Fenton-Glynn

This Element provides an accessible introduction to the

contemporary philosophy of causation. It introduces the
reader to central concepts and distinctions (type vs token
causation, probabilistic vs deterministic causation, difference-
making, interventions, overdetermination, pre-emption) and
to key tools (structural equations, graphs, probabilistic causal Philosophy of Science
models) drawn upon in the contemporary debate. The aim is
to fuel the reader’s interest in causation, and to equip them
with the resources to contribute to the debate themselves.
The discussion is historically informed and outward-looking.
‘Historically informed’ in that concise accounts of key historical
contributions to the understanding of causation set the stage

Causation

Causation
for an examination of the latest research. ‘Outward-looking’
in that illustrations are provided of how the philosophy of
causation relates to issues in the sciences, law, and elsewhere.
The aim is to show why the study of causation is of critical
importance, besides being fascinating in its own right.

About the Series Series Editors

This series of Elements in Philosophy of Jacob Stegenga
Science provides an extensive overview University of

Luke Fenton-Glynn
of the themes, topics and debates which Cambridge
constitute the philosophy of science.
Distinguished specialists provide an
up-to-date summary of the results of
current research on their topics, as well
as offering their own take on those topics
and drawing original conclusions.

Cover image: BSIP / Contributor/

Universal Images Group/Getty Images ISSN 2517-7273 (online)
https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press ISSN 2517-7265 (print)
https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press
 Elements in the Philosophy of Science
edited by
Jacob Stegenga
University of Cambridge

CAUSATION

Luke Fenton-Glynn
University College London

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 University Printing House, Cambridge CB2 8BS, United Kingdom
One Liberty Plaza, 20th Floor, New York, NY 10006, USA
477 Williamstown Road, Port Melbourne, VIC 3207, Australia
314–321, 3rd Floor, Plot 3, Splendor Forum, Jasola District Centre,
New Delhi – 110025, India
103 Penang Road, #05–06/07, Visioncrest Commercial, Singapore 238467

Cambridge University Press is part of the University of Cambridge.

It furthers the University’s mission by disseminating knowledge in the pursuit of
education, learning, and research at the highest international levels of excellence.

www.cambridge.org
Information on this title: www.cambridge.org/9781108706636
DOI: 10.1017/9781108588300
© Luke Fenton-Glynn 2021
This publication is in copyright. Subject to statutory exception
and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.
First published 2021
A catalogue record for this publication is available from the British Library.
ISBN 978-1-108-70663-6 Paperback
ISSN 2517-7273 (online)
ISSN 2517-7265 (print)
Cambridge University Press has no responsibility for the persistence or accuracy of
URLs for external or third-party internet websites referred to in this publication
and does not guarantee that any content on such websites is, or will remain,
accurate or appropriate.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation

Elements in the Philosophy of Science

DOI: 10.1017/9781108588300
First published online: June 2021

Luke Fenton-Glynn
University College London
Author for correspondence: Luke Fenton-Glynn, [email protected]

Abstract: This Element provides an accessible introduction to the

contemporary philosophy of causation. It introduces the reader to central
concepts and distinctions (type vs token causation, probabilistic vs
deterministic causation, difference-making, interventions,
overdetermination, pre-emption) and to key tools (structural equations,
graphs, probabilistic causal models) drawn upon in the contemporary
debate. The aim is to fuel the reader’s interest in causation, and to equip
them with the resources to contribute to the debate themselves. The
discussion is historically informed and outward-looking. ‘Historically
informed’ in that concise accounts of key historical contributions to the
understanding of causation set the stage for an examination of the latest
research. ‘Outward-looking’ in that illustrations are provided of how the
philosophy of causation relates to issues in the sciences, law, and
elsewhere. The aim is to show why the study of causation is of critical
importance, besides being fascinating in its own right.

Keywords: causation, counterfactual theories of causation, probabilistic

causation, regularity theories of causation, causal modelling

© Luke Fenton-Glynn 2021

ISBNs: 9781108706636 (PB), 9781108588300 (OC)
ISSNs: 2517-7273 (online), 2517-7265 (print)

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Contents

1 Introduction 1

2 Regularity Theories of Causation 3

3 Interlude: Some Important Background 13

4 Counterfactual Theories of Causation 29

5 Probabilistic Causation 55

6 Conclusion 82

References 85

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 1

1 Introduction
We humans take a great interest in causation. Causal knowledge helps us to
understand, predict, and influence the world around us. A baby quickly comes
to realise that pushing a button on her toy causes it to play a song; an adult
exploits her knowledge of the effects of chamomile to sooth the baby’s teething
gums. Because of the close relation between causation, understanding, pre-
diction, and control, natural and social scientists devote significant time and
resources to investigating causal questions. Among other things, they ask or
have asked: What are the causes of cancer? Of climate change? Of anoma-
lies in the orbit of Uranus? What caused the extinction of the dinosaurs? The
First World War? The 2007-8 financial crisis? Trump’s election? The Covid-19
outbreak? What causes mental health problems? Crime? Price inflation?
Causation is of importance to psychology: if we want to understand how
humans learn and reason, we need to understand their capacity for causal learn-
ing and reasoning. It’s of importance in AI: if we want computers and robots
to learn as well as (or better than!) humans, and to manipulate the world as (or
more!) effectively, we need to programme them to be able to acquire causal
knowledge and to use it.
Causation is also closely tied to questions of moral and legal responsibility.
In a landmark UK legal case, the wife of the late Arthur Fairchild successfully
sued Glenhaven Funeral Services1 over her husband’s death from mesotheli-
oma – a type of lung cancer caused by asbestos exposure. In order to establish
the company’s responsibility for Fairchild’s fatal illness, it was of course nec-
essary to establish that there was a causal link between something they’d
done – namely negligently expose Fairchild to elevated levels of asbestos –
and the illness itself.
Despite its ubiquity and importance, it’s surprisingly difficult to say exactly
what causation is. Difficult questions about the fundamental nature of the
world – especially those that don’t readily admit of empirical resolution – nat-
urally attract the attention of philosophers. But causation isn’t only of intrinsic
philosophical interest. Greater theoretical clarity on its nature has had signifi-
cant payoffs in the sciences and in law. And, close to home for philosophers, it
has payoffs in virtue of the fact that causation plays a role in key theories of a
variety of philosophically interesting phenomena including (but not limited to)
reference, perception, decision, knowledge, inference, action, and explanation.
It shouldn’t be thought that work on the theory of causation is the exclusive
preserve of philosophers. Much important theoretical work has been done by

1 Fairchild v Glenhaven Funeral Services Ltd ([2002] UKHL 22; [2003] 1 AC 32).

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 2 Philosophy of Science

computer scientists, economists, statisticians, legal scholars, psychologists and

others – reflecting the broad, interdisciplinary importance of a better under-
standing of causation. Fortunately, these days, there’s significant interaction
between theorists in these various disciplines, which has enriched our collective
understanding of causation. A prime example of the payoffs of this cross-
disciplinary interaction is the theory of causal modelling that we’ll examine
as part of Sections 4 and 5.
This Element examines some of the progress that’s been made in understand-
ing the nature of causation as well as some of the unresolved challenges. Since
the literature on causation is large, this Element is, of necessity, a selective
introduction. It focuses on three broad traditions within the theory of causa-
tion: the regularity, counterfactual, and probabilistic approaches. Perhaps the
most contentious omissions are the process approach – which seeks to analyse
causation in terms of causal processes with the latter understood, on the most
promising such account, as the world-lines of objects that possess conserved
quantities (Dowe 1992, 2000; Salmon 1994, 1997) – and the New Mechanist
approach – exemplified by Machamer et al. (2000) and Glennan (2017), among
many others. I’ve made this choice, not because I don’t think that understand-
ing causal processes and mechanisms is of vital importance (I do!), but because
I don’t think that these are causal bedrock: I think there are relations of causa-
tion that are more fundamental than the notion of a causal process or mechanism
and that an adequate understanding of processes and mechanisms will require
an adequate understanding of these more fundamental causal relations.2 Regu-
larity, counterfactual, and probabilistic approaches are attempts to understand
these fundamental causal relations.
This view is contentious as process theorists and some New Mechanists
think that the fundamental causal relation(s) can be understood in terms of
processes or mechanisms. For example, Dowe (2000, 90) seeks to define
causal interactions in terms of processes, while Glennan (1996) suggests that
causation – at least outside the domain of microphysics – might be analysed
in terms of mechanisms. However, process theories have been plagued by the
problem of distinguishing genuine causal processes from world-lines that don’t
correspond to processes without falling back on an appeal to some more basic
causal relation, perhaps understood in terms of counterfactual dependence (see,
e.g., Hitchcock 1995, 2009; Choi 2002).
As regards mechanistic approaches, there’s certainly no consensus among
New Mechanists that the notion of mechanism is prior to that of causation. As

2 The notion that there might be more than one fundamental causal relation is taken up in Section
3.4.2.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 3

Craver and Tabery (2019) note, ‘Mechanists have disagreed with one another
about how to understand the cause in causal mechanism. … Four ways of
unpacking [it] have been discussed: conserved quantity accounts, mechanistic
accounts, activities accounts, and counterfactual accounts’.
I’ve already said that it’s doubtful that the conserved quantity approach
can yield an understanding of the fundamental causal relation(s). The activ-
ities approach, on the other hand, is a primitivist approach (see Craver and
Tabery 2019), whereas we’ll be examining accounts that seek a deeper under-
standing of causation. Meanwhile, the counterfactual approach is one that
we’ll be exploring in Section 4. Finally, the mechanistic account – as advo-
cated by Glennan (1996) – is regressive. The proposal is that causal con-
nections that may seem basic at (say) the biological level can be understood
in terms of mechanisms at the chemical level, and those at the chemical
level in terms of mechanisms at the physical level. There’s thus a hierarchy
of mechanisms. The concern, though, is that this hierarchy bottoms out at
the level of fundamental physics at which level we have causings that can’t
be mechanistically understood. Again, this favours the view that there are
fundamental causal relations in terms of which mechanisms can ultimately
be understood. The regularity, counterfactual, and probabilistic approaches
seem the most promising approaches to understanding these basic causal
relations.

2 Regularity Theories of Causation

2.1 Hume
Though Western theorising about causation dates back at least to Aristotle
(Physics 195 a 4–14; Metaphysics V.2), David Hume (1739, 1748) is rightly
considered the father of the modern tradition of attempts to understand that
relation. Hume is standardly interpreted as advocating a regularity theory of
causation.
Specifically, according to Hume (1739 I.iii.2), causes occur temporally prior
to their effects, and are either contiguous with them in space and time or else
connected to them by a contiguous ‘chain’ of causation. For example, a person
may break a window by throwing a rock at it despite the fact that the throw
occurs a short interval of time before and a few metres away from the breaking
because, once the rock is thrown, it traces a continuous trajectory until it hits
and shatters the window. The window’s breaking is caused by the throw via
this ‘chain’ (there’s no ‘action at a distance’), with the position and momentum
of the rock at each stage on its trajectory being caused by its previous positions
and momenta and by the throw itself, and with the window’s shattering being

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 4 Philosophy of Science

caused by the prior states of the rock all the way back until we get to the throw
itself.
But we don’t have a case of causation just any time an event occurs prior
to and contiguously with another. Towards the end of the movie Saving Pri-
vate Ryan, in a defiant last stand, Captain Miller repeatedly fires his pistol at a
German Tiger tank. The bullets are of course completely incapable of piercing
the tank’s armour. Down to his last bullet, Miller points his gun and shoots at
the tank at the very moment the tank is blown to pieces by a bomb dropped by
a US P-51 aircraft. The impact of Miller’s bullet is immediately prior to, and
contiguous with, the explosion of the tank. Yet it’s the bomb and not the bullet
that causes the tank to explode.
Fortunately, Hume’s account doesn’t imply that the bullet impact was a
cause. That’s because, in addition to priority and contiguity, Hume adds a
third requirement: constant conjunction. For Hume, for an event c to be a
cause of an event e, it must be the case that events like c are always fol-
lowed contiguously by events like e.3 This criterion excludes the impact of
Miller’s bullet from counting as a cause of the tank’s explosion. That’s because
events like the former aren’t always followed by events like the latter. Indeed,
Miller had already fired his gun at the tank five times prior to firing his last
bullet: the impact of none of these previous five bullets was followed contig-
uously by an explosion. So Hume’s analysis yields the correct verdict about
this case.
Although Hume doesn’t say this, it’s tempting to think that not any old con-
stant conjunction can ground a causal relation, but rather one might wish to
require that the constant conjunction be entailed by the laws of nature. This
avoids problems such as the following. Suppose there exists an extremely
rare isotope, call it ‘unobtanium-352’. Only one atom of this isotope ever
exists. Suppose this atom happened to decay on the afternoon of Novem-
ber 19, 1863, immediately before Lincoln delivered the Gettysburg Address
and contiguously with it. Now, for any type T of event of which the Gettys-
burg Address is a member (‘famous speeches’, say), it’s true that all cases
of unobtanium-352 decay are followed by events of type T. Nevertheless, it
clearly doesn’t follow that the decay of the unobtanium-352 atom was a cause
of the Gettysburg Address.4 A sophisticated regularity theory can avoid this

3 We could, on Hume’s behalf, distinguish direct from indirect causation, with c counting as a
direct cause of e iff c is prior to, and contiguous with, e and events like c are always followed
contiguously by events like e. Indirect causation would then be understood in terms of chains
of (i.e. ordered sequences of events that stand in relations of) direct causation.
4 Note that even if we take ‘constant conjunction’ to require a multiple instances, we’re still
liable to get ‘accidental’ constant conjunctions that aren’t apt to underwrite causal relations
(see Armstrong 1983, 15–17).

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 5

conclusion by pointing out that the fact that all instances of unobtanium-352
decay are followed by events of type T isn’t entailed by the laws of nature
(rather, it’s an instance of an accidental regularity).5

2.2 Mill
An apparent problem with Hume’s account is that sometimes we seem to have
causation without constant conjunction. John Stuart Mill pointed this out in
making the following observation:

It is seldom, if ever, between a consequent and a single antecedent that …

invariable sequence subsists. It is usually between a consequent and the sum
of several antecedents … . In such cases it is very common to single out only
one of the antecedents under the denomination of Cause, calling the oth-
ers merely Conditions. Thus, if a person eats of a particular dish, and dies
in consequence … people would be apt to say that eating of that dish was
the cause of death. There needs not, however, be any invariable connexion
between eating of the dish and death; but there certainly is, among the cir-
cumstances which took place, some combination or other on which death is
invariably consequent: as, for instance, the act of eating of the dish, com-
bined with a particular bodily constitution … . The real Cause, is the whole
of these antecedents … . (Mill 1843, III.v.3)

Mill’s idea, then, is that constant conjunction (‘invariable sequence’) rarely

obtains between a single earlier event (‘antecedent’) and a single later event
(‘consequent’). Taking Mill’s example, it could easily happen that someone
dies from eating a particular dish (because, say, it contains an allergen) but
that others who eat it survive. Mill thinks that those who die have features
that distinguish them from those who don’t. For instance, they may have a
severe allergy to an allergen in the dish. In this case, the constant conjunction

5 A couple of comments are worth making regarding this proposed appeal to laws of nature. First,
whether one regards it as marking a departure from a pure regularity theory of causation will
depend upon one’s preferred metaphysics of laws. It will not be a departure if one adopts a regu-
larity theory of laws. Whilst sophisticated regularity theories of laws – such as the Best System
Analysis (Lewis 1994) – regard laws as regularities, they don’t count just any old regularity as a
law of nature. Thus, for instance, it’s to be hoped that they wouldn’t count an unobtainium-325
decay/famous speech regularity as a law.
Second, for the appeal to laws of nature to be satisfactory, it will presumably be necessary
that a wide range of regularities outside the domain of fundamental physics (e.g. ‘aspirin con-
sumption is followed by pain relief’) are entailed by laws of nature. That’s because it’s clear that
our causal claims extend to such domains. There’s an extensive philosophical literature on the
status of generalisations outside of fundamental physics. For overviews of important aspects of
this literature, see Cat (2017) and Reutlinger et al. (2019). Thanks to an anonymous referee for
encouraging me to say something about both of the foregoing points.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 6 Philosophy of Science

is between eating the dish and (e.g.) having a severe peanut allergy (as Mill
puts it: having ‘a particular bodily constitution’) and death.
Mill thinks that, properly speaking, in such a case it’s the combination of
the severe peanut allergy with the eating of the dish that’s the cause of death.
However he observes elsewhere (Mill 1843, III.v.3) that, in ordinary talk, we
often single out just one of the factors in such a combination as the cause
and regard the others as mere ‘conditions’. Specifically, he claims that we’re
inclined to pick out ‘events’ or ‘changes’ like the eating of the dish as causes
and treat long-standing ‘states’ like the possessing of the peanut allergy as mere
‘conditions’.

2.3 Hart and Honoré on the Cause/Condition Distinction

As noted in the Introduction, the study of causation is a truly interdisciplinary
endeavour and the profound contributions of the legal scholars H. L. A. Hart
and Tony Honoré to our understanding of it is the perfect illustration of this. One
of their contributions was to further investigate the relationship between those
factors that we pick out as ‘causes’ and those that we label mere ‘conditions’.
According to them, this distinction is one ‘to which common sense adheres in
face of the demonstration that cause and conditions are “equally necessary”
if the effect is to follow’ (Hart and Honoré 1959, 33). In giving an account
of the grounds on which we make such distinctions, they emphasise the close
connection between causation and explanation.
Hart and Honoré (1959, 35) point out that the sorts of effect that tend to
pique our interest – and lead us to seek causes to explain – are abnormal
events. In Mill’s example, this was the sudden death of a person. Previ-
ous examples that we’ve used include the shattering of a window and the
explosion of a tank. Hart and Honoré consider the example of the out-
break of a fire. They point out that, in order to explain an abnormal event,
some abnormal factor typically needs to be cited. For example, in order to
explain why a building was destroyed by fire, emphasising the presence of
oxygen or flammable material wouldn’t typically be that helpful, whereas
pointing out that a lit cigarette was dropped would be much more so. The
reason is that oxygen and flammable material were always present, but what
we need for an explanation is to know what made the difference between
this occasion, on which the building burned down, and all the previous
times when it didn’t (Hart and Honoré 1959, 35). It’s the abnormal factor,
the dropping of the cigarette, that’s the difference-maker and hence we’re
liable to call it the ‘cause’ while treating the others as ‘mere conditions’
(Hart and Honoré 1959, 35).

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 7

Normal factors will often be what Mill described as ‘states’ – ongoing or

permanent factors – while abnormal factors will often be ‘changes’ or ‘events’.
For instance, the presence of oxygen and flammable material are typically rel-
atively permanent states, while the dropping of a cigarette is a ‘change’. But
Hart and Honoré show that things aren’t that simple. They give the following
example:

If a fire breaks out in a laboratory or in a factory, where special precautions

are taken to exclude oxygen during part of an experiment or manufacturing
process, since the success of this depends on safety from fire, there would
be no absurdity at all in such a case in saying that the presence of oxygen
was the cause of the fire. The exclusion of oxygen in such a case, and not its
presence, is part of the normal functioning of the laboratory or factory … .
(Hart and Honoré 1959, 35)

Suppose that the manufacturing process involves the frequent production of

sparks. Then, it seems, we might pick out the presence of oxygen as a cause
even if the leak occurred some time before a spark was produced and the fire
started. If that’s right, then we would appear to have a case where the relatively
longstanding state (the presence of oxygen) is picked out as the cause, while
the change (the spark) might be treated as a mere condition. This would suggest
that the distinction we draw between causes and conditions tracks the abnor-
mal/normal distinction rather than the event/state distinction where these two
come apart.

