal fork, known as a multi-causal fork, is particularly useful when we are dealing with indeterministic causality. As we have seen, 19th and early 20th century scientific medicine, as it was developed by Pasteur, Koch, and others, used a deterministic notion of causality. An attempt was made to show that each disease had a single cause, which was both necessary and sufficient for the occurrence of that disease. So, for example, tuberculosis was caused by a sufficiently large number of tubercle bacilli in the patient. This could be called mono-causality. From the 1950s on, however, medicine had to introduce an indeterministic notion of causality. This went hand in hand with explaining diseases as caused by the conjunction of several causes acting together. We can illustrate this by developing our earlier example of smoking causing lung cancer. Since not all smokers get lung cancer, we might postulate that those who do develop the disease have a genetic susceptibility for it. We now have two causes which operate: smoking and genetic susceptibility. So, if we use indeterministic causality mono-causality is no longer adequate. We usually have to postulate several causes, and this could be called multi-causality. The various different causes, which act together to produce the disease, will be called causal factors. The use of multi-causality, or several causal factors, leads to the introduction of multi-causal forks, as shown in Figure 13.2. Here Z, a disease, has a finite number n of causal factors X1, X 2, … , X n. There are many examples of multi-causal forks in contemporary medicine. One which we have already discussed is heart disease. Since 1960, quite a number of causal factors for heart disease have been discovered, as a result of the pioneering work

FIGURE 13.2 

n-Pronged multi-causal fork.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Copyright © 2018. Routledge. All rights reserved.

Causal networks 

209

of Ancel Keys, the Framingham study (see Levy and Brink, 2005), the work of Bradford Hill and his group on smoking, and so on. Causal factors now generally accepted include: smoking, high blood pressure, unhealthy diet, high blood cholesterol, overweight/obesity, and diabetes. In the last few years, investigations have begun into possible genetic causal factors. Any disease with a number of indeterministic causal factors can be represented by a multi-causal fork as shown in Figure 13.2. Note, however, that this type of fork is different from the conjunctive and interactive forks considered by Reichenbach and Salmon, and illustrated in Figure 13.1. In Figure 13.1, a single cause C has different effects. In a multi-causal fork, a number of different causes combine to produce a single effect. The arrows in the two types of fork go, so to speak, in different directions. Reichenbach does, however, mention multi-causal forks in his 1956 classic. He calls them forks “open towards the past, constituted by a common effect” (p. 159). However, he does not discuss them in detail. It is perfectly legitimate to illustrate the casual factors in heart disease by the type of multi-causal fork shown in Figure 13.2. Here, Z = heart disease, and we could put X1 = smoking, X 2 = high blood pressure, and so on. However, we know that there are some important relations between the causal factors. For example, unhealthy diet is itself a causal factor for both high blood cholesterol and overweight/obesity; while overweight/obesity is a causal factor for diabetes. We can represent these causal relationships in a more complicated causal network as shown in Figure 13.3. Figure 13.3 illustrates the great advantage of the causal network diagrams. These show in a clear fashion the relationships between the various causal factors which, acting together, give rise to heart disease. My aim in Part III of this book is not, however, to build a causal model for heart disease, but to explore the relations between causality and probability. To do so, I will still focus on the causation of heart disease, but, instead of considering all the factors which have been identified, confine myself to the two most important. This means that we can use a two-pronged multi-causal fork, as shown in Figure 13.4. If Z = heart disease, the two most important causal factors are, arguably, smoking and unhealthy diet. “Unhealthy diet” is a little vague, however, and we can make it more precise by identifying it with eating a lot of fast food. Eating fast food is the principal form of unhealthy eating in the West and, increasingly, in the rest of the world. It might be objected that there are other forms of unhealthy eating. For example, someone could have an unhealthy diet consisting entirely of traditional gourmet French cuisine. However, such cases are rare, and a focus on fast food seems perfectly justified. I will therefore take X = Smoking and Y = Eating Fast Food. The resulting two-pronged causal fork is then shown in Figure 13.5. Let us now consider the Markov condition in relation to the causal networks of Figures 13.4 and 13.5. Applied to the two-pronged multi-causal fork of Figure 13.4, the Markov condition states that X should be independent of Y.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Copyright © 2018. Routledge. All rights reserved.

210  Causal networks

FIGURE 13.3 

Causal network for heart disease.

However, this is not satisfied in our heart disease example of Figure 13.5, since smoking is not independent of eating fast food. The two are correlated. This means that Pearl would not regard the causal network of Figure 13.5 as constituting a genuine causal model, since, for Pearl, a genuine causal model must satisfy the Markov condition. In Chapter 15, I will argue that this requirement of Pearl’s should be rejected, and that the causal model shown in Figure 13.5 is perfectly satisfactory. Part of the argument which will be used for this conclusion concerns the evidence which supports the claim that eating fast food causes heart disease. I will therefore examine this evidence in the next section, and then make use of the conclusions reached in Chapter 15 (see section 15.3). The best way to approach the evidence for the claim that eating fast food is

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Copyright © 2018. Routledge. All rights reserved.

Causal networks 

FIGURE 13.4 

Two-pronged multi-causal fork.

FIGURE 13.5 

Two-pronged multi-causal fork for heart disease.

211

unhealthy is to compare fast food with the kind of diet which Ancel Keys and his wife Margaret advocated in their ‘Eat Well and Stay Well’ project. This will be done in the next section (13.3).

13.3 The rise of fast food, and the failure of the ‘Eat Well and Stay Well’ project Information about fast food can be obtained from Eric Schlosser’s classic 2001 book on the subject: Fast Food Nation. What the All-American Meal is Doing to the World. Schlosser describes the rise of fast food in the period 1970 to 2001 as follows (2001, p. 3): In 1970, Americans spent about $6 billion on fast food; in 2001, they spent more than $110 billion. Americans now spend more money on fast food than on higher education, personal computers, computer software, or new

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

212  Causal networks

cars. They spend more on fast food than on movies, books, magazines, newspapers, videos, and recorded music – combined. He adds (2001, p. 6):

Copyright © 2018. Routledge. All rights reserved.

The typical American now consumes approximately three hamburgers and four orders of french fries every week. There is an irony here, because the period 1970 to 2001 was one in which surprising discoveries about what constituted a healthy diet were made and received overwhelming empirical confirmation. One might have expected that this would have led to a switch towards healthier eating, but the opposite happened. There was an increasing consumption of fast food which was definitely unhealthy according to the new, empirically confirmed, science of nutrition. It is interesting to compare this to what happened in the case of cholera. As we saw in Part I of the book, there was a long scientific debate before it was accepted that cholera was caused by the vibrio cholerae. However, once the medical researchers did accept this hypothesis, and the conclusion that the disease could be prevented by filtering drinking water, the filtration of water was implemented straight away. It would have been very surprising if, once the dangers of unfiltered drinking water became known, more unfiltered drinking water was used than before. Yet this is more or less what happened in the case of diet. 1970 is in fact a reasonable date for the beginning of the new science of nutrition, since it saw the publication of the five-year results of the Seven Countries Study (Keys, 1970). These results showed that diet did have an enormous effect on the prevalence of coronary heart disease, and that diets low in saturated fat were much healthier than those which contained a great deal of saturated fat. Subsequent investigations have all confirmed this conclusion, while adding more information about the constituents of a healthy diet. The discovery that meat, milk and cheese, if eaten in large quantities, were unhealthy foods was a most surprising one. It contradicted a long tradition that these foods – particularly meat – were among the most nourishing and that their consumption would bring health and strength. These ideas go back to the beginnings of Western civilization. Plato in The Republic considers what food his ruling class (the guardians) will need to eat when the go on military campaigns. He writes (III 403–404): What of their food? …  That, at any rate, may be, learnt from Homer. As you know, he gives his heroes on their campaign no fish for their meals, although they are on the shores of the Hellespont, and no boiled meat; nothing but roast, which is convenient for soldiers …  in the body, variety breeds maladies and simplicity health. Now Homer was writing an epic and Plato imagining a utopia. So it is unlikely that a diet consisting entirely of roast meat bore much relation to actual army Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Copyright © 2018. Routledge. All rights reserved.

Causal networks 

213

rations, even of the most wealthy, in ancient Greece. However, the passage shows that both Homer and Plato considered roast meat the best food for giving health and strength, and fish to be distinctly inferior. In fact, as Ancel Keys’ researches showed, the opposite is true. Similarly, the Pythagorean school in ancient Greece forbade the eating of beans which are in reality a very healthy food. This is a good illustration of how commonly held ideas can be shown to be wrong by careful empirical scientific investigation. As was shown in section 6.1, Ancel Keys reached his conclusion about the value of a diet low in saturated fat about the mid-1950s. His seven countries study, begun in 1958, was designed to confirm this hypothesis which indeed it did. Ancel Keys, always the activist, was anxious to apply this new hypothesis in order to reform the American diet, and make Americans healthier and less liable to coronary heart disease (CHD). At the same time, however, Ancel Keys was a man who enjoyed his food, and his wife was an excellent cook. He realized that it was not enough simply to urge Americans to eat less meat, cheese, and eggs. It was necessary to provide an alternative diet which was just as good to eat and just as satisfying of hunger. This was the origin of the ‘Eat Well and Stay Well’ project. He and his wife designed a series of recipes which were as tasty as traditional recipes, but which contained less saturated fat. These recipes together with an account of the scientific findings concerning the dangers of saturated fat were published as a book by Ancel and Margaret Keys with the title: ‘Eat Well and Stay Well’ in 1959. There was a second edition in 1963. When the first results of the seven countries study were analysed and published in 1970, Keys and his wife modified the ‘Eat Well and Stay Well’ project in the light of the findings. These findings showed that of all the cohorts, the Cretans had the lowest rate of CHD and were generally the healthiest. As we pointed out in section 6.4, this was something of an anomaly, since the Cretans, though they had a low median blood cholesterol level, did not have such a low blood cholesterol level as some other cohorts, notably the Japanese cohorts. Yet the Cretans turned out to be the healthiest and most long-lived of any of the cohorts. Keys concluded that the Cretan diet probably contained other elements which were protective against coronary heart disease and other illnesses, but that it was not known what these elements were. He and his wife therefore modelled their ‘Eat Well and Stay Well’ project more on the Cretan diet, and produced a new version of their book in 1975, this time entitled: ‘How to Eat Well and Stay Well. The Mediterranean Way’. Another factor which may have led the Keys to emphasize the traditional Mediterranean diet rather than the traditional Japanese diet is that the former is obviously closer to the American diet and traditional diets in Northern European countries such as Britain. However, since 1975, Americans and Europeans have become more familiar with international cuisines, so that a contemporary version of the ‘Eat Well and Stay Well’ project could include some Japanese as well as Mediterranean dishes. The Keys’ ‘Eat Well and Stay Well’ project got off to a good start. The first two editions of their book were best sellers, and they used the proceeds to purchase a

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Copyright © 2018. Routledge. All rights reserved.

214  Causal networks

villa in Southern Italy overlooking the sea near the coastal town of Pioppi to the south of Naples. Here Ancel and Margaret spent much of their later life, no doubt researching the delights of Mediterranean cuisine. The ‘Eat Well and Stay Well’ project was also helped initially by a high-profile case. Eisenhower was President of the United States from 1953 to 1961, but he had considerable health problems. From his time as a cadet at West Point, he had been a heavy smoker – often smoking two or three packs a day. In 1949, however, he succeeded in giving up. Still in 1955, when President, he had a severe heart attack. The cardiologist who treated him was an admirer of Ancel Keys, and he put Eisenhower on a very strict low-fat diet, which the president followed to the letter. He recovered from his heart attack, and went on to a second term. In fact, he was healthier in his last three years of office than before. This naturally gave a boost to Ancel Keys who appeared on Time magazine’s cover in 1961. Despite these initial successes, however, the rise of fast food made the ‘Eat Well and Stay Well’ project a failure. The diet in America and Northern Europe did indeed change, but in exactly the opposite direction from that recommended by Ancel and Margaret Keys. Of course, the Keys were very far from being the last word in nutrition, and research in the subject continued from the 1970s. Broadly we can say that subsequent research confirmed Ancel Keys’ ideas about saturated fat, but extended and supplemented his views in a number of ways. As we saw in sections 8.5 and 8.6, laboratory research in the period from 1979 to the late 1990s elucidated the mechanism by which smoking accelerated the formation of atherosclerotic plaques. The key factor was that smoking increased oxidative stress. Conversely, however, this result showed that decreasing oxidative stress should protect against atherosclerosis. Now fruit and vegetables contain many vitamins and other compounds (polyphenols) which are anti-oxidants. This gives a mechanism whereby the consumption of fruit and vegetables protects against atherosclerosis, and this protection is confirmed by statistical investigations. Here we have the basis of the recommendation to eat five-a-day of fruit and vegetables. This also provides an explanation of the Cretan anomaly in the seven countries study. Although the Cretans had slightly higher blood cholesterol levels than the Japanese cohorts, they had less coronary heart disease, and better overall health and longevity. This can be explained by their consumption of fruit, vegetables, wine, and olive oil. The Cretans may also have benefited from some consumption of meat and dairy products. In fact, there are definite advantages in a limited consumption of meat and dairy products, since the older view that these contained valuable nutrients was not altogether wrong. They are both good sources of protein. Meat supplies iron, and dairy products calcium which is needed to avoid bone weakness and osteoporosis. The downside of course is that they also contain harmful saturated fat. To some extent the problem can be avoided for dairy products by using skimmed milk and yogurt made from skimmed milk, and for meat by cutting off the fat and grilling rather than frying. However, milk and dairy products inevitably contain some saturated fat. The best strategy seems to be to continue with a low consumption of these items, and balance their harmful effect with a high consumption of fruit and vegetables.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Causal networks 

215

Further investigations have shown that fruit and vegetables contain ­compounds which protect the arteries by mechanisms other than that of acting as antioxidants. One result, which has, understandably, attracted a lot of attention, is the finding that moderate consumption of red wine protects against heart disease. This has led to attempts to isolate the ingredient in red wine, which has this protective effect. Following this line of research, Roger Corder was led to the conclusion that the relevant ingredient was procyanidin. He writes (2007, pp. 47–8):

Copyright © 2018. Routledge. All rights reserved.

Our initial research – published in Nature in 2001 - …  focused on ­edothelin-1 because research over the previous decade had established that this substance plays a key role in heart disease. Endothelin-1 is a potent vasoconstrictor – it narrows blood vessels and therefore raises blood pressure – and triggers processes leading to atherosclerosis. …  analysis of a wide variety of wines has shown that the ability of any given wine to modify endothelin-1 synthesis correlates with its procyanidin concentration. This is a mechanism different from that of the anti-oxidants. Corder supplemented his laboratory work by epidemiological surveys in Sardinia (2007, pp. 91–5) and France (2007, pp. 99–103). His claim is that the regions with greatest longevity are those with most procyanidin in their wine. However, his epidemiological work is not of the same quality as his laboratory work. Corder calls his book The Wine Diet, but actually this stress on wine is a bit misleading, since he shows that procyanidins are to be found in as great a quantity in a variety of fruits as in wine. An apple for example contains as much procyanidin as a glass of procyanidin-rich red wine (Corder, 2007, p. 61). In Corder’s view this justifies the old saying that an apple a day keeps the doctor away. Procyanidins are also to be found in raspberries, blackberries, strawberries, and nuts – particularly walnuts. Nuts were a part of the traditional Mediterranean diet, and their protective value has been shown in a recent randomized controlled trial carried out in Spain (Estruch et al., 2013). This trial was conducted on a sample of individuals who were at risk of cardiovascular disease (2013, p. 1280): Eligible participants were men (55 to 80 years of age) and women (60 to 80 years of age) with no cardiovascular disease at enrollment, who had either type 2 diabetes mellitus or at least three of the following major risk factors: smoking, hypertension, elevated low-density lipoprotein cholesterol levels, low high-density lipoprotein cholesterol levels, overweight or obesity, or a family history of premature coronary heart disease. 7447 people satisfying these criteria were enrolled, of whom 57% were women. They were assigned randomly to one of three diets: a Mediterranean diet supplemented with extra-virgin olive oil, a Mediterranean diet supplement with mixed nuts, or a low-fat control diet. They were followed for a median of 4.8 years,

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

216  Causal networks

and major cardiovascular events were recorded (myocardial infarction, stroke, or death from cardiovascular causes). Both the extra olive oil and the extra nuts diet resulted in a relative risk reduction of these major cardiovascular events of about 30%. Of course, I have given just a few examples of ongoing research into nutrition, but the general results are clear enough. A healthy diet is one which contains a little, but not very much, meat and dairy products and so is low in saturated fat. It contains a great deal of fruit and vegetables, and some nuts. It contains fish, particularly oily fish. Consumption of eggs should be limited. The healthiest form of oil is olive oil, and so on. These results are very well confirmed empirically. The Russo-Williamson thesis is satisfied, and the results are supported both by evidence of mechanism, and by statistical evidence (both epidemiological surveys and controlled trials). The principle of strength through combining applies to the evidence. All this is well illustrated by the examples of research which have been given. I will next argue that given these strongly confirmed findings, fast food definitely constitutes an unhealthy diet which is a causal factor for heart disease. Although fast food was in its infancy then, Keys and Keys mention a typical fast food lunch and urge their readers not to eat it (1963, p. 163):

Copyright © 2018. Routledge. All rights reserved.

we plead for a little consideration of the nutrient content of what you can eat for lunch. The classic example of what not to eat is the popular “Coke” …  and hamburger, providing most of the calories as saturated fats and a good share of the remainder as “empty calories” in sugar. Actually, the typical fast food meal has changed (for the worse) since 1963. In London there are hundreds of fast food restaurants, many of which offer to deliver meals to your home. Some research of the menus suggests that a typical fast food meal nowadays consists of a cheeseburger, French fries (not mentioned by the Keys), and a soda drink, followed by a dessert. The dessert menu of a wellknown chain contained the following items (referred to as “treats”): chocolatey donut, blueberry muffin, sugar donut, triple chocolate cookie, apple pie, and toffee sundae. The chocolatey donut is shown as being covered with chocolate and having chocolate in the hole of the donut. As this is likely to be very tempting for a lover of fast food, let us select it as the dessert of our paradigm fast food meal. How healthy is this meal? Well, the hamburger patty contains saturated fat, and there will be more in the melted cheese. The donut of the dessert is likely to contain more saturated fat. Altogether the meal is heavy in saturated fat and so unhealthy. But this is far from the end of the story. Curiously in the seven countries study, Keys did not find that body mass and fatness were causal factors for coronary heart disease. This is probably because no one in his cohorts reached the levels of overweight/obesity which are common today. Now what is most likely to lead to excessive weight gain is the consumption of high energy

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Causal networks 

217

Copyright © 2018. Routledge. All rights reserved.

density foods, because these deliver a large number of calories in a small and easily digested form. Conversely to avoid weight gain, it is desirable to eat low energy density foods, which contain a good deal of fibre and bulk. These fill one up without delivering a large number of calories. The French fries, which do not seem to have been a standard part of the offering in 1963, are a classic high energy density food. They are cut very thin which means that they absorb a lot of oil when fried and so become more laden with calories. So, they are a dish conducive to weight gain. They are also typically very salt, and salt consumption leads to high blood pressure, which is another causal factor in heart disease (see Figure 13.3). Sweet foods such as the chocolatey donut are also high energy density foods, as are sweet drinks. It is interesting that the Keys refer to a sugary drink as “empty calories”. By this they mean that such drinks add to calories without providing other valuable nutrients such as would be provided by eating a piece of fruit. As regards the calorie content of a fast food lunch, Sanders and Bazalgette (1991, p. 319) give an interesting comparative estimate. They consider first a traditional British Sunday lunch consisting of roast lamb, roast potatoes, peas, and canned peaches and custard. Such a meal was the high spot of the week, and we might well expect it to contain more calories than the usual lunch. Sanders and Bazalgette estimate its calories at 650, and its fat at 27 g. They then contrast this with a typical fast food meal, which they take as consisting of beef burger, chips, fruit pie, and a can of cola. This comes out at 1450 cals and 66 g of fat. In other words, the fast food meal has 2.23 x the calories of the traditional Sunday lunch, and 2.44 x the fat. The situation with a typical fast food menu of today may again be worse than it was in 1991. Cheeseburgers are now more common, and the chocolatey donut is likely to have more calories than the fruit pie. There has also been the phenomenon of “super sizing” which Schlosser describes as follows (2001, p. 241): Over the past forty years in the United States, per capita consumption of carbonated soft drinks has more than quadrupled. During the late 1950s the typical soft drink order at a fast food restaurant contained about eight ounces of soda; today a “Child” order of Coke at McDonald’s is twelve ounces. A “Large” Coke is thirty-two ounces – and about 310 calories. In 1972, McDonalds’s added Large French Fries to its menu; twenty years later, the chain added Super Size Fries, a serving three times larger than what McDonalds’s offered a generation ago. Super Size Fries have 610 calories and 29 grams of fat. At Carl’s Jr. restaurants, an order of CrissCut Fries and a Double Western Bacon Cheeseburger boasts 73 grams of fat – more fat than ten of the chain’s milk shakes. The important theoretical point here is that there are three causal paths which lead from unhealthy diet to heart disease, as can be seen in Figure 13.3. The first is: unhealthy diet →  high blood pressure →  heart disease. Here the unhealthy component in fast food is salt which is found in particularly large quantities in French

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Copyright © 2018. Routledge. All rights reserved.

218  Causal networks

fries, and potato crisps (or chips in American English). The second is: unhealthy diet →  high cholesterol →  heart disease. Here the principal culprit is saturated fat which is found in large quantities in fast food. The third is unhealthy diet → overweight/obese →  heart disease. This path can also go via Diabetes, which is itself a most unpleasant disease, as in: unhealthy diet →  overweight/obese →  diabetes →  heart disease. The chief culprit in fast food for this pathway is energy dense foods, which include savoury foods such as French fries, but also, and particularly, sweet foods and sugary drinks. At this point we should mention what could be called the “statin strategy”. Statins are an excellent drug for lowering blood cholesterol level. If anyone has a high cholesterol level, and is unable (or unwilling) to lower this level by changes in diet, then he or she would be very well advised to take statins on a regular basis. This suggests the general strategy of eating fast food but taking statins to counter its ill-effects. Is this a good strategy? Well, for those who are so addicted to fast food that they can’t give it up, the strategy is indeed a good one. However, it is hardly to be recommended for a whole population. To begin with, statins, though very safe drugs, do, like all drugs, have some negative side effects. Secondly, however, the statin strategy only blocks the second of the three paths from unhealthy diet to heart disease. It leaves the first and third paths completely open. So, statins are not the whole answer to the problem of unhealthy eating. So far, I have not attempted a general characterization of fast food, but only given a paradigm example, the cheeseburger, French fries, soda, and chocolatey donut meal. Fast food is, however, much more general than the hamburger restaurant. Pizzas have evolved from quite a healthy Neapolitan dish with only a little cheese as flavouring, to the offerings of typical chains which have toppings rich in saturated fat. Seemingly innocuous coffee chains are in reality purveyors of unhealthy fast food. A latte and snack consisting of a muffin or a Danish pastry contains a great deal of saturated fat, and is very high in calories. Sandwich chains are not much better. My own investigations have shown that most sandwiches on offer are high in saturated fat, and, from the range on display it is usually not possible to find more than one or two, which are low in saturated fat. With care, it is possible to eat reasonably healthily in some sandwich chains, but one has to be very careful. As well as restaurants, one should include the ready meals in supermarkets. Of course, there is no reason why fast food should be unhealthy, and some fast food, such as sushi, might indeed be healthy; but such cases are the exception rather than the rule. So Ancel and Margaret Keys’ ‘Eat Well and Stay Well’ project has been a failure, while the alternative ‘Eat Fast Food and Get Ill’ has triumphed in the USA, Britain, and increasingly in the rest of the world. Even the populations of areas noted in their past for healthy eating, such as Japan, Greece, and Southern Italy, have begun to consume fast food, and eat less healthily than they did before. So, the project has failed at a social level, but it has succeeded for some individuals, such as Ancel Keys and his wife. Ancel Keys live to be 100 and his wife Margaret to be 98 enjoying delicious Mediterranean cooking in their villa overlooking the sea in Southern Italy, and never consuming any fast food.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Copyright © 2018. Routledge. All rights reserved.