2.4 Mackie
As we saw, although Mill acknowledges that in ordinary talk we distinguish
between causes and conditions, he thinks that the ‘real Cause’ is the combina-
tion of all the factors needed to bring about the effect. We also saw that Hart
and Honoré seek to account for why we distinguish between causes and condi-
tions in terms of our explanatory interests, noting that these various factors may
be equally necessary for the effect. Especially if we think that the event/state
distinction does not always track the cause/condition distinction (as Hart and
Honoré’s example suggests), we might conclude that the cause/condition dis-
tinction isn’t really a metaphysical distinction at all but rather one that’s to be
accounted for by a suitable pragmatics of causal talk. It’s tempting to impute
this view to David Lewis when he says:

We sometimes single out one among all the causes of some event and call it
‘the’ cause, as if there were no others. Or we single out a few as the ‘causes,’
calling the rest mere ‘causal factors’ or ‘causal conditions.’ … We may select
the abnormal or extraordinary causes, or those under human control, or those

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 8 Philosophy of Science

we deem good or bad, or just those we want to talk about. I have nothing
to say about these principles of invidious discrimination. I am concerned
with the prior question of what it is to be one of the causes (unselectively
speaking). (Lewis 1973a, 558–9)

We’ll return to Lewis’s account of causation in Section 4. For now it’s worth
noting that, even if we agree that there’s no deep metaphysical distinction
between causes and conditions (or indeed between the factor that we might
on some occasion be inclined to pick out as ‘the’ cause of some effect, and the
rest of that effect’s causes), we needn’t follow Mill in taking the conjunction
of all the factors requisite to produce an effect as the real cause. An alternative
approach – as suggested in the passage from Lewis – is that we allow that each
of the factors in question counts as a cause, so that a given effect may have a
plurality of causes.
The latter is the approach of Mackie, whose well-known account of causation
perhaps marked the zenith of the regularity approach.6 Mackie’s account can
be introduced by means of his own example:

Suppose that a fire has broken out in a certain house …. Experts … con-
clude that it was caused by an electrical short-circuit at a certain place. …
Clearly the experts are not saying that the short-circuit was a necessary con-
dition for this house’s catching fire at this time; they know perfectly well
that … the overturning of a lighted oil stove, or any one of a number of
other things might, if it had occurred, have set the house on fire. Equally,
they are not saying that the short-circuit was a sufficient condition for this
house’s catching fire; for if the short-circuit had occurred, but there had been
no inflammable material nearby, the fire would not have broken out …. In
what sense, then, is it said to have caused the fire? At least part of the answer
is that there is a set of conditions … including the presence of inflammable
material, the absence of a suitably placed sprinkler … which combined with
the short-circuit constituted a complex condition that was sufficient for the
house’s catching fire – sufficient, but not necessary, for the fire could have
started in other ways. Also, of this complex condition, the short-circuit was
an indispensable part: the other parts of this condition, conjoined with one
another in the absence of the short-circuit, would not have produced the fire.
… In this case, then, the so-called cause is … an insufficient but necessary
part of a condition which is itself unnecessary but sufficient for the result. …
[L]et us call such a condition (from the initial letters of the words italicized
above), an inus condition. (Mackie 1965, 245)

6 This isn’t to denigrate subsequent accounts within the regularity tradition, including the excel-
lent contributions of Strevens (2007) and Baumgartner (2013). But these can be regarded as
attempts to revive the tradition after a prolonged period in which it has been out of favour. It’s
fair to say that it currently remains a minority approach.
https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press
 Causation 9

Clearly Mackie’s account makes heavy use of the notions of necessary and
sufficient conditions. The reason it’s usually classed as a regularity theory is
that necessary and sufficient conditions can themselves be understood in terms
of regularities.7 For instance, we might say that the set of conditions comprising
the short-circuit, the presence of inflammable material, the presence of oxygen,
the absence of a sprinkler, etc. is sufficient for the fire iff whenever such a
constellation of factors co-occurs, a fire always ensues. Likewise, we might
say that the short-circuit is a necessary (or non-redundant) element of this set
of conditions iff it’s not the case that whenever the remaining conditions in the
set co-occur then fire ensues.8
With the notions of ‘necessity’ and ‘sufficiency’ so interpreted, we can view
Hume as taking causes to be sufficient for their effects (and possibly necessary
too, if ‘constant conjunction’ is taken to cut both ways, so that not only does a
cause never occur without its associated effect, but an effect never occurs with-
out its associated cause), while Mill agrees that real causes are sufficient for
their effects but takes them to typically be complex (e.g. the complex condition
comprising the short-circuit, the presence of oxygen, the presence of inflam-
mable material, and the absence of a sprinkler system, etc.). Mackie, on the
other hand, allows non-redundant elements of such sets (e.g. the short-circuit)
to count as causes.

2.5 Problems with the Regularity Theory

Perhaps the most important problems that have been raised against the reg-
ularity approach – and which have contributed to its decline in popularity in
recent years – are (i) the problem of the direction of causation; (ii) the problem
of probabilistic causes; (iii) the problem of common causes. We’ll see in later
sections that other theories of causation aren’t immune to these problems, but
they’re often thought to be particularly intractable for the regularity approach.

2.5.1 The Direction of Causation

Causation is almost always supposed to be an asymmetric relation: if a causes

b, then b doesn’t also cause a. One might object that feedback loops provide a
counterexample: depression causes someone to drink, but alcohol itself acts as
a depressant. The rejoinder is that, while a person’s depressed state at time t1

7 Though, drawing upon the point made in Section 2.1, we may wish to understand them in terms
of lawfully entailed regularities.
8 Actually, Mackie himself proposes that necessity and sufficiency only indirectly be understood
in terms of regularities: specifically, he proposes to interpret necessity and sufficiency in terms
of counterfactuals, with the counterfactuals understood in terms of regularities (Mackie 1965,
253–5). So one might take Mackie’s account to be a sort of hybrid regularity/counterfactual
theory. Counterfactual theories will be examined in Section 4.
https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press
 10 Philosophy of Science

might cause her to drink at time t2 , which in turn causes her depressed state at
time t3 , what we’d never have is a person’s depressed state at t1 causing her to
drink at t2 with her drinking at t2 causing her depression at t1 . The point is that,
where a and b are particular events or states that obtain at specific times, it can’t
be the case both that a causes b and that b causes a. This point is sometimes
put by saying that token causation (that is, the causal relation between token –
i.e. particular, dated – events or states) is asymmetric. We’ll have more to say
about token causation – and its distinction from what’s sometimes known as
type causation – in Section 3.3.
One might wonder whether it’s really entirely impossible that a should be
a token cause of b and b a token cause of a. The General Theory of Relativ-
ity (GTR) allows for the possibility of what are known as ‘closed time-like
curves’ – paths in space-time of positive distance that lead from a given space-
time point p back to p that could be traversed without ever travelling at or above
the speed of light. Suppose object o traverses such a path from p to p and goes
through some point q on its way. Then the state of o at p (its position and veloc-
ity) is presumably a cause of its state at q and its state at q a cause of its state
at p. But, whilst GTR allows that this is a possibility, we don’t have reason to
think that closed time-like curves in fact exist in our universe.
The trouble for regularity theorists is that they face a challenge in accounting
for why we don’t have bi-directional causation even in quoditian cases. That’s
because necessity and sufficiency are two sides of the same coin: if a is suffi-
cient for b then b is necessary for a. Putting it in regularity-theoretic terms, if
events like a are always accompanied by events like b, then there are no a-like
events without b-like events. This means that a challenge arises for the accounts
of Hume and Mill when we have causes that are both necessary and sufficient
for their effects. It also turns out that often when a is an inus condition of b,
then b is also an inus condition of a.
This difficulty isn’t necessarily fatal to the regularity theory. What it shows
is that some extra element needs to be added to the analysis to distinguish cause
from effect. Hume takes this extra element to be time: he requires that the cause
be earlier in time than the effect. One potential objection to Hume’s approach
is that it’s not clear how to reconcile it with the possibility of closed causal
loops of the sort seemingly allowed by GTR. Mackie was reluctant to rule
out backwards-in-time causation, and so instead took the cause to be the event
that becomes fixed first (Mackie 1980, ch. 7). While, ordinarily, earlier events
become fixed before later ones (because becoming a past event is the usual
way in which an event becomes fixed), Mackie thought that in special circum-
stances (of the sort that would allow for retro-causation) a later event might
become fixed prior to an earlier one. Yet it’s not clear that Mackie’s proposal

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 11

allows the sort of bi-directional causation that appears to be compatible with

GTR, since it’s at best unclear in what sense, in such a case, each of the causal
relata could become fixed prior to the other.
But, even if the problem of distinguishing causes from effects can be
overcome, there are other troubling objections to the regularity theory.

2.5.2 Probabilistic Causes

Consider the following example, given by Elizabeth Anscombe (1971, 24)9 :

Probabilistic Bomb

“[A] bomb is connected with a Geiger counter, so that it will go off if the
Geiger counter registers a certain reading; whether it will or not is not
determined, for it is so placed near some radioactive material that it may
or may not register that reading. … [T]here would be no doubt of the
cause of the reading or of the explosion if the bomb did go off.”

According to quantum mechanics (at least on orthodox interpretations), the

decay of radioactive isotopes is an irreducibly probabilistic process. For
instance, there’s no feature that distinguishes those tritium atoms that decay
within 12.32 years (the half-life of tritium) from those that don’t. Each tritium
atom simply has a fundamental chance of 0.5 of doing so. Thus if we place a
chunk of tritium next to the Geiger counter described in Probabilistic Bomb,
it’s not determined by the state of the tritium when it’s placed, nor even this
together with the state of the entire universe at that time, whether sufficient
atoms will decay so that enough ˇ particles are emitted within a short enough
time that the Geiger threshold is reached and the bomb explodes. Rather, there’s
simply a certain probability that this will happen.
Nevertheless, as Anscombe notes, it’s clear that if the Geiger threshold is
reached and the bomb explodes, then the placing of the tritium near the Gei-
ger is a cause. Correspondingly, the person who placed the material there –
perhaps together with the person who set the whole contraption up – would
presumably be held legally and morally responsible for the explosion and its
consequences. This poses an insurmountable problem for the regularity ana-
lyst for, despite being a cause, the placing of the tritium isn’t sufficient, even
when taken together with any or all of the circumstances that obtain at the time
it’s placed, for the bomb’s explosion. The best the regularity theorist can do
is restrict the scope of the analysis, offering it as an analysis of deterministic

9 A similar example is given by the physicist Richard Feynman (1965, 147).

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 12 Philosophy of Science

causation, and allowing that probabilistic causation must be analysed in other

terms. We’ll examine accounts of probabilistic causation in Section 5.

2.5.3 Common Causes

Perhaps the most serious problem of all for regularity analyses – and the one
that prompted Mackie himself (Mackie 1980, ch. 3) to abandon a pure regular-
ity theory of causation10 – is that of common causes. A well-known example
is given by Hans Reichenbach (1971, 193): namely of a fall in atmospheric
pressure which results in (i) a certain barometer predicting a storm; and (ii) a
storm.
The structure of this example is illustrated in Figure 1, with a representing
the fall in atmospheric pressure, b the barometer’s prediction of a storm, and s
the storm itself, and with the ‘arrows’ representing causal relations.11 Event a
is a ‘common cause’ of events b and s.
This sort of structure poses problems for the regularity approach. Suppose
the fall in atmospheric pressure is sufficient for the storm, or at least there’s a set
of conditions X in combination with which it’s sufficient for the storm. Suppose
also that the barometer is in perfect working order so that it will only indicate
a storm if there’s been a fall in atmospheric pressure. Then the barometer’s
prediction is an inus condition of the storm: the set of conditions comprising
the barometer reading, the barometer being in perfect working order, and X, is

Figure 1 Common Causation

10 Likewise, the problem is one that Strevens (2007) seeks to deal with via an account that, while
incorporating elements of a Mackie-inspired regularity analysis, also appeals to a notion of
‘causal connection’. Though Strevens (2007, 110) suggests that this notion of ‘causal connec-
tion’ might potentially be understood via a ‘process theoretic’ account of causation (of the sort
briefly discussed in the Introduction of this Element) or a nuanced counterfactual approach, he
refrains from endorsing any specific analysis of the notion.
11 This representation isn’t to be confused with the graphs of causal models that will be introduced
in Sections 4 and 5. Rather, it’s merely an intuitive representation of the causal relations that
obtain in this case.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 13

sufficient for the storm,12 while the latter two of these conditions aren’t on their
own sufficient. Requiring that causes temporally precede their effects doesn’t
help since barometers are prediction devices and therefore, if the barometer is
any good, it will signal a storm prior to the storm’s occurrence.
In the face of these difficulties, the regularity approach has fallen out of
favour. In Sections 4 and 5 we’ll explore alternatives to it. But first we’ll
pause to discuss some important general questions about the nature of cau-
sation and our methodology in investigating it. It’ll be easier to do this now
we’ve embarked on our discussion of theories of causation, since this will lend
greater concreteness to the discussion.

3 Interlude: Some Important Background

Before examining alternatives to the regularity approach, it’s worth making
explicit some points that have hitherto been in the background, and making
some further observations and distinctions regarding causation. Many of the
issues discussed in this section are more general than the question of whether
this or that particular theory of causation is best, but will also augment our
understanding of the theories discussed in this Element.

3.1 Is Causation Really a Relation? What Are Its Relata?

So far, we’ve spoken as though causation is a relation and that’s indeed the
view taken by almost everyone who has theorised about it. Thus, for instance,
if the short-circuit caused the fire, then presumably the short-circuit stands
in a causal relation to the fire. On the assumption that causation is a rela-
tion, we may discuss the formal properties of this relation. For instance, we’ve
already briefly discussed whether it’s asymmetric – whether (necessarily) for
all x and y, if x causes y, then it’s not the case that y causes x – or whether
it’s simply that bidirectional causation is something unfamiliar to our everyday
experience. In Section 4, we’ll see that there’s also a question of whether cau-
sation is transitive: whether (necessarily) for all x, y, and z, if x causes y and
y causes z, then x causes z. A related question is whether causation is irreflex-
ive – whether (necessarily) for all x, it’s the case that x doesn’t cause x. This is
related because if causation isn’t asymmetric (i.e. there’s some x and y such that
x causes y and y causes x) but is transitive, then it follows that it’s not irreflexive

12 This means that Mill’s account yields an incorrect result about the causation of the storm.
Hume’s account, on the other hand, gives incorrect results whenever we have a common cause
c that’s necessary for an effect e1 and sufficient for an independent effect e2 , for this will mean
that e1 is sufficient for c and therefore e2 .

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 14 Philosophy of Science

(from the fact that x causes y and y causes x, we can infer by transitivity that
x causes x).
Also on the assumption that causation is a relation, we might ask what the
relata of the causal relation are. The standard answer is events, where the lat-
ter category is broadly construed to include what might be considered ‘states’
or ‘conditions’ (such as the presence of oxygen). Thus, for instance, David-
son (1967) and Kim (1973a) advance considerations in favour of the view that
causation relates events, on a reasonably broad construal of what counts as an
‘event’.
There’s a hitch, though, with the view that causation is a relation with events
as its relata. That hitch is how to make sense of statements like ‘the crop failed
because there was no rain’, or ‘the absence of rain caused the crop to fail’, which
seems a legitimate paraphrase. The difficulty is that it’s not clear that an absence
of an event is itself an event. Moreover, we can’t just happily acknowledge that
and simply allow that causation is a relation that may have events or absences
as its relata. The reason is that it’s not obvious that absences are particulars that
could stand in such a relation to events or to other absences (Beebee 2004, cf.
Mellor 1995, 163–5).
What are our options? One is to deny that causation is a relation at all. This
option is pursued by Mellor (1995). Mellor (1995, 156) regards the basic form
of causal truth as ‘E because C’, where ‘C’ and ‘E’ are facts, with facts under-
stood as true sentences (Mellor 1995, 161). Facts aren’t particulars and so aren’t
apt to be the relata of a real relation. Moreover, in cases like ‘the crop failed
because there was no rain’, Mellor (1995, 163–5) claims there are no particu-
lars that serve as truthmakers for C and E that could themselves stand in a real
relation of causation.
A second option, discussed by Lewis (1986b, 190–3, 2004b, 281–2), would
be to insist that absences are particulars and therefore apt to stand in a causal
relation. Lewis discusses two versions of this view: (i) there are negative events
(events whose essence is the absence of something); (ii) there are (positive)
events that are contingently absences. To illustrate the idea behind (ii), turn-
ing right at a junction might be one way for a car to fail to turn left; another
way for it to fail to turn left might be by its continuing straight on. If the car
in fact turns right, then it’s turning right might be taken to be the absence of
its turning left, but only contingently because it’s possible that another event
(the car’s continuing straight on) should have instead been the absence of its
turning left.
The trouble with (i) is that negative events would be strange sorts of things,
and it’s not obvious how to reconcile them with standard accounts of the ontol-
ogy of events (e.g. Kim 1973a, Lewis 1986a). The trouble with (ii) is that

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 15

there don’t always seem to be events of the right sort to contingently serve
as absences. For instance, what positive event could contingently serve as the
absence of rain? The presence of sun? But sun tends to be beneficial to crops
(so its presence, unlike the absence of rain, wouldn’t explain a crop failure)
and, in any case, its presence isn’t incompatible with rain (fortunately for rain-
bow fans). The presence of a cloudless sky? But this is really an absence of
clouds.
A third option (see, e.g., Beebee 2004) is to maintain that causation is a rela-
tion, but admit that absences aren’t particulars that can stand in such a relation,
and so deny that statements like ‘the crop failed because there was no rain’ have
a causal relation as their truthmaker. If one nevertheless wishes to maintain that
such statements can be true, then one takes on the burden of giving a semantics
for them that doesn’t appeal to the existence of a corresponding causal relation.
Lewis (2004b, 282–3), for instance, appeals to the obtaining of certain counter-
factuals. For instance, the statement ‘the absence of rain caused the crop to fail’
is made true by the truth of the counterfactual ‘if there had been rain, then the
crop wouldn’t have failed’ (which, Lewis would claim, doesn’t itself require
a causal relation as its truthmaker – Lewis 1979). A regularity theorist, on the
other hand, might take this statement to be made true by the regularity that,
whenever there’s healthy crop growth (in the absence of a sprinkler system,
etc.), it’s preceded by rain. On the other hand, an advocate of a probabilistic
analysis (of the sort to be discussed in Section 5) might take it to be made true
by the fact that rain raises the probability of healthy crop growth.
In what follows, I’ll adopt the standard view that causation is a relation and
remain neutral between the second and third option for dealing with absences.
But, setting aside the problem of absences, are there particulars other than
events (construed liberally) that could serve as the relata of a causal relation?
One might think that objects could. For instance, if a brick is thrown at a win-
dow, which shatters, it sounds perfectly acceptable to say ‘the brick caused the
window to break’. Yet the brick is an object, not an event.
It is, however, tempting to think that event causation is fundamental. That’s
because, whenever it sounds felicitous to say ‘object o was a cause of e’, it
seems there’s some event c that o participated in such that it’s apt to say that ‘o
was a cause of e in virtue of o’s participation in c’ and also that ‘c was a cause
of e’. The brick caused the window to break, but it did so by hitting the window,
and the brick hitting the window is an event which itself caused the window to
break. On the other hand, whenever it sounds felicitous to say ‘a was a cause of
object o’ (e.g. ‘the collision of the Indo-Australian and Eurasian plates caused
the Himalayas’), an apt paraphrase seems to be ‘a was a cause of o’s coming
into existence’, and the coming into existence of o is an event.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 16 Philosophy of Science

Likewise, we commonly speak of agents as causes (agents might be regarded

as a certain sort of object). For example, one might felicitously say ‘Aisha
broke the window’. But again it seems that Aisha broke the window by par-
ticipating in an event – the throwing of the brick – that caused the window
to break. Some, however, have argued that there are instances of agent causa-
tion that are irreducible to event causation (see, e.g., O’Connor 2002). Often
such arguments are motivated by attempts to account for freedom of the will:
the thought being that irreducible agent causation allows agents to originate
causal chains, whereas if agents caused things merely in virtue of participat-
ing in events, then they wouldn’t exhibit genuinely free action, since those
events would themselves be answerable to a chain of prior causes. But the view
that there’s irreducible agent causation is highly controversial, and won’t be
pursued here.
Finally, some take variables or their values to be (among) the relata of the
causal relation.13 There’s a certain ambiguity in the notion of a ‘variable’. On
the one hand, it’s sometimes used to refer to a particular piece of mathematical
formalism. In this sense, ‘X’ can be used as a variable, with ‘x’ as one of its
values. On the other hand, ‘variable’ can be used to refer to the sort of thing
representable by the mathematical formalism. In this sense, the speed of a car
is a variable, with 90mph being one of its possible values. It’s the sort of thing
represented by the formalism rather than the formalism itself that’s taken by
some to be a potential causal relatum.
Where variable values are taken to be the relata14 this view seems compat-
ible with the view that the causal relata are events (when ‘event’ is construed
liberally, to include states/conditions). After all, such things as the car’s trav-
elling at 90mph can be seen as events (in this liberal sense). This view doesn’t
escape the issues surrounding absences, however. For instance, we might take
the amount of rain that falls in some location during a particular crop season
to be representable by a variable R, with R D 0 corresponding to no rain and
positive values of R corresponding, say, to the total rainfall in millimetres. We
might also represent the success of a certain crop with a variable C that, for
example, takes C D 0 if the crop dies, but with positive values of C corre-
sponding, say, to the height in centimetres that the crop achieves by harvest. In
that case, adopting the language of variables, it would be perfectly reasonable

13 This view is particularly associated with the causal modelling approach to causation to be
explored in sections 4.6 and 5.3–5.6. Pearl (2009), Spirtes et al. (2000), and Halpern (2016)
are prime examples of this approach.
14 As we’ll see in Section 4.6.5, taking the variables themselves to be the relata might sometimes
be appropriate depending on the variety of causation at issue.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 17

– if, in fact, R D 0 and C D 0 – to say that R D 0 was a cause of C D 0. We’re then

confronted again with the question of exactly what makes such a claim true.