Causal networks 

219

The general picture of how eating habits have changed worldwide in the last 50 or so years is certainly a gloomy one, and the question arises as to whether it is possible to change the situation. Could something like the Keys’ ‘Eat Well and Stay Well’ be revived in an updated form, and perhaps be implemented more successfully this time? It seems worth discussing this question briefly. It might first be objected that Ancel and Margaret Keys were just elitist, and that their ideas could not be applied to the mass of the population. “It was all very well for them” it might be said “to relax in their villa by the shores of the Mediterranean, but the less well off in the USA or Britain simply could not afford the kind of food they ate, let alone a villa in the sun.” There is some truth in this point. In the USA and Britain, the cheapest foods are indeed unhealthy. One does not have to be very rich to afford a healthy diet containing a good deal of fruit vegetables and fish, but probably this is beyond the means of the very poorest. However, it must be remembered that this is not an intrinsic law, but a result of the way agriculture and the economy are organized. When Keys began his work, the workers in Naples ate more healthily than the rich, because a healthy diet was then cheaper than one with a great deal of meat and cheese (section 6.1). In the seven countries study the healthiest diet was to be found in Crete which was then one of the poorest countries in Europe, and among a rather lower-class section of the population. Once agriculture is mechanized and geared to meat and dairy production, then the cheapest types of food become the worst and fattest cuts of meat, and the lowest quality cheeses. These are precisely the foods which are highest in saturated fat. However, there are, apart from the quality of the food produced, major objections to a mechanized agriculture geared to meat and dairy production. Such an agriculture depends heavily on fossil fuels, and the animals emit carbon dioxide as well. So, the net effect is damage to the environment and an increase in global warming. Besides, the factory farming of animals is a process which rightly horrifies most people.1 What is needed anyway is a switch to an agriculture which produces more fruit and vegetables and less meat and dairy. If the public were eating less meat and dairy, they could afford to pay more for it, and this would enable animals to be reared in a more humane manner. Meat would become, as it was in the past, something eaten only two or three times a week, but, in compensation, it would be much more tasty. A more sustainable and environmentally friendly agriculture of this type would provide exactly the type of food needed for healthy eating. It might still be objected that fast food was destined to win because it is intrinsically more palatable than healthier diets. It is argued that the fat in fast food gives it a pleasant feel in the mouth, that fast food is a “comfort food” which is more appealing that healthier foods. Such arguments are not very convincing. Nowadays the inhabitants of large cities can sample the cuisines of other countries in a variety of restaurants; but, even as recently as 50 years ago, and for centuries before that, people would eat only the cuisine of their own country or region. Moreover, these cuisines differed very much in taste, cooking, and the

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Copyright © 2018. Routledge. All rights reserved.

220  Causal networks

type of food consumed. Chinese, European, Japanese, and Indian cuisines were all very different, and, within a region such as Europe, there were further differences. Italian, French and English cooking differed in many ways. Now what was the result of this situation? It can be summarized by saying that everyone thought that the cuisine of their own region was the best and that foreign cuisines were much inferior. This surely shows that humans are very adaptable, and that everyone likes best the kind of food and cooking which they habitually eat. It is thus very questionable whether one type of food (fast food say) is intrinsically more appealing to humans than another (the traditional Mediterranean diet say). Let us compare our paradigm fast food meal (cheeseburger, French fries, etc.) with the kind of meal which Ancel and Margaret Keys might have consumed in their villa overlooking the Mediterranean. This paradigm healthy meal could be something like the following: (i) spaghetti with fresh tomato and basil sauce; (ii) grilled swordfish steak, with a dressing of olive oil and parsley, and served with Calabrese aubergines and peppers; (iii) fresh peaches. (i) and (ii) would be served with Italian bread without butter or other spread, but very useful for mopping up the sauces, and, of course, water and a glass of wine to drink. It cannot be denied that a habitual eater of fast food would not find this meal to his or her taste. It would probably seem insipid and unsubstantial, and the taste of the aubergines might be regarded as disagreeable. Conversely, however, the habitual eater of traditional Mediterranean food is likely to regard our paradigm fast food meal with horror and disgust, as something quite uneatable. The thought of a greasy hamburger and sickly sweet soda and donut might turn such a person’s stomach. This is another illustration of the rule that people like whatever food they habitually eat. Given this rule, however, it is surely sensible for people to eat a diet which has been shown to be healthy, for, after eating it for a while, they will get to like it better than any other food, and they will avoid the illnesses to which unhealthy eating leads. If this point is accepted, it might still be objected that the same rule shows that it is very difficult to get people to change the food they eat, and this will make the spread of healthy eating very difficult since fast food is now the established way of eating in many countries. This point is of course correct, but, it must be remembered that fast food was at one time something new. Hamburgers, pizzas, and Danish pastries were almost unknown in Britain 60 years ago, but they have all become staples. The fast food industry did manage to change people’s eating habits. So, by studying the methods they used to bring about this change, we can get ideas about how a change to healthier eating could be brought about. One basic principle used in selling fast food was that of targeting children; and this principle is clearly very sound. Adults usually continue to like the kind of food they ate when they were children. Schlosser has a good account of how this worked in practice (2001. p. 47): The McDonald’s Corporation now operates more than eight thousand playgrounds at its restaurants in the United States. Burger King has more

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Causal networks 

221

than two thousand. A manufacturer of “playlands” explains why fast food operators build these largely plastic structures: “Playlands bring in children, who bring in parents, who bring in money.” As American cities and towns spend less money on children’s recreation, fast food restaurants become gathering spaces for families with young children. Every month about 90 percent of American children between the ages of three and nine visit a McDonald’s. The seesaws, slides, and pits full of plastic balls have proven to be an effective lure. “But when it gets to brass tacks,” a Brandweek article on fast food notes, “the key to attracting kids is toys, toys, toys.” The fast food industry has forged promotional links with the nation’s leading toy manufacturers, giving away simple toys with children’s meals and selling more elaborate ones at a discount. The major toy crazes of recent years – including Poké mon cards, Cabbage Patch Kids, and Tamogotchis – have been abetted by fast food promotions. A successful promotion easily doubles or triples the weekly sales volume of children’s meals. The chains often distribute numerous versions of a toy, encouraging repeat visits by small children and adult collectors who hope to obtain complete sets.

Copyright © 2018. Routledge. All rights reserved.

These are very ingenious ways of selling fast food. No doubt they were the product of the most creative and intelligent minds of the business community of the time. Naturally they could be easily adapted to persuading children, and future adults, to eat healthily rather than unhealthily.2 We have already mentioned “super sizing” as a technique used by fast food companies in an attempt to get their customers to eat more. Pollan (2006) gives an interesting account of how this new method for increasing sales was devised. David Wallerstein was the inventor of super sizing, and as Pollan explains (p. 105): Wallerstein …  in the fifties and sixties …  worked for a chain of movie theaters in Texas, where he labored to expand sales of soda and popcorn – the high-markup items that theaters depend on for their profitability. …  Wallerstein tried everything he could think of to goose up sales – twofor-one deals, matinee specials – but found he simply could not induce customers to buy more than one soda and one bag of popcorn. He thought he knew why: Going for seconds makes people feel piggish. Wallerstein discovered that people would spring for more popcorn and soda – a lot more – as long as it came in a single gigantic serving. Thus was born the two-quart bucket of popcorn, the sixty-four-ounce Big Gulp …  Now comes the irony in the story, for, when Wallerstein went to work for McDonald’s in 1968, he was at first unable to persuade the company’s boss, Ray Kroc, that super sizing would increase sales. It seems that Kroc could not initially believe that people would behave so irrationally. However, Wallerstein eventually succeeded in convincing Kroc, and the rest is history. Wallerstein’s discovery has most of the characteristics of great creative innovations. The discoverer wrestles

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Copyright © 2018. Routledge. All rights reserved.

222  Causal networks

with an apparently intractable problem, and tries solution after solution, all of which fail. Then there is a sudden and crucial insight, which leads to success; but this insight is counter-intuitive, and the discoverer has great difficulty in persuading people that his method really works. All these characteristics are to be found in Semmelweis’ key innovation described in section 10.3. Wallerstein’s crucial insight was used by the fast food industry to encourage people to eat more food. Now, if we want to promote healthy eating, we need to encourage people to eat less food. However, Wallerstein’s insight can easily be adapted for this purpose. We simply have to serve smaller portions, since people are on the whole reluctant to appear greedy by ordering a second portion. Examples like these show that a change from fast food to healthier eating could be promoted by exactly the same methods which were used to encourage people to eat fast food in the first place. Consequently, a change from fast food to healthier eating is by no means impossible to achieve. That concludes the present section on fast food, and I will now return to the main theme of Part III of the book. This section has, I would claim, established that the causal law: eating fast food → heart disease (→ = causes) is very well confirmed empirically, and can consequently be used as a basis for action. Naturally I have not presented all the evidence for this conclusion in detail, but enough has been given to show that this causal law is well-supported by all the various types of evidence – epidemiological surveys, controlled trials, and evidence of mechanism. The Russo-Williamson thesis is satisfied, and the principle of strength through combining results in augmented empirical confirmation when the various types of evidence are put together. Exactly the same can be said of the causal law: smoking →  heart disease. The case was made in sections 8.4, 8.5, and 8.6. It follows that in the causal network of Figure 13.5, the two causal claims are strongly confirmed empirically. The figure does not show the probability distributions associated with the network, but the empirical evidence establishing the causal links also provides data from which the probabilities in particular situations can be estimated. These points will be important when, in Chapter 15, I criticize Pearl’s claim that the causal network of Figure 13.5 is not a genuine causal model. For the moment, let us provisionally accept that it is not only a genuine causal model, but also one, which is sufficiently strongly confirmed empirically to be the basis for action. This raises the question of what deductions regarding action we can make from this model. I will consider this question in the next section.

13.4  The Hesslow counter-example revisited At first sight it seems easy to make deductions concerning desirable actions on the basis of a causal model such as that illustrated in Figure 13.5. In fact, cardiologists routinely recommend to patients’ strategies for preventing heart disease on the basis of the knowledge embodied in Figure 13.5. A cardiologist would typically advise a patient to give up smoking, and to replace a fast food diet by healthy eating, which, according to the standard medical specifications, consists of cutting down on meat and dairy products, increasing consumption of fruit and vegetables, and so

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Causal networks 

Copyright © 2018. Routledge. All rights reserved.

FIGURE 13.6 

223

Multi-causal fork for Hesslow example.

on. Moreover, if the patient is not strong-willed enough to give up both smoking and fast food, then the cardiologist would advise him or her to give up at least one. All this advice seems intuitively correct, but, as I will now show, we cannot in general draw such conclusions from a causal model like that illustrated in Figure 13.5, because of cases like the Hesslow counter-example. The Hesslow counter-example was described in section 12.2, and we can now illustrate it by a two-pronged multi-causal fork like that of Figure 13.5. This is shown in Figure 13.6. On the basis of the causal model of Figure 13.5, it seems perfectly reasonable, in order to avoid heart disease, to give up smoking. Giving up smoking lowers the probability of getting heart disease. If we reason in the same way on the basis of the causal model of Figure 13.6, it would seem reasonable, in order to avoid thrombosis, to give up taking the pill. However, in the circumstances described in the Hesslow counter-example, giving up the pill increases the probability of getting pregnant, which carries with it a greater risk of thrombosis than it produced by taking the pill. It follows that giving up taking the pill, instead of lowering the probability of a thrombosis, actually increases it. This creates a tricky situation. We have to find some way of making deductions from models such as multi-causal forks which allows the valid conclusion that giving up smoking reduces the probability of heart disease, while blocking the incorrect conclusion that giving up taking the pill in the Hesslow situation reduces the probability of thrombosis. How can this be done? The first step in my view is to examine more closely the interpretation of probability in causal models such as multi-causal forks. This will be carried out in the next chapter.

Notes 1 A very interesting account of the present day agricultural system in the USA is to be found in Pollan (2006, particularly Part I, pp. 15–119). Pollan shows clearly that this system is very damaging to the environment, and unsustainable long-term. For example, he describes a fast food meal which he, his wife and his son consumed, and

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

224  Causal networks

Copyright © 2018. Routledge. All rights reserved.

calculates that the production of its food calories (2006, p. 117): “took at least ten times as many calories of fossil energy, the equivalent of 1.3 gallons of oil.” (I am very grateful to David Teira for drawing my attention to this book, and indeed giving me a copy.) 2 It should be stressed, however, that a healthy diet for a child is not the same as that for an adult, since children are growing. Fast food, however, is just as unhealthy for children as it is for adults. Indeed, it is probably even more unhealthy, as is suggested by the increasing rates of childhood obesity and childhood onset of type II diabetes.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

14

Copyright © 2018. Routledge. All rights reserved.

HOW SHOULD PROBABILITIES BE INTERPRETED?

The problem of this chapter is how to interpret the probabilities which appear in causal models in medicine. To simplify the discussion without losing too much generality, I will confine myself to two-pronged multi-causal forks, such as are illustrated in Figure 13.4. The smoking, fast food, and heart disease example (Figure 13.5) and the Hesslow counter-example (Figure 13.6) are instances of this type of model. In the figures, probabilities are not shown, but they are implicit. In Figure 13.4, it is assumed that the variables X, Y, Z are random variables which have a joint probability distribution. In specific cases like the smoking, fast food, and heart disease example (Figure 13.5), the probabilities can be estimated from statistical evidence, such as that to be found in the table given in section 8.4. As a final simplification, assume that X, Y and Z are bivariate variables taking the values 0 or 1. So people are classified into smokers (X = 1) or non-smokers (X = 0), and similarly in the other cases. This is obviously a simplification, since in the earlier table of results regarding smoking and heart disease, the doctors were classified into various groups according to the amount that they smoked. However, as our project is to study the relations between causality and probability, it seems best to take a simple, though quite realistic, case first, and then to consider the effects of adding more complexities later. I will refer to the two-pronged multi-causal fork with binary variables as our simple model. In trying to link the causes in this model with probabilities, I will follow the probabilistic causality tradition initiated by Good (1959). This is based on what I called the causality probability connection principle or CPCP. This was formulated in section 12.2. In the context of our simple model, it takes the form:

If X causes Z, then P ( Z = 1| X = 1) > P ( Z = 1| X = 0 ) (*)

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

226  How should probabilities be interpreted?

In the case of smoking and heart disease, it becomes: If smoking causes heart disease, then P(heart disease | smoking) >  P(heart disease | non-smoking) Of course, this sounds eminently reasonable, but, as we saw in sections 12.2 and 13.4, (*) is liable to counter-examples such as the Hesslow counter-example. Our problem is to put restrictions on (*) so that, with these restrictions, it can be used to draw valid conclusions and the counter-examples are avoided. Our first step in trying to formulate such restrictions, is to consider what interpretation should be given to the probabilities in the causal model and in (*). I will give a brief survey of the various interpretations of probability in the next section.

Copyright © 2018. Routledge. All rights reserved.

14.1  Interpretations of probability1 Probability and statistics are now used constantly within scientific theories and for the evaluation of evidence. Unfortunately though, there are still considerable controversies about how this use should be carried out. The good news is that everyone agrees on the mathematical theory of probability. This is based on a set of axioms introduced by the Russian mathematician Kolmogorov in 1933. Since then, almost all mathematicians have accepted the Kolmogorov axioms for probability. Yet although everyone agrees on the formal mathematical theory, this theory can be interpreted in a number of different ways. There are different meanings of probability, and, in a particular application one needs to be clear about the sense of probability being used. It is here that controversies can and do arise. Ian Hacking (1975) has memorably described probability as being “Janusfaced”. This is after the ancient Roman god “Janus” whose name is used for our month of January. Janus had two faces looking in opposite directions, and this is true also of probability. One face of probability is the objective face. This refers to probabilities which exist in the world quite independent of human beings. Consider for example the probability of a particular isotope of uranium disintegrating in the course of a year. This probability has a definite value which has been found by physicists. It had this value before there were human beings, and will continue to have this value if we destroy ourselves in some catastrophe. It is a constant of nature independent of us. The other face of probability is the epistemic face, from the Greek epistē mē , meaning “knowledge”. This is the kind of probability we use when we say that there is a high probability of the big bang theory being true, given present evidence. It is the kind of probability which is used to analyse human knowledge, and so the kind of probability which would not exist if human beings ceased to exist. I will give an example of an epistemic theory in a moment, but let us start with objective theories. There are two main objective interpretations, namely (i) the frequency theory, and (ii) the propensity theory. I will deal with these in turn. The frequency theory of probability was developed by Venn at Cambridge in the 19th century, and by an Austrian philosopher and mathematician (Richard von Mises),

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Copyright © 2018. Routledge. All rights reserved.

How should probabilities be interpreted?  227

who was associated with the Vienna Circle. Von Mises published a well-known account of his theory in 1928. According to von Mises, probability theory is a mathematical science like mechanics, but, instead of dealing with the motions and states of equilibrium of bodies and the forces which act on them, it treats (1928, p. 11): “problems in which either the same event repeats itself again and again, or a great number of uniform elements are involved at the same time.” Von Mises called his repetitive events and mass phenomena: collectives. The simplest examples of collectives, come from games of chance, where we deal, for example, with a long sequence of tosses of a particular coin. However, we can have examples from the natural sciences, such as the case of radioactive isotopes mentioned already. Suppose we have a large collection of radioactive atoms, which are all specific isotopes of uranium. This would be a collective in von Mises’ sense. Now at each member of a collective a particular attribute occurs. For example, in a collective of coin tosses, each member is either heads or tails. In the collective of radioactive atoms, each atom either disintegrates in the course of a year or remains intact. Let us call a particular attribute, which may or may not occur at each member of a collective, A. Suppose in the first n members of the collective, e.g. the first n tosses of the coin, A occurs m(A) times. Then m(A)/n is called the frequency of the attribute in the collective. We now define the probability of A as the limiting frequency of m(A)/n as n tends to infinity. The frequency theory of probability is quite plausible, and still widely held. Yet there are difficulties in it. For example, probabilities are only attributed to collectives, but don’t we also want to attribute probabilities to the individual members of the collective? Consider the collective of atoms of a specific radioactive isotope of uranium. Don’t we want to attribute the probability of disintegrating within a year to each individual atom, and not just to a collection of these atoms? This problem led to the development of the propensity interpretation of probability, which I will now consider. The basic idea of the propensity theory is due to the American philosopher Charles Peirce in 1910. Peirce identified the probability of a particular die giving a specific result with what he called a “would-be”, remarking rather wittily (1910, p. 79): the die has a certain “would-be” …  quite analogous to any habit that a man might have. Only the “would-be” of the die is presumably as much simpler and more definite than the man’s habit as the die’s homogeneous composition and cubical shape is simpler than the nature of the man’s nervous system and soul. Popper took up this idea (1957a) where he introduced the propensity theory of probability. On this approach we attribute to a particular radioactive atom a propensity to disintegrate within a year. This propensity is a probability. Probabilities are not identified with frequencies, but may manifest themselves in the appearance of patterns of frequencies, just as a man’s habit, e.g. drinking too much, may manifest in various patterns of behaviour, such as visiting bars frequently, etc.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

228  How should probabilities be interpreted?

This gives rise to a problem. If probabilities are these ­theoretical propensities, how do we link probabilities to the frequency data which we observe? Popper’s approach is that we can conjecture any values we like for these propensities, but we must test out such conjectures against frequency data to try to refute or falsify our conjectures. Popper’s approach seems to me correct, but it does give rise to a serious problem, for, in a strict logical sense, probability statements are not falsifiable by frequency data. To see this, let us take a simple example. Suppose we are tossing a bent coin, and postulate that the tosses are independent, and that the probability of heads is p. Let Prob(m/n) be the probability of getting m heads in n tosses. Then we have

Copyright © 2018. Routledge. All rights reserved.



Prob ( m / n ) = n Cm pm (1 − p)

n −m



So, however long we toss the coin (that is, however big n is) and whatever number of heads we observe (that is, whatever the value of m), our result will always have a finite, non-zero probability. It will not be strictly ruled out by our assumptions. In other words, these assumptions are (Popper, 1934, p. 146) “in principle impervious to strict falsification.” Popper’s answer to this difficulty consists in an appeal to the notion of methodological falsifiability. Although, strictly speaking, probability statements are not falsifiable, they can nonetheless be used as falsifiable statements, and in fact they are so used by scientists. Popper’s approach has been strongly vindicated by one of the main approaches to statistics – what is called classical statistics. Classical statisticians are constantly applying one or other of a battery of statistical tests. Now, whenever they do so, they are implicitly using probability hypotheses, which from a strictly logical point of view are unfalsifiable, as falsifiable statements. The procedure in any statistical test is to specify what is called a “rejection region”, and then regard the hypothesis under test (H say) as refuted if the observed value of the test statistic lies in this rejection region. Now there is always a finite probability (called the “significance level” and usually set at around 5%) of the observed value of the test statistic lying in the rejection region when H is true. Thus, H is regarded as refuted, when, according to strict logic, it has not been refuted. This is as much as to say that H is used as a falsifiable statement, even though it is not, strictly speaking, falsifiable, or, to put the same point in different words, that methodological falsifiability is being adopted. Thus, what underlies classical statistics is an objective notion of probability, and methodological falsifiability. Let us now turn to an example of an epistemic theory of probability, namely the subjective theory. Subjective probabilities represent the degree of belief of a particular person (Mr B say) in the occurrence of a specific event (E say). To illustrate this, let us consider a family, Mr and Mrs B with their children Master B and Miss B, who are on holiday. The weather one evening looks threatening and they are considering whether it will rain the next day which might ruin their planned boat trip. Each member of the family has a different view of the

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

How should probabilities be interpreted?  229

Copyright © 2018. Routledge. All rights reserved.

question. Mr B, who inclines towards pessimism, has a high degree of belief that it will rain. Master B, his son, who is happy-go-lucky by nature, has a low degree of belief. The two women of the family have intermediate positions. So, each family member has a different degree of belief in the event of rain the next day. Now it might be objected that while this is quite realistic at an everyday level, the mathematical theory of probability can hardly apply here, because the degrees of belief held by the various family members are qualitative in character and cannot be measured exactly. However, the mathematical calculus of probability deals with numerical probabilities. This is certainly a major difficulty, but it was overcome by the independent work of Ramsey (1926) and De Finetti (1937). Essentially Ramsey and De Finetti suggested that Mr B’s degree of belief in the occurrence of E could be measured by the rate at which he would be prepared to bet on E (his so-called betting quotient on E) in a specially designed betting situation. Mr B has, however, to choose his betting quotients carefully because of the possibility of what is called a Dutch book. A cunning opponent (Ms A say) makes a Dutch book against Mr B, if, given Mr B’s betting quotients, she chooses stakes and the direction of bets so that Mr B loses whatever happens. Obviously, Mr B will want to choose his betting quotients so that a Dutch book cannot be made against him. Ramsey and De Finetti showed independently that Mr B can avoid the possibility of a Dutch book against him if and only if he chooses his betting quotients to satisfy the standard axioms of the mathematical theory of probability. This striking and famous result, now known as the Ramsey-De Finetti theorem, provides a foundation for the subjective interpretation of probability. It shows that an individual’s degrees of belief can be measured as betting quotients and that these betting quotients should be chosen to satisfy the ordinary axioms of probability. That concludes my very brief sketch of some of the interpretations of probability. Let us next see how they can be applied in the context of causal models.