3.2 How Many Relata?

Assuming causation is a relation, we might ask how many relata it has. The
answer that most naturally springs to mind is ‘two’. When one says ‘c is a
cause of e’, ‘… is a cause of …’ is a two-place predicate which might natu-
rally be taken to pick out a binary relation (specifically, a relation that obtains
between pairs of events) in the world. Yet not everyone takes causation to be
binary. For instance, Hitchcock (1996) argues that it’s a ternary relation with
the relata being three events: two actual and one merely possible. The idea is
that causal relations are best described by claims of the form ‘c rather than c0
is a cause of e’. For instance, consider the interrogative sentence: Did the car’s
travelling at 60mph cause the crash? Arguably this is unanswerable because the
question asked by this sentence is underdetermined. By contrast, consider the
pair of sentences: Did the car’s travelling at 60mph rather than 40mph cause
the crash? and Did the car’s travelling at 60mph rather than 80mph cause the
crash? These seem to ask more determinate questions and are more tractable.
Arguably, then, the problem with the first sentence is that it asks did c cause
e simpliciter, but there are no facts about whether c caused e simpliciter, only
facts about whether c rather than c0 caused e for various possible alternatives,
c0 to c. The second and third sentences ask determinate questions because they
make clear what causal fact is at issue.
Others (e.g. Schaffer 2005, 2013; Northcott 2008) go further and argue that
causation is a quaternary relation which finds expression in sentences of the
form ‘c rather than c0 caused e rather than e0 ’. To illustrate, suppose Mo can
take no dose, one dose, or two doses of medicine and that doing so will lead
to no recovery, slow recovery, or quick recovery, respectively. Then claims
of the following form seem entirely cogent: ‘taking one dose rather than two
doses caused Mo to recover slowly rather than quickly’. More generally, it has
been suggested (Schaffer 2005, 327–9; Northcott 2008) that we should take the
contrasts to c and e to be (possibly singleton) sets of alternatives. For instance,
it seems sensible to ask whether the car’s travelling at 60mph rather than less
than 60mph caused the crash, but of course the contrast here appears to concern
a set of specific alternatives to the car’s travelling 60mph.
It is of course true that we often make claims of the form ‘c caused e’ with
no contrasts to c or e specified. Those who take causation to be contrastive
on the cause and/or effect side respond that such sentences manage to express
propositions because context supplies the contrasts. For instance, if an assertion

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 18 Philosophy of Science

such as “drinking lots of wine last night caused my hangover” is in most con-
texts evaluable as true without explicitly specified contrasts, that’s because
in most contexts the implicit contrast to my drinking all that wine is some
sort of default behaviour like my not drinking alcohol (and the implicit con-
trast to my having a hangover is just my not having a hangover). But perhaps
in rare contexts the implicit contrast might be different so that the sentence
expresses a false proposition. For example, suppose that last night I was par-
ticipating in an initiation ceremony where I was forced to choose to drink
either lots of wine or lots of whisky. Then the implicit contrast to my drink-
ing lots of wine is my drinking lots of whisky, so that the sentence can be
considered elliptical for “drinking lots of wine rather than lots of whisky last
night caused my hangover (rather than my not having a hangover)”, which is
presumably false.

3.3 Token vs Type

A distinction can be drawn between token (or ‘singular’) causal claims and
generic (or ‘type’) causal claims. Examples of generic causal claims include
smoking causes cancer or quantitative easing causes house price rises. Exam-
ples of token causal claims include Jane’s smoking caused her cancer or
the Bank of England’s 2009–12 quantitative easing programme caused the
London house price bubble that began in 2012. Token causal claims con-
cern individual events (such as Jane’s smoking and her developing cancer),
while generic causal claims concern types of event (such as smoking and
cancer).
Scientists are interested in assessing the truth both of token causal claims –
the Chicxulub impact caused the K-Pg Extinction event, the Big Bang caused
the Cosmic Microwave Background, the financial crisis of 2007–8 caused the
Great Recession – and generic causal claims – smoking causes cancer, loose
monetary policy causes inflation, abnormal protein folding is a cause of aller-
gies. Lawyers are principally interested in the truth-values of token causal
claims (this act caused that harm).
A fairly common view among philosophers is that type causal claims are
some sort of generalisation over token causal claims. As Lewis (1973a, 558)
puts it: ‘Presumably [type causal claims] are quantified statements involv-
ing causation among particular events (or non-events)’. This view yields a
parsimonious metaphysics of causation, since it implies that there’s no sui
generis relation of type causation in addition to token causation. Still, as Lewis
(1973a, 558) admits, type causal claims don’t admit of a straightforward anal-
ysis in terms of the standard quantifiers of first order logic. One plausible

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 19

suggestion is that the quantifier involved is a generic operator (Carroll 1991,

Eagle 2015).15
But the metaphysical priority of token causation isn’t a completely uncon-
troversial matter, and what view one takes on this issue is likely to be in part
informed by what one thinks is the best theoretical approach to causation. For
instance, on the regularity approach, token causation is understood in terms of
associations between event types. But then, for the regularity theorist, it seems
natural simply to interpret such associations as type-level causal relations. By
contrast, as we’ll see in Section 4, it’s natural for advocates of counterfactual
theories of causation to take token causation to be the more basic notion.16
Although the type vs token distinction is one of the most commonly drawn
and important when it comes to varieties of causal claim, it’s not exhaustive
of those that can be distinguished. For instance, Woodward (2003) defines a
number of different causal relations that might obtain between variables. As
we’ll see in Section 4.6.5 and Section 5.6, it’s doubtful that any sort of causal
relation between variables falls neatly into the ‘token’ or ‘type’ categories, and
indeed there may be a case for treating such relations as sui generis, even if not
entirely unrelated to token and type causation.

3.4 Aims and Method

When engaged in almost any task, it’s helpful to ask what, exactly, are we trying
to achieve? An answer to this will help us to frame an appropriate method for
reaching our goals. The task of theorising about causation is no different.
There are a variety of sometimes subtly distinct aims that one might have in
theorising about causation. For instance, one might be engaged in a reductive
or a non-reductive project, one might be engaged in something like ‘conceptual
analysis’, or (to the extent that this differs) one might be concerned with cau-
sation as it is ‘out there in the world’. If engaged in conceptual analysis, then
one might be concerned with analysing the layperson’s concept of causation or
(insofar as these differ) the scientific or legal concept of causation. Also, one
might be concerned with the project of ‘descriptive’ analysis – examining the
contours of the concept as it actually is – or prescriptive/revisionary analysis
– seeking to outline a more useful and/or more coherent concept of causation.
If investigating causation ‘in the world’, one might be concerned with what
causation contingently is in our world (Dowe 2000, ch. 1), or with what it is as

15 Approaches to the semantics of the generic operator are discussed by, for example, the
contributors to Carlson and Pelletier (1995).
16 And, as we’ll see in Section 5.6, the question of the relation between token and type causation
is somewhat complex if one adopts a probabilistic approach.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 20 Philosophy of Science

a matter of metaphysical necessity (i.e. across all possible worlds). Examining

these various sorts of project, and situating the accounts of causation examined
in this Element within them is the task for this section.17

3.4.1 Concepts vs Metaphysics

Hume famously distinguished between ‘our idea of causation’ and causation
‘in the objects’ (in the world) (Hume 1739, I.iii.14). It’s tempting to regard
investigation of the nature of the former as corresponding to the project of
conceptual analysis, and investigation of the latter as a metaphysical project.
Unfortunately, things aren’t quite so straightforward for reasons that will be
described at the end of this section.
Hume’s own primary focus was on the idea of causation. The extent of the
implications he took his discussion to have for the nature of causation ‘in the
objects’ is a matter of some controversy (see Read 2014). Hume was partic-
ularly interested in the aetiology of our idea of causation. This was part of a
more general project of showing that all human ideas (concepts) ultimately
derive from experience. Hume acknowledged that causation was a tricky case.
Central to our idea of causation, he thought, is the idea that there is a neces-
sary connection between cause and effect. Yet Hume claimed that when we
observe interactions that we take to be causal – for example, of one billiard
ball impacting upon, and imparting motion in, a second – we don’t observe any
such necessary connection (Hume 1739, I.iii.13). Ultimately, Hume traces our
idea of necessary connection to a certain introspective experience: our alleged
experience, upon observing a particular event, of feeling ‘a determination of
the mind’ (i.e. an involuntary compulsion) to expect its usual effect(s).
The example of Hume makes it clear why an investigation of the idea of
causation isn’t automatically, or in any direct way, an examination of the
phenomenon as it exists in the world. For, although necessary connection is
(according to Hume) a key element of our concept of causation, ‘necessity is
something, that exists in the mind, not in the objects’ (Hume 1739, I.iii.14), and
when we speak of necessary connections in the world, the word ‘necessity’ is
‘wrongly apply’d’ (Hume 1739, I.iii.1; italics original). For Hume, causation
in the objects can be captured simply by defining a cause as ‘[a]n object [viz.
event] precedent and contiguous to another, and where all the objects resem-
bling the former are plac’d in like relations of precedency and contiguity to

17 Of necessity, the discussion here is brief. Readers are referred to Paul and Hall (2013, esp. ch.
2, sect. 3) for a more extensive discussion of the goals and methods of a range of contemporary
approaches to caussation.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 21

those objects, that resemble the latter’ (Hume 1739, I.iii.14). This, of course, is
Hume’s regularity account of causation.18
Hume, as well as other British empiricists (notably John Locke), were
important influences in the later analytic tradition which gave the method of
conceptual analysis a central role in philosophy. We’ve already seen that there
is some reason to think that analysis of the concept of causation isn’t auto-
matically an investigation of causation as it exists in the world. Still, the two
projects aren’t entirely unrelated. If one gave an account of causation in the
world that bore little resemblance to our concept of causation, then one could
reasonably be accused of changing the subject. For example, this would be a
fair charge if one asserted that causation ‘in the objects’ were identical to mere
correlation. After all, it’s a common aphorism that correlation doesn’t imply
causation: barometer readings don’t affect the weather, for instance.
There’s a certain amount of give and take here, however. Perhaps nothing ‘in
the objects’ could perfectly answer to our concept of causation. This would be
so if, for instance, there are no necessary connections in nature, but Hume is
right that the concept of causation involves the idea of ‘necessary connection’
as a key component. One response would be to simply deny that there is such
a thing as causation ‘in the world’. Yet if there is something in the world that
answers near enough to our concept of causation, or if there’s something in the
world that our concept of causation tracks with at least a reasonable degree of
systematicity (for instance, Hume thought that we were systematically willing
to apply our concept of causation in cases in which there’s a confluence of
priority, contiguity, and constant conjunction), then it may be reasonable to
call this thing ‘causation’.
If one regards causation ‘in the world’ as diverging from our concept of
causation, then it might be natural to suggest a revision to our ordinary con-
cept so that it better corresponds to causation in the world or so that it better
corresponds to something that it would be useful for our concept of causation
to track. This is one reason why one might endorse a ‘revisionary’ or ‘prescrip-
tive’ account of the concept of causation, as opposed to a purely descriptive
one. There are other (perhaps related) reasons for endorsing a revisionary anal-
ysis. For example, one might think that the ordinary concept of causation is
somehow imprecise or even incoherent and that we would be better served by
a concept that was coherent and precise. We’ll see in Sections 4 and 5 that

18 It’s a matter of controversy whether Hume thought that this ‘definition’ captures all that causa-
tion is in the world or merely all that we can know about causation as it is in the world – see
(Read 2014). The former thesis seems fairly clearly the one argued for in (Hume 1739). There
is, however, some evidence that he rows back from this in (Hume 1748).

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 22 Philosophy of Science

especially some of the more formal accounts of causation (e.g. those appealing
to structural or probabilistic models) arguably involve a degree of precision
that our ordinary concept of causation lacks.19 Indeed, as we’ll see, some of
those who have developed such accounts (e.g. Spirtes et al. 2000, Hitchcock
2001b, Woodward 2003, Pearl 2009) define overtly technical causal notions –
for example, ‘total’, ‘contributing’, ‘direct’ causation – which, though arguably
useful, are sometimes only loosely related to the causal concept(s) of the lay-
person. Nevertheless, several of these authors also discuss the notion of what
they call ‘actual causation’ (Woodward 2003, sect. 2.7; Pearl 2009, ch. 10),
which is taken to be more-or-less the ordinary person’s token causal concept.
When theorising about causation, it can often be helpful to have in mind the
question of what the point is in our having a concept of causation, or perhaps
a concept of this or that sort of causation (e.g. a concept of type causation and
a concept of token or actual causation). This is perhaps particularly obvious if
one is in the business of developing a prescriptive analysis of causation – since
a consideration of what practical or cognitive needs are served by a concept (or
concepts) of causation will help guide the development of a concept that better
meets those needs. But it is also the case even when it comes to developing a
descriptive analysis: an understanding of why we have a concept (or concepts)
of cause can help us to more successfully delineate the contours of the con-
cept(s) that we do have, just as an understanding of the purpose or function of
a carburetor can help us to understand (and perhaps improve) its design.
Mellor (1995, 58–66, ch. 7) argues that causation has several important
‘connotations’, including that effects are evidence for their causes, that causes
explain their effects, and that causes are means for bringing about their effects
as ends. The connections to evidence, explanation, and practical reasoning can
well explain why we have a concept or concepts of causation and should inform
our theorising about it. For instance, if one were to devise a theory of causation
on which it were difficult or impossible to see why – on such a theory – causes
are evidence for, explanatory of, or means to, their effects, then this would be
at least prima facie evidence that we were barking up the wrong tree. It turns
out that this is a genuine issue for many accounts of causation. For instance,
each of the broad traditions that we consider in this Element – the regularity,

19 This is also true of some of the less technical accounts. For instance, we’ll see that Lewis’s
counterfactual analysis of causation implies that overdeterminers are not causes in cases of so-
called symmetric overdetermination. Lewis (1986b, 208) believes that there is a ‘lack of firm
common-sense judgements’ regarding whether overdeterminers are genuine causes and says
that when, as in cases of overdetermination, ‘common sense falls into indecision or controversy
… then … [w]e can reasonably accept as true whatever answer comes from the analysis [of
causation] that does best on the clearer cases’.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 23

counterfactual, and probabilistic traditions – has, at its core, a simple insight:

that causes are constantly conjoined with their effects, or that effects (counter-
factually) depend upon their causes, or that causes raise the probability of their
effects. These basic insights tend to be relatively conducive to seeing how our
concept of causation serves its purpose: if a cause is constantly conjoined with
its effect, then it’s good evidence for that effect (and vice versa); if an effect
depends upon its cause, then the cause appears to be a means (or at least a part
of a means) to its effect; if a cause raised the probability of its effect, then it
would seem that it helps explain why the effect happened, etc.
The trouble is that these ‘insights’ aren’t quite right. As we’ve seen, not all
causes are constantly conjoined with their effects (e.g. probabilistic causes).
We’ll also see that not all effects counterfactually depend upon their causes,
nor do all causes raise the probability of their effects. Thus credible philo-
sophical accounts of causation are more subtle than simply equating causation
with constant conjunction, counterfactual dependence, or probability-raising.
Yet this greater complexity tends to come associated with a greater difficulty
of explaining why these accounts capture a notion that serves the functions that
we suppose our concept(s) of causation to have. Whilst we may be able to see,
roughly at least, why an event c upon which another event e depends is a means
to e, or why c can help explain e if c is a probability-raiser of e, it’s rather more
difficult to see why c should count as a means to e if e merely bears to c the
transitive closure of the counterfactual dependence relationship (see Section
4.1) or if c is merely a de facto probability-raiser of e (Section 5.4). It therefore
seems that if one is to properly justify such an account, one incurs the burden of
showing how it captures a concept that does the work our concept of causation
does. This is an issue taken up in Section 5.7.
A final point worth making before leaving the topic of conceptual and met-
aphysical analysis is that there is a certain amount of blurring of the lines in
contemporary philosophy. These days when a philosopher offers an ‘analy-
sis’ of a concept like ‘causation’ or ‘knowledge’ it’s rare for the analysans
to be couched in terms that would be particularly easy for the layperson
to understand. For example, popular contemporary analyses of knowledge
invoke technically-defined notions of ‘safety’ and/or ‘sensitivity’. Analyses
of causation invoke such things as the transitive closure of the relation of
non-backtracking counterfactual dependence. It’s not particularly plausible
that such analyses are articulating the structure implicit in the ordinary lan-
guage user’s concept of causation. Rather, they’re engaged in something like
the project of ‘simply’ providing truth-conditions for causal claims. Argua-
bly at least, it’s a short step from the latter project to making claims about
the metaphysical truthmakers for causal claims. Insofar as this characterisation

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 24 Philosophy of Science

of the contemporary project of ‘conceptual analysis’ is correct, criticisms of

such analyses as obscurum per obscurius may sometimes be unfair. They’re
not intended to be anything like dictionary definitions, but are perhaps bet-
ter regarded as preludes to metaphysical accounts. Critiques may therefore be
best directed at the claims of the extensional adequacy of the truth-conditions
and/or the attractiveness of the corresponding metaphysical picture of
causation.20
The fact that Hume’s investigation of the ‘idea of causation’ is an attempt to
articulate the structure of the ordinary concept is the reason I hesitated previ-
ously to equate it with the modern project of conceptual analysis. There’s also
a reason for caution about equating his project of investigating ‘causation in
the objects’ with metaphysical analysis. This is that, when Hume defines cau-
sation ‘in the objects’, it’s not entirely clear whether he means to be giving an
account of what causation objectively is as a matter of metaphysical necessity
(i.e. across all possible worlds) or what it objectively is contingently (i.e. in our
world, and perhaps in worlds that share some important feature with our world
– such as having the same laws of nature).
Most of the major traditions in the contemporary philosophy of causation
seek the former sort of account. But not all do. For instance, Dowe (2000, ch.
1) explicitly advances his account of causation – according to which causation
consists of the transference of conserved quantities – as what he calls an ‘empir-
ical analysis’: an analysis of what causation contingently is in the actual world.
He doesn’t wish to deny that there might be causation in worlds in which there
are no conserved quantities, or in which there are no conserved quantities of
the sort required to underwrite causal relations.
The fact that many philosophers focus on what might be called ‘metaphysi-
cal analysis’ rather than ‘empirical analysis’ can help to explain aspects of the
structure of the debate. For example, we have seen that regularity accounts of
causation have been criticised for being unable to handle probabilistic causa-
tion. It also turns out that some probabilistic accounts of causation are criticised
for not being able to handle deterministic causation! If both deterministic and
probabilistic causation are genuine possibilities, then an account that can han-
dle only one or the other won’t make for a satisfactory metaphysical analysis.
But, for instance, if it turns out (contrary to the prevailing assumption) that all

20 Thus, for instance, though the notion of ‘non-backtracking counterfactual dependence’ may be
obscure as far as the ordinary competent language user is concerned, if this relation is grounded
in some asymmetry of physics – as Lewis (1979) and Dunn (2011) have argued – then giv-
ing extensionally adequate truth-conditions for causal claims in terms of non-backtracking
counterfactual dependence potentially paves the way for an overall simplification of our
metaphysics.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 25

causation in our world is deterministic then a regularity analysis might still be

advanced as an empirical analysis.21

3.4.2 Causal Proliferation?

It has already been noted that ordinary language users make both type and
token causal claims. So prima facie it looks like we’ll need to make room
for the existence of at least two causal notions (albeit it may turn out that
one is conceptually or metaphysically reducible to the other). It’s plausible
that the notions of type and of token causation serve slightly different uses.
If we recall Mellor’s ‘connotations’ of causation, the notion of type causation
seems particularly useful when it comes to prediction or means–ends reason-
ing, while token causation is closely tied to explanation. Thus, for instance,
to explain the COVID-19 outbreak we investigate its token causes. Part of
the reason this is important is that presumably there are features of what-
ever the token causes were that are generalisable. Thus, for instance, if one
of the token causes turns out to have been an instance of bat–livestock inter-
action, this might help to confirm a type-causal claim about the risks of such
interactions which will hopefully help us to reduce human virus outbreaks in
future.
In any case, it looks as though we have, and have use for, at least two causal
notions. But it’s also an interesting question whether there might be, or be
grounds for, still further proliferation of causal concepts. Regarding actual cau-
sation, Pearl (2009, 309) points out that ‘[h]uman intuition is extremely keen
in detecting and ascertaining this type of causation’. To the extent that intu-
itions about type causation are less firm, it might be that there’s no canonical
variety of type-causation that’s picked out by our ordinary thought and talk.
This perhaps gives us the flexibility to define various type-causal notions as
and when they are useful without doing violence to causal intuition and with-
out being susceptible to accusations of changing the subject. In this context,
it’s interesting to note that when Woodward (2003, sect. 2.3) characterises a
variety of technical causal notions – ‘direct’, ‘total’, and ‘contributing’ cause
– he regards them each as varieties of type causation, and distinguishes them
from actual causation. As we’ll see in Sections 4 and 5, these technical notions
are useful in analysing relations between variables in complex causal systems.

21 In fact, it might be that deterministic and probabilistic causation coexist in some worlds (perhaps
at different scientific ‘levels’), and that ours is such a world. For relevant discussion, see Albert
(2000), Loewer (2001), Ismael (2009), Werndl (2011, 2263), Emery (2015), and List and Pivato
(2015, sect. 9). Potentially, then, there’s an interesting discussion to be had about the legitimacy
of seeking a philosophical account of causation that applies only to certain domains (e.g. levels)
even within a world.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 26 Philosophy of Science

And perhaps one can also usefully define technical causal notions at the token-
level provided that one is careful to distinguish these from our notion of actual
causation.22
But what would be particularly interesting, given our firm intuitions about
actual causation, is if there were some inconsistency in our thought or talk about
it that might motivate a distinction between two or more concepts of actual
causation. There are at least two ways in which this might occur. First, it might
be that the conception of actual causation in play varies across contexts. For
instance, it might be that lawyers or scientists or historians draw upon a dif-
ferent conception of actual causation from one another or from ordinary folk.
Second, it might be that there’s an inconsistency in our thought or talk about
actual causation even within a context.
The latter is a view argued for in some detail by Hall (2004) who thinks
that the ordinary folk concept of actual causation bifurcates into a concept of
‘dependence’ and a concept of ‘production’. In many cases, production and
dependence align. For example, a short circuit might produce a fire, with the
occurrence of the latter also depending upon the occurrence of the former.
Sometimes, though, they come apart: if two short circuits occur simultaneously,
each sufficient for the fire, then whilst both may play a role in the fire’s produc-
tion, the occurrence of the fire might depend on neither (since in the absence of
each, the other short circuit would have still been enough to produce the fire).
This is an instance of so-called ‘symmetric overdetermination’ (a phenomenon
that will be discussed in more detail in Section 4.4). To the extent that we feel
conflicted about whether each of the short circuits was a cause in such a case,
the breaking of the normal association of production and dependence might be
one explanation for this feeling of conflict.
There are also prima facie reasons for thinking that the conception of actual
causation varies across contexts. For instance, a simple counterfactual ‘but for’
test for actual causation is preponderant in the law (“But for the short circuit,
would the fire have occurred?”). In apparently stark contrast, there’s consid-
erable controversy about the legitimacy of using counterfactual reasoning to
establish causal claims in history (see, e.g., Evans 2013). Still, this prima facie
divergence may only be that. When application of the ‘but for’ test produces
counterintuitive consequences, judges tend to vary the test for causation.23

22 For instance, Hitchcock (2007, 503–4) introduces a technically defined notion of ‘token causal
structure’ which he clearly distinguishes from the ordinary notion of actual causation. And, as
we’ll see in Section 4.6.5, Woodwardian notions of ‘direct’, ‘total’, and ‘contributing’ cause
can also be defined for the token level.
23 Examples include R v Dyson [1908 2 KB 454 (CA)], Kingston v Chicago & N.W. Ry [191 Wis.
610, 211 N.W. 913 (1927)], McGhee v National Coal Board [1972 3 All E.R. 1008, 1 W.L.R.
1], and Fairchild v Glenhaven Funeral Services Ltd [2002 UKHL 22].
https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press
 Causation 27

Indeed, the so-called ‘plain meaning rule’ of statutory interpretation24 dictates

that lawyers ought not to depart from the ordinary language meaning of terms
(including causal terms) employed in statute provided that these are precise and
unambiguous. On the other hand, despite the fact that many historians overtly
repudiate counterfactuals, it has been pointed out that “[h]istorians … use
counterfactuals regularly though implicitly when they assign necessary causes,
and sometimes in assigning degrees of importance to causes” (Tucker 1999,
265). It’s therefore not clear that there’s any hard evidence that either lawyers
or historians operate with a concept of actual causation that’s distinct from the
folk concept.
Surveying the usage of token causal concepts across the sciences is clearly
far too big an undertaking to pursue here, but there are some prima facie reasons
to be skeptical that scientific usage differs dramatically from usage in ordinary
life. One reason is that much of our ordinary reasoning about causes is clearly
proto-scientific. When I try to figure out what caused the tripping of a circuit
breaker in my house, my investigation seems to be at least broadly continu-
ous with scientific investigation of token causes. Indeed, Hitchcock and Knobe
(2009, 591) provide an interesting conjecture about the purpose of the ordinary
notion of actual causation: namely, ‘the concept of actual causation enables us
to pick out appropriate targets for intervention’. By this they have in mind that
the ordinary concept tends to pick out as causes those factors that could have
been manipulated to avoid or to influence the effect. (In this instance, prevent-
ing my toddler from tipping her apple juice onto a power block would have done
the trick!) If this is right, then not only does it connect actual causation – which
we previously noted seems closely tied to the ‘explanatory’ connotation of cau-
sation – with the connotation concerning control, it suggests that our ordinary
concept of actual causation serves a purpose that is also useful to scientists,
which would make continuity between the ordinary and scientific concepts
unsurprising.

3.4.3 Reduction
Reductivity is often seen as a virtue of philosophical analyses. For those seek-
ing a metaphysical reduction of causation, success would – other things being
equal – allow for a simplification of their overall metaphysic. For example, if
causation could be reduced to constant conjunction with the latter interpreted

24 This ‘rule’ is a convention followed by the courts of England and Wales, with similar conven-
tions also adopted in other jurisdictions. It was articulated in the Sussex Peerage Case [1844; 11
Cl and Fin 85] as follows: ‘If the words of the Statute are in themselves precise and unambig-
uous, then no more can be necessary than to expound those words in that natural and ordinary
sense.’
https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press
 28 Philosophy of Science

simply in terms of tokens of event types standing in regular spatio-temporal

relations of priority and contiguity, then – since it is likely that spatio-temporal
relations, as well as event types and tokens, will feature elsewhere in our meta-
physical theory anyway (and not just in our account of causation) – our overall
metaphysic will be simplified since causation will not need to feature as a basic
posit of it.
To the extent that a conceptual analysis seeks to elucidate the truth-conditions
for causal statements as a prelude to our investigation of what the metaphys-
ical truth-makers might be, we might hope that those truth-conditions can be
given without invoking causal notions, for otherwise it seems that our result-
ing metaphysical account may end up being non-reductive. On the other hand,
if, like Hume, we’re seeking a ‘conceptual analysis’ that articulates the struc-
ture of the ordinary person’s concept of causation, we might favour a reductive
account to the extent that – like Hume – we think there’s a prima facie mys-
tery about that concept that might be dispelled by analysing it into non-causal
constituents.
Yet in spite of the fact that many philosophers of causation have pursued
reduction, some (e.g. Woodward 2003, 104–7) have persuasively argued that
analyses of causation can be informative even if they are non-reductive: that
is, even if the analysans isn’t entirely free of causal notions. To see how this
might work, consider, by way of analogy, the following two biconditionals:

1. S knows p iff S knows p.

2. S knows p iff S knows that one knows that p.

To assert the first biconditional would be to be guilty of the most blatent

unilluminating circularity. By contrast, whilst the second biconditional obvi-
ously couldn’t serve as a reductive analysis of knowledge (since the verb
‘to know’ appears on both sides of the biconditional), it nevertheless makes
an interesting and substantive assertion about knowledge.25 For instance, it
says something about knowledge that (if true) clearly distinguishes it from
propositional attitudes like hope, regret, etc.26
The regularity theories examined in Section 1 appear to fall into the cate-
gory of reductive analyses: causation is analysed in terms of (lawful) regularity,
which itself doesn’t appear to be an inherently causal notion. Later in this Ele-
ment, however, we’ll consider accounts (such as the ‘interventionist’ approach

25 This biconditional is implied by the (controversial) KK principle together with the factivity of
knowledge.
26 I don’t claim that knowledge is itself a propositional attitude, though this is a view defended by
Williamson (2000, ch. 1).

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 29

discussed in Section 4.6.6) that are non-reductive in nature. In evaluating these

accounts it is important to have in mind the question of whether they are
illuminating even if they aren’t reductive.

4 Counterfactual Theories of Causation

Despite the fact that they’re usually associated with the regularity approach to
causation, both Mill and Hume suggest a close connection between causation
and counterfactuals. Thus Mill says, ‘[I]f a person eats of a particular dish, and
dies in consequence, that is, would not have died if he had not eaten it, people
would be apt to say that eating of that dish was the cause of death’ (Mill 1843,
III.v.3; italics mine). Mill treats the italicised counterfactual as synonymous
with the causal claim that the person died in consequence of eating the dish.
Similarly, Hume at one point (Hume 1748, sect. VII) offers a counterfactual
definition of causation. It has been pointed out (Lewis 1973a) that counter-
factual definitions in fact aren’t equivalent to standard regularity-theoretic
definitions.
The counterfactual definition implicit in the aforementioned Mill passage is
equivalent to the ‘but for’ test for causation in the law: a cause is an event but
for which the effect wouldn’t have occurred. The seminal philosophical work
developing a counterfactual theory of causation is (Lewis 1973a). As we’ll
see in the next subsection, the counterfactual analysis developed there is more
sophisticated than the ‘but for’ criterion.
Before proceeding, it’s worth pointing out that counterfactual approaches
standardly focus on token causation. This is natural, since it makes good sense
to talk about the (actual and counterfactual) occurrence and non-occurrence
of token events (e.g. Jane’s smoking and Jane’s developing lung cancer). By
contrast, it’s not clear that it makes sense to talk about the (actual and counter-
factual) occurrence and non-occurrence of event types (e.g. smoking and lung
cancer) unless this is simply understood in terms of the occurrence and non-
occurrence of instances of those types, in which case we’re back to talking
about token events.
A second point to note before proceeding is that we’ll assume determin-
ism in this section. That is, we’ll assume that given the state of the world at
one time, the laws of nature determine how it will be at all future times. We’ll
see in Section 5 that the counterfactual approach can be generalised to cases
where determinism doesn’t hold. But things will be simpler if we focus on the
deterministic case for now.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 30 Philosophy of Science

4.1 Lewis’s 1973 Analysis of Causation

Let c and e be token events, and let O.c/ and O.e/ be the propositions c
occurs and e occurs, respectively. Then, where € expresses the counterfac-
tual connective ‘If it had been that … , then it would have been that … ’, Lewis
(1973a, 562–3) says that e causally depends27 upon c iff the following two
counterfactuals hold:

(i) O.c/ € O.e/:

(ii) :O.c/ € :O.e/

The first of these expresses the counterfactual ‘if it had been the case that c
occurred, then it would have been the case that e occurred’ (for short: ‘if c had
occurred, then e would have occurred’); the second expresses the counterfac-
tual ‘if it hadn’t been the case that c occurred, then it wouldn’t have been the
case that e occurred’ (for short: ‘if c hadn’t occurred, then e wouldn’t have
occurred’).28
Lewis (1973a, 1973b, 1979) gives counterfactuals a possible worlds seman-
tics. According to him, a counterfactual A € C is true iff either there are no
worlds in which A (i.e. A is necessarily false) or there’s some world in which
A&C that’s more similar (‘closer’) to our own, actual world (hereafter ‘@’),
than is any world in which A&:C. Lewis (1973b, 14–15) subscribes to a prin-
ciple known as Strong Centering, according to which @ is more similar to itself
than any other possible world is to it. Given his semantics for counterfactuals,
this has the immediate implication that any counterfactual with a true anteced-
ent has the same truth-value as its consequent. This means that, where c and
e are actual events (as they must be if they’re to be contenders to stand in a
causal relation in the actual world), the aforementioned counterfactual (i) is
automatically true. This means that whether e causally depends upon c hinges
upon whether counterfactual (ii) is true.
Lewis, however, doesn’t equate causation with causal dependence between
events.29 Rather, he takes this merely to be a sufficient condition for causation.

27 Sometimes I’ll say ‘counterfactually depends’ rather than ‘causally depends’ to express this
relation in what follows. Lewis reserves ‘counterfactually depends’ for the relation between the
propositions O.e/ and O.c/ that obtains when the following pair of counterfactuals are true. I’ll
be more lax in also using it to describe the relation between the corresponding events too.
28 Following Lewis, I’m calling both subjunctive conditionals that have true antecedents and those
with false antecedents ‘counterfactuals’. Lewis gives a uniform semantics to both types of
subjunctive.
29 Strictly speaking, we should say ‘distinct’ events. To give Kim’s (1973b, 571) example: ‘[i]f I
had not written ‘r’ twice in succession, I would not have written “Larry” ’. Yet writing ‘r’ twice
in succession isn’t a cause of writing ‘Larry’ but rather an essential part of it.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 31

Instead, causation is equated with the transitive closure (or ‘ancestral’) of

causal dependence. That is, c is a cause of e iff there’s a sequence of actual
events c, d1 , …, dn , e such that e causally depends upon dn , which causally
depends upon … d1 , which causally depends upon c (Lewis 1973a, 563).
Lewis (1973a, 563) calls such a sequence a ‘causal chain’. The case where
e causally depends upon c is the special case of a ‘one link’ causal chain.
Lewis identifies causation with the ancestral of causal dependence because he
takes causation to be transitive. Yet, as Lewis is aware, c can be connected
to e via a causal chain without its being the case that e causally depends
upon c. This happens in cases of what Lewis (1986b, Postscript E) calls early
preemption.
Before discussing early preemption, it’s worth noting that the question
of whether causation is transitive is contested (see, e.g., Hall 2000, 2004,
Lewis 2004a, Hitchcock 2001a, Sartorio 2005). There are apparent coun-
terexamples. Suppose a terrorist plants a bomb on a train but, fortunately,
the bomb is spotted and the train safely evacuated. The bomb’s being
spotted counterfactually depends upon and is caused by its being planted.
The survival of all the passengers counterfactually depends upon and is
caused by the evacuation. Yet it’s at best dubious whether the bomb’s
being planted causes the passengers to survive. Lewis’s account in terms
of the ancestral of counterfactual dependence implies that it does. And
Lewis can’t readily drop his commitment to transitivity by identifying cau-
sation with counterfactual dependence rather than its ancestral because, as
we’ll see, this would deprive him of the ability to handle cases of early
preemption.

4.2 Early Preemption

Cases of preemption are cases in which, in the absence of a cause of some par-
ticular effect, another event (the ‘preempted alternative’) would have brought
about that effect instead. Lewis (1986b, Postscript E) distinguishes between
‘early preemption’ and ‘late preemption’. In cases of early preemption, as
Lewis (1986b, 200) puts it, ‘the process running from the preempted alterna-
tive is cut off well before the main process running from the preempting cause
has gone to completion’. By contrast, in cases of late preemption, this ‘backup
process’ isn’t cut off before the effect itself occurs.
The following is an example of early preemption based on one given by
Edgington (1997, 420). An example of late preemption will be considered in
the next subsection.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 32 Philosophy of Science

Early Preemption

Alice and Bob are hunting deer. They have one gun between them and are
down to their last bullet. They spot a deer. Both are ace shots and the shot
is a simple one. If Alice doesn’t shoot, then Bob will. Alice shoots. The
bullet hits the deer and the deer dies.

In Early Preemption, although Alice’s shot is a cause of the deer’s death, the
deer’s death doesn’t counterfactually depend upon her shot: but for her shot,
Bob would have shot and so the deer would still have died. Bob’s intention to
shoot if Alice doesn’t can be considered the preempted backup. Had Alice not
shot, this would have initiated a process (involving Bob’s shooting) that would
itself have resulted in the deer’s death. However, such a process is stopped
in its tracks before the deer’s death (which is why this counts as a case of
early preemption). Specifically, as soon as Alice shoots, Bob is prevented from
shooting.
Cases of preemption pose a problem for regularity theories. Alice’s shooting
together with other prevailing conditions such as the shot being an easy one,
the gun’s being in good working order, etc., suffices for the deer’s death (recall
that, for now, we’re assuming determinism). The trouble is that Bob’s inten-
tion to shoot if Alice doesn’t, together with his being an excellent shot, Alice’s
being an excellent shot, the shot being an easy one, the gun’s being in good
working order, etc., is also a sufficient condition (it guarantees that, one way
or the other – by means of Alice’s shooting or Bob’s shooting – the deer’s death
will come about). Bob’s intention is a non-redundant element of this sufficient
condition (after all, the condition doesn’t include Alice’s shooting). So Bob’s
intention is an inus condition of the deer’s death. Yet Bob’s intention isn’t a
cause.30
Returning to Lewis’s account, observe that, despite the fact that Bob’s inten-
tion means there’s no counterfactual dependence of the deer’s death on Alice’s
shot, Lewis would claim there’s a chain of counterfactual dependence connect-
ing the two, and this is why Alice’s shot counts as a cause of the deer’s death.
Specifically, consider the event comprising the speeding bullet’s presence at the
mid-point between Alice and the deer. Call this event m. Lewis’s (1973a, 567,
1986b, 200–1) suggestion would be that there’s a two-link chain of counterfac-
tual dependence connecting Alice’s shot and the deer’s death, with the deer’s

30 Although he develops an account of causation that draws upon the notion of an inus condition,
Strevens (2007, esp. 5.4) recognises the need to supplement the notion of an inus condition with
a notion of ‘causal production’ (which he doesn’t define in regularity-theoretic terms) in order
to deal with cases like this.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 33

death counterfactually dependent upon m, and m counterfactually dependent

upon Alice’s shot. That one of these links holds is fairly obvious: it’s intu-
itive that, if Alice hadn’t shot, then m wouldn’t have obtained (there might
instead have been a speeding bullet mid-air between Bob and the deer). Lewis
would also claim that, if m hadn’t obtained then the deer wouldn’t have died.
Specifically, Lewis claims that the nearest worlds in which m doesn’t occur
are worlds in which Alice still shoots (but perhaps shoots off target) and so
Bob doesn’t shoot. We therefore shouldn’t be tempted to think that had m not
occurred, then that would have been because Alice didn’t shoot, so Bob would
have shot, and the deer would have died anyway. The latter sort of reasoning
relies on what (Lewis 1979, 456) calls ‘backtracking’ counterfactuals (which
assert that if the state of the world at some later time had been different, then
its state at some earlier time would have been different – e.g. if the later event
m hadn’t occurred, then the earlier event of Alice’s shooting wouldn’t have
occurred) and claims that a correct counterfactual semantics, at least for the pur-
poses of evaluating causal claims, is one that rules backtracking counterfactuals
false.
It’s this prohibition on backtracking counterfactuals that’s also Lewis’s
proffered solution to two of the key difficulties for the regularity theory exam-
ined in Section 2.5: namely, the problem of distinguishing cause from effect
(considered in Section 2.5.1) and the problem of the spurious implication
of causation between independent effects of common causes (discussed in
Section 2.5.3). Lewis (1973a, 564–5) claims that the counterfactual analysis
of causation gets the direction of causation right because effects counter-
factually depend upon their causes (or counterfactually depend upon events
that counterfactually depend upon … their causes), but (in virtue of non-
backtracking) causes don’t counterfactually depend upon their effects (or coun-
terfactually depend upon events that counterfactually depend upon … their
effects).
Now consider the example of common causation discussed in Section 2.5.3
in which the fall in atmospheric pressure is a cause both of the barometer
reading and the storm. According to Lewis (1973a, 565–7), his counterfac-
tual analysis avoids the spurious implication that the barometer reading is a
cause of the storm because (in virtue of non-backtracking) it’s false that if
the barometer hadn’t predicted a storm, the atmospheric pressure wouldn’t
earlier have fallen. Rather it would still have fallen and so the storm would
still have occurred. So the storm neither counterfactually depends on the
barometer reading nor is connected to the barometer reading by a chain
of counterfactual dependence that goes via the earlier fall in atmospheric
pressure.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 34 Philosophy of Science

Lewis (1979) provides a detailed story about why his possible worlds seman-
tics for counterfactuals implies the ‘non-backtracking’ result.31 His approach
has, however, been criticised by several authors (Elga 2001; Woodward 2003,
ch. 3.6; Kment 2006). A now-popular alternative approach to providing a non-
backtracking semantics is the so-called ‘interventionist’ approach (Goldszmidt
and Pearl 1992; Meek and Glymour 1994; Woodward 2003), which we’ll
examine in Section 4.6.6. But however we understand the semantics for coun-
terfactuals, let’s grant that Lewis’s analysis correctly handles cases of early
preemption (even if it does, rather controversially, commit to causal transitiv-
ity). Still the account runs into trouble in cases of late preemption, as Lewis
(1986b, 203–5) himself admits.

4.3 Late Preemption

Edgington’s example can be modified to give a case of late preemption.

Late Preemption

Alice and Bob are hunting deer. They each have a gun and each gun has
plenty of ammunition. They spot a deer. If Alice doesn’t shoot or if Alice
shoots but fails to kill the deer, then Bob will shoot and kill the deer. Alice
shoots and kills the deer.