14.2  Interpreting the probabilities in causal models So how should probabilities in causal models be interpreted? Here there are differences of opinion. Pearl has always interpreted the probabilities in causal networks and causal models subjectively as degrees of belief (see Pearl [1988] and Pearl [2000]). Most supporters of subjective probability adopt the Bayesian approach to statistical inference, which is an alternative to classical statistics, and Pearl is no exception to this rule. His research was motivated by the Bayesian approach (see Pearl [1982], [1985a] and [1985b]), and he invented the term: “Bayesian network” for some of the causal networks he considered (see Appendix 2). The approach adopted in this book is different. Pearl’s “Bayesian networks” can be reinterpreted so that the probabilities which occur in them are objective, and so that they can be handled by the methods of classical statistics rather than those of Bayesianism. It will be shown, with examples, in Chapter 15, how this can be done. In the present chapter, I will concentrate on how the

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

230  How should probabilities be interpreted?

probabilities in causal models can be interpreted objectively. This approach was already s­ uggested in Neapolitan (1990). As we saw in section 14.1, there are two objective interpretations of probability – the frequency and the propensity. I prefer the propensity interpretation, but this at once poses a problem since there are many different versions of the propensity approach. An account of some of these different versions is to be found in my 2000 work and also in a more recent paper (2016a). The views expressed in the 2016 are somewhat different from those I held in 2000. This change was due to some very helpful comments I received from Alan Há jek. Without going into all the complexities of the issues involved here, I will now show how to interpret the probabilities involved in causal models in terms of my own preferred version of the propensity theory. It is expounded in detail in my 2016a paper, and the main points of the theory can be summed up in the following four propositions:

Copyright © 2018. Routledge. All rights reserved.

1. Probabilities are associated with the outcomes of a set of repeatable conditions S. 2. To say that the probability of A given S is p [P(A | S) = p] is to claim that the conditions S have a propensity equal to p to produce A on any repetition of S. 3. Propensities satisfy the Kolmogorov axioms, so that the propensity interpretation is an interpretation of probability as it is used in the standard mathematical theory of probability. 4. Propensities are connected to frequencies, not by a definition as in the frequency theory of probability, but by adopting a principle of methodological falsifiability which is designed to underpin the standard statistical tests. The propensity interpretation, as defined by the above four statements is an objective interpretation of probability. What ensures objectivity is the association of propensities with repeatable conditions. This means that it is possible to test whether any value p assigned to a probability agrees with experience or not. In principle all one needs to do is to repeat the conditions a large number of times and see whether the observed long-run frequency is close to p or not. Condition (3) ensures that one can do the necessary probability calculations here, while condition (4) ensures that a standard statistical test can be applied to check whether the observed frequency agrees with the value p postulated for the probability, or whether there is a significant divergence between the two. This version of the propensity theory is the version which is most similar to von Mises’ frequency theory. It is possible therefore to take some ideas from von Mises’ frequency theory, transfer them to our long-run propensity theory, and then apply them to the problem of linking causality and probability. In order to carry out this plan, I will begin by making a few remarks about von Mises’ frequency theory of probability.2 As we saw in section 14.1, von Mises uses the concept of collectives. Collectives are long sequences of attributes, and the probability of an attribute

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Copyright © 2018. Routledge. All rights reserved.

How should probabilities be interpreted?  231

in a collective is defined as the limit of the frequency with which that attribute appears in the collective. An attribute may, however, appear in a number of different collectives, and have a different probability in each of these collectives. We can illustrate this situation by a couple of examples. Consider first a collective C1 say consisting of a long sequence of throws of a standard fair die. The basic attributes here are 1, 2, 3, …  , 6, and P(5 | C1) = 1/6. However, we can define a new collective as follows. We take the original collective C1 but form a subsequence by selecting only those members of C1 which have an odd attribute, i.e. 1, 3, or 5. Call this new collective C2, then P(5 | C2 ) = 1/3. Another, quite similar, example is the following. Suppose we are considering the probability of a 40-year-old man living to be 41. Let us take the collective C1 to be the collective of all 40-year-old men. A statistical investigation will give us the value of P(living to be 41 | C1) – p say. But now let us consider the collective C2 say, consisting of all 40-year-old Japanese men. Since the Japanese are noted for their longevity, we may presume that P(living to be 41 | C2 ) has a value q say, where q >  p. In general then, changing the collective changes the probability, and this led von Mises to propound the principle that probability (in his frequency sense) is relative to a collective. Now in our version of the propensity theory, probabilities are associated with repeatable conditions rather than collectives. However, von Mises’ argument goes through exactly as before, and we need only alter the above principle to: probability (in the propensity sense) is relative to a set of repeatable conditions. Yet in order to connect the discussion within the propensity theory even more closely to earlier discussion within the frequency theory, I propose the following strategy. It is always possible to consider the repeatable conditions of the propensity theory extensionally by taking a sequence of repetitions of those conditions. Thus, instead of the condition “ …  is a 40-year old man”, we can consider “the set of 40-year old men”; instead of the condition “…  is a 40-year old Japanese man”, we can consider “the set of 40-year old Japanese men”; and so on. To do this, I propose to use the term “reference class” in the following precise sense. A reference class is associated with a set of repeatable conditions S, and consists of a long sequence of repetitions of S. Reference classes are collectives in von Mises’ sense, but not all collectives in von Mises’ sense are reference classes.3 In our version of the propensity theory, it is perfectly legitimate to associate probabilities with reference classes, and, by doing so, we can carry over to the propensity theory many principles formulated within the frequency theory. In particular the key principle now becomes: propensity is relative to a reference class. To resolve our problem about connecting causality to probability, the essential move is to apply this principle to (*). To do so, we have to recognize that the same principle applies to causality in the generic sense as to propensity, that is to say we have the principle: causality is relative to a reference class. Consider, for example, the case of a medication M say, which has no side effects for the average patient, but produces vomiting and dizziness in diabetic patients. We then have that, in the reference class of diabetic patients, “M causes vomiting and dizziness” is true,

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Copyright © 2018. Routledge. All rights reserved.

232  How should probabilities be interpreted?

but that in the reference class of non-diabetic patients, it is false. This example can easily be generalized, and so the reference class principle can be extended from propensity to causality. In particular it applies to both sides of the suggested Causality Probability Connection Principle (*). We can now ask the question: “For which reference class or reference classes does (*) hold?” Because we are applying the propensity interpretation, the variables X, Y, Z are all associated with some underlying set S of repeatable conditions. However, Hesslow’s counter-example shows that (*) does not in general hold for the reference class S. This is because the causal effects of X on Z are disturbed by the causal effects of Y on Z, given that X and Y are themselves causally linked. This suggests that, in order to make manifest the causal effect of X on Z, we have to hold Y fixed. Since we are dealing with the simple case of binary variable, we have just two cases Y = 0 and Y = 1. My proposal then is to divide S into two disjoint reference classes S & (Y = 0) and S & (Y = 1), and to claim that (*) holds for each of these two reference classes, but not necessarily for S itself.4 The division of a reference class into two disjoint reference classes can be illustrated by our earlier example of the reference class consisting of a sequence of throws of a standard fair die. This reference class can be divided into the reference class consisting of those throws which have an odd result, and the reference class consisting of those throws which have an even result. In Hesslow’s example, the underlying reference class S say consists of a set of young women living with male partners in a situation in which taking the pill is the only contraceptive method available. This can be divided into the following two disjoint reference classes: S & (Y = 0), i.e. the set of those who do not become pregnant in the time period under consideration, and S & (Y = 1), i.e. the set of those who do become pregnant. Now in both these reference classes, (*) holds. Consider S & (Y = 0), this is the set of young women who do not become pregnant, those who take the pill (X = 1) have a higher probability of getting a thrombosis than those who do not (X = 0) because of the side effects of the pill. Consider next S & (Y = 1), this is the set of young women who do become pregnant. In this case it might be objected that there are no members of this set who take the pill (X = 1). However, I would reply that the pill is unlikely to be 100% effective, and that, because of the side effects of the pill, someone who both took the pill and became pregnant is likely to have a higher probability of getting a thrombosis than someone who became pregnant without having taken the pill. So (*) again holds. Our proposal does therefore resolve the problem created by the Hesslow counter-example. Both this problem and the suggested solution are rather closely related to Simpson’s paradox, as I will show in the next section.

14.3  Simpson’s paradox The problem generated by the Hesslow counter-example is, as Galavotti observes in her 2010 work, quite similar to that of Simpson’s paradox. To see this, let us consider the most famous example of Simpson’s paradox, which concerned

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

How should probabilities be interpreted?  233

Copyright © 2018. Routledge. All rights reserved.

the seeming occurrence of gender discrimination in admission to Berkeley’s ­g raduate school. Let us take as our initial reference class S, the set of applicants to this graduate school. These applicants were either male or female, and their application was either successful or unsuccessful. Statistics showed that, in the reference class S, P( success | male) >  P(success | female). It therefore looked as if gender-discrimination in favour of males was at work. However, in practice, applicants applied to particular departments and each department conducted its own admission procedure. So, if there were gender-discrimination, it would have to be at the departmental level. Now suppose the departments were D1, D2, …  , Dn. Then we can partition our reference class S into n sub-reference classes S & Di where 1 ≤  i ≤  n. Statistics showed that, for each of these sub-reference classes, P(success | male)   P(E | ¬ C) or equivalently P(E | C) >  P(E).2 This is a mistake, however, because (Pearl, 2011, p. 715): “the relationship ‘raises the probability of ’ is counterfactual (or manipulative) in nature, and cannot, therefore, be captured in the language of probability theory.” In order to express this relationship, we need to use the language of the do-calculus, introduced by Pearl, which goes beyond probability theory. As Pearl himself says (2011, p. 715): The way philosophers tried to capture this relationship, using inequalities such as P ( E|C) > P ( E) was misguided from the start – counterfactual ‘raising’ cannot be reduced to evidential ‘raising’ or ‘raising by conditioning’. The correct inequality, according to the structural theory …  should read: P ( E |do ( C ) ) > P ( E )

Copyright © 2018. Routledge. All rights reserved.

where do(C) stands for an external intervention that compels the truth of C. The conditional probability P(E|C) …  represents a probability resulting from a passive observation of C, and rarely coincides with P(E|do(C)). So Pearl’s main idea seems to be that the problem of probability raising is solved by replacing P(E|C) by P(E|do(C)). One might expect him therefore to go on to show that the counter-examples to probability raising such as Hesslow’s counterexample can be eliminated by making this move. However, Pearl does not do this, and does not mention any of the well-known counter-examples to probability raising in his 2011 article.We will continue our exposition of Pearl’s own argument in a moment, but first it seems interesting to see how his suggestion applies to the Hesslow counter-example as formulated in sections 12.2 and 13.4. As a word of warning we should say that this investigation might not be accepted as legitimate by Pearl because our formulation of section 13.4 uses multi-causal forks, which, because they are non-Markovian, he does not accept as valid. Still the investigation is not without interest, and we will accordingly carry it out. So, let us replace the version of the causality probability connection principle (CPCP), given at the beginning of Chapter 14, namely (*) by (**)

(

)

(

)

If X causes Z,then P Z = 1| do ( X = 1) > P Z = 1| do ( X = 0 ) (**)

As before our underlying reference class consists of a set of young women all with male partners in a situation in which the only method of contraception is the pill. Before, we imagined that we were simply observing which of the women took Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Pearl’s alternative approach  245

Copyright © 2018. Routledge. All rights reserved.

the pill, so that the ordinary probabilistic conditioning P(Z = 1 | X = 1) seemed appropriate. However, we could instead imagine a situation, in some very authoritarian country, in which an active intervention was made by the government. Some women, perhaps those who, according to the government, have wrong political opinions or ethnic character, would be forced under police supervision to take the pill [do (X = 1)], while others, whose children it was thought would be more useful to the state, would be prevented from using the pill [do(X = 0)]. In this new situation, the counter-example which applied to (*) would apply in just the same manner to (**). Those who were forced to take the pill would have a lower probability of getting thrombosis than those who were prevented from taking the pill, because the latter would have a much higher probability of becoming pregnant and so having pregnancy induced thrombosis. It follows from this that use of the do-calculus on its own does not solve the conundrums of CPCP. Pearl must be making some further assumptions – which of course is the case. Let us now examine what these further assumptions are. The key further assumption is that, when we are dealing with indeterministic causes, we should use a structural causal model. These models are described in Pearl (2000), and they are of two types. The first type, which could be called observable, consists of a set of observable parameters, which are connected to their parents by a functional equation involving an error or disturbance term. These error or disturbance terms are assumed to be independent, and from this it follows that the Markovian assumption is satisfied so that the network is a Bayesian network. But what about cases where the Markovian assumption is known not to be satisfied – for example Salmon’s interactive forks? Pearl proposes to deal with these by introducing latent, unobservable variables. Referring to the parents of a variable X i as PA i, Pearl writes (2000, p. 44): If a set PAi in a model is too narrow, there will be disturbance terms that influence several variables simultaneously and the Markov property will be lost. Such disturbances will be treated explicitly as “latent” variables …  . Once we acknowledge the existence of latent variables and represent their existence explicitly as nodes in a graph, the Markov property is restored. So, Pearl’s second type of structural causal model could be called latent, because it involves latent, unobservable variables as well as observable variables. He thinks that any significant, but apparently non-Markovian, causal model can be reduced to a latent structural causal model. There might indeed be some non-Markovian models, which could not be so reduced, but Pearl does not think they would be of any use. As he says (2000, p. 62): we confess our preparedness to miss the discovery of non-Markovian causal models that cannot be described as latent structures. I do not consider this loss to be very serious, because such models – even if any exist in the macroscopic world – would have limited utility as guides to decisions.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Copyright © 2018. Routledge. All rights reserved.

246  Pearl’s alternative approach

Now the multi-causal forks, which we described earlier, are not in general structural causal models. They involve only observable variables, but the Markov assumption is not always satisfied. So, Pearl would suggest that such models be reduced to structural causal models by, for example, introducing unobservable latent variables. We will consider how this might be done in a moment, but let us now return to Pearl’s solution to the problem of whether causes raise the probabilities of their effects. Essentially his approach is that we should formulate the problem within a structural causal model, and we can then calculate the value of CE = P(y | do(x)) – P(y). Sometimes it will be greater than zero and sometimes not. However, no additional assumption along the lines of CPCP needs to be made. In the case of observable structural causal models, the calculation of CE can definitely be carried out. As Pearl says (2011, p. 717): “The solution follows immediately from the identification of causal effects in Markovian models … ” Pearl is a bit more cautious in the case of latent structural causal models. He writes (2011, p. 717): “The solution is less obvious when P is defined over a proper subset W of V, where {V – W} represents the unmeasured variables.” However, he thinks that there are results which (2011, p. 717) “reduce this problem to algorithmic routine.” Such then is Pearl’s proposed solution to the problem of whether causes raise the probabilities of their effects.We will now present some criticisms.The main difficulty in Pearl’s approach seems to me to be his assumption that, whenever we are handling indeterministic causes, we should do so by introducing a structural causal model. In some cases, of course, structural causal models may be quite appropriate, but, in other cases, it might be simpler and easier to use different kinds of causal model, such as an interactive fork, or a multi-causal fork. Multi-causal forks are very simple causal models, which apply in a straight-forward way to well-known examples of the use of indeterministic causality in medicine, such as the causal factors of heart disease. They can be handled quite easily. So why should they be banned? As we know, Pearl would reply that the use of such non-Markovian models is unnecessary, because they can easily be replaced with Markovian models by adding latent variables. In the next section we will consider how this might be done in the case of our simple model of a multi-causal fork giving causal factors for heart disease (see Figure 13.5).

15.3  Restoring the Markov condition by adjusting the model In section 15.1, we described three methods, which could be used to restore the Markov condition in networks for which the Markov condition failed. These were: (1) eliminating variables, (2) adding arrows, and (3) adding variables. Method (1) was illustrated by the colonoscopy example of Figure 15.1. Let us now examine how, using such methods, we might try to restore the Markov condition in the case of two-pronged multi-causal forks, as illustrated by Figure 13.4. For multi-causal forks, method (1), i.e. eliminating variables, does not seem appropriate. So let us try method (2). We can than add an arrow joining X to Y, as shown in Figure 15.2.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Pearl’s alternative approach  247

Copyright © 2018. Routledge. All rights reserved.

FIGURE 15.2 

Adding arrow to multi-causal fork.

Now this does seem appropriate for the Hesslow example. If we apply it, we transform Figure 13.6 into Figure 15.3, by adding the additional postulate that taking the pill has a causal influence on getting pregnant. Now this additional postulate is plainly correct, since taking the pill prevents pregnancy, and prevention is of course a form of causal influence. Indeed, if our main aim had been to analyse this example, there would have been no need to introduce a non-Markovian model. The Markovian model of Figure 15.3 is clearly satisfactory. However, our main aim was not in fact to analyse this example, but rather the heart disease example of Figure 13.5. The point of introducing the non-Markovian model of the Hesslow example, as shown in Figure 13.6, was to provide a simple and vivid illustration of the pitfalls of non-Markovian models. Obviously, anyone advocating non-Markovian models should find some way of bringing their pitfalls to light, so that these pitfalls can be avoided. Our main claim is that a non-Markovian causal model is suitable for the heart disease example of Figure 13.5. So, the key question is whether the Markov condition can be restored for this model. Of course, once again, we could try adding an arrow as in Figure 15.2. In this case, this would amount to making the claim that smoking causes the eating of fast food. Such a claim, in contrast to the corresponding claim in the Hesslow case, seems to have little evidence in its favour. Smoking and eating fast food are indeed correlated, but does the first cause the second? It is just possible that having a dose of nicotine may cause a craving for the consumption of food high in salt, sugar, and saturated fat, but there is little physiological evidence for such a causal pathway, or for other causal pathways which would justify an arrow of causal influence joining smoking to eating fast food (or vice versa). In this case, then, it seems we should try what is anyway Pearl’s preferred method, and introduce a latent variable U (= unobservable) between X and Y, as in Figure 15.4. Now there would be no mathematical difficulty involved in such a move. The problems, which arise here, have an empirical, or scientific, character. In general, problems of this kind arise as soon as we start using unobservable

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Copyright © 2018. Routledge. All rights reserved.

248  Pearl’s alternative approach

FIGURE 15.3 

Adding Arrow in Hesslow example

FIGURE 15.4 

Adding latent variable to multi-causal fork.

variables in modelling some observable phenomenon. We can illustrate these problems by considering a simple case involving ordinary rather than random variables. Suppose we are modelling some observable quantity y, and, to do so, take into account n observable variables x1, x 2, …  , x n. We then postulate the ­following model

y = f ( x1 , x 2 , …, x n ) (15.1)

Unfortunately, when this model is tested out, it is found that its predictions differ quite dramatically from observation. However, not at all daunted, we decide to adjust the model by adding a latent or unobservable variable u.

y = f ( x1 , x 2 , ¼, x n ) + u (15.2)

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Pearl’s alternative approach  249

Since u is unobservable, we postulate that its value is given by y – f(x1, x2, …  , x n). We then conclude that the adjusted model (15.2) now agrees exactly with observation. Obviously, such a procedure would be pseudo-science rather than science. We give this example not in order to argue that the use of latent or unobservable variables is always wrong. On the contrary, there are many examples in which the use of such variables has proved fruitful. Our aim is rather to warn of dangers associated with the use of such variables. They can all too easily lead to converting empirical testable models into pieces of pure mathematics. In order to avoid such a danger, some steps are needed. The unobservable variable U should be interpreted in some way so that it can be checked whether there is anything, which corresponds to U in the real world. If possible, some independent method of measuring U should be developed. Then, most of important of all, any claims of the form “U causes X” should be tested against evidence to see whether they really hold or not. Merely writing down such claims without bothering to test them against data is to carry out fiction or fantasy rather than science.3 Similar warnings against the dangers of the use of unobservable variables are to be found in Korb et al. (2004). In the context of a critique of determinism, they consider the model

Z = a1 X + a 2 Y + U

Copyright © 2018. Routledge. All rights reserved.

and comment as follows (p. 324): But, Z is not a strict function of any of X or Y or the combination of the two: there is a residual degree of variation, described by U. U is variously called the residual, the error term, the disturbance factor, etc. Whatever it’s called, once we add it into the model, the model is deterministic, for Z certainly is a function – a linear function, of course – of the combination of X, Y and U. Does this make the physical system we are trying to model with the equation (or, Bayesian network) deterministic? Well, only if as a matter of fact U describes a variable of that system. Since as a matter of actual practice U is typically identified only in negative terms, as what is “left over” once the influences of the other parents of Z have been accounted for, and since in that typical practice U is only ever measured by measuring Z and computing what’s left over after our best prediction using X and Y, it is simply not plausible to identify this as a variable of the system. Bearing these potential dangers in mind, let us look at the result of introducing a latent or unobservable variable U into our simple heart disease model. This is illustrated in Figure 15.5. The first step in dealing with the model of Figure 15.5 is to try to find some interpretation of the unobservable variable U. One possibility would be to interpret U as a measure of the psychological disposition to go for immediate gratifications without regard for any long-term negative consequences. Such a disposition might

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

250  Pearl’s alternative approach

Copyright © 2018. Routledge. All rights reserved.

FIGURE 15.5 

Adding latent variable in heart disease example.

be described as “improvidence”. It seems plausible that improvident people might enjoy the pleasures of smoking and eating fast food without taking account of the long-term negative consequences on their health. But while such an account sounds reasonable enough at a common-sense level, it is by no means easy to establish that there really is such a psychological disposition and to find some way of measuring it. Suppose, however, we manage to overcome these problems, we have still got to establish empirically that U causes smoking, and U causes eating fast food. Now, in general, it is by no means easy to establish relations of indeterministic causality in medicine. Think of the case of “smoking causes lung cancer”. This is now generally accepted, but for decades it was a highly controversial claim, and it took a great deal of evidence to convince the community as a whole that it was true. The problems of establishing causal relations in medicine were discussed in Part II, Chapters 5 to 10. Here it was argued that there were two types of evidence, namely statistical evidence in human populations, and evidence of mechanism. It was further argued that the Russo-Williamson thesis, with a few qualifications, was valid, that is to say that normally both statistical evidence and evidence of mechanism are needed to establish a causal hypothesis in medicine. Further different types of evidence need to be combined using the principle of strength through combing. In short, it is a complicated business to establish that an indeterministic causality relation holds in medicine. Despite this complexity, such relations can be established in some cases. Consider again the non-Markovian causal model of Figure 13.5. Here the main causal links are well established by a great deal of scientific empirical evidence. A sketch of the evidence which establishes that smoking causes heart disease was given in sections 8.4, 8.5, and 8.6, and a similar sketch for the evidence which establishes that eating fast food causes heart disease was given in section 13.3. In both cases the Russo-Williamson thesis was satisfied. There was both statistical evidence and evidence of mechanism, and these were combined according to the principle of strength through combining. As a result, both these causal laws are accepted by mainstream medical opinion. Moreover, the relevant evidence provides a mass of data from which the probabilities used in the model can be estimated. Contrast this with the model of Figure 15.5. Here a long empirical scientific investigation

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Pearl’s alternative approach  251

would be needed to establish the postulated causal links “U causes smoking” and “U causes eating fast food”. Moreover, in order to show that the Markov condition was indeed satisfied in this model, we would need to test whether smoking and eating fast food were independent given U. As our earlier discussion of the network of Figure 15.1 showed, this cannot be taken for granted a priori. Of course, it is an easy matter from a purely mathematical point of view to write down any number of latent variables, link these with arrows to observable variables, and postulate that the Markov condition is always satisfied. Once this is done, probabilities within the model can be calculated using standard techniques. However, such a procedure, though mathematically rigorous, may not be satisfactory from an empirical-scientific point of view.To guide our actions, we should use causal models, which have been rigorously tested, and are well-corroborated by evidence. It is not satisfactory to use mathematically elegant models, which have little empirical confirmation. In the examples we are considering, the non-Markovian causal model of Figure 13.5 has been rigorously tested and is strongly corroborated by evidence. For the Markovian model of Figure 15.5 to reach the same level of empirical corroboration, a great deal of empirical-scientific research would be needed. So we see the situation as one of a trade-off.Those who favour the simpler mathematics of the Markovian model of Figure 15.5, would be forced to carry out a great deal of empirical scientific work. Those, who favour the non-Markovian model, can avoid this empirical scientific work, which has already been done, but they are forced to tackle the more complicated mathematical problem of trying to handle the non-Markovian case. Can this mathematical problem be solved? In the next chapter (16), I will show that it can.