It’s worth emphasising how Late Preemption differs from Early Preemption.
In Early Preemption, the backup process (the one via which Bob would have
brought about the deer’s death if Alice hadn’t shot) is cut short early in pro-
ceedings. Once Alice shoots, there’s no possibility of Bob’s shooting (because
there are no bullets left in their one rifle). Since this backup process is cut short
early, the effect (the deer’s death) depends upon the later stages of the process
via which Alice’s shot brings about the effect: for instance, upon m (cf. Lewis
1986b, 200). This allows us to complete a chain of causal dependence
In contrast, in Late Preemption the backup process isn’t cut short until the
effect itself (the death) occurs. That’s because it’s only the effect itself that
prevents Bob from shooting. This means that the effect (the death) doesn’t
counterfactually depend upon any part of the process via which Alice’s shot
brought it about (cf. Lewis 1986b, 203–4). For instance, the death doesn’t coun-
terfactually depend upon the presence of a bullet at the midpoint between Alice

31 In fact, Lewis (1979, 475) claims that his semantics entails the falsity of backtracking coun-
terfactuals only where certain physical asymmetries that ordinarily correspond to the direction
of time obtain. He thinks that if these physical asymmetries broke down – as they might on a
closed time-like curve – then backwards causation might occur and that his semantics would
entail the truth of the corresponding backtracking counterfactuals.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 35

and the deer since, if the bullet hadn’t reached that point, it wouldn’t have hit
the deer and so (seeing that the deer was initially unscathed) Bob would have
shot and killed it.
The counterfactual reasoning from the absence of the bullet at the mid-point
to Bob’s shooting and the deer’s death isn’t in this case the sort of ‘backtracking’
reasoning that Lewis takes to be illicit. That’s because, in Late Preemption,
the bullet’s absence would mean that it wouldn’t (later) hit the deer and so
Bob (later) would have shot. By contrast, in Early Preemption, since Bob was
prevented from shooting by Alice’s very act of shooting, we need to reason
backwards from the bullet’s absence to Alice’s (earlier) not shooting to reach
the conclusion that Bob would have shot and killed the deer.
So invoking transitivity doesn’t solve the problem posed by late preemption.
Lewis’s 1973 account of causation therefore doesn’t handle late preemption
correctly: it fails to diagnose the preempting cause as a cause. Can a counter-
factual account of causation be given that gets cases of late preemption right?
Before seeking to answer that question, let’s briefly consider another sort of
case that arguably poses problems for Lewis’s 1973 analysis.

4.4 Symmetric Overdetermination

Symmetric overdetermination occurs where there are two or more events that
(in the circumstances in which they occur) are ‘enough’ to bring about some
effect so that, in the absence of either one, the effect would still have occurred.
This means that the effect doesn’t counterfactually depend upon either. Unlike
preemption cases – which are also cases where the effect is in a sense ‘overde-
termined’ – symmetric overdetermination cases are those where there’s no
distinguishing the overdeterminers into ‘preempting cause’ and ‘preempted
backup’: both have equally good claim to the title ‘cause’. The following is
an example.

Symmetric Overdetermination

Alice and Bob are hunting deer. Each has a loaded gun. They spot a deer.
They both shoot, with their bullets piercing the deer’s heart simultane-
ously. Each bullet alone would have been sufficient to bring about the
deer’s death.

In Symmetric Overdetermination, the deer’s death doesn’t counterfactually

depend upon Alice’s shot since, in the absence of Alice’s shot, Bob’s shot would
still have been enough to kill the deer. Nor is there a chain of counterfactual
dependence connecting Alice’s shot to the deer’s death. In particular, no part
of the process connecting Alice’s shot to the deer’s death (e.g. the presence of

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 36 Philosophy of Science

a speeding bullet at the mid-point between Alice and the deer) is such that the
deer’s death depends on it: after all, the process initiated by Bob’s shot was com-
plete and would have sufficed to bring about the deer’s death even had some
part of the Alice-process been missing. Exactly analogous reasoning shows that
the deer’s death neither counterfactually depends upon, nor is connected by a
chain of counterfactual dependence to, Bob’s shot.
Lewis (1986b, 194–5, 207, 211–12) claims that the intuition that symmetric
overdeterminers are causes isn’t strong. He therefore thinks it’s not a signifi-
cant strike against his counterfactual analysis of causation that it doesn’t deliver
the result that they are. But not everyone agrees with this assessment (cf. Hitch-
cock 2007, 522) and there are cases in which courts have recognised symmetric
overdeterminers as causes.32 Nevertheless, since late preemption cases are the
most uncontroversial counterexamples to Lewis’s 1973 account of causation,
we’ll focus on what might be done to deal with these cases, though we’ll see that
some of the potential solutions may help with symmetric overdetermination too
insofar as this is seen as problematic for the counterfactual approach.33

4.5 The Fragility Response

A natural response to the problem posed by late preemption for counterfac-
tual theories of causation (a similar response could be adopted for dealing with
early preemption cases) would be to say that the event caused by the preempting
cause isn’t the same as the event that would have been caused by the preempted
alternative. In Late Preemption, for instance, one might say that the death
that the deer actually died is different from the death that would have occurred
had Bob shot and killed the deer. For one thing, because Bob gives Alice the
first opportunity to shoot and kill the deer, a death of the deer at Bob’s hands
would have been later than the death that actually occurred. So perhaps the
actual effect does indeed counterfactually depend upon the preempting cause
after all. This response might also work in at least some symmetric overdeter-
mination cases. For instance, in Symmetric Overdetermination, perhaps the
actual death of the deer, due to the two bullets simultaneously piercing its heart,

32 E.g. Kingston v Chicago & N. W. Railway [191 Wis. 610, 211 N.W. 913 (1927)].
Note that regularity theories have better prospects for counting overdeterminers as causes since
each overdeterminer is (in the circumstances) sufficient for the effect.
33 There are varieties of ‘overdetermination’ that differ in structure from the ones we’ve dis-
cussed in the last three subsections (e.g. Moore 2009, 418; Kroedel 2008, 130–1) but which
we lack space to discuss here. One particularly well-known variety of overdetermination not
discussed in the main text is so-called ‘trumping preemption’ (Schaffer 2004). However, sev-
eral philosophers (e.g. McDermott 2002, 89–92, 97; Halpern and Pearl 2005, 873–5; Hitchcock
2007, 512n) take the view that ‘trumping preemption’ is really just an instance of symmetric
overdetermination, and so has limited independent diagnostic value.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 37

is somewhat different in manner from the death that would have occurred had
only one bullet been fired.
The response under consideration invokes what Lewis (1986b, 204) calls
the modal ‘fragility’ of the effect: the notion that a counterfactual version
of the effect (e.g. the death of the deer at Bob’s hands) isn’t to be identi-
fied with the actual effect if it is even slightly different in time or manner
from the actual effect. Lewis doesn’t avail himself of this solution because
he rejects extreme standards of event fragility (the adopted standards would
have to be fairly extreme if this solution were to work for all cases of late
preemption, including ones in which the preempted alternative would have
brought about a very similar and only very slightly later version of the effect).
The key reason for his doing so is that adopting extreme standards of fragility
appears to yield spurious cases of causation on the counterfactual approach.
For example, such things as the precise angle at which the deer was stand-
ing or even its precise blood pressure might have made a difference to the
precise time and/or manner of its death, even if they were entirely irrelevant
to whether or not it died. Yet it’s tempting to think that such factors are not
causes.34

4.6 De Facto Dependence and Causal Modelling

Perhaps the most influential recent attempt to modify the counterfactual
approach to causation to deal with some of the challenges described here
appeals to a notion that (Yablo 2002, 2004) calls de facto dependence. This
is a sort of latent counterfactual dependence that’s revealed by holding fixed
certain features of the actual world.
Consider Early Preemption. In this example, the deer’s death doesn’t
straightforwardly counterfactually depend upon Alice’s shot. But rather than
appeal to the existence of a chain of counterfactual dependence a la Lewis
(and hence – as we’ve seen – commit to the rather dubious thesis that causation
is transitive), the de facto dependence account says that the status of Alice’s
shot as a cause is revealed by the truth of the following counterfactual: if Alice
hadn’t shot, and Bob still hadn’t shot, then the deer wouldn’t have died. Bob’s
non-shooting is a feature of the actual world (or rather the world in which the

34 Lewis (2004b) later came to advocate a counterfactual analysis of causation that appeals to event
fragility in a subtler way. His guiding idea is that a cause ‘influences’ (i.e. makes a difference to)
the timing and manner of its effect in a way that non-causes – such as preempted alternatives
– don’t. Though we lack the space to discuss the details here, this later account suffers from
serious counterexamples, as Bigaj (2012) shows, and has not widely caught on.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 38 Philosophy of Science

scenario described in Early Preemption plays out) and the de facto depend-
ence approach tells us to hold this feature fixed in seeking latent counterfactual
dependence of the deer’s death on Alice’s shot.35
The counterfactual that reveals this de facto dependence of the deer’s death
on Alice’s shot has a more complex antecedent than those to which Lewis
appeals in his analysis of causation. As well as counterfactually varying the
putative cause (Alice’s shooting), the antecedent serves to hold fixed a certain
feature of the actual situation (Bob’s non-shooting). This feature was something
that depended upon the putative cause (because Bob intended to shoot if Alice
didn’t) and therefore would have been absent under the simple counterfactual
supposition that the cause didn’t occur.
Hitchcock (2001a, 275) articulates this point well when he points out that
Lewis – in virtue of supposing that the counterfactuals relevant to analysing
causation don’t backtrack – takes history prior to the occurrence of the event
c to be implicitly held fixed (in virtue of the semantics for the counterfac-
tual) when evaluating :O.c/ € :O.e/. But Lewis (1979) takes the relevant
semantics to allow foretracking, so that events lying to the future of c are
allowed to vary under the counterfactual supposition that :O.c/. Yet, as Hitch-
cock (2001a, 275) observes,’ ‘The necessity of foretracking from c to e does
not prevent us from entertaining a counterfactual of the form ‘If c had not
occurred, but e had occurred anyway, then …’. Let us call a counterfactual
of this sort an explicitly nonforetracking or ENF counterfactual.’ In order
to give a full-fledged account of causation in terms of ENF counterfactuals/
de facto dependence, we need an account of which features of the actual world
should be held fixed. Several authors (e.g. Hitchcock 2001a, 2007; Halpern
and Pearl 2005; Halpern and Hitchcock 2015; Halpern 2016; Weslake 2020)
have appealed to causal models (more precisely: structural equation models)
in giving such an account.

4.6.1 Structural Equation Models

A structural equation model (SEM), M comprises a set V of variables and a
set E of structural equations. The values of the variables (or sometimes – as
we’ll see in Section 4.6.5 – the variables themselves) are taken to be, or at least
to represent, potential causal relata. For example, if one wanted to model Early
Preemption, one might include in one’s model a binary variable A that takes
the value A D 1 if Alice shoots and A D 0 otherwise; a binary variable BI that

35 Though we won’t go into the detals, Hitchcock (2001a) demonstrates that such an approach
doesn’t commit one to causal transitivity.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 39

takes BI D 1 if Bob intends to shoot iff Alice doesn’t and BI D 0 otherwise;36

a binary variable B that takes B D 1 if Bob shoots and B D 0 otherwise; and
a binary variable D that takes D D 1 if the deer dies, and D D 0 if the deer
survives.
The structural equations in a model represent the way in which the model’s
variables depend upon one another. Specifically, each of the variables in the
variable set V appears on the left-hand side (LHS) of exactly one structural
equation in the equation set E: the structural equation ‘for’ the variable in ques-
tion. The variables in V comprise two mutually exclusive subsets: a set U of
exogenous variables, the values of which don’t depend upon the values of any
other variables in the model, and a set Y of endogenous variables, the val-
ues of which do depend upon the values of other variables in the model. For
example, taking the variable set of our model representing Early Preemption
to be fA; BI; B; Dg where the variables have the interpretation given above, D
is an endogenous variable, since what happens to the deer depends upon what
Alice and Bob do. Variable B is also an endogenous variable, since whether
Bob shoots depends upon whether Alice shoots and also upon Bob’s intention.
Variable A, on the other hand, is an exogenous variable since what Alice does
doesn’t depend either upon what Bob intends to do or what he in fact does.37
Likewise BI is an exogenous variable because Bob’s intention to shoot if Alice
does not doesn’t depend upon what Alice does.
The structural equation for each endogenous variable in a SEM shows how
its value depends upon the values of other variables in the variable set. More
precisely, where X and Z are variables in the variable set V of a model, X
appears on the right-hand side (RHS) of the equation for Z just in case there’s
some assignment of values to the other variables in V such that the value of
Z depends upon that of X if we assume those other variables are held fixed
at the assigned values. Less formally, a structural equation expresses how the
variable on the LHS depends on those variables in the model upon which it
immediately depends: if Z immediately depends upon X and also depends on Y,
but only depends on Y because Y has a bearing on the value of X, then the struc-
tural equation for Z expresses its value as a function of X but not Y. In the case

36 For simplicity, I’ll assume that the alternative to Bob’s having this intention is his not intending
to shoot under any circumstances. However, if we wanted to model the possibility that (e.g.)
Bob has the intention to shoot come what may, we could deploy a multi-valued variable.
37 Of course, Alice’s shooting might depend upon Bob’s not having shot even earlier, but if we
wanted to represent whether Bob shot even earlier, we’d need an additional variable in our
model. (Variable B can’t serve this function, since B represents a state that depends on A. Given
that this is a token causal scenario, A itself doesn’t depend upon the state represented by B. If
it did, we’d have a causal loop.) In general, whether a variable is exogenous or endogenous
depends on what other variables we include in our model.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 40 Philosophy of Science

of an exogenous variable, W, no variable appears on the RHS of the equation

for W. Rather, the equation simply takes the form W D w , where w is the
actual value of W. Any variables that appear on the RHS of the equation for a
variable V are known as the ‘parents’ of V; V is a ‘child’ of theirs. The notions
of ‘ancestor’ and ‘descendant’ are defined in terms of the transitive closures of
parenthood and childhood respectively.
Thus, in our model of Early Preemption, the structural equations for A and
BI just state their actual values, A D 1 and BI D 1, because A and BI are
exogenous. Since the value of B (immediately) depends upon the value of A
and the value of BI (and not the value of D: no backtracking!), the equation
for B expresses its value just as a function of A and BI. Specifically, the equa-
tion is B D MinfBI; 1 Ag which encodes the fact that Bob shoots (B D 1)
iff he has the intention (BI D 1) and Alice doesn’t shoot (A D 0). Finally,
the structural equation for D expresses its value as a function of A and B, but
not of BI. That’s because D immediately depends upon A and B, but doesn’t
immediately depend upon BI. This is evidenced by the fact that there’s no com-
bination of values of A and B such that the value of D depends upon that of
BI when A and B are held fixed at those values: specifically, once we know
whether Bob shoots, what he earlier intended has no further bearing on the
deer’s death. (By contrast, it’s easy to verify that there are values of A and BI
– e.g. A D 0 and BI D 1 – such that the value of D depends upon the value
of B when A and BI are held fixed at those values, and that there are values
of B and BI – e.g. B D 0 and BI D 1 – such that the value of D depends
upon the value of A when B and BI are held fixed at those values.) The struc-
tural equation for D is, specifically, D D MaxfA; Bg, reflecting the fact that
D takes value D D 1 (i.e. the deer dies) iff either A D 1 (Alice shoots) or
B D 1 (Bob shoots). To summarise, our model ‘EP’ of Early Preemption is as
follows:

Variables: fA; BI; B; Dg

Equations: A D 1; BI D 1; B D MinfBI; 1 Ag; D D MaxfA; Bg

The solution to these structural equations is A D 1, BI D 1, B D 0, and D D 1.

That is, Alice shoots, Bob intends to shoot iff Alice doesn’t, Bob doesn’t shoot,
and the deer dies.
Since structural equations encode information about counterfactual depend-
ence (specifically, about how the variable on the LHS counterfactually depends
upon the variables on the RHS), they differ from algebraic equations. That’s
because counterfactual dependence isn’t symmetric – it doesn’t ‘backtrack’
(at least not on the sort of counterfactual semantics that we need if we’re to

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 41

successfully analyse causation in terms of counterfactuals) – and so algebraic

rearrangement of structural equations won’t generally preserve truth.
A SEM can be given a graphical representation by taking the variables in V
as the nodes of the graph and drawing a directed edge (‘arrow’) from a variable
Vi to a variable Vj (Vi ; Vj 2 V) just in case Vi is a parent of Vj according to
the structural equations in E. Implementing this rule for EP yields the graph
depicted in Figure 2 (which follows a summary of the model EP). A ‘directed
path’ can then be defined as an ordered sequence of variables hVi ; Vj ; : : : ; Vk i,
such that there’s a directed edge from Vi to Vj , and a directed edge from Vj to
… Vk (i.e. directed paths run from variables to their descendants). The directed
paths in the graph depicted in Figure 2 are thus hA; Di, hBI; Bi, hA; Bi, hB; Di,
hBI; B; Di, and hA; B; Di.

EP

Variables: fA; BI; B; Dg

Interpretation: A D 1=0 if Alice shoots/doesn’t shoot; BI D 1=0 if Bob
intends/doesn’t intend to shoot iff Alice doesn’t shoot; B D 1=0 if Bob
shoots/doesn’t shoot; D D 1=0 if the deer dies/survives.
Equations: A D 1; BI D 1; B D MinfBI; 1 Ag; D D MaxfA; Bg

Figure 2: Early Preemption

One can evaluate a counterfactual of the form Vi D vi & : : : &Vk D vk €

Z D z with respect to a SEM by replacing the equations for Vi ; : : : ; and Vk with
the equations Vi D vi ; : : : ; and Vk D vk (thus rendering each of Vi ; : : : ; and
Vk exogenous), while leaving all other equations intact. The result is a new set
of equations. The counterfactual holds just in case, in the solution to the new
set of equations, Z D z. This ‘equation replacement’ method gives us a means
to evaluate even those counterfactuals whose truth or falsity isn’t implied by
any single equation in the model alone (Hitchcock 2001a, 283): for example,
counterfactuals saying how the value of a variable would differ if the values of
its grandparents were different.38

38 As we’ll discuss further in Section 4.6.6, the equation replacement method for evaluating coun-
terfactuals is sometimes taken as modelling what would happen if the variables in the antecedent

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 42 Philosophy of Science

So, for example, if one wants to evaluate the ENF counterfactual A D

0&B D 0 € D D 0 – which is the one that advocates of the de facto depend-
ence approach take to be indicative of the causation of D D 1 by A D 1 in
Early Preemption – with respect to our model EP, then one needs to replace
the equation that EP gives for A (i.e. A D 1) with the equation A D 0 and
the equation that EP gives for B (i.e. B D MinfBI; 1 Ag) with the equation
B D 0 while leaving the equations that EP gives for BI (i.e. BI D 1) and D (i.e.
D D MaxfA; Bg) intact. In the solution to this new set of equations, D D 0,
hence the counterfactual is true relative to EP.

4.6.2 A Structural Equations Analysis of Causation

A simple and elegant proposal for analysing causation in terms of SEMs is

given by SEA (for Structural Equations Analysis). This analysis is closely
related – but not identical – to the analysis given by Hitchcock (2001a, 286–7).

SEA

Where x and x0 (x ¤ x0 ) are possible values of the variable X and y and y0

(y ¤ y0 ) are possible values of the variable Y, then X taking value X D x
rather than X D x0 is a cause of Y taking value Y D y (rather than Y D y0 )
iff X D x and Y D y are the actual values of X and Y and, in some apt SEM,
there’s a path P from X to Y such that, where W is the set of variables in
the model that don’t lie on P and W D w represents the variables W
taking their actual values, then (according to the model) if X had taken the
value X D x0 and W had taken W D w (i.e. the off-path variables had
taken their actual values), then Y would have taken the value Y D y0 .