Copyright © 2018. Routledge. All rights reserved.

Notes 1 On pp. 51–57 of his 2005 work, Williamson is actually discussing exceptions to what he calls the Causal Markov Condition, which is defined on p. 50, and is distinguished from the Markov Condition, which is defined on p. 15. In this book, I am using the term ‘Markov condition’ to cover both what Williamson calls the ‘Markov condition’ and what he calls the ‘Causal Markov Condition’. So on pp. 51–57 of his 2005 work, Williamson is indeed discussing what are exceptions to the Markov condition in the sense of the term used in this book. The differences here are purely terminological. 2 This applies to my own formulation of the causal probability connection principle (CPCP) given in section 12.2, and to the version of it for our simple model (*) given at the beginning of Chapter 14. 3 This remark may seem exaggerated. Yet there are many learned and highly mathematical papers published in leading journals which devote themselves to constructing models for fantasy examples such as “A spell cast by Merlin caused the prince to turn into a frog.” In our view such work is valueless. Clearly in the real world spells do not cause princes to turn into frogs. Why should causality, as imagined in such a fantasy world, have anything to do with causality in the real world? Causal modellers should devote themselves to genuine scientific examples, which arise in the real world. There are a rich variety of these, and there is consequently no need to bring in the consideration of purely imaginary examples.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

16 EXTENSION OF THE ACTIONRELATED THEORY TO THE INDETERMINISTIC CASE

Copyright © 2018. Routledge. All rights reserved.

16.1  Mathematical formulation of the problem1 For simplicity, we will confine ourselves to the simple causal network illustrated in Figure 13.4. Our problem is the following. We have a disease Z which is known to have two indeterministic causes X and Y. It would seem sensible in such a case to try to avoid Z by eliminating through our actions at least one of X and Y, and preferably both. It would seem to be a good strategy for doctors to advise such a course of action. However, as the pregnancy, contraceptive pill, thrombosis example shows, this advice would not always be correct. Can we then formulate some mathematical condition on X, Y and Z such that the “common-sense” advice of eliminating at least one of X and Y is in fact correct advice? It will be assumed that X, Y and Z are random variables defined on some underlying probability space Ω . So the joint distribution {X, Y, Z} is defined, and so also are the marginal distributions {X,Z}, {Y,Z}, etc. This makes our network a probabilistic network as well as a causal network, but note that it need not be a Bayesian network. For it to be a Bayesian network, the Markov condition would have to be satisfied, and, in this simple case, the Markov condition is that the random variables X and Y are independent. This is not the case in the two examples we considered in Figures 13.5 and 13.6. Smoking and eating fast food are not independent, and neither are taking the pill and getting pregnant. Regarding our simple model, two theorems will be proved, which are due to Aidan Sudbury (Sudbury’s Theorems). For the purpose of Theorem 1, we will make the further assumption that X, Y, and Z are binary variables taking the values 0 or 1. Z is assumed to be a disease, and Z = 1 means “having the disease”, while Z = 0 means “managing to avoid the disease”. We are also assuming that X and Y are indeterministic causes of Z. The question now is: “what probabilistic assumptions does this allow us to make?” Here we adopt the version of the causal

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Extension of the action-related theory  253

probability connection principle (CPCP), which was argued for in Chapter 14. This amounts to assuming that, if we set one of the variables X, Y to an arbitrary value, then a positive value of the other will increase the probability of getting the illness. We can state it mathematically by defining

Zij =

def

P ( Z = 1| X = i, Y = j) for i, j = 0,1

The assumption made on the basis of X, Y being causal factors is then the following: Z11 > Z01

Z11 > Z10 Z01 > Z00



Z10 > Z00

Copyright © 2018. Routledge. All rights reserved.

We will call this the causal factor assumption in the binary case. What we want to prove, to justify the strategy of eliminating at least one of the causal factors in order to reduce the probability of getting the disease, is the following:

P ( Z = 1| X = 1) > P ( Z = 1| X = 0 ) (16.1)



P ( Z = 1| Y = 1) > P ( Z = 1| Y = 0 ) (16.2)

(16.1) and (16.2) both seem to hold in the smoking, fast food example, but (16.1) fails in the contraceptive pill, pregnancy example. So (16.1) and (16.2) do not follow from the causal factor assumption. Can we formulate a mathematical condition, which, if it is added to the causal factor assumption will ensure that (16.1) and (16.2) follow? As will be shown in Appendix 3, if we assume that X, Y are independent (the Bayesian network case), then (16.1) and (16.2) do follow. However, this independence assumption does not hold in the various examples we are considering. Sudbury’s Theorem 1 shows that, if we assume P(Y = 1 | X = 1) >  P(Y = 1 | X = 0), then (16.1) follows, and of course if we simply reverse X, Y in this condition, we get (16.2). Now the condition P(Y = 1 | X = 1) >  P(Y = 1 | X = 0) is a very reasonable one. It clearly fails in the contraceptive pill, pregnancy case, since there we obviously have P(Y = 1 | X = 1) < P(Y = 1 | X = 0), whereas it plausibly holds in the smoking, fast food case, and this could be checked empirically. In order to prove Sudbury’s Theorem 2, we drop the assumption that X and Y are binary variables. The binary variable assumption amounts to classifying individuals as smokers or non-smokers, or as fast food eaters or non-fast food eaters. However, this is obviously inadequate when trying to relate smoking or fast food

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

254  Extension of the action-related theory

eating to heart disease. In such an investigation, it is obviously very important to consider the quantity of tobacco that an individual smokes, or the quantity of fast food, which he or she consumes. Indeed, in the statistics about smoking and lung cancer given in section 8.3 and in those about smoking and heart disease given in section 8.4, the effects of smoking different amounts of tobacco were considered. It is generally held to be important to take into account the so-called “dose relation” in assessing causality. In the case of smoking, the dose relation would be how the probability of getting the disease varies with the quantity smoked. For these reasons, it is better to take X and Y to be continuous random variables taking nonnegative values. The causal factor assumption can be given in a form appropriate to continuous random variables X and Y. The condition P(Y = 1 | X = 1) >  P(Y = 1 | X = 0) now becomes a strong version of positive correlation between X, Y. Under these assumptions Sudbury’s Theorem 2 is proved. The full mathematical statement and proofs of Sudbury’s theorems is given in Appendix 3. In the next section, I will give an informal statement of these results, and discuss their significance.

Copyright © 2018. Routledge. All rights reserved.

16.2  Informal statement of Sudbury’s theorems Sudbury’s theorems are stated and proved in Appendix 3 in a mathematically precise fashion, and in this section I will consider them in a more qualitative fashion. The conditions under which the theorems are proved could be stated roughly and informally somewhat as follows. Let us say that two indeterministic causes X and Y are associated if, in the binary case, the presence of one increases the probability of the presence of the other, and if, in the continuous case, they are strongly positively correlated. Two indeterministic causes X and Y are opposed if, in the binary case, the presence of one decreases the probability of the presence of the other, and if, in the continuous case, they are negatively correlated. The conditions of being associated or opposed are quite intuitive, and it would be easy to check from data whether they held in a particular case. If X and Y are associated indeterministic causes of Z, the two theorems show that it is a good strategy, in order to avoid Z, to make sure that either X or Y or both do not occur, or that their effects are blocked if they do occur. There is, however, an objection to the claim that the results of our theorems justify the avoidance strategies we have described.2 One of the results of theorem 1 is that

P ( Z = 1| X = 1) > P ( Z = 1| X = 0 ) (16.3)

This is taken to justify the strategy of trying to avoid the disease Z by setting X = 0, i.e. giving up smoking. However, it could be objected that the inequality (16.3) might hold, but eliminating X will not reduce the probability of getting the disease. Let X = 1 represent yellow teeth, and Z = 1 lung cancer. The inequality (16.3) is satisfied, since people with yellow teeth tend to be smokers, and hence

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Copyright © 2018. Routledge. All rights reserved.

Extension of the action-related theory  255

have higher rates of lung cancer. But whitening your teeth will not reduce your probability of getting lung cancer. This is not, however, a counter-example to the claims we have made, since, we are assuming not just inequality (16.3), but also that X causes Z, and tooth colour is not a cause of lung cancer. This alleged counterexample is instructive, however, because it reinforces the familiar point that statistical claims on their own are often not action guiding, and one may need causal assumptions as well as statistical ones to justify actions. It is indeed a fundamental characteristic of causes that they are action-related. In the next section I will generalize from this to produce an extension of the action-related theory to the indeterministic case. But before doing so, I will conclude this section with a few further remarks on the significance of Sudbury’s theorems. What has been shown is that it is possible to develop a non-Markovian causal model for a well-known medical situation, that this model is both well corroborated empirically, and mathematically tractable, and that, in particular, we can draw action-guiding conclusions from the model. In the mathematical theories of causal networks so far developed, there has been an almost exclusive focus on cases in which the Markov condition is satisfied. So one of the important features of Sudbury’s theorems is that they show that interesting results can be obtained in at least some cases in which the Markov condition is dropped. It seems likely that more interesting results covering this situation could be obtained in the case of more complicated networks – for example in the case of multi-causal forks with more than two indeterministic causes. However, in pointing to a possible use of non-Markovian models, I do not want to make the dogmatic claim that it is impossible to obtain similar results using Markovian models. I have done no more than pose a challenge to the advocates of the exclusive use of Markovian models to solve the problem just dealt with using their preferred scheme. However, as stressed above in section 15.3, a satisfactory solution requires that the model used is both mathematically tractable and empirically well corroborated. Our non-Markovian model of Figure 13.5 satisfies both these conditions, and any satisfactory Markovian model would have to satisfy them both as well. I would also like to recommend the type of example just considered as a very suitable field of study for causal modellers. Heart disease is still the number one killer in most developed countries. Medical science has made considerable advances in its study, and these depend on the use of indeterministic causality and multi-causal forks. The very important Framingham study, which has carried out investigations into heart disease continuously since 1948 (see Levy and Brink, 2005), has provided a whole mass of data concerning possible causal factors of heart disease. Yet strange to say, there have been very few attempts to create causal models for this data. One of the rare and admirable exceptions is Korb et al. (2004). 3 Instead of causal models, only traditional statistical models have been employed on the Framingham data. Surely the development of causal, and hence action guiding, models here would be a step forward.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

256  Extension of the action-related theory

Copyright © 2018. Routledge. All rights reserved.

16.3 Extension of the action-related theory to the indeterministic case In Part I of the book, I developed the action-related theory of causality for the deterministic case.This is an important case for medicine, because it covers the 19th and early 20th century development of the germ theory of disease, and also some more recent cases, such as McArdle disease, discussed in section 10.2. However, a great deal of modern medicine uses indeterministic causality, and so it is very desirable to extend the action-related theory to this case, which I will now do. I will begin with a reminder that this book deals only with generic (or type) causality, that is to say with causal laws which apply in a number of different cases. Single-case (or token) causality is obviously very important in clinical practice, where a doctor has to discover what is causing the symptoms of a particular patient. However, this will not be considered in the present book, which is concerned with theoretical medicine rather than clinical practice. As was pointed out in section 14.2, a generic causal law of the form A causes B is relative to a set of repeatable conditions S which define the reference class to which the law applies. I gave as an example, the case of a medication M which has no side effects for the average patient, but produces vomiting and dizziness in diabetic patients. We then have that, in the reference class of diabetic patients, “M causes vomiting and dizziness” is true, but that in the reference class of non-diabetic patients, it is false. Now suppose that a particular effect Z say has a number of different indeterministic causes A, B, C, …  . We suppose further that these causes are arranged in a causal network. The variables of the network are considered to be random variables with a joint distribution. Probabilities are interpreted objectively as propensities, and are consequently relative to a set of repeatable conditions S, which define a reference class. The same set of repeatable conditions S applies to the causal claims of the network. This defines the causal model. This causal model must be tested out empirically in a rigorous fashion, and only accepted if it is strongly confirmed by a variety of data. We can suppose that this data enables reasonable estimates of the probabilities involved to be made. Our problem is to show how actions can be based on an acceptable model of this sort. Let us consider a particular indeterministic cause (A say) and the causal link A →  Z in the model. Let us suppose that A causes Z in a positive fashion rather than preventing Z. We can only base actions on A →  Z if we can prove that:

P ( Z | A ) > P ( Z |~ A ) (16.4)

Suppose it is possible to prove (16.4). Then if we want to produce Z (a productive action), we should instantiate A, because this raises the probability of Z occurring. If we want to prevent Z occurring (an avoidance action), then we should prevent A occurring, or block its usual consequences, because this lowers the probability of Z. Note that these action-guiding inferences depend not just on the probability relation (16.4), but on this in conjunction with the causal law A →  Z.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Extension of the action-related theory  257

We can see that this is a straightforward generalization of the deterministic case. In the deterministic case, P(Z | A) = 1, because if A occurs, then Z always follows. So, leaving aside the case where P(Z |~A) also equals 1, (16.4) holds. If A is a necessary condition for Z, then P(Z |~A) = 0, and so, leaving aside the case where P(Z | A) = 0, (16.4) holds. The theory, which emerges, is therefore quite like Good’s original project (see section 12.2) in several respects. Probabilities are interpreted as propensities, and the cases where the cause A is sufficient [necessary] for the effect Z correspond to probabilities of P(Z | A) = 1 [P(Z |~A) = 0]. However, it obviously differs from Good in other respects. There is no attempt to define causality in terms of probability, and, in order to avoid counter-examples like that of Hesslow, the causality probability connection principle (CPCP) has to be stated in a restricted form. In effect, (16.4) is not true in general, but only when the other causes in the model are given fixed values. With this restricted version of CPCP, (16.4) is not always true, but it may be possible to prove (16.4) is some cases, as Sudbury’s theorems show. In the indeterministic case, we cannot infer immediately what actions are justified on the basis of the causal relations established. We have to prove a theorem like (16.4) first. It is worth noting, however, that this is a probabilistic result, which will be proved using only the standard mathematical calculus of probability. This theory of indeterministic causality does not therefore involve a mathematization of causality as such, but only a use of the mathematical theory of probability. Claims of causality are a qualitative addition to mathematical probability, and are not involved in mathematical calculations. However, the qualitative addition of causality is nonetheless of the greatest importance, because it shows that actions can be justified which would not be justified by probabilistic claims alone. This is why it is important, though difficult, to go beyond correlation and establish causality.

Copyright © 2018. Routledge. All rights reserved.

Notes 1 In the mathematical work for this book, I was greatly helped by two friends of mine, who are mathematicians, specializing in probability and statistics, namely Christian Hennig of University College London, and Aidan Sudbury of Monash University. Aidan Sudbury proved Theorems 1 and 2. The theorems and their proofs were first published in Gillies and Sudbury (2013). In early discussions, it was Christian Hennig who made the key points, which led to the mathematical formulation of the problem in section 16.1, and he also made the interesting comment on the conditions of Theorem 1, which has been included in Appendix 3. However, Christian Hennig’s philosophy of probability and statistics is rather different from the one in this book. A general view of his philosophy is to be found in Hennig (2010), and he is at the moment developing it in more detail. 2 This objection was made, by an anonymous referee, to an earlier version of Gillies and Sudbury (2013). I have quoted the objection, more or less verbatim, from the referee’s report. 3 A Google search for causal models for the Framingham data produced only Korb et al. (2004). Of course, I may have missed some other papers, but there cannot be many of these.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

Copyright © 2018. Routledge. All rights reserved. Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-17 01:44:07.

APPENDIX 1

Copyright © 2018. Routledge. All rights reserved.

Example of a simple medical intervention which is not an intervention in Woodward’s sense

In his formal definition (2003, p. 98) Woodward introduces an intervention variable I for X with respect to Y where X causes Y. The definition also considers another variable Z which causes Y. Using this terminology let us describe the following simple medical intervention. A doctor has a patient with very high blood pressure. She decides between two possible interventions. The first of these (I say) operates on a variable X, whose value consists of the patient taking by mouth a pill containing a certain quantity q of a drug d. I then consists in the doctor fixing the value of X by prescribing a pill with particular values for q and d. The second possible intervention (I’ say) consists of the doctor asking a nurse to give the patient an injection of a certain quantity q of drug d. I’ fixes the value of a variable Z which specifies q and d in the injection. Let us suppose that the doctor decides in favour of intervention I and prescribes the pill. The patient takes the pill and his blood pressure Y drops to a safe level. Is this an intervention in Woodward’s sense? The answer seems to be “no” for several reasons. To begin with clause 4 of the definition of IV runs (Woodward, 2003, p. 98): “I is (statistically) independent of any variable Z that causes Y and that is on a directed path that does not go through X.” Our variable Z satisfies the condition of the clause, but is not independent of I. If it were independent, then P(I = 1 | Z ≠  0) would equal P(I = 1), or, to translate this formula into ordinary language, the doctor would be just as likely to prescribe the drug if the patient had already received an injection. This is obviously false. There is also a problem about clause 3. This begins (Woodward, 2003, p. 98): “Any directed path from I to Y goes through X.” It is saying, in the jargon of causal networks, that the only causal factor, which reduces the patient’s blood pressure, is the biochemical action of the pill. But let us suppose that the patient is very agitated when he enters the doctor’s consulting room. However, the doctor is very charming and reassuring. As she gives the

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:25.

260  Appendix 1

Copyright © 2018. Routledge. All rights reserved.

patient his prescription, she says with a smile: “Just take this pill, and you’ll soon feel as right as rain.” Through the action of the placebo effect, this reassurance already reduces the patient’s blood pressure before he has swallowed the pill. The action of the placebo effect here contradicts Woodward’s clause 3, and this is another reason for supposing that the medical intervention just described is not an intervention in Woodward’s sense. Now the notion of intervention in medicine is given a pretty clear meaning by medical practice. This is in accordance with the Wittgensteinian view that meaning is use in a social practice. Would it not therefore be better to adopt this ordinary sense of intervention when analysing causality in medicine rather than giving a very complex definition of intervention, which diverges from the ordinary sense of the word? In my view, the complexity of Woodward’s definition of intervention is neither necessary nor desirable. It is true that there are some situations in which complicated clauses, like those to found in the definitions of IV and IN, do need to be considered. Suppose we are considering a complicated causal model of some situation. In this model, a variable Y has several causes X, Z1, …  Zn say, which are connected to Y in a complicated network. To estimate the influence which X exerts on Y, we may have to plan a complicated intervention, which ensures that the value of Y is caused largely by the value to which X is set, and only to a small degree, if at all, by the values of the remaining causes Z1, …  Zn. If such an intervention is possible, it would depend on a series of complicated conditions being satisfied. Such conditions are needed, however, only for specific interventions in particular circumstances. They do not need to be part of a general definition of intervention. In a terminology which Woodward himself uses, “ham-fisted” interventions are just as much interventions as “finegrained and surgical” interventions. Actually, Woodward comes close to recognizing this point in his recent (2014) paper, in which he says (p. 706): If Y changes under a manipulation of X, but this manipulation affects Y via a route that does not go through X, we have a badly controlled or illdesigned experiment for the purposes of determining whether X causes Y, an experiment that is confounded by Z. This is true enough, but surely the manipulations (or interventions) in illdesigned experiments are still manipulations (or interventions). Why should we limit manipulations (or interventions) to those in well (or perhaps perfectly) designed experiments? Moreover, somewhat ham-fisted interventions may be quite adequate for practical purposes. In our example of the doctor prescribing the pill to reduce the patient’s blood pressure, it does not matter if the fall in the patient’s blood pressure is caused both by the placebo effect and by the bio-chemical action of the drug in the pill. We do not in this instance need a fine-grained and surgical intervention, which reduces or eliminates the placebo effect.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:25.

Appendix 1  261

Copyright © 2018. Routledge. All rights reserved.

For these reasons, I will in this book, unless I am referring specifically to Woodward’s theory, use the terms “intervention” and “manipulation” in their usual sense as given by their use in social practice. This means that interventions and manipulations are always human actions, and that possible interventions and manipulations are those, which humans could carry out, given the contemporary levels of technology.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:25.

APPENDIX 2

Copyright © 2018. Routledge. All rights reserved.

Mathematical terminology

A network or net is a directed acyclic graph. The nodes or vertices of a network are variables, which are denoted by capital letters, e.g. X, Y, Z, A, B, … . If an arrow joins two nodes of a network A, B (see Figure A2.1), then A is said to be a parent of B, and B is said to be a child of A. Children, children of children, etc. of A are said to be descendants of A. If an arrow joining any two nodes A, B (see Figure A2.1) of a network means that A has a causal influence on B, then the network is said to be a causal network. If the set of variables of a network, X1, X 2, … , X n say, are random variables all defined on some underlying probability space and so having a joint distribution, then the network is said to be a probability network. The Markov condition is satisfied for a node A of a network, if A, conditional on its parents, is probabilistically independent of any other set of nodes in the network not containing any of A’s descendants. A probability network in which every node satisfies the Markov condition is said to be a Bayesian network. In a Bayesian network, the parents of a node are said to screen it off from the other nodes of the network except its descendants. If a causal network is also a probability network, it is said to be a causal probability network, or causal model. When the term “causal network” is used in this book with no further qualification, it will be assumed to be a causal probability network.

FIGURE A2.1 

A is a parent of B, and B is a child of A.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:32.

Appendix 2  263

Note that when Pearl introduced the term Bayesian network or Bayes ­network in his 1985b work, he used it to refer to causal Bayesian networks. In fact Pearl wrote (1985b, p. 330): Bayes networks are directed acyclic graphs in which the nodes represent propositions (or variables), the arcs signify the existence of direct causal influences between the linked propositions, and the strengths of these influences are quantified by conditional probabilities.

Copyright © 2018. Routledge. All rights reserved.

We here, following a later convention, defined Bayesian networks purely probabilistically, so that the arrows in a Bayesian network need not represent causal influences. However, when the term “Bayesian network” is used in this book with no further qualification, it will be assumed to be a causal Bayesian network. A causal model, in which every node satisfies the Markov condition, is said to be a Markovian causal model. A causal model, in which at least one node does not satisfy the Markov condition, is said to be a non-Markovian causal model.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:32.

GLOSSARY OF MEDICAL TERMS

If a word appears in italics, this indicates that there is a separate entry for it in the glossary. Acetylated:  A compound is acetylated if an acetyl group (formula C2H3O) is

added to it.

Acetyl-salicylic acid:  This is the chemical name for aspirin.

Copyright © 2018. Routledge. All rights reserved.

Acupuncture:  This was part of traditional Chinese medicine, according to which qi (a kind of energy) flowed through channels in the body. If the flow of qi became unbalanced, pain or illness resulted, and these could be cured by inserting needles into appropriate points in the channels through which qi flowed. AIDS:  This is an abbreviation of Acquired Immune Deficiency Syndrome. This

is a disease in which the patient’s immune system is gradually destroyed making him or her liable to acquire fatal infections. It is caused by infection by HIV (Human Immunodeficiency Virus). Amino acid:  Twenty different amino acids are the building blocks of all the proteins

in the human body. Amino acid residues are joined in a sequence to form a protein. Analgesic:  Reducing the sensation of pain. Painkiller. Aneurysm:  If the walls of a blood vessel weaken and balloon out, this is said

to be an aneurysm. Aneurysms are liable to rupture causing bleeding, and can also be the site of a blood clot (thrombosis). They are often caused by atherosclerotic plaques, but other causes are possible. Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:37.

Glossary of medical terms  269

Angina (pectoris):  Severe chest pains caused by atherosclerotic plaques in the c­oronary arteries. These pains typically come on as a result of physical exertion which increases the heart’s oxygen requirements. Anti-oxidants:  These are substances which inhibit the process of oxidation. The process of the formation of atherosclerotic plaques starts with the oxidation of LDL. So anti-oxidants in the body, such as vitamin C and vitamin E protect against atherosclerosis. Antipyretic:  Reducing fever, from Greek: pyro = fire. Anthrax:  This is a disease mainly of cattle and sheep, though it can occur in

humans as well. It is caused by the bacillus anthracis. Arteriosclerosis:  The older name for what is now usually called atherosclerosis.

Copyright © 2018. Routledge. All rights reserved.

Atherosclerosis:  Atherosclerosis (from Greek: athera = gruel, and sclerosis = hardening) describes the formation of plaques in the walls of arteries. These restrict the flow of blood in the artery and can result in a blood clot (thrombosis) which closes the artery completely. Atherosclerosis results in a variety of medical conditions depending on the location of the plaques. Angina pectoris occurs if the coronary arteries are partially blocked. The complete blockage of a major coronary artery by a blood clot (thrombosis) results in a heart attack (myocardial infarction). The blockage of an artery supplying the brain results in a stroke, while the restriction of the arteries supplying the legs and feet results in peripheral vascular disease, such as intermittent claudication which consists of pains in the legs when walking which disappear after a short rest. Atherosclerosis also causes aneurysms. The conditions which result from atherosclerosis are very varied, but the underlying cause is always the same. Atherosclerotic plaques :  These are the plaques which from in the artery walls in atherosclerosis. The plaques begin when LDL is oxidized, which leads to the formation of foam cells. Bacillus (plural –i):  Bacilli are rod-shaped bacteria. Bacillus anthracis:  This is the pathogenic bacillus which causes anthrax in ani-

mals and humans. It is one of the largest and thickest of the bacteria, and could be seen easily through the optical microscopes available by the middle of the 19th century. An important feature of this bacillus is its ability to form spores which are very resistant to destruction, but which turn back into the normal form of the bacillus when they enter the blood stream of an animal. Bacterium (plural –a):  Bacteria are unicellular organisms which multiply by

division. They are invisible to the naked eye, but nearly all can be seen through Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:37.

270  Glossary of medical terms

an optical microscope which magnifies x1000. They are classified by their shape into bacilli, which are rod-shaped, and (ii) cocci, which are spherical. Most bacteria are harmless, or even beneficial. However, a few of them are pathogenic bacteria, and cause diseases. Beta-carotene:  Beta-carotene is an organic chemical, strongly coloured

r­ed-orange, which is found in many fruits and vegetables, notably carrots. In the body it can be transformed into vitamin A, which is important for maintaining health, particularly good vision. Binding site:  Binding sites, or receptors, are areas on the surfaces of mac-

rophages, or other biochemical entities, which enable them to attach themselves to other biological molecules or organisms. Blood cholesterol level :  This is a measure of the amount of cholesterol which circulates in the blood stream in the form of lipoproteins Burkitt’s lymphoma:  This is a cancer first identified in 1958 by Denis Burkitt (1911–1993) when he was working in equatorial Africa. It commonly occurred in the jaws of African children. It was the first human cancer which was found to be caused by a virus, in this case the Epstein-Barr virus. CaCer:  Abbreviation of cancer of the cervix, or cervical cancer. Capillaries:  These are the smallest of all blood vessels and form the connection

between arteries and veins. Carcinogenic:  Having the property of causing cancer.

Copyright © 2018. Routledge. All rights reserved.

Case control studies:  These are epidemiological studies in which a sample of

people with a particular disease are compared with a control sample of people who do not have the disease. Cell membrane:  This is a thin semi-permeable barrier that separates the inte-

rior of a cell from the outside environment. Cerebral infarction:  An infarct is a localized area of dead tissue which has

resulted from an obstruction of the blood supply to that area. So a cerebral infarction is the formation of an infarct in the cerebrum, or principal part of the brain, due to an obstruction of the blood supply. Cervical cancer:  This is cancer of the cervix, or neck of the womb. It is the second most common cancer in women world-wide. CFTR gene:  A defective version of this gene causes cystic fibrosis.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:37.

Glossary of medical terms  271

CHD:  Abbreviation of coronary heart disease. Childbed fever:  This is a severe fever which can affect mothers a few days after they have given birth. Before antibiotics, it was often fatal, resulting in death in two or three days. It is also known as puerperal fever. Chloride ions:  Salt is sodium(Na)chloride(Cl). When dissolved it separates

into positively charged sodium ions Na+ and negatively charged chloride ions Cl-. The movement of chloride ions round the body is important for proper functioning. In the disease cystic fibrosis, the transport of chloride ions in and out of cells does not occur as it should, and this is responsible for many of the symptoms of the disease. Cholera:  This is a disease whose symptoms are severe vomiting and diarrhoea. If untreated, it kills about half the sufferers within a few days. Koch established that the disease was caused by a bacterium: vibrio cholerae. Cholesterol:  Cholesterol is a lipid, which is a vital component of many bodily

structures, including cell membranes, hormones, digestive juices, and nervous tissues. Because of its importance, it is carried round the body in the blood stream in the form of lipoproteins. However, too much cholesterol can result in the formation of atherosclerotic plaques, which are composed of cholesterol and calcified connective tissue. Cholesterol esters:  These are chemicals formed by a combination of cholesterol

with another lipid, a fatty acid. Cells store cholesterol in the form of cholesterol esters, which are less toxic than free cholesterol.

Copyright © 2018. Routledge. All rights reserved.

Cholesterin:  An older name for cholesterol. Clinical trials:  These are trials designed to study the effectiveness of treatments.

They are normally carried out on groups of patients who volunteer to take part. Cohort studies:  These are observational epidemiological studies in which a

group of people, who are either considered as a single cohort, or divided into several different cohorts, are followed for a number of years with a view to studying a particular disease D. The participants are examined at regular intervals and various of their characteristics are measured and noted. An attempt is made to relate these characteristics to those who contract the disease D. Cohort studies are normally prospective studies in which the participants are initially free of the disease D, but in which some of them contract D in the course of the study. However, retrospective studies are also possible, based on historical data. Colon:  This is the lower part of the digestive system. It consists of a long, coiled

tube-like organ which terminates in the anus. Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:37.

272  Glossary of medical terms

Colonoscopy:  This is an examination of the inner lining of the colon. A typi-

cal method is to insert a small camera at the end of a thin tube through the anus, and to steer it along the colon. The camera transmits images of the interior of the colon which can be seen on a screen. A colonoscopy enables ulcers, polyps, tumours, and other problems with the colon to be detected. Confounder:  If a correlation is established between two variables A and B, the question often arises whether A causes B. In some cases, there is another variable C such that C causes B, and A has little or no causal effect on B, although it is correlated with B. C is then known as a confounder, because it confounds the supposed causal relation between A and B. For example, heavy drinking is strongly correlated with lung cancer, but this relation is confounded by heavy smoking. The confounder (heavy smoking) is the real cause of lung cancer. Heavy drinking is correlated with heavy smoking and so with lung cancer, but probably has little causal influence on the occurrence of lung cancer. Controlled trials:  These are clinical trials in which the patients receiving a par-

ticular treatment are compared with a control group of patients who do not receive the treatment. If the treatment is a new one, the control group may receive the standard existing treatment, or the control group may receive no treatment, or they may receive a placebo, for example an inactive sugar pill which looks like the medication used in the main treatment. Placebo trials can be single-blinded, in which case the participants do not know whether they are receiving the medication under test or the placebo. They can also be doubleblinded, in which case neither the patients nor the doctors administering the test know who is receiving a placebo and who the medication. Coronary arteries:  These are the arteries which supply blood to the heart.

Copyright © 2018. Routledge. All rights reserved.

Coronary heart disease:  This is disease caused by atherosclerosis affecting the coro-

nary arteries. One form is angina pectoris which occurs when atherosclerotic plaques reduce the blood flow to the heart. This results in chest pains, especially on exertion. A more severe form is a heart attack (myocardial infarction) which occurs when an advanced atherosclerotic plaque causes a blood clot (thrombosis) which closes off a coronary artery. Cystic fibrosis:  This is a genetic disease caused by a faulty gene inherited from

both parents. The lungs and digestive system become clogged with thick sticky mucus which prevents them working properly. The disease is progressive, and currently only about half the sufferers live beyond 40. Death is usually from lung failure. DBS:  Abbreviation of deep brain stimulation. Deep brain stimulation:  This is a therapeutic technique in which an electrode

is inserted into the patient’s brain. This electrode can, by means of an outside Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:37.

Glossary of medical terms  273

control, produce an electrical stimulation of a part of the brain, and this in turn can relieve the symptoms of some neurological conditions such as Parkinson’s disease. Diabetes mellitus:  This is a disease in which blood glucose levels are not properly controlled and can rise too high for prolonged periods. Early symptoms are frequent urination, thirst, dehydration, tiredness, and lethargy. If left untreated, more serious symptoms can develop. Blood glucose level is normally controlled by insulin produced in the pancreas. In Type I diabetes, the pancreas fails to produce enough insulin. In Type II diabetes, the patient develops insulin resistance, and cells fail to respond properly to the insulin produced. Type II diabetes often begins later in life, but can often be treated by changes in diet and medication. Type I diabetes nearly always requires insulin injections. Diabetes requiring insulin:  These are the cases of diabetes mellitus which require injections of insulin as part of the treatment. Diphtheria:  This begins with a sore throat or fever. Characteristically a grey or white patch develops in the throat. Complications include fatty degeneration of the heart and damage to peripheral nerves which can cause paralysis. If untreated, the disease can cause death in about half of the children affected, either because of the complications or because the airway is occluded. The disease is caused by infection with a bacillus: corynebacterium diphtheriae. DNA:  This is an abbreviation of DeoxyribonoNucleic Acid, a molecule with a double helix structure which carries the genetic instructions used in the growth, development, functioning and reproduction of all known living organisms and many viruses.

Copyright © 2018. Routledge. All rights reserved.

EBM:  Abbreviation of evidence based medicine. ECG:  This is short for ElectroCardioGram, a measurement of the heart’s rhythm and electrical activity. Sensors are attached to the skin, and these record electrical signals from the heart. Endorphins:  These are opioids which are produced by the brain and central

nervous system. They act as natural painkillers. Endoscopy:  From the Greek endon = within, skopein = look, this consists

of looking inside the body by means of an instrument (an endoscope) which is inserted into an organ. The endoscope has a camera at its tip which transmits pictures to a screen outside. Endothelin-1:  This is a potent vasoconstrictor which narrows blood vessels and raises blood pressure. Substances in the diet which restrict the production of endothelin-1 are therefore thought to be helpful in preventing heart disease.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:37.

274  Glossary of medical terms

Endothelium:  This is the lining of blood vessels and lymphatic vessels. It con-

sists of a thin layer of cells, known as endothelial cells. Enkephalins:  These are opioids which occur naturally in the body, and have a

function similar to endorphins. Enzymes:  A substance which facilitates a biochemical reaction without itself

undergoing any permanent change. Epidemiology:  The study of health issues through the collection of statistics in populations and groups. Epithelium:  The layer of cells lining cavities and organs in the body. For

example, the bronchial epithelium is the lining of the bronchi, the tubes which conduct air into the lungs. Epstein-Barr virus:  This virus was discovered when it was isolated from cells

from Burkitt’s lymphoma by Anthony Epstein (1921–) and his PhD student Yvonne Barr in 1964. This virus causes two quite different diseases, namely (i) Burkitt’s lymphoma, a facial cancer found among African children, and (ii) infectious mononucleosis (or glandular fever), a flu-like disease of teenagers and young adults which can be long lasting, though there is usually a complete recovery without treatment.

Copyright © 2018. Routledge. All rights reserved.

Evidence based medicine:  This is a movement which was begun in 1992 by

Gordon Guyatt and his associates in McMaster University in Canada. Its aim is to improve medical decisions by making sure that they were based on good quality evidence. Evidence based medicine (EBM) arranges evidence into a hierarchy. At the top are randomized controlled trials, then observational studies, while the observations of individual doctors are either omitted altogether or put very low in the hierarchy. EBM has been a very influential movement, but it has also been subject to many criticisms. For example, it has been criticized for an undue emphasis on statistical evidence, and the relative neglect of evidence of mechanism obtained from laboratory experiments. F2-isoprotanes:  These are chemicals found in the body, whose quantity pro-

vides a measure of oxidative stress.

Fat:  Fat in the body is stored mainly in the form of triglycerides, which consist of three fatty acids attached to a glycerol molecule. Triglycerides constitute an energy store for the body, and are used up in exercise. Fatty acids are divided into (i) saturated, which contain no double bonds between carbon atoms, (ii) monounsaturated, which contain just one double bond between carbon atoms, and (iii) polyunsaturated, which contain more than one double bond between carbon atoms. It was one of Ancel Keys’ most important discoveries that these different types of fat in

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:37.

Glossary of medical terms  275

the diet have different effects on blood cholesterol level. Saturated fats raise this level, polyunsaturated fats lower it, while monounsaturated fats leave it about the same. Foam cells:  The first stage in the formation of atherosclerotic plaques is the appearance of fatty streaks in the intima, the part of the wall of the artery consisting of the endothelium or inner lining of the artery, and a section just underneath it. These fatty streaks are mostly composed of foam cells, so-called because the excess lipid in them (mainly cholesterol esters) is stored in large droplets that give the cells a foamy appearance under the microscope. Gangrene:  This consists of the death of cells in areas of the body due to

reduced blood supply to those areas. It can occur in the feet and legs as the result of peripheral vascular disease, that is the formation of atherosclerotic plaques, perhaps leading to thrombosis, in the arteries supplying the feet and legs. Genes:  These are the basic physical and functional units of heredity. They are

composed of DNA, and contain instructions for making proteins. Glucose:  This is a simple sugar which circulates in the blood, and acts as a source of energy. Glycogen:  This consists of several glucose molecules joined together. It is the

form in which glucose is stored in the body, and is broken down into glucose when this is required. HDL:  High density lipoprotein. High levels of HDL are associated with lower

Copyright © 2018. Routledge. All rights reserved.

risk of atherosclerosis. Heart attack:  This is an obstruction of one of the major arteries feeding the heart, usually due to a blood clot (thrombosis) forming at the surface of an artery affected by advanced atherosclerosis. The obstruction causes the death of a portion of the heart muscle. In 25–35% of cases, a first heart attack is fatal. Helicobacter pylori:  This is a bacterium found in the stomach, which is thought

to be a cause of peptic ulcers. Herpes Simplex Virus Type 2:  This is a virus which infects humans and is transmitted by sexual intercourse. It is correlated with cervical cancer, and from about 1970 to about 1985 was thought to be the cause of cervical cancer. However, this turned out to be false. Histopathology:  This is the microscopic examination of tissue to study whether it contains any manifestations of disease. HIV:  Human Immunodeficiency Virus, considered to be the cause of AIDS.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:37.

276  Glossary of medical terms

HPV:  Abbreviation of Human Papilloma Virus. HSV2:  Abbreviation of Herpes Simplex Virus Type 2. Human Papilloma Virus: A virus which infects humans, and is now considered to be the cause of cervical cancer. Hydrogen peroxide:  This is a strong oxidizing agent. It is dangerous if found

in the body because the oxidation of LDL and other bodily substances can result in disease. Hyper-:  From Greek hyper, meaning over much, above measure. It is used as a

prefix for a substance to indicate that the quantity of that substance in the body is so high as to constitute a threat to health. Hypercholesterolemia:  Over high cholesterol level. Hyperlipidemia:  Over high level of lipids. Hypertension:  Over high blood pressure. Infarction:  This is tissue death caused by inadequate blood supply to the affected area. Inflammation:  This is an immune response which causes the affected part of the body to become red, swollen, hot, and often painful. Insulin:  This is a chemical which is produced by the pancreas, and which regu-

Copyright © 2018. Routledge. All rights reserved.

lates the level of glucose in the blood. If the pancreas fails to produce enough insulin, the result is the disease diabetes. Integrase:  This is an enzyme produced by HIV, which enables the genetic ­m aterial of the virus to be integrated into the DNA of the infected cell. Intermittent claudication:  This consists of muscle pains in the legs brought

about by walking and relieved by rest. It is caused by atherosclerotic plaques in the arteries supplying blood to the legs. Intima:  The part of the wall of the artery consisting of the endothelium or inner

lining of the artery, and a section just underneath it. Ischaemic heart disease:  This is a synonym for coronary heart disease. The name comes from ischaemia, which means a restriction of the blood supply to tissues. LDL:  Low density lipoprotein, the major carrier of cholesterol in the blood. High

LDL levels are associated with increased risk of atherosclerosis. Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:37.

Glossary of medical terms  277

Leukocytes:  From the Greek leukos = white, kutos = cell, these are the white cells of the immune system which protect the body against infectious diseases and foreign invaders. Lipids:  These are substances which have a very limited solubility in water, but dis-

solve easily in alcohol, ether, or other organic solvents. They include cholesterol, and fat. Lipoids:  “Lipoid” is usually a synonym for lipid. Lipoprotein: Since cholesterol and other fats do not dissolve in blood serum, they

are stabilized for transportation by being packaged in particles called lipoproteins. Lumen:  The inside space of a tubular structure in the body, such as an artery, or the colon. Lycopene:  A powerful anti-oxidant found in tomatoes, carrots, and other fruits and vegetables. Macrophages:  From the Greek: makros = large, phagin = eater, these are

white blood cells which engulf and digest cellular debris, invading microbes, cancer cells, and anything else which does not seem to be a healthy body cell. McArdle Disease:  This is a rare disease first identified by a British doctor,

Brian McArdle (1911–2002) in Guy’s Hospital, London in 1951. The symptoms are fatigue and muscle pain after a small amount of exertion. Its cause is the lack, for genetic reasons, of an enzyme (phosphorylase) which is needed to break down the muscle’s store of glycogen into glucose.

Copyright © 2018. Routledge. All rights reserved.

Metabolic:  Metabolism refers to the chemical transformations which occur

within the cells of living organisms and are needed to sustain life. These chemical reactions are organized into metabolic pathways in which one chemical is transformed into another in a sequence of steps facilitated by a sequence of enzymes. Monocytes:  These are white blood cells with a single nucleus which are the precursors of macrophages. When a monocyte reaches a place where action is needed, it can transform into a macrophage. Monounsaturated fat:  This is a type of fat, which has one and only one double bond between its carbon atoms. Consumption of this type of fat leaves blood cholesterol levels largely unaffected. Olive oil is the most familiar source of monounsaturated fat. Mycobacterium tuberculosis: The bacillus which causes tuberculosis. Myo-:  From the Greek muos = muscle, a prefix which indicates that what fol-

lows relates to the muscles. Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:37.

278  Glossary of medical terms

Myocardial infarction:  This is the medical term for heart attack. It is an infarc-

tion, i.e. tissue death due to inadequate blood supply, of part of the myocardium, i.e. heart muscle. It is usually caused by a blood clot (thrombosis) in one of the arteries supplying the heart, and this in turn is usually brought about by advanced atherosclerotic plaques in the artery. In some cases, myocardial infractions are preceded by angina pectoris, but quite often they occur with no previous warning. Myocardium:  The heart muscle. Myopathy:  Disease of the muscles. Myophosphorylase:  The enzyme phosphorylase in the muscles which is needed

to transform stores of glycogen into glucose which can be used to provide energy. Osteomyelitis:  This is a bone infection. Its occurrence in children is usually a result of the strong blood supply to the growing bones – particularly the long bones of the arms and legs. This bloodstream can carry to the bone pathogenic microbes, which, in 90% of cases, are staphylococcus aureus. Osteoporosis:  This is a disease which causes bones to become weak and brittle, making them liable to fractures. One of the factors which leads to osteoporosis is a low calcium intake in the diet, especially if this is combined with low levels of vitamin D. Oxidative stress:  Reactive oxygen species or ROS are powerful oxidizing agents,

such as superoxide and hydrogen peroxide. In large quantities, they can cause damage to cells unless countered by antioxidants. The preponderance of such species produces what is known as oxidative stress.

Copyright © 2018. Routledge. All rights reserved.

Para-amino-salicylic acid:  A variant on aspirin (acetyl-salicylic acid) which

acts as a tuberculostatic agent. Parkinson’s disease:  A progressive neurological disease producing tremors, rigidity, slowness of movement, and difficulty with walking. It is caused by the death of brain cells which results in a failure to produce enough dopamine for normal functioning. PAS:  Abbreviation of para-amino-salicylic acid. Pathogenic bacteria:  Bacteria which cause diseases. Peripheral vascular disease:  This is disease of the arteries other than those

which supply the brain and heart. It is usually caused by the formation of atherosclerotic plaques in the artery. It characteristically affects the arteries of the legs, producing symptoms such as intermittent claudication, i.e. leg pain when walking Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:37.

Glossary of medical terms  279

which resolves with rest. In extreme cases it can result in gangrene, and the need for amputation. Phenylalanine:  This is an amino acid which is essential for humans, but which cannot be synthesized in the body, and so must be obtained from the diet. It is found in milk. Phosphorylase:  This is an enzyme needed to convert glycogen stored in the body to glucose which can be used to provide energy for the muscles. Polyunsaturated fat:  This is fat which contains more than one double bond between carbon atoms. Consumption of polyunsaturated fats lowers blood cholesterol level. Polyunsaturated fats are found in vegetable oils, such as corn oil and sunflower oil, and also in fish oils. Procyanidin:  This is a chemical found in apples, raspberries, strawberries, nuts

(particularly walnuts), and some red wines. It is thought to be ­protective against heart disease by inhibiting the production of the ­vasoconstrictor endothelin-1. Prospective studies:  These are epidemiological studies related to a particular disease D. A group of participants who are free of D at the beginning of the study are followed for a number of years, and measurements of various of their characteristics are carried out at regular intervals. An attempt is then made to find significant correlations between those who get D and some of their measured characteristics. Protease:  An enzyme which breaks the long chainlike molecules of proteins

Copyright © 2018. Routledge. All rights reserved.

into shorter fragments. Protein(s):  These are large biomolecules composed of one or more long chains of amino acid residues. Protein folding/misfolding :  Proteins consist of sequences of amino acid

residues, but as they form, they fold into shapes which are important for their biological function. If a protein misfolds, it can no longer carry out its proper function, and may produce a disease process instead. Protein synthesis:  A protein is formed from information in a gene in the

DNA. Usually this is a two-step process, in which RNA is synthesized from a DNA template (transcription), and then the amino acids are assembled from the RNA (translation). Puerperal fever:  Another name for childbed fever. It was so-called because it

often struck mothers during the puerperium. The fever is often severe, and, before antibiotics, could be fatal in two or three days. Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:37.