There’s quite a bit to unpack in this analysis. First, it analyses a contrastive

notion of causation. As noted in Section 3.2, taking causation to be contrastive
seems necessary when dealing with multi-valued variables. When dealing with
binary variables, the contrast is obvious and so mention of it can be suppressed
(and often will be in what follows). Second, the reference to ‘apt’ models.
We’ve seen what a SEM is, but what makes such a model apt? We’ll return
to this issue in the next subsection, but let’s assume for now that the model EP
is apt (spoiler: it passes all tests for appropriateness that have been suggested
in the literature) and see how SEA gives the verdict that A D 1 (rather than

of the counterfactual were set to the specified values by means of ‘interventions’ (Woodward
2003, 98). Glynn (2013) discusses the extent to which the equation replacement method could
alternatively be seen as modelling a version of Lewis’s (1979) counterfactual semantics.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 43

A D 0) causes D D 1 (rather than D D 0): that is, that Alice’s shooting (rather
than not) causes the deer to die (rather than survive).
To see this, focus on the path hA; Di in EP. SEA tells us that A D 1 is
a cause of D D 1 if it’s true that, when all variables that lie off this path
are held fixed at their actual values, if A had taken the value A D 0, then
D would have taken the value D D 0. In EP, there are two variables that
lie off the path hA; Di: namely BI and B, which have BI D 1 and B D 0 as
their actual values. Thus the fact that A D 1 is a cause of D D 1 is indicated
by the truth relative to EP of the ENF counterfactual A D 0&BI D 1&B D
0 € D D 0.39 The truth of this ENF counterfactual relative to EP can be veri-
fied by employing the equation replacement method described in the previous
subsection.
Analysis SEA also yields the correct verdict that BI D 1 isn’t a cause of
D D 1 (confirming that it’s ‘merely’ a preempted backup). To see this, con-
sider the sole directed path from BI to D in EP: hBI; B; Di. The sole variable
that lies off this path is A, the actual value of which is A D 1. For BI D 1
to count as a cause of D D 1 according to SEA, it would be necessary that
the following counterfactual hold in EP: BI D 0&A D 1 € D D 0. How-
ever it’s easy to verify via the equation replacement method that this isn’t
the case.
Analysis SEA can also potentially help with Late Preemption,40 an example
which – as we saw – Lewis struggles to deal with. To model Late Preemption it
will be helpful to recall that, since Bob first waits to see whether Alice succeeds
in killing the deer, the deer would have died slightly later if Alice hadn’t killed
it. Suppose the deer would have died by 1:00pm if Alice had shot, but only by
1:01pm if Alice hadn’t shot. Thus we can adopt the following model (‘LP’) of
Late Preemption.41

39 I previously said that A D 0&B D 0 € D D 0 is the counterfactual that the advocate of the
de facto dependence approach takes to be indicative of causation of D D 1 by A D 1. And so it
would be, according to SEA, if our model only included variables A, B, and D. But what I said
previously was a slight simplification in that, once BI is included in our model, it too must be
held fixed according to SEA’s general recipe for revealing the sort of de facto dependence that’s
indicative of causation. In this instance, holding fixed BI – although harmless – isn’t doing any
real work. It’s the holding fixed of B at B D 0 that’s the key to revealing latent dependence of
D D 1 on A D 1.
40 The treatment suggested here follows that of Halpern and Pearl (2005, 863–5) and Hitchcock
(2007, 524–9).
41 It’s worth emphasising how the interpretation of variable B differs in LP as compared to EP.
In LP, B D 1 represents Bob’s intention to shoot iff Alice doesn’t kill the deer (rather than
iff Alice doesn’t shoot). As in the case of EP, I assume for simplicity that the alternative to
Bob’s having the intention he does is for him to have no intention to shoot the deer under any
circumstances.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 44 Philosophy of Science

LP

Variables: fA; BI; B; D1W00 ; D1W01 g

Interpretation: A D 1=0 if Alice shoots/doesn’t shoot; BI D 1=0 if Bob

intends/doesn’t intend to shoot iff Alice doesn’t kill the deer; B D 1=0 if
Bob shoots/doesn’t shoot; D1W00 D 1=0 if the deer is/isn’t dead by 1:00pm;
D1W01 D 1=0 if the deer is/isn’t dead by 1:01pm.

Equations: A D 1; BI D 1; D1W00 D A; B D MinfBI; 1 D1W00 g; D1W01 D

MaxfD1W00 ; Bg

Figure 3: Late Preemption

Assuming LP is apt then, according to SEA, A D 1 is a cause both of

D1W00 D 1 and D1W01 D 1. First consider the sole path between A and D1W00 :
hA; D1W00 i. There are three variables that lie off this path: BI, B, and D1W01 with
the actual values BI D 1, B D 0, and D1W01 D 1. SEA counts A D 1 as a
cause of D1W00 D 1 in virtue of the truth of the following counterfactual rela-
tive to LP: A D 0&BI D 1&B D 0&D1W01 D 1 € D1W00 D 0. The truth
of this counterfactual is easily verified using the equation replacement method.
This exhibits the fact that the equation replacement method corresponds to a
non-backtracking way of evaluating counterfactuals. In particular, it doesn’t
validate the reasoning that if the deer had been dead at 1:01pm, then it must
have been dead at 1:00pm. Indeed it doesn’t validate this reasoning even on
the assumption that neither Alice nor Bob shot.
When it comes to the relationship between A D 1 and D1W01 , consider the path
hA; D1W00 ; D1W01 i. There are two variables – BI and B – that lie off this path and
their actual values are BI D 1 and B D 0. SEA implies that A D 1 is a cause of
D1W01 D 1 if the latter depends upon the former when we hold fixed BI and B at
their actual values: that is, if it’s true that A D 0&BI D 1&B D 0 € D1W01 D
0. That this is true can again be verified via the equation replacement method.
Although the event of the deer’s death isn’t straightforwardly represented in
LP – though one possibility would be to take it to correspond to the disjunction
D1W00 D 1 _ D1W01 D 1 (cf. Hitchcock 2007, 525) – following Hitchcock (2007,

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 45

525), we might suppose that ‘we will have adequately captured our intuitions
about the case if we can show that’ A D 1 is a cause of both D1W00 D 1 and
D1W01 D 1.
It’s also easy to see that BI D 1 isn’t a cause of D1W00 D 1 or D1W01 D 1
according to SEA.42 First, since there’s not even a directed path from BI to D1W00
(reflecting the fact that Bob is going to give Alice until at least 1:00pm to kill
the deer before himself shooting), SEA straightforwardly doesn’t count BI D 1
as a cause of D1W00 D 1. Second, though there’s a directed path from BI to D1W01
– namely, hBI; B; D1W01 i – holding fixed the off-path variables A and D1W00 at
their actual values – A D 1 and D1W00 D 1 – doesn’t reveal de facto dependence
of D1W01 D 1 upon A D 1: it’s false that BI D 0&A D 1&D1W00 D 1 € D D 0,
as can again be verified via the equation replacement method.
Although SEA deals well with Early Preemption and at least reasonably
well with Late Preemption, it arguably runs into difficulties with Symmetric
Overdetermination. Before demonstrating this and discussing whether there
might be an alternative SEM analysis of causation that fares better, we ought
to return to the question – deferred earlier – of what makes for an apt SEM.

4.6.3 Apt Models

Several criteria for model appropriateness have been suggested in the literature.
Specifically, the following criteria for the variables of such a model have been
suggested:

1. (Partition) The values of each of the variables in the variable set ought
to form a partition (i.e. a set of mutually exclusive and jointly exhaustive
alternatives) (Halpern and Hitchcock 2010, 397–8; Blanchard and Schaffer
2015, 182).
2. (Independence) The values of distinct variables in the variable set shouldn’t
be logically or metaphysically related (Hitchcock 2001a, 287; Halpern and
Hitchcock 2010, 397).
3. (Naturalness) The values of the variables ought to represent reasonably
natural and intrinsic states of affairs (cf. Blanchard and Schaffer 2015, 182).

Since analyses like SEA take all and only values of distinct variables to be
potential causal relata, (Partition) ensures that we don’t thereby miss actual
causal relations because they obtain between the values of a single variable,

42 At least SEA doesn’t regard BI D 1 as a cause of D1W00 D 1 or D1W01 D 1 relative to the

model LP. Yet SEA existentially quantifies over apt models so BI D 1 will count as a cause
simpliciter of D1W00 D 1 or D1W01 D 1 if it’s a cause relative to at least one apt model. In the
next subsection, we’ll discuss the extent to which we have reason to think that models relative
to which BI D 1 counts as a cause of D1W00 D 1 or D1W01 D 1 are inapt.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 46 Philosophy of Science

(Independence) ensures that we don’t mistake stronger-than-causal relations

for causal relations, and (Naturalness) ensures that unnatural or non-intrinsic
states of affairs do not get counted as causes and effects (justification for such
a restriction on causal relata is given by Lewis 1986b, 190, 263).43 Three other
conditions of model appropriateness are also fairly standard:

4. (Veridicality) The model should entail only true counterfactuals (Hitchcock

2001a, 287; 2007, 503)
5. (Serious Possibilities) The model shouldn’t represent possibilities “that we
consider to be too remote” (Hitchcock 2001a, 287; cf. Woodward 2003,
86–91; Weslake 2020)
6. (Stability) It shouldn’t be possible to overturn causal verdicts reached rela-
tive to the model by moving to a model (satisfying Partition, Independence,
Naturalness, Veridicality, and Serious Possibilities) that’s richer in the sense
of having a variable set that’s a superset of that of the original model (or per-
haps by including variables that are fine-grainings of those in the original
model) (cf. Halpern and Hitchcock 2010, 394–5; Halpern 2015; Hitchcock
2007, 503).

The motivation for (Veridicality) should be obvious: if the model entails false
counterfactuals (via the equation replacement method described in Section
4.6.1), then it’s not a reliable guide to genuine causal relations. (Serious Pos-
sibilities) is needed to avoid a variety of counterintuitive results (see, e.g.,
Hitchcock 2001a, 287; Woodward 2003, 86–91). Most straightforwardly, it
helps avoid the overgeneration of cases of causation by absence. For instance,
JFK’s death presumably counterfactually depends upon Paul McCartney’s fail-
ure to get in the way of the oncoming bullets. But most of us would be reluctant
to say that the latter is a cause of the former. This might be explained by the fact
that we view McCartney’s getting in the way as a non-serious possibility. (Seri-
ous Possibilities) ensures that no appropriate model of the events surrounding

43 The fact that (Naturalness) refers to ‘reasonably’ natural and intrinsic states of affairs introduces
a certain amount of vagueness into the account, but is necessary because demanding perfect
naturalness would presumably rule out high-level states of affairs (i.e. the states – such as price
inflation and asbestos exposure – of concern to the non-fundamental sciences) as causes and
effects, which would be undesirable. Since any account of causation will need a condition like
this, the SEM approach doesn’t suffer a relative disadvantage in this regard.
If absences are unnatural states of affairs (e.g. if they’re disjunctions of positive events, cf.
Lewis 1986b, 189–93), we might replace (Naturalness) with the requirement that each variable
has at most one value representing such a state of affairs if we wish to allow (as we will in the
following) that absences of reasonably natural events should count as causes and effects, and
also that positive events that are causes and effects may be represented by binary variables that
have the absence of these events as one of their values.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 47

JFK’s death includes a variable representing whether McCartney gets in the

way. Consequently no such model will treat his failure to do so as a cause.
Adopting (Serious Possibilities) results in a certain amount of vagueness and,
perhaps, subjectivity. But traditional accounts of causation – which don’t appeal
to causal models – also stand in need of appeal to serious possibilities because,
for instance, McCartney’s failure to get in the way is an inus condition of JFK’s
death and the latter counterfactually depends upon the former (cf. Collins 2000,
esp. 229; Woodward 2003, 86–8; McGrath 2005).44
The idea behind (Stability) is that an appropriate model is a sufficiently rich
representation of causal reality that moving to a richer representation by includ-
ing additional variables representing further factors (and/or by representing
already-represented factors in more detail) wouldn’t reveal causal verdicts to
be spurious.45 This requirement just reflects the obvious fact that we don’t want
to make causal judgements on the basis of a model that’s too impoverished to
accurately reflect causal reality.
It should be easy for the reader to verify that the models EP and LP, as well
as the models to be considered in the remainder of this section, satisfy (Parti-
tion), (Independence), and (Veridicality) (where the latter is assessed relative to
the worlds of the scenarios that they model) and that they don’t represent overly
gerrymandered or extrinsic states, thus satisfying (Naturalness). As already
noted (Serious Possibilities) is vague, but most participants in the literature
accept standards for what counts as a ‘serious possibility’ that don’t rule either
EP or LP as inapt. Specifying exactly what these standards amount to is tricky:
after all, there are combinations of possible values for the variables that figure
in EP and LP that correspond to intuitively weird possibilities – e.g. that Alice
and Bob shoot but the deer doesn’t die or, weirder still, in the case of LP, that
the deer is dead at 1:00pm but not at 1:01pm. But let’s assume for the sake of
argument there’s a way of cashing out (Serious Possibilities) so that (as advo-
cates of the SEM approach seem to desire) models like EP and LP don’t count
as violating it, but (say) a model of JFK’s assassination that includes a variable
corresponding to McCartney’s jumping in the way of the bullets does.46

44 If one regards appeal to ‘serious possibilities’ in an account of causation as too problematic, one
option would be to simply admit that, e.g., McCartney’s failure to get in the way was a cause
of JFK’s death. If so, as Blanchard and Schaffer (2015, 198) point out, (Serious Possibilities)
‘may be reinterpreted, not as an aptness constraint on models, but as a descriptive psychological
claim about which causal models are most readily available to us when we form our causal
judgements’.
45 (Stability) thus renders the notion of an appropriate model relative to the causal claim being
evaluated.
46 For further discussion, see Hitchcock (2001a, 295–8).

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 48 Philosophy of Science

Convincing ourselves that (Stability) is satisfied is also something of a dif-

ficult task: it doesn’t seem that there can be any proof of the non-existence
of unmodelled variables whose inclusion would overturn causal verdicts. This
could be seen as a drawback of the SEM approach. On the other hand, one
might argue that the discovery of relevant but previously unmodelled variables
is precisely the sort of thing that should lead us to reconsider our causal verdicts
and that, as any scientist will tell you, it’s just a fact of life that we can never
attain total certainty that all critical variables have been taken into account. In
any case, given the information we’re provided with in Early Preemption and
Late Preemption, it doesn’t seem that there’s any way of enriching EP or LP
to undermine the causal verdicts that we reached. In what follows, I’ll occa-
sionally comment on cases where there seem to be obvious enrichments of the
models that we consider and evaluate whether causal verdicts are overturned
when these enrichments are made.
Of course one thing that ought to trouble the advocate of the structural equa-
tions approach is if there are systematic biases in the models that occur to
us. This could potentially render our judgements about the likely stability of
a given model unreliable. There are at least two potential sources of bias that
have been pointed out in the literature. First, Glymour et al. (2010) have pointed
to an excessive focus on simple models with small numbers of variables. Such
a focus likely reflects the cognitive limitations of humans. Second, it might
be that our causal presuppositions influence our model construction (Paul and
Hall 2013, 4.2.8–9).47 Whilst we don’t have space to do justice to this topic
here, it seems a rich vein for further research that is liable to be interdiscipli-
nary both in its nature (since the input of specialists in psychology and machine
learning would appear to be crucial) and in its implications (since it’s relevant
to an assessment of the objectivity of model-building in the sciences).

4.6.4 A Problem for SEA

Even if SEA fares well with respect to Early Preemption and Late Pre-
emption, it runs into difficulties with Symmetric Overdetermination. The
trouble is that, in this case, there doesn’t appear to be any feature of the
actual situation that we can hold fixed to reveal de facto dependence. The
following model (‘SO’) would seem to be a reasonable one for Symmetric
Overdetermination:

47 Paul and Hall focus in particular upon the influence that such presuppositions have on our judge-
ments about relations between variables (i.e. what structural equations we judge to be true). This
is liable to affect our judgements about (Veridicality) as well as (Stability).

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 49

SO

Variables: fA; B; Dg

Interpretation: A D 1=0 if Alice shoots/doesn’t shoot; B D 1=0 if Bob

shoots/doesn’t shoot; D D 1=0 if the deer dies/survives.

Equations: A D 1, B D 1, D D MaxfA; Bg

Figure 4: Symmetric Overdetermination

There’s only one path from A to D and only one variable off this path, namely
B. So the test for causation appealed to by SEA requires that D D 1 de facto
depends upon A D 1 when B is held fixed at its actual value B D 1. But there’s
no such dependence: the deer’s death doesn’t counterfactually depend upon
Alice’s shot when we hold fixed Bob’s shot. The falsity of the counterfactual
A D 0&B D 1 € D D 0 can be verified via the equation replacement method
(described in Section 4.6.1) with respect to SO. Strictly analogous considera-
tions show that SEA yields the verdict that B D 1 isn’t a cause of D D 1 either.
This is a problem to the extent that we regard symmetric overdeterminers as
causes.
The most obvious way to try to deal with Symmetric Overdetermination
is to liberalise the analysis by allowing that we may sometimes vary features
of the actual situation in seeking latent dependence of effect on putative cause.
This is an approach taken by Hitchcock (2001a, 289) and Halpern and Pearl
(2005), inter alia. For instance, were we allowed to suppose that Bob didn’t
shoot (and hold this fixed by including it in the antecedent of the relevant ENF),
then we’d recover counterfactual dependence of the deer’s death on Alice’s
shooting.
But we don’t want to be too liberal: if we’re allowed to vary just any feature
of the actual situation, we’ll end up generating spurious cases of causation. For
instance, consider the intention of Bob to shoot if Alice doesn’t kill the deer in
Late Preemption. If we hold fixed the non-actual fact that Alice didn’t shoot

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 50 Philosophy of Science

(by including it in the antecedent of the relevant ENF), then the deer’s death
comes to counterfactually depend upon Bob’s intention. Yet Bob’s intention
isn’t a cause of the deer’s death.
Briefly, the way this problem is standardly tackled (Hitchcock 2001a, 289;
Halpern and Pearl 2005, 853–5) is by noting that, in Late Preemption, if
Alice hadn’t shot (A D 0), then the causal path between Bob’s intention and
the deer’s death would have been crucially different: Bob would have shot
rather than not doing so (i.e. B would have taken B D 1 rather than B D 0
in LP). In contrast, in SO, setting A D 0 doesn’t make any difference to the
causal pathway between Bob’s shooting and the deer’s dying (and this is so
even if we enriched SO by interpolating variables between B and D, such as
one representing the presence of the speeding bullet mid-air between Bob and
the deer). Setting A D 0 is therefore permitted in the latter case but not the
former.

4.6.5 Token vs Type Again

It was remarked at the outset of this section that counterfactual analyses of cau-
sation have typically been directed at understanding token causation. Having
said that, when it comes to approaches that invoke variables, it’s possible to ana-
lyse a family of causal relations that aren’t in themselves type or token causal
relations, but in terms of which it might be possible to analyse type causation:
namely, a family of relations of causation between variables as opposed to the
(actual) values of those variables. Causal relations between variables are dis-
cussed quite extensively by Pearl (2009), Spirtes et al. (2000), and Woodward
(2003).
Woodward (2003) distinguishes various relations of causation between vari-
ables. Direct causation (Woodward 2003, 55), which is an inherently model-
relative notion of causation, is defined by him so that X counts as a direct cause
of Y in a model iff X is a parent of Y. On the other hand, X is defined as a total
cause of Y iff there’s a possible change to the value of X that makes a difference
to the value of Y (Woodward 2003, 51).48 This isn’t an inherently model-relative
notion of causation.
Finally, X is defined as a contributing cause of Y relative to a model iff there’s
a path P from X to Y and some possible setting of the off-path variables (which
doesn’t have to be their actual values) such that, when the off-path variables are

48 Formally: X is a total cause of Y iff there’s a pair of values x, x0 (x ¤ x0 ) of X (neither of which

need be the actual value) and a pair of values y, y0 (y ¤ y0 ) of Y such that, had X taken X D x,
then Y would have taken Y D y and had X taken X D x0 , then Y would have taken Y D y0 .

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 51

held fixed at that setting, there’s a possible change to the value of X that makes
a difference to the value of Y (Woodward 2003, 57).49 Though this definition is
model-relativised, one might define X to be a contributing cause of Y simpliciter
iff X is a contributing cause of Y relative to at least one apt model.50
These definitions of various causal relations that might obtain between vari-
ables don’t invoke the actual values of the cause and effect variables. This
sets them apart from definitions of actual causation (i.e. the token causal rela-
tion that has been of central philosophical focus). Still, they don’t necessarily
capture type-causal notions either (cf. Hitchcock 2007, 503–4). Specifically,
there’s no requirement that the variables in question represent type-level phe-
nomena. For example, according to these definitions, A counts as a direct, total,
and contributing cause of D1W00 in our model LP of Late Preemption. Yet A
and D1W00 represent quite specific things: namely, whether Alice shoots when
she has the chance, and whether the deer is dead at 1pm. In general, where
variables represent particulars such as token events, causation between such
variables isn’t ‘type causation’ in the usual sense. Hitchcock (2007, 503–4)
suggests that we might in such cases regard these relations as part of what he
calls the ‘token causal structure’ of a scenario, but not as instances of actual
causation.51
Still, it’s of course possible to use variables to represent things that are non-
particular. For instance, consider the Arrhenius Equation (AE) in chemistry.
The AE shows how the rate of a chemical reaction depends upon temperature.
It can be formulated as follows:
Ea
k D Aexp RT : (AE)

49 Formally: X is a contributing cause of Y iff there’s a pair of values x, x0 (x ¤ x0 ) of X (neither

of which need be the actual value), a pair of values y, y0 (y ¤ y0 ), and an assignment of values
to all off-path variables such that (i) had X taken X D x and had the off-path variables taken
the assigned values, then Y would have taken Y D y and, (ii) had X taken X D x0 and had the
off-path variables taken the assigned values, then Y would have taken Y D y0 .
50 Woodward (2003, 51, 57) provides probabilistic versions of his definitions of total and
contributing cause, which will be considered in Section 5.6.
51 Of course, where the variables in our model concern particulars, we could build on Woodward’s
definitions to define a variety of ‘token causal relations’. For instance if x and y were required
to be the actual values of X and Y in the definition of ‘total cause’ given in Footnote 48, then
we’d have a ‘token causal relation’ that was defined in terms of straightforward counterfactual
dependence (though explicitly incorporating contrastivity). Still, as should be clear in light of
the foregoing subsections, none of these definitions (even when restricted to focus on the actual
values of X and Y and the requirement that the variables concern particulars) seems particularly
well aligned with ‘actual causation’: the token causal relation that’s been the subject of so much
philosophical interest (as well as interest from lawyers, psychologists, and computer scientists)
and that seems to be the central token causal concept of scientists and laypeople.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 52 Philosophy of Science

Here k is the reaction rate (the amount of reaction product produced per litre of
the reaction solution per second), A and Ea are constants whose values depend
upon the nature of the reactants, R is the universal gas constant, and T is absolute
temperature. The AE captures the fact that small increases in temperature result
in dramatic increases in reaction rate. AE is a single-equation SEM where k and
T are the variables.
Relative to AE, temperature T counts as a direct, total, and contributing cause
of the reaction rate k by Woodward’s definitions. Moreover, the variables in
AE don’t concern particulars. They are general in the sense that the AE can be
applied to any reaction rather than being limited to a single one. As applied to
variables like these, Woodward’s notions of causation between variables can
reasonably be thought of as capturing type-level relations.