280  Glossary of medical terms

Puerperium:  The period of about six weeks after childbirth during which the

mother’s reproductive organs return to their normal pre-pregnancy condition. R.R.:  Abbreviation of resistance ratio. Randomization:  This is a process whereby membership of a particular group is decided randomly, e.g. by tossing a coin or rolling a die. Randomized controlled trials:  These are controlled trials in which patients are assigned to the control group or one of the treatment groups by a random process such as tossing a coin or rolling a die. RCT:  Abbreviation of randomized controlled trial. Reactive oxygen species:  These are powerful oxidizing agents containing

oxygen, such as hydrogen peroxide or superoxide. Resistance ratio:  This is a measure of the resistance which pathogenic bacteria

have developed to an antibiotic. It is the ratio of the minimum concentration of the antibiotic to which the resistant pathogenic bacteria are sensitive to the corresponding figure for a standard non-resistant strain of the pathogenic bacterium. Retinylstearate:  This is a compound of vitamin A, and acts as an antioxidant, protecting LDL from oxidation. Retrospective studies:  These, like prospective studies, are epidemiological stud-

Copyright © 2018. Routledge. All rights reserved.

ies relating to a particular disease in which a group is followed of a number of years. In contrast to prospective studies, the time period in which the group is followed is before the present, and the information about the group is obtained from historical records. RNA:  An abbreviation for RiboNucleic Acid. This is an intermediate sub-

stance through which DNA makes proteins. The instructions on the DNA are transformed into RNA, which then assembles the amino acids and constructs the protein. ROS:  An abbreviation for reactive oxygen species. Saturated fat:  This is fat which contains no double bonds between carbon

atoms. Consumption of saturated fat raises blood cholesterol level. It is found mainly in animal products such as butter, cheese, cream, and meat. Serum cholesterol level:  This is a measure of the amount of cholesterol which

circulates in the blood serum in the form of lipoproteins. It is an alternative expression for blood cholesterol level. Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:37.

Glossary of medical terms  281

Streptomycin:  This is an antibiotic discovered in 1944 in America by Schatz, Bugie, and Waksman. It destroys the bacilli of tuberculosis and other pathogenic bacteria. It cannot, however, be used on its own as a cure for tuberculosis, because, in the time it takes to eliminate the bacilli, some of them acquire resistance. The first effective cure for tuberculosis used a combination of streptomycin with other anti-tubercular agents. Stroke:  This is an event which results in the death of cells in the brain. Its

consequences depend on which part of the brain is affected. There are two main types of stroke. The first is ischaemic, due to lack of blood flow to an area in the brain. This in turn is usually caused by atherosclerotic plaques in the arteries supplying blood to the brain. These can result in a blood clot (thrombosis) which blocks an artery. The second type of stroke is haemorrhagic, due to bleeding in the brain. Superoxide:  This is an oxygen molecule with one extra electron, making it a negatively charged free radical. Superoxide can cause oxidative damage to cells in the body unless countered by antioxidants. T4 lymphocyte:  T lymphocytes are white cells which are found in the lymph and blood, and which form an important part of the immune system. The T4 lymphocyte is a favoured target of HIV when this virus attacks the immune system. Thalidomide:  This is a drug discovered by the German pharmaceutical com-

Copyright © 2018. Routledge. All rights reserved.

pany, Chemie Grü nenthal in 1954. It was an effective sedative and sleeping pill, and some animal tests suggested that it would be a completely safe. However, after it was put on the market in 1957, it led to horrifying birth defects in babies when taken by pregnant women. Thrombosis:  The formation of a clot in a blood vessel. This characteristically occurs in arteries in places where they are badly affected by atherosclerotic plaques. Tubercle bacillus:  A common name for mycobacterium tuberculosis, the bacillus

which causes tuberculosis. Tuberculosis:  This is characteristically a chronic infection lingering for months

and sometimes years. It most commonly affects the lungs (pulmonary tuberculosis), but can affect any organ of the body. Indeed, one form of the disease (miliary tuberculosis) affects almost every organ of the body. Koch established in 1882 that tuberculosis was caused by a bacillus, mycobacterium tuberculosis. However, an effective cure for tuberculosis was not found until the early 1950s. Tuberculostatic agent:  This is an agent which inhibits the growth of the tuber-

cle bacillus. Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:37.

282  Glossary of medical terms

Typhoid:  This is a systemic infection caused by the bacillus, salmonella typhi.

Untreated the fever last about three to four weeks and kills about 10% of the sufferers. Vibrio cholerae:  This is the bacterium which causes cholera. Virus:  This is a small infectious agent which replicates only inside the living

cells of other organisms. Vitamin:  This is an organic compound which is required in small amounts for normal growth and health. Deficiency of a vitamin results in the development of a specific deficiency disease, which can be cured or prevented only by that vitamin. Vitamin C:  This is a water-soluble vitamin found in citrus fruits, tomatoes, red

peppers, and potatoes. Vitamin C deficiency causes scurvy. Vitamin C is an antioxidant, and its presence in the bloodstream protects against atherosclerosis. Vitamin E:  This is a fat-soluble vitamin, which is an anti-oxidant, and so pro-

tects against atherosclerosis. Xanthoma cells:  This is an old name for what are now normally called foam

Copyright © 2018. Routledge. All rights reserved.

cells.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:37.

Copyright © 2018. Routledge. All rights reserved.

REFERENCES

Abela, George S. (2003) Atherosclerosis as an Inflammatory Arterial Disease “Dé jà vu”? American College of Cardiology. Current Journal Review, Jul/Aug, pp. 23–25. Anitschkow, Nikolai (1933) Experimental Arteriosclerosis in Animals. In E. V. Cowdry (ed.) Arteriosclerosis, New York: Macmillan, Chapter 10, pp. 271–322. Anitschkow, N. and Chalatow, S. (1913) Ü ber experimentelle Cholesterinsteatose und ihre Bedeutung fü r die Entstehung einiger pathologischer Prozesse, Zentralblatt fü r allgemeine Pathologie und pathologische Anatomie, 24, pp. 1–9. Anscombe, G. E. M. (1971) Causality and Determination. In Ernest Sosa and Michael Tooley (eds.) Causation, Oxford Readings in Philosophy, Oxford University Press, 2011, pp. 88–104. Aquinas, St Thomas Summa Theologiae. Austin, J. L. (1956) A Plea for Excuses, Proceedings of the Aristotelian Society, 57, pp. 1–30. Bacon, Francis (1620) Novum Organum. English translation in R.L. Ellis and J. Spedding (eds.) The Philosophical Works of Francis Bacon, Routledge, 1905, pp. 212–387. Bernard, Claude (1865) An Introduction to the Study of Experimental Medicine. English translation, New York: Dover, 1957. Broadbent, Alex (2011) Inferring Causation in Epidemiology: Mechanisms, Black Boxes, and Contrasts. In Phyllis McKay Illari; Federica Russo; and Jon Williamson (eds.) Causality in the Sciences, Oxford University Press, Chapter 3, pp. 45–69. Broadbent, Alex (2013) Philosophy of Epidemiology. Basingstoke, UK: Palgrave Macmillan. Brock, Thomas, D. (1988) Robert Koch. A Life in Medicine and Bacteriology. 2nd Edition, Washington, DC: ASM Press, 1999. Brynner, Rock and Stephens, Trent (2001) Dark Remedy. The Impact of Thalidomide and its Revival as a Vital Medicine. New York: Basic Books. Burton, Robert (1621) The Anatomy of Melancholy. Reprint of the 1651–2 Edition, Oxford: Benediction Classics. Buzzoni, Marco (2014) The Agency Theory of Causality, Anthropomorphism, and Simultaneity, International Studies in the Philosophy of Science, 28(4), pp. 375–395. Campaner, Raffaella (2012) Philosophy of Medicine. Causality, Evidence and Explanation. Bologna: Archetipolibri.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:43.

Copyright © 2018. Routledge. All rights reserved.

284 References

Campaner, Raffaella and Galavotti, Maria Carla (2007) Plurality in Causality. In Peter Machamer and Gereon Wolters (eds.) Thinking about Causes. From Greek Philosophy to Modern Physics, University of Pittsburgh Press, pp. 178–199. Reprinted in Campaner, 2012, pp. 63–82. Carter, K. Codell (1985) Koch’s Postulates in Relation to the Work of Jacob Henle and Edwin Klebs, Medical History, 29, pp. 353–374. Carter, K. Codell (1987) Essays of Robert Koch. Translated into English by K. Codell Carter, New York: Greenwood Press. Carter, K. Codell (2003) The Rise of Causal Concepts of Disease. Case Histories. Aldershot, UK: Ashgate. Carter, K. Codell (2012) The Decline of Therapeutic Bloodletting and the Collapse of Traditional Medicine. New Brunswick, NJ: Transaction Publishers. Cartwright, Nancy (1979) Causal Laws and Effective Strategies. Reprinted in How the Laws of Physics Lie, Oxford University Press, 1983, pp. 21–43. Cartwright, Nancy (1989) Nature’s Capacities and their Measurement. Oxford University Press. Cartwright, Nancy (1995) False Idealisation: A Philosophical Threat to Scientific Method, Philosophical Studies, 77, pp. 339–352. Cartwright, Nancy (2001) What is Wrong with Bayes Nets?, The Monist, 84, pp. 242–264. Clarke, Brendan (2011) Causality in Medicine with Particular Reference to the Viral Causation of Cancers. PhD Thesis, University College London. Clarke, Brendan (2012) Causation in Medicine. In Wenceslao J. Gonzalez (ed.) Conceptual Revolutions: From Cognitive Science to Medicine, A Coruñ a, Spain: Netbiblio, pp. 181–194. Clarke, Brendan (2017) Discovery in Medicine. In Miriam Solomon; Jeremy R. Simon; and Harold Kincaid (eds.) The Routledge Companion to Philosophy of Medicine, New York and London: Routledge, pp. 285–295. Clarke, Brendan; Gillies, Donald; Illari, Phyllis; Russo, Federica; and Williamson, Jon (2013) The Evidence that Evidence-Based Medicine Omits, Preventive Medicine, 57, pp. 745–747. Clarke, Brendan; Gillies, Donald; Illari, Phyllis; Russo, Federica; and Williamson, Jon (2014) Mechanisms and the Evidence Hierarchy, Topoi, 33, pp. 339–360. Clarke, Brendan, and Russo, Federica (2017) Causation in Medicine. In J. Marcum (ed.) The Bloomsbury Companion to Contemporary Philosophy of Medicine, London: Bloomsbury, pp. 297–322. Clarke, Brendan, and Russo, Federica (2018) Mechanisms in Medicine. In S. Glennan and P. Illari (eds.) Routledge Handbook of Mechanisms, New York: Routledge, pp. 319–331. Coleman, William (1987) Koch’s Comma Bacillus: The First Year, Bulletin of the History of Medicine, 61, pp. 315–342. Collingwood, R. G. (1938) On the So-Called Idea of Causation, Proceedings of the Aristotelian Society, 38, pp. 85–112. Collingwood, R. G. (1939) An Autobiography. Oxford University Press. Collingwood, R. G. (1940) An Essay on Metaphysics. Oxford: Clarendon Press, Part IIIc: Causation, pp. 285–337. Corder, Roger (2007) The Wine Diet. London: Sphere, Paperback Edition, 2014. Daniels, M. and Bradford Hill, A. (1952) Chemotherapy of Pulmonary Tuberculosis in Young Adults. An Analysis of the Combined Results of Three Medical Research Council Trials, British Medical Journal, 31 May, pp. 1162–1168.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:43.

Copyright © 2018. Routledge. All rights reserved.

References 

285

Darden, Lindley (2013) Mechanisms Versus Causes in Biology and Medicine. In H.-K. Chao; S.-T. Chen; and R. L. Millstein (eds.) Mechanism and Causality in Biology and Economics, Dordrecht, the Netherlands: Springer, chapter 2, pp. 19–34. Dayton, Seymour; Pearce, Morton Lee; Goldman, Hubert; Harnish, Alan; Plotkin, David; Shickman, Martin; Winfield, Mark; Zager, Albert; and Dixon, Wilfrid (1968) Controlled Trial for a Diet High in Unsaturated Fat for Prevention of Atherosclerotic Complications, The Lancet, 2, pp. 1060–1062. Dayton, Seymour, Pearce, Morton Lee; Hashimoto, Sam; Dixon, Wilfrid; and Tomiyasu, Uwamie (1969) A Controlled Clinical Trial of a Diet High in Unsaturated Fat in Preventing Complications of Atherosclerosis, Circulation, 40(1), Supplement II, pp. 1–63. De Finetti, B. (1937) Foresight: Its Logical Laws, Its Subjective Sources. English translation in H. E. Kyburg and H. E. Smokler (eds.) Studies in Subjective Probability. New York: Wiley, 1964, pp. 93–158. De Renzy, A. C. C. (1884) The Extinction of Cholera Epidemics in Fort William, The Lancet, 2, pp. 1043–1044. Debré , Patrice (1994) Louis Pasteur. English translation by Elborg Forster, Baltimore, MD: Johns Hopkins University Press, 1998. Dingler, H. (1938) Die Methode der Physik. Munich: Reinhardt. Doll, Richard and Peto, Richard (1976) Mortality in Relation to Smoking: 20 Years’ Observations on Male British Doctors, British Medical Journal, 2, pp. 1525–1536. Dowe, P. (1992) Process Causality and Asymmetry, Erkenntnis, 37, pp. 179–196. Dummett, Michael (1964) Bringing about the Past. Reprinted in Truth and Other Engimas, London: Duckworth, 1978, pp. 333–50. Eells, Ellery (1991) Probabilistic Causality. Cambridge University Press. Esterbauer, Hermann; Striegl, Georg; Puhl, Herbert; Oberreither, Sabine; Rothender, Martina; El-Saadani, Mohammed; and Jü rgens, Gü nther (1989) The Role of Vitamin E and Carotenoids in Preventing Oxidation of Low Density Lipoproteins, Annals of the New York Academy of Sciences, 570, pp. 254–267. Estruch, Ramón; Ros, Emilio; Salas-Salvadó, Jordi; Covas, Maria-Isabel; Corella, Dolores; Arós, Fernando; Gómez-Gracia, Enrique; Ruiz-Guitié r rez, Valentina; Fiol, Miquel; Lapetra, José ; Lamuela-Raventos, Rosa Maria; Serra-Majem, Lluí s ; Pintó, Xavier; Basora, Josep; Muñ oz, Miguel, Angel; Sorlí , José , V.; Martí nez, José , Alfredo; and Martí nez-Gonzá lez, Miguel, Angel (2013) Primary Prevention of Cardiovascular Disease with a Mediterranean Diet, The New England Journal of Medicine, 368(14), 4 April, pp. 1279–1290. Evans, Alfred S. (1993) Causation and Disease. A Chronological Journey. New York: Plenum Medical Book Company. Evans, Richard J. (1987) Death in Hamburg. Society and Politics in the Cholera Years. New York: Penguin, 2005. Florey, M. E. (1961) The Clinical Application of Antibiotics. Volume II Streptomycin and other Antibiotics Active against Tuberculosis. Oxford University Press. Fortin, Martin; Dionne, Jonathan; Pinbo, Geneviè ve; Gignac, Julie; Almirall, José ; and Lapointe, Lise (2006) Randomized Controlled Trials: Do They Have External Validity for Patients with Multiple Comorbidities? Annals of Family Medicine, 4(2), pp. 104–8. Freudenheim, Jo L.; Ritz, John; Smith-Warner, Stephanie A.; Albanes, Demetrius; Bandera, Elisa V.; van den Brandt, Piet A.; Colditz, Graham; Feskanich, Diane; Goldbohm, R. Alexandra; Harnack, Lisa; Miller, Anthony B.; Rimm, Eric; Rohan, Thomas E.; Sellers, Thomas A.; Virtamo, Jarmo; Willett, Walter C.; and Hunter,

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:43.

Copyright © 2018. Routledge. All rights reserved.

286 References

David J. (2005) Alcohol Consumption and Risk of Lung Cancer: A Pooled Analysis of Cohort Studies, American Journal of Clinical Nutrition, 2005, 82, pp. 657–667. Galavotti, Maria Carla (2010) Probabilistic Causality, Observation and Experimentation. In W. J. Gonzalez (ed.) New Methodological Perspectives on Observation and Experimentation in Science. A Coruñ a, Spain: Netbiblo, pp. 139–155. Gasking, Douglas (1955) Causation and Recipes, Mind, 64, pp. 479–487. Gillies, Donald (2000) Philosophical Theories of Probability. London: Routledge. Gillies, Donald (2005a) An Action-Related Theory of Causality, British Journal for the Philosophy of Science, 56, pp. 823–842. Gillies, Donald (2005b) Hempelian and Kuhnian Approaches in the Philosophy of Medicine: The Semmelweis Case, Studies in the History and Philosophy of Biological and Biomedical Sciences, 36, pp. 159–181. Gillies, Donald (2008) How Should Research be Organised? College Publications. Gillies, Donald (2011) The Russo-Williamson Thesis and the Question of Whether Smoking Causes Heart Disease. In Phyllis McKay Illari; Federica Russo; and Jon Williamson (eds.) Causality in the Sciences, Oxford University Press, Chapter 6, pp. 110–125. Gillies, Donald (2014) Selecting Applications for Funding: Why Random Choice is Better Than Peer Review, RT. A Journal on Research Policy and Evaluation, 2, pp. 1–14. Gillies, Donald (2016a) The Propensity Interpretation. In Alan Há jek and Christopher Hitchcock (eds.) The Oxford Handbook of Probability and Philosophy, Oxford University Press, pp. 406–422. Gillies, Donald (2016b) Establishing Causality in Medicine and Koch’s Postulates, International Journal of History and Philosophy of Medicine, 6, pp. 1–13. (Open Access Journal). Gillies, Donald and Sudbury, Aidan (2013) Should Causal Models Always be Markovian? The Case of Multi-Causal Forks in Medicine, European Journal for Philosophy of Science, 3(3), pp. 275–308. Glennan, Stuart (1996) Mechanisms and the Nature of Causation, Erkenntnis, 44, pp. 49–71. Glennan, Stuart (2002) Rethinking Mechanistic Explanation, Philosophy of Science, 69, pp. S342–S353. Glennan, Stuart (2017) The New Mechanical Philosophy. Oxford University Press. Goldstein, Joseph L.; Ho, Y. K.; Basu, Sandip K.; and Brown, Michael S. (1979) Binding Site on Macrophages that Mediates Uptake and Degradation of Acetylated Low Density Lipoprotein, Producing Massive Cholesterol Deposition, Proceedings of the National Academy of Sciences USA, 76(1), pp. 333–337. Good, I. J. (1959) A Theory of Causality, British Journal for the Philosophy of Science, 9, pp. 307–310. Good, I. J. (1961) A Causal Calculus I, British Journal for the Philosophy of Science, 11, pp.305–318. Reprinted in I. J. Good, Good Thinking. The Foundations of Probability and Its Applications. University of Minnesota Press, 1983, pp. 197–217. Good, I. J. (1962) A Causal Calculus II, British Journal for the Philosophy of Science, 12, pp. 43–51. Reprinted in I. J. Good, Good Thinking. The Foundations of Probability and Its Applications. University of Minnesota Press, 1983, pp. 197–217. Gradmann, Christoph (2005) Laboratory Disease. Robert Koch’s Medical Bacteriology. English translation by Elborg Forster, Baltimore, MD: Johns Hopkins University Press, 2009. Gradmann, Christoph (2014) A Spirit of Scientific Rigour: Koch’s Postulates in Twentieth-Century Medicine, Institut Pasteur. Microbes and Infection, 16, pp. 885–892.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:43.

Copyright © 2018. Routledge. All rights reserved.

References 

287

Guyatt, G. H.; Oxman, A. D.; Vist, G. E.; Kunz R.; Falck-Ytter, Y.; Alonso-Coello, P.; Schü nemann, H. J.; and GRADE Working Group (2008) GRADE: An Emerging Consensus Rating Quality of Evidence and Strength of Recommendations, British Medical Journal, 336, pp. 924–926. Haavelmo, T. (1943) The Statistical Implications of a System of Simultaneous Equations, 11, pp. 1–12. Reprinted in D. F. Hendry and M. S. Morgan (eds.) The Foundations of Econometric Analysis, Cambridge University Press, 1995, pp. 477–490. Hacking, I. (1975) The Emergence of Probability. A Philosophical Study of Early Ideas about Probability, Induction and Statistical Inference. Cambridge University Press, Paperback Edition, 1984. Harden, Victoria (1992) Koch’s Postulates and the Etiology of AIDS: An Historical Perspective, History and Philosophy of the Life Sciences, 14, pp. 249–269. Hare, Ronald (1963) An Outline of Bacteriology and Immunity. 2nd Edition, London: Longmans. Hennig, C. (2010) Mathematical Models and Reality – A Constructivist View, Foundations of Science, 15, pp. 29–48. Henriksen, Tore; Mahoney, Eileen M.; and Steinberg, Daniel (1981) Enhanced Macrophage Degradation of Low Density Lipoprotein Previously Incubated with Cultured Endothelial Cells: Recognition by Receptors for Acetylated Low Density Lipoproteins, Proceedings of the National Academy of Sciences USA, 78(10), pp. 6499–6503. Hesslow, Germund (1976) Discussion: Two Notes on the Probabilistic Approach to Causality, Philosophy of Science, 43, pp. 290–292. Hill, A. B. (1965) The Environment and Disease: Association or Causation? Proceedings of the Royal Society of Medicine, 58, pp. 295–300. Hill, A. B. (1990) Memories of the British Streptomycin Trial, Controlled Clinical Trials, Vol. II, pp. 77–79. Howick, Jeremy (2011a) The Philosophy of Evidence-Based Medicine. Chichester, UK: Wiley-Blackwell. Howick, Jeremy (2011b) Exposing the Vanities – and a Qualified Defense – of Mechanistic Reasoning in Health Care Decision Making, Philosophy of Science, 78(5), pp. 926–940. Hume, David (1738) A Treatise of Human Nature. Everyman Edition, Volume 1, London: J. M. Dent, 1961. Humphreys, P. (1985) Why Propensities Cannot be Probabilities, The Philosophical Review, 94, pp. 557–570. Ignatowski, A. (1909) Ü ber die Wirkung des tierischen Eiweisses auf die Aorta und die parenchymatö sen Organe der Kaninchen, Virchows Archiv fü r pathologische Anatomie und Physiologie und fü r klinische Medizin, 198(2), pp. 248–270. Illari, Phyllis, McKay (2011) Mechanistic Evidence: Disambiguating the RussoWilliamson Thesis, International Studies in the Philosophy of Science, 25(2), pp. 139–157. Illari, Phyllis, McKay; Russo, Federica; and Williamson, Jon (eds.) (2011) Causality in the Sciences. Oxford University Press. Illari, Phyllis, McKay and Williamson, Jon (2012) What is a Mechanism? Thinking about Mechanisms across the Sciences, European Journal for Philosophy of Science, 2(1), pp. 119–135. Illari, Phyllis, Mckay and Russo, Federica (2014) Causality. Philosophical Theory Meets Scientific Practice. Oxford University Press. Jü rgens, Gü nther; Hoff, Henry F.; Chisolm, Guy M. III; and Esterbauer, Hermann (1987) Modification of Human Serum Low Density Lipoprotein by Oxidation – Characterization and Pathophysiological Implications, Chemistry and Physics of Lipids, 45, pp. 315–336.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:43.

Copyright © 2018. Routledge. All rights reserved.

288 References

Kalra, Vijay K.; Ying, Yong; Deemer, Kathleen; Natarajan, Rama; Nadler, Jerry L.; and Coates, Thomas D. (1994) Mechanism of Cigarette Smoke Condensate Induced Adhesion of Human Monocytes to Cultured Endothelial Cells, Journal of Cellular Physiology, 160, pp. 154–162. Kant, Immanuel (1781/7) Critique of Pure Reason. English translation by Norman Kemp Smith, London: Macmillan, 1958. Kant, Immanuel (1821) Vorlesungen ü ber die Metaphysik. Erfurt: Kessinger. Kaptchuk, Ted J. (2002) Acupuncture: Theory, Efficacy, and Practice, Annals of Internal Medicine, 136(5), 5 March, pp. 373–383. Keys, Ancel (1957) Diet and the Epidemiology of Coronary Heart Disease, Journal of the American Medical Association, 164(17), pp. 1912–1919. Keys, Ancel (ed.) (1970) Coronary Heart Disease in Seven Countries, Supplement I to Circulation Vols. XLI And XLII, April, pp. 1–198. Keys, Ancel (1980) Seven Countries. Harvard University Press. Keys, Ancel (1983) From Naples to Seven Countries – A Sentimental Journey. In Progress in Biochemical Pharmacology, Vol 19, Karger: Basel, pp. 1–30. Keys, Ancel (1990) Recollections of Pioneers in Nutrition: From Starvation to Cholesterol, Journal of the American College of Nutrition, 9(4), pp. 288–291. Keys, Ancel with the collaboration of Olaf Mickelsen, E. v. O. Miller, and Carleton B. Chapman (1950) The Relation in Man between Cholesterol Levels in the Diet and in the Blood, Science, 2899 (21 July), pp. 79–81. Keys, Ancel; Anderson, Joseph, T.; Grande, Francisco (1957) Prediction of SerumCholesterol Responses in Man to Changes in Fats in the Diet, The Lancet, 16 November, pp. 959–966. Keys, Ancel; Kimura, Noboru; Kusukawa, Akira; Bronte-Stewart, B.; Larsen, Nils and Keys, Margaret, Haney (1958) Lessons from Serum Cholesterol Studies in Japan, Hawaii and Los Angeles, Annals of Internal Medicine, 48, pp. 83–94. Keys, Ancel and Keys, Margaret (1963) Eat Well and Stay Well. 2nd Revised Edition, Garden City, NY: Doubleday & Company, Inc. Keys, Ancel and Keys, Margaret (1975) How to Eat Well and Stay Well. The Mediterranean Way. Garden City, NY: Doubleday & Company, Inc. Kim, J. H. and Pearl, J. (1983) A Computational Model for Combined Causal and Diagnostic Reasoning in Inference Systems, Proceedings of the 8th International Joint Conference on AI (IJCAI-85), pp. 190–3. Kistler, Max (2013) The Interventionist Account of Causation and Non-Causal Associa­ tion Laws, Erkenntnis, 78, pp. 65–84. Koch, Robert (1876) The Etiology of Anthrax, Founded on the Course of Development of the Bacillus Anthracis, English translation in Carter, 1987, pp. 1–17. Koch, Robert (1878) Investigations of the Etiology of Wound Infections, English translation in Carter, 1987, pp. 19–56. Koch, Robert (1881) On the Etiology of Anthrax, English translation in Carter, 1987, pp. 57–81. Koch, Robert (1882) The Etiology of Tuberculosis, English translation in Carter, 1987, pp. 83–96. Koch, Robert (1884a) Lecture at the First Conference for Discussion of the Cholera Question, English translation in Carter, 1987, pp. 151–170. Koch, Robert (1884b) On Cholera Bacteria, English translation in Carter, 1987, pp. 171–177. Koch, Robert (1893) Wasserfiltration und Cholera, Gesammelte Werke, 2/1, pp. 183–206.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:43.

Copyright © 2018. Routledge. All rights reserved.

References 

289

Kolmogorov, A. N. (1933) Foundations of the Theory of Probability. 2nd English Edition, New York: Chelsea, 1956. Korb, K.B.; Hope, L.R.; Nicholson, A.E.; and Annick, K. (2004) Varieties of Causal Intervention, Pacific Rim International Conference on AI’04, pp. 322–31. Laker, Mike (2012) Understanding Cholesterol. Family Doctor Publications in association with the British Medical Association. Lauritzen, S.L. and Spiegelhalter, D.J. (1988) Local Computations with Probabilities on Graphical Structures and their Application to Expert Systems (with discussion), Journal of the Royal Statistical Society B, 50(2), pp. 157–224. Lehr, Hans-Anton; Frei, Balz; and Arfors, Karl-E. (1994) Vitamin C Prevents Cigarette Smoke-Induced Leukocyte Aggregation and Adhesion to Endothelium in vivo, Proceedings of the National Academy of Sciences USA, 91 (August), pp. 7688–7692. Levy, Daniel and Brink, Susan (2005) A Change of Heart. Unraveling the Mysteries of Cardiosvacular Disease. New York: Vintage Books. McArdle, B. (1951) Myopathy Due to a Defect in Muscle Glycogen Breakdown, Clinical Science, 10(1), pp. 13–33. Macfarlane, Gwyn (1984) Alexander Fleming. The Man and the Myth. London: Chatto & Windus. Machamer, Peter, Darden, Lindley, and Craver, Carl F. (2000) Thinking about Mechanisms, Philosophy of Science, 67(1), pp. 1–25. Malmros, Haqvin (1950) The Relation of Nutrition to Health. A Statistical Study of the Effect of the War-time on Arteriosclerosis, Cardiosclerosis, Tuberculosis and Diabetes, Acta Medica Scandinavica. Supplementum, 246, pp. 137–153. Marshall, B. J.; Armstrong, J. A.; McGechie, D. B.; and Clancy, R. J. (1985) Attempt to Fulfil Koch’s Postulates for Pyloric Campbylobacter, Medical Journal of Australia, 142, pp. 436–439. Mayerl, Christina; Lukasser, Melanie; Sedivy, Roland; Niederegger, Harald; Seiler, Ruediger; and Wick, Georg (2006), Atherosclerosis Research from Past to Present – on the Track of Two Pathologists with Opposing Views, Carl von Rokitansky and Rudolf Virchow, Virchows Archiv, 449, pp. 96–103. Menotti, Alessandro; Kromhout, Daan; Blackburn, Henry; Fidanza, Flaminio; Buzina, Ratko & Nissinen, Aulikki (1999) Food Intake Patterns and 25-year Mortality from Coronary Heart Disease: Cross-Cultural Correlations in the Seven Countries Study, European Journal of Epidemiology, 15, pp. 507–515. Menzies, P. and Price H. (1993) Causation as a Secondary Quality, British Journal for the Philosophy of Science, 44, pp. 187–204. Mommaerts, W. F. H. M.; Illingworth, Barbara; Pearson, Carl M.; Guillory, R. J.; and Seraydarian, K. (1959) A Functional Disorder of Muscle Associated with the Absence of Phosphorylase, Proceedings of the National Academy of Sciences of the United States of America, 45(6), pp. 791–7. Morrow, Jason D.; Frei, Balz; Longmire, Atkinson W.; Gaziano, J. Michael; Lynch, Sean M.; Shyr, Yu; Strauss, William E.; Oates John A.; and Roberts, L. Jackson  II (1995) Increase in Circulating Products of Lipid Peroxidation (F2-Isoprostanes) in Smokers. Smoking as a Cause of Oxidative Damage, The New England Journal of Medicine, 332(18), pp. 1198–1203. MRC (1948) Streptomycin Treatment of Pulmonary Tuberculosis, British Medical Journal, 30 October, pp. 769–782. MRC (1949) Treatment of Pulmonary Tuberculosis with Para-Aminosalicylic Acid and Streptomycin: Preliminary Report, British Medical Journal, 31 December, p. 1521.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:43.

Copyright © 2018. Routledge. All rights reserved.

290 References

MRC (1950) Treatment of Pulmonary Tuberculosis with Streptomycin and Para-AminoSalicylic Acid, British Medical Journal, 11 November, pp. 1073–1085. MRC (1952) The Prevention of Streptomycin Resistance by Combined Chemotherapy, British Medical Journal, 31 May, pp. 1157–1162. Neapolitan, R.E. (1990) Probabilistic Reasoning in Expert Systems. New York: John Wiley. Pearl, J. (1982) Reverend Bayes on Inference Engines: A Distributed Hierarchical Approach, Proceedings of the National conference on AI (ASSI-82), pp. 133–6. Pearl, J. (1985a) How to do with Probabilities What People Say You Can’t, Proceedings of the Second IEEE Conference on AI Applications, Miami, Florida, pp. 6–12. Pearl, J. (1985b) Bayesian Networks: A Model of Self-Activated Memory for Evidential Reasoning, Proceedings of the Cognitive Science Society, Ablex, pp. 329–34. Pearl, J. (1986) Fusion, Propagation and Structuring in Belief Networks, Artificial Intelligence, 29, pp. 241–88. Pearl, J. (1988) Probabilistic Reasoning in Intelligent Systems. Networks of Plausible Inference. San Mateo, CA: Morgan Kaufmann. Pearl, J. (2000) Causality. Models, Reasoning, and Inference. Cambridge University Press. Pearl, J. (2011) The Structural Theory of Causation. In Phyllis McKay Illari; Federica Russo; and Jon Williamson (eds.) Causality in the Sciences, Oxford University Press, Chapter 33, pp. 697–727. Pearson, Carl H.; Rimer, David G.; and Mommaerts, Wilfried F.H. (1961) A Metabolic Myopathy due to Absence of Muscle Phosphorylase, The American Journal of Medicine, 30(4), pp. 502–17. Peirce, C.S. (1910) Notes on the Doctrine of Chances, Reprinted in Essays in the Philosophy of Science, The American Heritage Series, Bobbs-Merrill, 1957, pp. 74–84. Pettenkofer, Max von (1884) Cholera, The Lancet, 2, pp. 769–771, 816–819, 861–864, 904–905, 992–994, 1042–1043, 1086–88. Phillips, B.; Ball, C.; Sackett D.; Badenoch, D.; Straus, S.; Haynes, B.; and Dawes, M.. (2001) Oxford Centre for Evidence-Based Medicine Levels of Evidence, Oxford: CEBM. Available from www.cebm.net/?o=1021. Plato The Republic. English translation by F. M. Cornford, Oxford: Clarendon Press, 1966. Pollan, Michael (2006) The Omnivore’s Dilemma. The Search for a Perfect Meal in a Fast-Food World. London: Bloomsbury, Paperback Edition, 2007. Popper, K. R. (1934) The Logic of Scientific Discovery. 6th revised impression of the 1959 English translation, Hutchinson, 1972, Popper, K. R. (1957a) The Propensity Interpretation of the Calculus of Probability, and the Quantum Theory. In S. Kö r ner (ed.) Observation and Interpretation, Proceedings of the Ninth Symposium of the Colston Research Society, University of Bristol, pp. 65–70 and pp. 88–9. Popper, K. R. (1957b) The Aim of Science, Ratio, 1(1), pp. 24–35. Quotation from the reprint as Chapter 5 of Objective Knowledge, Oxford University Press, 1972, pp. 191–205. Popper, K. R. (1959) The Propensity Interpretation of Probability, British Journal for the Philosophy of Science, 10, pp. 25–42. Popper, K. R. (1963) Conjectures and Refutations. London: Routledge and Kegan Paul. Popper, K. R. (1990) A World of Propensities. Bristol, UK: Thoemmes. Poston, Robin N.; and Johnson-Tidey, Ruth R. (1996) Localized Adhesion of Monocytes to Human Atherosclerotic Plaques Demonstrated in Vitro, Implications for Atherogenesis, American Journal of Pathology, 149(1), July, pp. 73–80. Price, Huw (1992) Agency and Causal Asymmetry, Mind, 101, pp. 501–20.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:43.

Copyright © 2018. Routledge. All rights reserved.

References 

291

Price, Huw (2017) Causation, Intervention and Agency – Woodward on Menzies and Price. In Helen Beebee; Chris Hitchcock; and Huw Price (eds.) Making a Difference: Essays on the Philosophy of Causation, Oxford University Press, forthcoming. Psillos, Stathis (2002) Causation and Explanation. 2nd Edition, New York: Routledge, 2014. Ramsey, F.P. (1926) Truth and Probability. In Ramsey, 1931, pp. 156–98. Ramsey, F.P. (1931) The Foundations of Mathematics and Other Logical Essays. London: Routledge and Kegan Paul. Reichenbach, Hans (1956) The Direction of Time. Edited by Maria Reichenbach, University of California Press, 1971. Reiss, Julian (2009) Causation in the Social Sciences. Evidence, Inference, and Purpose, Philosophy of the Social Sciences, 39(1), pp. 20–40. Robertson, Thomas L.; Kato, Hiroo; Rhoads, George, G.; Kagan, Abraham; Marmot, Michael; Syme, S. Leonard; Gordon, Tavia; Worth, Robert M.; Belsky, Joseph L.; Dock, Donald S.; Miyanishi, Michihiro; Kawamoto, Sadahisa (1977) Epidemiologic Studies of Coronary Heart Disease and Stroke in Japanese Men Living in Japan, Hawaii and California. Incidence of Myocardial Infarction and Death from Coronary Heart Disease, The American Journal of Cardiology, 39, pp. 239–243. Russell, Bertrand (1913) On the Notion of Cause. Reprinted in Mysticism and Logic, London: Routledge, 1994, pp. 173–99. Russell, Bertrand (1957) Mr. Strawson on Referring, Mind, 66, July, pp. 385–389. Russo, Federica (2009) Causality and Causal Modelling in the Social Sciences. Measuring Variations. New York: Springer. Russo, Federica and Williamson, Jon (2007) Interpreting Causality in the Health Sciences, International Studies in the Philosophy of Science, 21(2), pp. 157–170. Sackett, David L.; Rosenberg, William M. C.; Gray, J. A. Muir; Haynes, R. Brian; Richardson, W. Scott (1996). Evidence Based Medicine: What It Is and What It Isn’t: It’s About Integrating Individual Clinical Expertise and the Best External Evidence, British Medical Journal, 312(7023), pp. 71–72. Salmon, Wesley (1978) Why Ask, “Why?”? An Inquiry Concerning Scientific Explanation. Reprinted in Salmon, 1998, pp. 125–141. Salmon, Wesley (1979) Propensities: A Discussion Review of D. H. Mellor The Matter of Chance, Erkenntnis, 14, pp. 183–216. Salmon, Wesley (1980) Probabilistic Causality. Reprinted in Salmon, 1998, pp. 208–232. Salmon, Wesley (1993) A New Look at Causality. Reprinted in Salmon, 1998, pp. 13–24. Salmon, Wesley (1994) Causality without Counterfactuals. Reprinted in Salmon, 1998, pp. 248–260. Salmon, Wesley (1998) Causality and Explanation. Oxford University Press. Sanders, Tom and Bazalgette, Peter (1991) The Food Revolution. New York: Bantam Press. Schlosser, Eric (2001) Fast Food Nation. What the All-American Meal is Doing to the World. 2nd Edition, New York: Penguin Books, 2002. Semmelweis, I. (1861) The etiology, concept, and prophylaxis of childbed fever. English translation by K. Codell Carter, University of Wisconsin Press, 1983. Solomon, Miriam (2015) Making Medical Knowledge. Oxford University Press. Spirtes, P.; Glymour, C.; and Scheines, R. (1993) Causation, Prediction and Search. New York: Springer Verlag. Steel, Daniel (2008) Across the Boundaries. Extrapolation in Biology and Social Science. Oxford University Press. Steinberg, Daniel (1995) Clinical Trials of Antioxidants in Atherosclerosis: Are We Doing the Right Thing? The Lancet, 346, 1 July, pp. 36–38.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:43.

Copyright © 2018. Routledge. All rights reserved.

292 References

Steinberg, Daniel (2007) The Cholesterol Wars. The Skeptics vs. the Preponderance of Evidence. San Diego, CA: Academic Press. Steinberg, Daniel; Parthasarathy, Sampath; Carew, Thomas E.; Khoo, John C; and Witztum, Joseph L. (1989) Beyond Cholesterol. Modifications of Low-Density Lipoprotein That Increase its Atherogenicity, The New England Journal of Medicine, 320(14), pp. 915–924. Steinbrecher, Urs P.; Parthasarathy, Sampath; Leake, David S.; Witztum, Joseph L.; and Steinberg, Daniel (1984) Modification of Low Density Lipoprotein by Endothelial Cells Involves Lipid Peroxidation and Degradation of Low Density Lipoprotein Phospholipids, Proceedings of the National Academy of Sciences USA, 81( June), pp. 3883–3887. Sucar, L. E.; Gillies, D. F.; and Gillies, D. A. (1993) Objective Probabilities in Expert Systems, Artificial Intelligence, 61, pp. 187–203. Suppes, Patrick (1970) A Probabilistic Theory of Causality. Amsterdam: North-Holland. Teira, David (2013) On the Impartiality of Early British Clinical Trials, Studies in History and Philosophy of Biological Sciences, 44, pp. 412–418. Thagard, Paul (1999) How Scientists Explain Disease, Princeton University Press. Paperback Edition. 2000. Trö h ler, Ulrich (2000) “To Improve the Evidence of Medicine”. The 18th Century British Origins of a Critical Approach. Royal College of Physicians of Edinburgh. Twardy, C. R. and Korb, K. B. (2004) A Criterion of Probabilistic Causality, Philosophy of Science, 71, pp. 241–62. Von Mises, R. (1928) Probability, Statistics and Truth. 2nd revised English Edition, London: Allen and Unwin, 1961. Vonka, Vladimí r (2000) Causality in Medicine: The Case of Tumours and Viruses, Philosophical Transactions of the Royal Society of London B, 355, pp. 1831–1841. Vonka, V.; Kaň k a, J.; Jelí nek, J.; Š ubrt, I.; Suchá nek, A.; Havrá n ková , A.; Vá chal, M.; Hirsch, I.; Domorá zková , E.; Zá vadová , H.; Richterová , V.; Ná prstková , J.; Dvoř á ková , V.; and Svoboda, B. (1984) Prospective Study on the Relationship Between Cervical Neoplasia and Herpes Simplex Type-2 Virus. I. Epidemiological Characteristics, International Journal of Cancer, 33(1), pp. 49–60. Vonka, V.; Kaň k a, J.; Hirsch, I.; Zá d adová , H.; Krč m á ř , M.; Suchá n ková , A.; Rezá cová , D.; Broucek, J.; Press, M.; and Domorá zková , E. (1984) Prospective Study on the Relationship Between Cervical Neoplasia and Herpes Simplex Type-2 Virus. II. Herpes Simplex Type-2 Antibody Presence in Sera Taken at Enrollment, International Journal of Cancer, 33(1), pp. 61–6. Vonka, Vladimí r and Hamš i ková , Eva (2007) Vaccines Against Human Papillomaviruses – A Major Breakthrough in Cancer Prevention, Central European Journal of Public Health, 15(4), pp. 131–139. Williamson, Jon (2005) Bayesian Nets and Causality. Oxford University Press. Williamson, Jon (2009) Probabilistic Theories. In Helen Beebee; Christopher Hitchcock; and Peter Menzies (eds.) The Oxford Handbook of Causation. Oxford University Press, Paperback Edition, 2012, pp. 185–212. Williamson, Jon (2010) In Defence of Objective Bayesianism. Oxford University Press. Wittgenstein, Ludwig (1953) Philosophical Investigations. 2nd Edition of English translation by Elizabeth Anscombe, Oxford: Blackwell, 1963. Woodward, James (2003) Making Things Happen: A Theory of Causal Explanation. Oxford University Press, Paperback Edition, 2005. Woodward, James (2014) A Functional Account of Causation; or, A Defense of the Legitimacy of Causal Thinking by Reference to the Only Standard That Matters –

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:43.

References 

293

Copyright © 2018. Routledge. All rights reserved.

Usefulness (as Opposed to Metaphysics or Agreement with Intuitive Judgment), Philosophy of Science, 81(5), pp. 691–713. Wright, Georg Henrik von (1973) On the Logic and Epistemology of the Causal Relation. In P. Suppes; L. Henkin; A. Joja; and Gr.C. Mosil (eds.) Logic, Methodology and Philosophy of Science IV, Amsterdam: North-Holland, Elsevier, pp. 293–312. Wright, Georg Henrik von (1974) Causality and Determinism, New York: Columbia University Press.

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:43.

Copyright © 2018. Routledge. All rights reserved.

INDEX

Abela, G. 101 acetylated 143 Acquired Immune Deficiency Syndrome (AIDS) 52–53, 186, 188 action(s) vii, xv, 4–5, 15, 17–18, 23–26, 28, 30–35, 37–39, 43, 55–56, 67, 78, 85, 88,199, 202, 222, 254–257, 261; avoidance vii, 25–26, 31–33, 55–56, 67, 85, 199, 254, 256; blocking 25–26, 31–32, 55, 254, 256; productive vii, 24, 33, 55–56, 67, 256 activities viii, 76–79, 82–84, 89 acupuncture 174–176 aesthetics 87 amino acid 84 analgesic 183 angina pectoris 20, 94, 107 animal experiment(s) 54, 134–135, 161 Anitschkow, N. viii, ix, 95–100, 104, 108, 116, 128–129 Anscombe, G. E. M. viii, 76, 79, 82, 85–88 anthrax 5, 46–47, 56, 79–80, 83, 91, 97, 186, 190 anthropomorphic 37 antibiotics 55, 92, 153, 155 anti-oxidants 144–145, 147, 214–215 antipyretic 183 antiseptics 46, 171, 173 Aquinas, St Thomas 26, 29 arteriosclerosis 93, 96, 105–106 aspirin (acetylsalicylic acid) 156–157, 177, 182–183

asymmetry of causality/causation vii, 11, 31, 40–43 atherosclerosis viii, ix, xiii, 93–100, 104, 106, 124–126, 128, 142–145, 147– 148, 214–215 atherosclerotic plaques xvi, 93–94, 96–98, 100, 128, 142–148, 214 Austin, J. 86–87 autopsies 92–94, 100, 128–129, 135 bacillus anthracis 46–47, 56, 80, 83 Bacon, F. 5, 17 bacterium (bacteria) xv, 5, 20–21, 26, 43–44, 47–48, 50–52, 58, 63–65, 98, 143, 155, 172, 176 Ball, C. 151 Bazalgette, P. 217 Behring, E. von 47 Bergami, G. 104–105 Bernard, C. 29, 96, 134–135, 152, 172, 180 Bernheim, F. 156 Berzuini, C. xiii beta-carotene 144 binding site 143 blood-letting 135, 180 blood/serum cholesterol level 95, 104–105, 107, 110–111, 128, 142, 209, 213–214, 218 Bradley, F.H. 22 Brink, S. 209 Broadbent, A. 101, 168, 172–173 Brock, T. 52–53, 58, 68

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:49.

Index  295

Copyright © 2018. Routledge. All rights reserved.

Brynner, R. 159–162 Bugie, E. 153 Burkitt’s lymphoma 197–198 Burton, R. 29, 178, 180, 182 Buzzoni, M. xiii, 29, 35 Campaner, R. xii, 7, 8, 163, 183, 186 Carter, K. Codell 2, 43, 50, 52, 180 Cartwright, N. xiii, 201–202, 233–234, 237–240 case control studies 151 causal factor(s) 110, 208–209, 217, 234, 238, 246, 253, 255 causal hypothesis(es) viii, 79, 81, 90, 94, 107, 111–112, 115–116, 125, 127–128, 131–134, 148–150, 152, 164, 171, 178, 181, 185–186, 189–191, 198, 250 causal model(s) x, 210, 222–223, 225, 229–230, 234, 237–240, 245–247, 250–251, 255–257, 262–263 causal monism 164 causal pluralism ix, 163–164 causality/causation iii, vii–viii, x–xiii, xv–xvi, 1–11, 13, 15, 17–44, 49–50, 54–55, 58–59, 67, 69, 71–75, 77–79, 81–82, 85, 88, 111–112, 129, 135–139, 148–149, 159, 172, 193, 195–197, 199–204, 207–209, 225, 230–236, 243, 246, 250–257, 260; action-related theory of vii–viii, x–xii, xv, 5, 7, 13, 17–18, 24, 26–28, 31, 33–35, 37–44, 50, 54–55, 67, 78–79, 82, 85, 88, 129, 195, 199, 204, 252, 255–257; AIM theories of vii, 4, 5, 7, 17–18, 21, 23–24, 28, 30–34, 36–38, 40, 42–43, 55, 67; deterministic vii, 2–5, 9–11, 25, 28, 43–44, 135, 195–197, 200, 208, 234, 256–257; Dowe-Salmon theory of viii, 71–75; generic 1, 28, 30, 32, 202, 231, 256; indeterministic vii, x–xi, 2–3, 5, 8–11, 43, 88, 195–197, 200, 203, 208, 234, 246, 250, 252, 255–257; mechanistic theories of vii, 4–7, 71, 73, 77–79, 81–82; mono- 208; multi- 208; probabilistic theories of vii, 4, 8–10; single-case/singular 1, 30, 32, 202, 256; token 1, 256; type 1, 256; viral xiii causality probability connection principle (CPCP) 13, 202–203, 207, 225, 232– 234, 236, 244–246, 251, 253, 257 cell membrane 84 cervical cancer (CaCer) x, xiii, 8, 88, 197–200

CFTR gene 83–84, 186 Chadwick, E. 58 Chain, E. 21 Chalatow, S. 96 Chang, H. xiii childbed fever 171–172 chloride ions 84 cholera viii, xv, 5, 44, 47–50, 53–54, 57–68, 91, 169, 171, 212 cholesterol xv–xvi, 94, 96–97, 99–100, 104–105, 108, 110–115, 118–122, 124–125, 128, 142, 144, 209–210, 215, 218 circularity vii, 38, 40, 74–75, 77–78, 88 Clarke, B. xii, 95, 121, 130–131, 162, 164, 166–167, 183, 197, 203 clinical trials ix, xv, 92, 102, 119, 121–122, 127, 129, 160 cohort study(ies) 91, 94, 108–109, 136, 151 Coleman, W. 49 collective(s) 227, 230–231, 235 Collingwood, R. G. vii, 18, 21–25, 30, 35, 138 colon endoscopy (colonoscopy) 237, 241–242, 246 comma bacillus viii, 48–50, 53–54, 57, 59–60, 65–66 common cause 112, 205–208, 236– 237, 241 confirmation viii, 81–82, 90, 108, 117, 121, 131–132, 134, 171, 189, 198, 211, 222, 251 confounder(s)/confounding viii, 113, 115–116, 121, 123–124, 126–127, 129, 131–132, 139, 148 constant conjunction 3, 4, 6, 19, 23, 138, 195 contagion(s) 45, 58–59, 171–173, 176, 181 controlled trials xvi, 91, 94 Corder, R. 215 Corfield, D. xiii coronary arteries 94 coronary heart disease (CHD) viii, xv–xvi, 92–94, 101–103, 105–107, 109–112, 115–117, 128–129, 131, 141–142, 196–197, 212–216 correlation(s) 60–62, 109, 111–112, 116–117, 124–125, 129–131, 137–139, 142, 148, 206–207, 236, 257 Craver, C. 73, 75–79, 82 cross-over study 120–122 cure(s) viii, x, 7, 21, 25, 29, 82, 90, 134, 150, 152–153, 159, 180, 182, 185–186 cystic fibrosis 83–84, 186

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:49.

296 Index

Copyright © 2018. Routledge. All rights reserved.

Daniels, M. 162 Darden, L. xiii, 73, 75–79, 82–84, 186 Davaine, C. 45–46 Dawid, P. xiii Dayton, S. ix, 124–126, 128 De Finetti, B. 229 De Renzy, A. 60–61 Debré , P. 20, 68 decoherence theory 78 deep brain stimulation (DBS) 7, 177, 183 deeper explanations x, 185, 188–189 degree of belief 13, 228–229 determinism 134–135 diabetes 122, 209–210, 215, 218, 224 diet(s) 103–104, 106–108, 110, 128, 131, 142, 196–197, 209–215, 217–218, 220 Dingler, H. 18, 29 diphtheria 50 disconfirmation viii, 90, 115, 117, 198 DNA 81, 187 do-calculus 18, 244–245 Doll, R. 136–137, 139, 141–142, 148 Dowe, P. 6, 71–72 ‘Eat Well and Stay Well’ project x, 211, 213, 218–219 EBM+ 162 ECGs 110–111 Eells, E. 201, 233–234 Ehrlich, P. 47 Einstein, A. 189 Eisenhower, D. 214 Emmerich, R. 66–67 endorphins 175 endothelin-1 215 endothelium 94, 97–98, 146–147 enkephalins 175 enzyme(s) 165–166, 187–188 epidemic(s) 48–49, 54–55, 57, 59, 61–62, 65–67 epidemiological study(ies)/survey(s) 94, 102, 129, 132, 136, 147, 196, 216, 222 Esterbauer, H. 144 Evans, A. 50, 52 Evans, R. 57, 59–60, 62, 67–68 evidence viii–ix, xv–xvi, 7, 24, 55–56, 67, 90–92, 94–98, 101–102, 108, 113, 119, 125, 127–136, 138–140, 148–149, 151– 152, 155, 158–159, 161–164, 166–168, 172–173, 175–176, 178–183, 190, 210, 216, 222, 225, 250; difference-making 167–168; interventional ix, 24, 55–56, 67, 92, 94, 102, 119, 125, 128–129, 138–140, 14 8–149; observational viii,

92, 94, 101, 108, 125, 128, 138–140, 148–149; of mechanism viii–ix, xv, 7, 90–92, 94–96, 98, 101, 127–130, 132–136, 149, 151–152, 155, 158–159, 161–164, 166, 168, 172–173, 178–183, 190, 216, 222, 250; statistical viii–ix, xv, 7, 24, 91–92, 101–102, 108, 113, 119, 125, 127–129, 133–136, 139–140, 148–149, 152, 158, 163–164, 166–168, 172–173, 175–176, 178–180, 216, 225, 250; types of viii, xv, 90, 94–95, 101, 130–132, 152, 158 evidence-based medicine (EBM) ix, 132, 136, 150–153, 157–158, 162; hard 151–152; soft 152 evidential pluralism 163–164 F2-isoprostanes 144 fast food x, 209–212, 214, 216–223, 225, 238, 247, 250–254 FDA 161 filtration 62–63, 65, 91–92, 212 Fisher, R. 123 Fleming, A. 21 Florey, H. 21 Florey, M.E. 155 foam cells 97, 100, 128, 142–146 forks x, 204–210, 223, 225, 236–239, 245–246, 255; 2 pronged 209–210, 223, 225; conjunctive x, 204–209, 236; interactive x, 204, 207–209, 237–239, 245–246; multi-causal x, 208–210, 223, 225, 246, 255; n pronged 208 Fortin, M. 122 Framingham study 209, 255, 257 frequencies, statistical 9, 11, 13, 227, 230 Freudenheim, J. 137 Gaff ky, G. 47, 50, 53, 66 Galavotti, M.C. xii, 7–8, 29, 163, 195, 197, 201, 232 Galileo 153, 189–191 gangrene 94 Gasking, D. vii, 18, 30–31, 34–35, 40–41 gene(s) 83–84, 115, 165, 186, 195 germ theory of disease 5, 43, 45–47, 58–59, 171–172, 176, 181, 195, 256 Gillies, Donald A. 18, 35, 67, 169, 171, 177, 203, 230, 235, 237, 241–242, 257 Gillies, Duncan F. 237, 241–242 Glennan, S. xiii, 73, 75, 77–78, 81 glucose 164–168 glycogen 165–166 Glymour, C. 239

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:49.

Index  297

Copyright © 2018. Routledge. All rights reserved.

gold standard of evidence 131–132, 152 Goldstein, J. 143 Gonzalez, W. xiii Good, I.J. 8–10, 12–13, 200–201, 207, 225, 257 Gradmann, C. 52–53, 66 Gray, J. 152 Green, T.H. 22 Guyatt, G. 151 Haavelmo, T. 237 Hacking, I. 226 Há jek, A. 230 Hamš i ková , E. 197–200, 203 Harden, V. 52 Hare, R. 48, 52 Harvey, W. 180, 182 Hegel, G. W. F. 22 Helicobacter pylori 48 Hennig, C. xiii, 235, 257, 267 Henriksen, T. 143 Hesslow counter-example x, 11–13, 202–203, 222–223, 225–226, 232, 234, 244, 247, 257 Hesslow, G. x, 11–13, 202–203, 222–223, 225–226, 232, 234, 244, 247, 257 high-density lipoprotein (HDL) 215 Hill, A.B. 123, 136–137, 139, 141, 153–154, 158, 162, 209 HIV 186–187 Hoffmann, F. 182 Homer 212–213 Howick, J. 150–151, 168–169, 172–173, 177, 181, 183–184 human learning 240 Hume, D. 2–4, 6, 23, 40, 111, 138 humours 178 Humphreys, P. 10–13, 18, 31, 201 Humphreys’ paradox 10–13, 18, 31, 201 hydrogen peroxide 144 hypercholesterolemia 110 Ignatowski, A. 96, 116 Illari, P. M. xii, 4, 77–78, 95, 130, 162 independence 206–207, 237, 241– 242, 253; conditional 206–207, 237, 241–242 intermittent claudication 94 intervention(s) vii, xi, 4–5, 17–18, 23–24, 28, 33–37, 40, 43, 55, 85, 92, 108, 129, 138, 150, 167–169, 173, 175, 244– 245, 259–261 intima 94, 97–98, 101, 146 ischaemic heart disease 141–142

Johnson-Tidey, R. 146 Jü rgens, G. 143 K Ration 103 Kalva, V. 146 Ka ň k a, J. 198 Kant, I. 2–5, 22, 29, 40 Kaptchuk, T. 174–175 Kelly, M. xii Kelsey, F. 161 Kepler, J. 189–191, 240 Keys, A. viii–ix, 95, 102–122, 124, 126, 128–129. 196, 209, 211, 213–214, 216, 218–220 Keys, M. 105, 211, 213–214, 216, 218–220 Kim, J. 204 Kistler, M. 42–43 Kitasato, S. 47 Koch, R. viii, xv, 2–3, 5, 43–58, 60–68, 79–80, 83, 90, 92, 96, 99, 135, 172, 186, 190, 195, 208 Koch’s postulates viii, 43–45, 48–55, 90, 135 Kolmogorov, A.N. 226, 230 Korb, K. 201, 234, 249, 255, 257 Kroc, R. 221 Kvasz, L. xiii, 67 Laker, M. 95, 108 Lauritzen, S. 12, 204 law(s) xv–xvi, 1, 3–5, 7, 19–21, 24–28, 30, 32–34, 36–38, 55, 74–75, 77, 83, 85, 121, 132, 149, 189, 199, 222, 250, 256; causal xv–xvi, 1, 3, 4–5, 7, 19–21, 24–28, 30, 32–34, 37–38, 55, 74–75, 83, 85, 121, 132, 149, 199, 222, 250, 256; functional 20, 27–28, 3 2–33, 37; probabilistic/statistical 20 Lehmann, J. 156 Lehr, H-A 144, 146 leukocytes 147 Levy, D. 209, 255 limitation(s) 121–122, 126–127, 132, 155, 159–160; sample 121–122, 126–127, 132, 160; time 121–122, 126–127, 132, 155, 159 linear regression 111–112 lipids 128, 142–143, 147 lipoids 97–98, 100 Lister, J. 44–47, 51 Loeffler, F. 47, 50, 52–53 Louis, P-C-A 135, 180 low-density lipoprotein (LDL) 143–145, 190, 215

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:49.

298 Index

Copyright © 2018. Routledge. All rights reserved.

lumen 97–98, 237, 241–242 lung cancer ix, xvi, 2, 10–11, 20, 33, 136–137, 139–142, 148–150, 196–197, 200, 208, 250, 254–255 lycopene 144 McArdle, B. ix, 164–168, 195, 256 McArdle disease ix, 164–168, 195, 256 Macfarlane, G. 21 Machamer, P. 73, 75–77, 79, 82 machine learning 239–240 macrophages 97–98, 100, 128, 142–145 McTaggart, J. M. G. 22 Malmros, H. 105–106 manipulation(s) vii, 4–5, 8, 17–18, 23, 28, 30, 33–35, 37, 260–261 Markov, A. A. 206 Markov condition x, 206–207, 209–210, 236–239, 241–243, 246, 251–252, 255, 262–263 Marshall, B. 48, 67 masking 130–131 Maxwell, J. C. 26–27 Mayerl, C. 101 MDC 82, 83–85, 89 mechanism(s) viii–ix, xii, xv, 6–8, 59, 69, 71, 73–85, 89, 91, 96–101, 108, 128–131, 133, 136, 139–142, 145, 147–148, 151, 155–156, 158–159, 161–162, 166–168, 173, 175–183, 185–186, 188–191; basic 78–80, 89, 91, 139, 181, 185; causal 98–99; causal theory of viii, 7, 71, 78–79, 81–83, 85; confirmed 140–141, 145, 147–148; definition(s) of viii, 7, 73–74, 76; linking ix, xii, 71, 73, 79–80, 82–83, 91, 96, 129, 131, 141, 148, 185–186, 189, 190–191; minimal 77–78; plausible 59, 140–142, 145, 147–148, 168, 175–176, 178–179, 181 mechanistic counter-example to RussoWilliamson thesis (RWT) 164, 168 mechanistic reasoning 151, 168 Medical Research Council (MRC) 153–158, 162 medicine iii, v, viii–ix, xii, xv, 1, 5, 8, 20–21, 23–24, 26–27, 32–33, 36–37, 41, 43–45, 55, 57, 67, 72–73, 75, 79–83, 85, 87–90, 92, 94–95, 121, 129, 133–136, 139, 148–150, 152–153, 159, 164, 172, 174– 175, 178, 180, 182, 185, 188–189, 191, 196–196, 208, 240, 243, 246, 250, 256, 260; alternative ix, 174; clinical 1; history of xv, 136, 153, 172, 178; scientific v, 23, 32, 57, 95, 135, 174–175, 180, 189; theoretical xv, 1, 5, 256

Mediterranean diet 213–215, 220 Menotti, A. 109, 118–119 Menzies, P. xiii, 18, 30–33, 35, 38–39 meridians 174–175 metaphysics/metaphysical 37, 174, 178 methodological falsifiability 228 Metschnikow, E. 96 miasma(s) 45, 58–59, 62–64, 170–173, 176, 181 Millar, J. 179 Mommaerts, W. 166–168 monocytes 143–147 monounsaturated fat 107, 121 Monti, M. 109 Moore, G. E. 22 Morrow, J. 144 mycobacterium tuberculosis 47 myocardial infarction 94, 105, 111–112, 118, 125, 146, 216 myocardium 94 myopathy 164, 166 Neapolitan, R. 12, 230 network(s) x, 12–13, 80, 84, 88–89, 203– 206, 209–210, 222, 229, 236–237, 239– 242, 245–246, 249, 252–253, 255–256, 259–260, 262–264; Bayesian 12, 203–204, 206, 229, 236–237, 239, 240–242, 245, 249, 252–253, 262–263; causal x, 12–13, 80, 88–89, 203–205, 209–210, 222, 229, 236, 240, 252, 255–256, 259, 262, 264; probability 89, 252, 262 Newton, I. 189–190 Nightingale, F. 58 node(s) 205–207, 262–263 Norell, C. xii ordinary language philosophy 86–88 osteomyelitis 20 osteoporosis 214 Otsuka, J. xiii, 235 oxidative stress 144, 147, 214 Oxman, A. 151 para-amino-salicylic acid (PAS) ix, 156–158, 176 Parkinson’s disease 7, 177, 183 Parkkinen, V.-P. xii Pasteur, L. 2–3, 20, 26, 43–46, 51, 56–57, 83, 186, 195, 208 pathological anatomy 92–94, 101 Pearl, J. x, xiii, 11–13, 18, 25, 138–139, 201, 203–204, 206–207, 210, 222, 229, 235–239, 241–247, 263 Pearson, C. 165, 167–168

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:49.

Copyright © 2018. Routledge. All rights reserved.

Index  299

peer review 177 Peirce, C. S. 227 peptic ulcers 48, 53 peripheral vascular disease 94 Peto, R, 137, 139, 141–142, 148 Pettenkofer, M. von 49, 57–64, 66–67, 90–91, 170–171, 181 phenylalamine 84 Phillips, B. 151 phosphorylase (myophosphorylase) 165–166 Plato 212–213 Pollan, M. 221, 223 polyunsaturated fat 107, 121 Popper, K. R. 9, 116, 190–191, 200, 227–228, 240 Poston, R. xiii, 146 pregnancy/pregnant 202, 223 Price, H. xiii, 18, 29–33, 35, 38–39 principle of interventional evidence 24, 55, 129, 140 principle of the common cause 207 probabilistic causality x, xiii, 195–196, 200–201, 203, 207, 225, 233–235, 243 probability iii, x, xiii, xvi, 3, 8–13, 18–19, 31, 43, 89, 133–134, 167, 193, 196–197, 200–207, 209, 222–223, 225–237, 242–243, 256–257; epistemic 226, 228; frequency theory of 226–227, 230–231; interpretation(s) of x, 9, 226–227, 230; objective 9, 11, 13, 202, 226, 228–230, 237, 242, 256; subjective 13, 228–229 probability raising principle 10–11, 200–201, 244 procyanidin 215 propensity(ies) 9–10, 13, 200–202, 226–228, 230–232, 256–257 prospective study(ies) 108, 136, 198 protein(s) 77, 84–85, 116–117, 187–188 protein folding/misfolding 84 protein synthesis 77 Psillos, S. 4 Puddu, V. 109 puerperal fever 169–172 puerperium 170 Pythagorean school 213 qi 174–175 quantity 71–72, 74, 76; conserved 71–72, 74, 76 invariant 72 Ramsey, F. 17–18, 229 Ramsey-De Finetti theorem 229 randomization 122–123, 127, 132

randomized controlled trials (RCTs) 91, 120, 122–124, 126, 128, 131–132, 139–140, 148–150, 152–155, 157–161, 173–176, 215 rare disease(s) 166–168 reactive oxygen species (ROS) 144–145, 147 reduction/reductionism 39–40, 73–75, 78 reference class 202, 231–233, 235, 244, 256 Reichenbach, H. 8, 71, 204–209, 236–237, 241 reinforced concrete 131 Reiss, J. 163–164 repeatable conditions 230–232, 256 research funding 176–177 resistance ratio (R.R.) 155, 157 retinylstearate 144 Robertson, T. L. 116 Rokitansky, Carl von 94, 98 Rosenberg, W. 152 Russell, B. vii, 18–19, 21–23, 32, 36–37, 87–88, 188–189 Russo, F. ix, xii, 4, 7, 24, 95, 111, 121, 132–136, 141–142, 148–150, 152, 157–159, 161–164, 167–168, 172–174, 176–184, 201, 216, 222, 250 Russo-Williamson thesis (RWT) ix, xii, 7, 132–136, 141–142, 148–150, 152, 157–159, 161, 163–164, 167–168, 172–174, 176–184, 216, 222, 250 Ryle, G. 86 Sackett, D. 151–152 Salmon, W. 5–6, 10, 71–73, 75–76, 81, 201–202, 207–209, 236–239, 245 Sanders, T. 217 saturated fat viii, 95, 107–108, 111–112, 115, 117–119, 121, 124, 126, 128, 131, 142, 196–197, 212–214, 216, 218–219, 247 Savage, L. J. 9 Schatz, A. 153 Scheines, R. 239 Schlosser, E. 211, 217, 220 Schrö d inger, E. 27 screen(s) off 206–208, 237, 262 Semmelweis, I. ix, 168–177, 181, 184, 222 seven countries study viii, 102, 108, 111, 113, 116, 118, 125–129, 214, 216, 219 Simboli, B. xiii, 29 Simpson’s paradox x, xvi, 13, 196–197, 232–233, 235 simultaneous causation 31, 40–43 slow sand filters/filtration 62, 64–65, 67

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:49.

300 Index

Copyright © 2018. Routledge. All rights reserved.

smoker/smoking ix, xvi, 2, 7, 10–11, 33, 110, 136–137, 139–142, 144–145, 147–150, 164, 190, 196–197, 200, 208–210, 214–215, 222–223, 225–226, 235, 238, 247, 250–254 Snow, J. 60, 64, 168–169, 177, 181, 184 Solano, J. xiii Solomon, M. 153 Spiegelhalter, D. 12, 204 Spirtes, P. 239 spores 46, 80, 186, 190 starvation 103–104, 106 statin(s) 218 statistical counter-example to RussoWilliamson thesis (RWT) 164, 168, 177, 183–184 Steel, D. 130 Steinberg, D. 95, 104, 143, 145–146 Steinbrecher, U. 143 Stephens, T. 159–162 Strawson, P. 87 strength through combining ix, xvi, 101, 126, 129–130, 132–133, 216, 250 streptomycin ix, 150, 153–159, 173, 176 stroke 93, 125 structural theory of causation 203 sublata causa vi, xiii, 25, 29, 43, 66, 85 Sucar, E. 237, 241–242 Sudbury, A. x–xi, xiii, xvi, 12, 204, 252–255, 257, 264–266 Sudbury’s theorem(s) x, xi, xvi, 12, 204, 252–255, 257, 264–266 super sizing 217, 221 superoxide 144 Suppes, P. 201, 237 T4 lymphocyte 186–187 Taussig, H. 160 Taylor, H. 109 Teira, D. xiii, 123, 224 Thagard, P. 67 thalidomide ix, 150, 158–162, 182 thrombosis 93–94, 146, 202, 223, 232 Trö h ler, U. 179–180 trypan blue 98, 101, 161 tubercle bacillus 47–48, 153, 155–156, 158, 176, 208 tuberculosis viii, 2, 5, 44, 47–48, 51, 53, 92, 99, 153–155, 158, 188, 208 tuberculostatic agent 156 Twardy, C. 201, 234

Tycho Brahe 240 typhoid 50, 59, 65, 67, 91 vaccine/vaccination 8, 25, 55–57, 83, 186, 199, 200 Vandenbroucke, J. xii variable(s) 12, 18, 20, 27, 34, 89, 111, 205, 225, 234, 238–239, 242–243, 245–249, 251–252, 256, 259, 260, 264–265; latent 238–239, 242–243, 245–249, 251; random 89, 205, 225, 248, 252, 256, 265 variation 111 Venn, J. 81, 226 Vesalius, A. 180 vibrio cholerae xv, 48, 177, 212 vicious circle 7, 78 viral causes of cancers 39, 203 Virchow, R. 46, 94, 98 virus(es) 2, 8, 26, 52, 88, 181, 186–188, 197–200; Epstein-Barr 197–198; herpes 197–200; papilloma 8, 88, 197, 199–200 Vist, G. 151 vitamin C 144–145, 147 vitamin E 144–145, 147 von Mises, R. 226–227, 230–231, 235 Vonka, V. vi, xiii, 26, 43, 52, 66, 197–200, 203 Vonka’s Thomist Maxim vi, 26, 43, 66, 199 Waksman, A. 153 Wallerstein, D. 221 Wallmann, C. xii, 68 Warren, R. 48 Whitehead, A.N. 22 Wilde, M. xii Williamson, J. ix, xii, xiii, 7, 8, 29, 77–78, 95, 132–136, 141–142, 148–150, 152, 157–159, 161–164, 167–168, 172–174, 176–184, 207, 216, 222, 241–242, 250–251 Windaus, A. 94 Wittgenstein, L. 21, 86, 88 Woodward, J. xi, 18, 31, 34–40, 42–43, 259–261 wound infections/suppuration 45, 47, 50 Wright, G. H. von 18, 35 X-rays 182 xanthoma cells 98 zur Hausen, H. 8, 199

Gillies, Donald. Causality, Probability, and Medicine, Routledge, 2018. ProQuest Ebook Central, http://ebookcentral.proquest.com/lib/sfu-ebooks/detail.action?docID=5582360. Created from sfu-ebooks on 2019-07-21 18:54:49.