4.6.6 Interventions

Before closing our discussion of the counterfactual approach it’s worth noting
that, with the notion of a SEM in hand, we’re now in a position to more precisely
characterise the notion of an ‘intervention’, which plays an important role in
some theorists’ thinking about the SEM approach to causation, and which we’ll
appeal to again in our discussion of probabilistic causation. Our characterisa-
tion will be a slightly simplified version of that given by Woodward (2003,
94–114).
On Woodward’s approach, interventions are characterised in terms of ‘inter-
vention variables’. The notion of an intervention variable can be understood in
terms of a three-place relation between variables: ‘I is an intervention variable
for X with respect to Y’, where X is the variable that’s the subject of the inter-
vention, and the intervention is one that can test whether Y depends upon X. We
can keep track of this using subscripts: a variable that’s a putative intervention
variable for X with respect to Y can be denoted IX;Y .
According to Woodward’s definition, IX;Y counts as an intervention variable
for X with respect to Y iff (a) IX;Y is a contributing cause of X (in the sense
defined in the previous subsection); (b) IX;Y acts as a ‘switch’ for X: that is,
there are values of IX;Y (call these ‘switching values’) such that when IX;Y takes
those values, the value of X no longer depends on the values of any variables
besides X that don’t lie intermediate on a directed path from IX;Y to X; (c) there’s
no directed path from IX;Y to Y that doesn’t go via X; (d) there’s no variable that’s
an ancestor of IX;Y that lies on a directed path to Y that doesn’t run via IX;Y (i.e.
there’s no ‘common cause’ of IX;Y and Y).52

52 It seems there’s something of a regress in the offing in the definition of an intervention vari-
able. As we’ll see, intervention variables are supposed to test what counterfactuals are true and

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 53

The idea, then, is that there are circumstances under which IX;Y makes a dif-
ference to the value of X (condition (a)) and indeed there’s a certain set of values
of IX;Y such that when IX;Y takes one among those values, the value of X doesn’t
depend upon those of any other variable besides IX;Y (except any that lie on a
path from IX;Y to X) (condition (b)). Moreover, IX;Y isn’t a variable that poten-
tially influences Y independently of its influence on X: any influence that IX;Y
has on Y goes by way of its influence upon X (condition (c)). Finally, there’s no
variable that’s a ‘common cause’ of IX;Y and Y so there’s no correlation between
IX;Y and Y that isn’t explained by the influence of IX;Y upon X (condition (d)).
The latter two conditions are designed to ensure that any changes in the value of
Y associated with changes in IX;Y are entirely due to changes that IX;Y produces
in X. This ensures that any such changes in the value of Y genuinely reflect its
dependence on X.
It’s consistent with the definition of an intervention variable that, when the
value of IX;Y lies outside the set of ‘switching values’, its value doesn’t make a
difference to that of X. In such circumstances, there might be apt models from
which this variable can be omitted (i.e. left as latent). Of course if the variable
is then ‘added’ to the model, this will require an adjustment to the structural
equations so that the equation for X now reflects its dependence upon IX;Y in
addition to its other parents in the model.
To illustrate, recall the example Early Preemption. In this scenario, the bul-
let’s passing the mid-point between Alice and the deer might be represented by
a variable M which takes M D 1 if the bullet indeed passes this mid-point and
M D 0 otherwise (see Figure 5). Now imagine a putative intervention on M
with respect to D (the variable representing the deer’s death) that comprises
someone’s placing an object that would intercept the speeding bullet before it
reached this mid-point. Whether such an object is placed might be represented
by a binary variable IM;D that takes value IM;D D 1 if one is or IM;D D 0
otherwise.
Consider whether IM;D is a genuine intervention variable for M with respect
to D, the variable representing whether the deer dies. IM;D acts as a ‘switch’
with respect to M in the sense that, when IM;D D 0, the value of M depends
upon the value of A – representing whether Alice shoots – but when IM;D D 1,
then M D 0 no matter what the value of A. Moreover, as is implied by its acting
as a switch, IM;D is a contributing cause of M (i.e. there are circumstances under

therefore what the structural equations are – and hence what the arrows in the graph are – in a
model satisfying (Veridicality). But don’t we need an intervention variable for the intervention
variable in order to establish what arrows emanate from IX;Y itself, and therefore whether IX;Y
satisfies (for instance) condition (c)? This issue is taken up in (Glynn 2013).

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 54 Philosophy of Science

Figure 5 An Intervention Variable for M with respect to D

which it makes a difference to the value of M). Now if no one actually places an
object to intercept the bullet (indeed, perhaps there’s no-one present who has
any intention or capacity to do so), then it would be entirely reasonable to omit
IM;D from our model altogether. However, once IM;D is included (as it is in the
model represented in Figure 5), it can be verified that it satisfies the remaining
conditions for an intervention variable for M with respect to D. Specifically,
there’s no directed path from IM;D to D that doesn’t go via M and there’s no
variable that’s an ancestor of IM;D that lies on a directed path to D that doesn’t
run via IM;D .53
Note, by contrast, that if we added a variable that was a parent of A (perhaps
representing whether someone shouts at Alice to stop her from shooting) to our
model then this wouldn’t count as an intervention variable for M with respect
to D (though – so long as such a shout wouldn’t scare off the deer – it may
count as an intervention variable for A with respect to D). That’s because there
would be a directed path from such a variable to D that bypasses M: namely
the one that runs via A and B.
When it comes to conditions (c) and (d) for an intervention variable, it’s of
course not enough that they be satisfied in just any old model that includes IX;Y ,
X, and Y in its variable set. Rather it will have to be the case that, no matter how
many variables we include in our model, then (provided the resulting model sat-
isfies (Veridicality), (Naturalness), etc.) conditions (c) and (d) still hold. That’s
because if there’s a path from IX;Y to Y that bypasses X that could be revealed by
the addition of more variables to the model, or if there are unmodeled common
causes of IX;Y and Y, then changes in the value of IX;Y won’t reliably test the
influence of X upon Y.

53 It might be objected that – if in fact there’s no one around with the intention and capacity to place
an object that would intercept the bullet – a model including IM;D violates (Serious Possibilities).
One possible response would be to claim that, while we may wish to think about interventions
(which may in some cases correspond to far-fetched possibilities) when considering whether
the counterfactuals entailed by a causal model are true, and while to help us do so we may wish
to expand the original model to include variables representing these interventions, it isn’t the
expanded model of which we need (Serious Possibilities) to hold, but only the original model.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 55

So far we’ve only defined the notion of an ‘intervention variable’. But the
definition of an intervention in terms of an intervention variable is simple
enough. Setting aside some subtleties discussed by Woodward (2003, 94–114),
we can think of an intervention on X with respect to Y as simply X’s taking
some value X D x as a result of an intervention variable IX;Y taking one of its
switching values. So, in Early Preemption, IM;D taking IM;D D 1 serves as
an intervention on M with respect to D that sets M D 0. The interventionist’s
proposal is that when evaluating counterfactuals for the purposes of analyzing
causation, we should suppose their antecedents are realised by interventions.
This is taken to be equivalent to adopting the ‘equation replacement’ method
of evaluating counterfactuals introduced in Section 4.6.1. That’s because both
methods involve a change to the target variable, X, that preserves the functional
relationship between X and Y as well as affecting neither the values of ances-
tors of Y that don’t lie on a pathway between X and Y nor any influence that
those ancestors have on Y that isn’t mediated by X.54 In the next section, we’ll
see that the notion of an intervention can also potentially help with analysing
probabilistic causation.55

5 Probabilistic Causation
David Hume – whose regularity theory was examined in Section 2 – wrote
at a time when Newtonian mechanics ruled the roost. While it has recently
been argued that Newtonian mechanics isn’t strictly deterministic,56 it can be
treated as such for most intents and purposes, and in any case doesn’t yield well-
defined non-trivial (i.e non-0 or 1) probabilities for the unfolding of events
over time. At a time when it was reasonable to think that the universe was
fundamentally deterministic, it might have seemed plausible to think that all
causes are, in the circumstances in which they occur, lawfully sufficient for
their effects.
But Newtonian mechanics is now known not to be strictly correct, and has
been supplanted by quantum mechanics (QM) and GTR. On the orthodox

54 The assumption that it’s possible to intervene upon each variable V in a model of a causal system
– thus overriding the usual functional dependence described by the equation for V – without
impacting the functional relationships described by any of the other equations in the model is
referred to as the assumption of ‘modularity’. For a defence of this assumption, see Hausman
and Woodward (1999); for criticism, see Cartwright (2002, 2004).
55 Note that, on the foregoing characterisation of an intervention, interventions needn’t be human
actions. Rather, natural processes can serve as interventions provided they have the formal char-
acteristics described in this section (see Woodward 2003, 103–4). So there’s nothing particularly
anthropocentric about adopting an interventionist semantics for the counterfactuals needed to
analyse causation.
56 See Earman (1986, esp. 29–40), Malament (2008, 803–4) and Norton (2008).

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 56 Philosophy of Science

interpretation of QM, and on several heterodox interpretations, the world is

fundamentally probabilistic. If one of these interpretations is correct, then there
might be causes that are neither sufficient for their effects nor elements of sets
of conditions that suffice for their effects. Rather, at least some causes might
(in the circumstances in which they occur) simply confer a certain probability
upon their effects. The example Probabilistic Bomb, given in Section 2.5.2 is
an example of causation involving quantum probabilities and was seen to pose
(insurmountable) difficulties for the regularity approach.
Cases of probabilistic causation don’t just pose problems for regularity anal-
yses, but also for the counterfactual analyses examined in the previous section.
To see this, suppose that, in Probabilistic Bomb, the bomb is itself an inher-
ently unstable device: it has some probability of spontaneously exploding even
if the threshold reading of the Geiger isn’t reached. Perhaps the bomb is a
nuclear device with a Uranium-235 core. There’s some (small) probability that
enough U-235 atoms undergo spontaneous ˛-decay to start a chain reaction
resulting in the bomb’s exploding. In such a case, it appears to be false that if I
hadn’t placed the radioactive material near the Geiger, the bomb wouldn’t have
exploded. After all, it might still have (spontaneously) exploded, since there
would have been a residual probability of its doing so. 57 Indeed it might still
have exploded at more-or-less the same time at which it did in fact explode
(and perhaps – depending on the workings of the bomb – in much the same
manner).58
Probabilistic Bomb might seem a contrived example: it’s a deliberate
attempt to show that quantum probabilities can have significant macroscopic
consequences. But it’s not just in fundamental physics that probabilities figure
heavily. Probabilities are also widespread throughout the higher-level sciences,
including statistical mechanics, genetics, ecology, chemistry, meteorology, and
economics, to name just a few. There are some deep and difficult scientific

57 As Lewis (1986b, 180–4) argues, because the probabilities involved here are (assuming ortho-
dox QM) objective, it’s simply wrong to think that, although the bomb would still have had
a certain (small) probability of exploding, there’s nevertheless some fact of the matter about
whether it would or wouldn’t have exploded in such circumstances.
58 One might object as follows (thanks to an anonymous referee for pressing me on this point).
Suppose that, for example, the bomb’s detonation mechanism consists in the shooting of one
piece of U-235 into another and that in the actual world the impact between the two pieces occurs
at 12 noon. Then, one might argue, if an insignificant number of U-235 atoms in the bomb’s
core in fact spontaneously decayed prior to 12 noon, then it’s true after all that, if I hadn’t placed
my radioactive material near the Geiger, then the bomb wouldn’t have exploded. A worry about
this response, however, is that there would still have been a chance of significant spontaneous
decay occurring just after 12 noon. One may be forced to accept reasonably stringent standards
of event fragility if one is to maintain that the resultant explosion is a different event from that
which occurred in the actual world and therefore that the explosion counterfactually depended
upon my action.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 57

and philosophical questions to be asked concerning (a) how, if at all, the

probabilities of these higher-level sciences relate to the probabilities of QM;
and (potentially relatedly) (b) to what extent these probabilities are objective,
as opposed to simply being introduced into scientific models and theories to
reflect our partial ignorance of the world.59 What seems clear however is that,
firstly, viewed through a macro-lens our world appears to be probabilistic and,
secondly, in light of the fact that probabilistic theories prevail at every level all
the way down to QM, it’s doubtful whether, even if we knew enough of the fine
details of the universe, we’d be able to uncover sufficient (or indeed necessary)
causes for each effect.
To illustrate, consider the causal relation between asbestos inhalation and
mesothelioma. The process by which the former brings about the latter is com-
plex, and there isn’t yet consensus on the details. But one study (Yang et al.
2010) suggests the following. Asbestos is toxic to human mesothelial cells
(which line the pleural cavities in the lungs). The damage that asbestos fibres do
to these cells results in the protein HMGB1 being released from them. The pres-
ence of extracellular HMGB1 leads macrophages (a type of white blood cell)
to release another protein, TNF-˛. TNF-˛ activates NF-B in damaged mes-
othelial cells. Activated NF-B helps some of these damaged cells to survive.
These damaged cells can then develop into cancer cells.
Assume that some picture like this is right. Then, even in the case of an
individual person, the causal process by which asbestos exposure can bring
about mesothelioma is probabilistic. It depends, for instance, upon the ‘ran-
dom walks’ followed by asbestos fibres experiencing Brownian motion in the
air and in the lungs. It depends upon the probabilistic process of osmosis (of
HMGB1 and TNF-˛), which comes under the purview of statistical mechan-
ics. And it depends upon molecular bonding processes (which are involved in
the stimulation of macrophages by HMGB1, and in the activation of NF-B by
TNF-˛), which are subject to QM probabilities.
It thus seems that we ought to make room in our theorising about causation
for probabilistic causation since it appears that at least some, and perhaps even
the vast majority, of causation in our universe is probabilistic. Many scholars
working in the regularity and counterfactual traditions explicitly acknowledge
that their accounts are designed only to handle the deterministic case (see, e.g.,
Lewis 2004a, 79–80; Baumgartner 2013, 87; Strevens 2007, 95; and Hitchcock
2001a, 275; 2007, 498). In what follows, we’ll explore some of the leading
accounts of probabilistic causation. In the course of doing so, we’ll see that the

59 These questions are discussed by, inter alia, Albert (2000), Loewer (2001), and Callender and
Cohen (2010).

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 58 Philosophy of Science

challenges confronting accounts of deterministic causation – such as preemp-

tion – have their probabilistic analogues, but that there are also novel challenges
that arise in the probabilistic case.

5.1 Accounts of Probabilistic Causation

A number of authors have proposed accounts of probabilistic causation that
appeal to a notion of probability-raising. Classic accounts include those of Sup-
pes (1970), Reichenbach (1971, see esp. 204), and Lewis (1986b, Postscript
B). In Probabilistic Bomb, for example, my placing the radioactive material
near to the Geiger raises the probability of the bomb’s exploding, even though
the bomb may have retained some residual probability of exploding (due to its
inherent instability) had I not done so.
There’s more than one way of understanding the notion of ‘probability-
raising’. The two most popular are in terms of (i) conditional probabilities;
and (ii) counterfactuals. The latter approach can thus be seen as an extension,
to the probabilistic case, of the counterfactual tradition of attempts to analyse
causation. But first let’s focus on the conditional probability approach.60 Let’s
also initially focus upon actual causation (of which Probabilistic Bomb is an
instance).61
Let c and e be token events with O.c/ and O.e/ being, respectively, the prop-
osition that c occurs and the proposition that e occurs. Then, in the conditional
probability sense, c’s raising the probability of e is understood in terms of the
following inequality:

P.O.e/jO.c// > P.O.e/j:O.c// (1)

This says that the probability of e’s occurring conditional upon c’s occurring is
greater than the probability of e’s occurring conditional upon c’s not occurring:
in Probabilistic Bomb, for example, the probability of the bomb’s explod-
ing conditional upon my placing the radioactive material near the Geiger is
greater than the probability of the bomb’s exploding conditional upon my not
doing so.
There are, however, at least three problems with understanding probability-
raising in terms of conditional probabilities if causation is, in turn, to be
understood in terms of probability-raising (see Lewis 1986b, 178–9). The first

60 This is the more traditional approach, adopted by Suppes (1970) and Reichenbach (1971),
among others.
61 Probabilistic type-causation is discussed in Section 5.6.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 59

is that it’s a straightforward consequence of the probability calculus that if

Inequality 1 holds, then so does:

P.O.c/jO.e// > P.O.c/j:O.e// (2)

Understood in terms of conditional probabilities, probability-raising therefore

doesn’t have the asymmetry that we associate with causation.
A second, related problem is that, understood in the conditional probability
sense, independent effects of a common cause typically raise the probability
of one another. Recalling the example of Section 2.5.3, the probability of a
storm is higher conditional upon a barometer predicting a storm than it is con-
ditional upon the barometer’s not doing so. The barometer would be useless if
its readings were not correlated with the weather in this way. But this corre-
lation doesn’t exist because the barometer reading is a cause of the storm, but
rather because an earlier fall in atmospheric pressure is very probable condi-
tional upon the barometer’s prediction of a storm and a storm is very probable
conditional upon a fall in atmospheric pressure.
One way of responding to these two problems is as follows. First, it appears
that the probabilities of an event evolve over time, at least when we con-
sider objective probabilities (Lewis 1986b, 91) of the sort that are presumably
relevant to analysing causation.62 For instance, in Probabilistic Bomb, the
probability of the bomb’s exploding once the Geiger threshold is reached is
higher than prior to its being reached. So we need some specification of which
time’s probabilities are the relevant ones for evaluating causal claims. A natural
suggestion is that we should consider the (conditional) probabilities that obtain
just before the putative cause occurs.
Now suppose that c is in fact a cause of e. Then, assuming that causes occur
prior to their effects, there should be a time t sufficiently close to e’s occurrence
by which c has already occurred. Once it has occurred, c has a probability 1 of
occurring. So, at t, c has a probability 1 of occurring. It follows via the probabil-
ity calculus that, at t, the probability of c occurring conditional either upon the
occurrence of e or upon the non-occurrence of e is also 1 (provided e doesn’t
have probability 1 by that time – an issue we’ll discuss shortly). But then e
doesn’t raise the probability of c according to the time t probabilities, which
are the relevant ones for assessing whether e is a cause of c.63 This approach

62 I’ll say more about the interpretation of probabilities in Section 5.6.

63 A residual concern is that there might not be a time prior to the occurrence of e by which all of e’s
causes have occurred. For instance, if time is continuous and there’s no unmediated action at a
temporal distance, then presumably e must have an infinite series of more and more temporally
proximate causes. So, no matter what time prior to the occurrence of e we choose, some of

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 60 Philosophy of Science

also helps with the problem of common causes. If, say, we take the probabilities
relevant to assessing whether the barometer reading was a cause of the storm to
be those that obtain immediately before the barometer indicates a storm, then
since the atmospheric pressure has already fallen by that time (and so has proba-
bility 1 of doing so), the barometer reading by then doesn’t raise the probability
of the atmospheric pressure falling and hence doesn’t raise the probability of
the storm.64
There is, however, a third problem with understanding probability-raising in
terms of conditional probabilities. This problem is more technical. Conditional
probabilities are standardly defined as a ratio:

P.A&B/
P.AjB/ D (3)
P.B/
But ratios are undefined when the denominator is equal to zero. Now suppose
that event c has probability 1 of occurring shortly before it does occur. Then,
at that time, P.:O.c// D 0 and so the conditional probability on the RHS
of Inequality 1 is undefined. The concern is thus that, if probability-raising is
understood in terms of conditional probabilities, we appear not to be able to
say that events that have probability 1 (shortly before their occurrence) raise
the probability of (and therefore cause) anything.
It’s possible that, in our world, no event has probability 1 before – even
shortly before – it occurs. But even if so, wouldn’t we like our probabilistic
analysis of causation to apply to deterministic worlds too? A plausible thought

e’s causes won’t yet have occurred by that time and might retain a probability less than 1. If
so, we’re liable to get the result that e raises the probability of those causes in the conditional
probability sense.
64 The solution under consideration in the last two paragraphs of the main text has appealed to
the notion that objective chance is time-relative. Certainly we often speak as though this is
so. For instance, as I write, it’s reasonable to say ‘the chance of average global temperatures
increasing by > 0:5ı C by 2050 is increasing each day the international community puts off
tough decisions’. But we can allow this without committing to the notion that time-relativity is
a fundamental feature of chance. For example, Hoefer (2007, 562) regards chances as relativised
to chance setups (e.g. a particular coin flip might be a chance setup relative to which the chance
of heads is 0.5) and not as in any fundamental way relativised to times (Hoefer 2007, sect. 3.2).
Commenting on David Lewis’s view of chances as time-relative, Hoefer (2007, 564–5) says that
‘[f]or Lewis, a non-trivial time-indexed objective probability Prt .A/ is, in effect, the chance of
A occurring given the instantiation of a big setup: the entire history of the world up to time t.’
(Hoefer himself makes the plausible allowance that objective chances may be defined given
setups that are more limited: for instance, there’s a well-defined chance of a silver atom’s being
deflected in a particular direction given that it’s fired through a Stern-Gerlach device.) One way
of construing this is to adopt an analysis of time-relativised objective probabilities according
to which Pt .A/ Ddef P.AjHt / where Ht is a proposition characterising the entire history of
the world through time t. That is, objective probability isn’t inherently time-relativised, but a
time-relativised notion of objective probability can be defined by conditioning the objective
probability distribution upon history up to the time in question.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 61

is that deterministic worlds are simply the special case where the set of each
event’s causes raises its probability all the way to 1. But in a deterministic
world each event has probability 1 of occurring (shortly before it occurs). So it
seems that no event could raise the probability of any other in the conditional
probability sense. Since there surely is causation in deterministic worlds (if
Bohmian mechanics turns out to be true then ours is a deterministic world – at
least at the fundamental level), a probabilistic analysis that invoked conditional
probabilities would fail to apply to some possible cases of causation.65
The counterfactual approach to probability-raising, advocated by Lewis
(1986b, 175–9), seemingly overcomes these difficulties. To see how the coun-
terfactual approach interprets probability-raising, suppose that c and e are
actual events and that, at a time t just after the occurrence of c, the (uncon-
ditional) probability of e’s occurring is Pt .O.e// D x (the subscript to the
probability function indicates that this is the objective probability distribution
that obtains at t). Then consider the following counterfactual:

:O.c/ € Pt .O.e// < x (4)

This says that, if c hadn’t happened, then the probability at time t of e occur-
ring would have less than x (the actual probability of e at time t). This is a
counterfactual understanding of what it is for c to raise e’s probability.
This approach has the capacity to overcome the problems associated with
the conditional probability approach (Lewis 1986b, 175–9). For, firstly, the fact
that :O.c/ has zero probability in the actual world doesn’t imply that there’s
no world in which it’s true, and doesn’t stand in the way of O.e/ having a well-
defined probability in such a world. For instance, in a deterministic world, it’s
plausible that, where c and e are actual events and c is a non-overdetermining
cause of e, then P.:O.c// D 0 and P.O.e// D 1 and yet, evaluated with respect
to such a world, it’s a true counterfactual that if c hadn’t occurred, then the
probability of e would have been 0.
Second, assuming counterfactuals don’t backtrack, this approach avoids the
problem of effects raising the probability of their causes and also the prob-
lem of common causes. That’s because, on the non-backtracking reading of
counterfactuals, it’s not true that (for example) if the barometer hadn’t said
there was going to be a storm, then the probability of the atmospheric pressure
having earlier fallen would have been lower. Rather, because the atmospheric
pressure has already fallen, its probability of doing so is 1 by a time t just after

65 At best it would seem that it could serve as an empirical analysis of causation, in the sense
discussed in Section 3.4.1.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 62 Philosophy of Science

the barometer says there’s going to be a storm, and it’s also 1 at t in the near-
est worlds (according to a similarity metric that implies non-backtracking) in
which the barometer doesn’t indicate that there’s going to be a storm. Moreo-
ver, because the barometer reading doesn’t raise the probability of the fall in
atmospheric pressure in the counterfactual sense, it doesn’t raise the probability
of the storm either.
As in the deterministic context, one can achieve the desired non-backtracking
results about counterfactuals by appealing to interventions (as an alternative to
Lewis’s similarity semantics). Though it will be easier to be formally precise
about the notion of an ‘intervention’ in the probabilistic case once we have
the notion of a probabilistic causal model – to be introduced in Section 5.3
– in hand, we can for the time being just think of interventions as exogenous
manipulations of variable values. As in the deterministic case, variables can be
used to represent the occurrence or non-occurrence of events.
Goldszmidt and Pearl (1992) introduce a special notation do.X D x/ to rep-
resent that the value of variable X is set to X D x by means of an intervention.
They then use P.
jdo.X D x// to represent the probability distribution that
would result from setting X to X D x by means of an intervention. Using that
notation, the interventionist counterfactual approach to probability-raising says
that – where C and E are binary variables which each have the possible values
1 and 0 – C D 1 raises the probability of E D 1 iff:

P.E D 1jdo.C D 1// > P.E D 1jdo.C D 0// (5)

Despite the resemblance in the notation, terms like P.E D 1jdo.C D 0// don’t
represent conditional probabilities. Rather, they represent counterfactual prob-
abilities: the probability for E D 1 that would obtain if C were intervened upon
to set C D 0. Thus Inequality 5 says that the probability for E D 1 that would
obtain if C were intervened upon to set C D 1 is higher than the probability of
E D 1 that would obtain if C were intervened upon to set C D 0.66 In virtue of
the non-backtracking character of interventionist counterfactuals, this inequal-
ity won’t obtain if the events represented by C D 1 and E D 1 are independent

66 Given that C D 1 must be the actual value of C if (the state represented by) C D 1 is to be an
actual cause of (the state represented by) E D 1, couldn’t we simply have the actual uncondi-
tional probability P.E D 1/ that obtains shortly after the occurrence of the state represented by
C D 1 figure on the LHS of Inequality 5 (as opposed to the probability for E D 1 that would
have obtained under an intervention setting C D 1)? The answer is ‘yes’ if one is consider-
ing whether C D 1 straightforwardly raises the probability of E D 1. However, as we’ll see
in the next subsection, some approaches to probabilistic causation appeal to a notion of latent
probability-raising that’s revealed by holding certain other variables fixed (by interventions).
On some approaches we’re allowed to hold common causes of C and E fixed at non-actual
values, in which case an intervention may well be needed to return C to its actual value.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 63

effects of a common cause or if E D 1 is the cause and C D 1 is the effect. For

instance, an intervention on the barometer reading with respect to the atmos-
pheric pressure or to the storm might consist of an experimenter grabbing the
barometer needle and twisting it so that it no longer points at the word ‘storm’
and instead points at ‘fair’. But, in such circumstances, the probability of the
fall in atmospheric pressure or of the storm wouldn’t be any lower.

5.2 Difficulties
Even when understood in a sense that rules out probability-raising between
independent effects of a common cause and probability-raising of causes by
their effects, there are (insurmountable) difficulties for any account that seeks
simply to identify causation with probability-raising. Cases of probabilistic pre-
emption can be used to illustrate that (even when understood in such a sense)
probability-raising is neither necessary nor sufficient for causation (Menzies
1989, 1996). An example is provided by the following variant on Probabilistic
Bomb.
Probabilistic Preemption

Someone (neither you nor I) has connected a Geiger counter to a bomb

so that the bomb will explode if the Geiger registers above a threshold
reading. I place a chunk of U-232, which has a half-life of 68.9 years
and decays by ˛-emission, near the Geiger. By chance, enough U-232
atoms decay within a short enough interval for the Geiger to reach the
threshold reading so that the bomb explodes. Unbeknownst to me, you’ve
been standing nearby observing. You have a chunk of Th-228, which has a
half-life of 1.9 years and decays by ˛-emission, which contains at least as
many atoms as my chunk of U-232. You’ve decided that you’ll place your
Th-228 near the Geiger iff I fail to place my U-232 there. Seeing that I
place my U-232 near the Geiger, you don’t place your Th-228 there.

In this case, my placing my U-232 near the Geiger is plausibly regarded as a

cause of the bomb’s exploding. Let M, D, Y, T, and E be binary variables which,
respectively, take value 1 if the following things occur (and 0 otherwise): I place
my U-232 near the Geiger; you decide to place your Th-228 near the Geiger
iff I don’t place my U-232 near the Geiger; you place your Th-228 near the
Geiger; the threshold Geiger reading is reached; the bomb explodes. Though
my act (M D 1) is a cause of the explosion (E D 1), the following inequality
holds:

P.E D 1jdo.M D 1// < P.E D 1jdo.M D 0// (6)

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 64 Philosophy of Science

That is, my placing my U-232 near the Geiger lowers the probability of the
bomb’s exploding. That’s because it strongly lowers the probability of your
placing your more potent Th-228 near the Geiger (your Th-228 is more potent
because Th-228 has a shorter half life than U-232 and your chunk of Th-228
contains at least as many atoms as my chunk of U-232). Probability-raising
(even in the interventionist counterfactual sense) is therefore unnecessary for
causation.
Probabilistic Preemption also illustrates the fact that probability-raising
(even in the interventionist counterfactual sense) isn’t sufficient for causation.
Specifically, notice that your decision (D D 1) to place your Th-228 near the
Geiger iff I don’t place my U-232 there raises the probability of the explosion.67
That is:

P.E D 1jdo.D D 1// > P.E D 1jdo.D D 0// (7)

Specifically, Inequality 7 holds provided there’s some chance that M D 0,

because D D 1 raises the probability that the bomb will still explode in the sce-
nario in which M D 0. Yet your decision (D D 1) isn’t a cause of the explosion
since you don’t place your Th-228 near the Geiger.68
Recall that one promising response to the fact that counterfactual depend-
ence isn’t necessary for causation in the deterministic case was to appeal to a
notion of de facto dependence, which we formalised with the help of SEMs.
It turns out that there’s an analogous response to the fact that probability-
raising isn’t necessary for causation: namely to appeal to a notion of de facto
probability-raising, that is, latent probability-raising that’s revealed by hold-
ing fixed certain features of the actual situation.69 In the present example,

67 At least if we assume that the alternative is for you to decide not to place your Th-228 near the
Geiger come what may. More will be said about causal contrastivity in the probabilistic case in
Section 5.4.
68 Cases of probability-raising without causation are discussed by Menzies (1989), Schaffer
(2001), and Hitchcock (2004), inter alia. Cases of causation without probability-raising are
discussed by Hesslow (1976), Rosen (1978), and Salmon (1980), inter alia.
69 Such an approach is taken by Hitchcock (2001b), Kvart (2004), and Fenton-Glynn (2017),
inter alia. Cartwright (1979, 423) endorses an an analogous approach to understanding type-
causation: specifically that a type-level factor C is a cause of a type-level factor E iff ‘C increases
the probability of E in every [population] which is otherwise causally homogenous with respect
to E’, where a situation is ‘causally homogenous’ with respect to E iff all other causal factors
for E are held fixed. Note that in the case of type causation it generally makes no sense to speak
of holding factors fixed ‘at their actual values’, since the values will vary across populations.
So Cartwright effectively requires that C increases the probability of E no matter what values
we hold these other factors fixed at. For a critique of this requirement, see Carroll (1991). Eells
(1991, 86–7) offers a similar definition of what he calls ‘positive causal relevance’, but also
defines ‘negative causal relevance’ (when C lowers the probability of E no matter what values
we hold these other factors fixed at), and ‘mixed causal relevance’ (e.g. when C raises the prob-
ability of E for some assignments of values to these other factors, but fails to raise it for others).

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 65

although my placing my U-232 near the Geiger doesn’t raise the probabil-
ity of the bomb’s exploding, latent probability-raising is revealed by holding
fixed the fact that you don’t place your Th-228 there. On the interventionist
approach, ‘holding fixed’ your not placing your Th-228 there means interven-
ing on the variable Y to keep it at Y D 0 whilst varying the value of M by
means of a further intervention. The result, captured by Inequality 8, is that
latent probability-raising is revealed:

P.E D 1jdo.Y D 0&M D 1// > P.E D 1jdo.Y D 0&M D 0// (8)

This inequality says that the probability of the explosion if interventions are
made to ensure that you don’t place your radioactive material near the Geiger
but I place mine there is higher than the probability of the explosion if inter-
ventions are made to ensure that neither of us places our radioactive material
near the Geiger.
Just as the features of the actual situation to be held fixed in seeking de facto
dependence in the deterministic case could be more clearly identified with the
aid of SEMs, so the features of the actual situation to be held fixed in seeking de
facto probability-raising can be more clearly identified with the help of prob-
abilistic causal models. Though the converse problem of probability-raising
without causation doesn’t have such a clear analogue in the deterministic case
– it’s much more plausible to take counterfactual dependence to be sufficient
for causation than it is to take probability-raising to be sufficient – it turns out
that appeal to probabilistic causal models opens up possible avenues for dealing
with the problem of probability-raising without causation too.

5.3 Probabilistic Causal Models

Adopting the interventionist framework discussed in the previous subsection,
a probabilistic causal model M ,70 can be thought of as comprising a set
of variables, V, together with a function, do.
/. The do.
/ function takes as
input a value of a variable in V – do.X1 D x1 / – or a conjunction of values
of different variables in V – do.X1 D x1 & : : : &Xn D xn / – and returns a
probability distribution over the possible values of the variables in V: the prob-
ability distribution that would obtain if X1 & : : : &Xn were set to the values
X1 D x1 & : : : &Xn D xn by interventions.

We’ll return to the question of whether (latent) probability-raising rather than, say, probability-
lowering should – as many accounts seem to assume – play a special role in our causal thinking
in Section 5.7.
70 I’ll use asterisks to indicate that a model is a probabilistic model, rather than a SEM.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 66 Philosophy of Science

We can construct a graphical representation of a probabilistic model by tak-

ing the variables in V as the nodes of the graph and drawing a directed edge
(‘arrow’) from a variable Xi to a variable Xj (Xi ; Xj 2 V) just in case there’s
some assignment of values to the other variables in V such that the value of Vi
makes a difference to the probability distribution over the values of Vj when
the other variables in V are held fixed by interventions at the assigned values.
To illustrate, consider Probabilistic Preemption. The key features of the
do.
/ function – to be explored in more detail in the following – can be inferred
from our statement of Probabilistic Preemption, whilst the remainder of our
model – which we might call PP  – is summarised as follows.

PP 

Variables: fM; D; Y; T; Eg

Interpretation: M D 1=0 if I do/don’t place my U-232 near the Geiger;

D D 1=0 if you do/don’t decide to place your Th-228 near the Geiger iff I
don’t place my U-232 near the Geiger; Y D 1=0 if you do/don’t place your
Th-228 near the Geiger; T = 1/0 if the threshold Geiger reading is/isn’t
reached; E D 1=0 if the bomb does/doesn’t explode.

Figure 6: Probabilistic Preemption

As an illustration of our rule for drawing arrows in a causal graph, consider

the arrow from M to T in Figure 6 (the graphical representation of PP  ). This
arrow exists because there’s an assignment of values to all of the other vari-
ables in our model – one such assignment is fD D 0; Y D 0; E D 0g – such
that if those variables are held fixed at the assigned values by interventions,
then intervening to change the value of M makes a difference to the probability
distribution over the values of T. That is:

P.T D 1jdo.D D 0&Y D 0&E D 0&M D 1//

¤ P.T D 1jdo.D D 0&Y D 0&E D 0&M D 0// (9)

That is, if interventions occur to ensure that (i) you don’t make the decision
that you do; (ii) you don’t place your radioactive material near the Geiger;

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 67

and (iii) the bomb doesn’t explode, then interventions changing whether or
not I place my radioactive material near the Geiger make a difference to the
probability that the threshold is reached.71 It’s key here that an intervention to
prevent the bomb exploding would be something – such as the severing of the
wire between the Geiger and the bomb – that doesn’t affect whether the thresh-
old is reached. This is needed to secure the relevant non-backtracking reading
of the counterfactuals. The justification for the inclusion of each of the other
arrows in Figure 6 is analogous to the justification for the inclusion of that
from M to T.
Now consider why there isn’t, for instance, an arrow from Y to E. That’s
because there’s no assignment of values to all other variables in the model
such that, when the variables are held fixed at those values by interventions,
intervening on Y makes a difference to the probability distribution over E. In
particular, any such assignment includes an assignment of a value to T but,
when the value of T is held fixed by an intervention (either at 1 or 0), then the
value of Y no longer has a bearing on the probability that E D 1. Next, consider
why there isn’t an arrow from D to M. This is simply because your decision
is completely irrelevant to my action and can’t be made probabilistically rele-
vant to it by holding fixed by interventions the other variables in the model.72
Finally, consider why there isn’t an arrow from E to T. This is because, as men-
tioned in the previous paragraph, an intervention on E would only count as such
if it doesn’t affect T. When the value of E is changed in such a way, its value

71 It’s important to say something about which time’s probabilities are in question. One suggestion
(assumed in what follows) is that, where Xi represents an event that occurs at t then, in evaluating
whether there’s an arrow from Xi to Xj , the probabilities to consider are those that would obtain
just after t if Xi had been intervened on and interventions had already occurred by t to fix the
values of all variables in the model besides Xj . The fact that some of those other variables
represent events that occur later than t doesn’t stand in the way of interventions fixing their
values having occurred prior to t. For example, the wire connecting the Geiger and the bomb
could already have been severed to ensure that the bomb doesn’t explode (an intervention on
the value of E) by the time I place my radioactive material near the Geiger.
One might wonder whether, in a probabilistic world, it’s reasonable to suppose that an inter-
vention occurring by t could deterministically fix the value of a variable E representing an event
occurring sometime later. In fact, it’s not strictly necessary to appeal to deterministic interven-
tions to get the results described in the main text. For instance, an intervention that (merely)
sufficiently lowered the probability of your placing your radioactive material near the Geiger
(even in the circumstance in which I don’t place mine there) would ensure that my action raises
the probability that the threshold is reached. Still, the complication of probabilistic interventions
isn’t an avenue we’ll pursue further here.
72 In the language of graph theory, Y acts as an ‘unshielded collider’ for D and M (Spirtes et al.
2000, 10): that is, there are arrows from each of D and M to Y but there’s no arrow connecting
D and M. While probabilistically independent variables can be rendered dependent by condi-
tioning upon an unshielded collider, they can’t be rendered dependent by intervening on an
unshielded collider.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 68 Philosophy of Science

has no bearing upon that of T (no matter what values we hold the other vari-
ables in the model fixed at). Each of the other arrow absences in Figure 6 has
a justification analogous to the justification for the absence of an arrow from Y
to E, from D to M, or from E to T.
Just as, in seeking to analyse deterministic causation in terms of SEMs
appeal to an ‘apt’ SEM is needed, so too in seeking an analysis of (prob-
abilistic) causation in terms of probabilistic models, the notion of an ‘apt’
probabilistic model is needed. The requirements (Partition), (Independence),
(Naturalness), and (Serious Possibilities) placed on SEMs carry across to prob-
abilistic models without need of modification (and with the same motivation).
(Veridicality) however needs some modification, or at least elaboration, to
make clear what it means for a probabilistic model to ‘entail only true coun-
terfactuals’. The requirement (Veridicality ) is what we need to substitute in
its place.

(Veridicality ) For any conjunction X1 D x1 & : : : &Xn D xn 73 (where X1 : : : Xn

are variables in the model and x1 : : : xn are possible values of
those variables) taken as an input, the probability distribution
P.
jdo.X1 D x1 & : : : &Xn D xn // yielded as output by the do.
/
function of the model should be the objective probability dis-
tribution over the variables in the model that would result from
interventions setting X1 D x1 & : : : &Xn D xn .74

The requirement (Veridicality ) is thus the requirement that the model (specif-
ically its do.
/ function) entail true counterfactuals about what the objective
probabilities would be under all combinations of values of the variables in
the model’s variable set. The models that I consider in what follows all sat-
isfy (Veridicality ). Specifically, in describing various causal scenarios, I’ll
stipulate what the objective probabilities would be under various counter-
factual suppositions about their variables. I shall then only describe mod-
els of these scenarios that entail the correct values for these counterfactual
chances.
The only requirement left to discuss now is (Stability). This can be applied
to probabilistic models in exactly the same form (and with the same motiva-
tion) as it was applied to SEMs in Section 4.6.3, except that its reference to
(Veridicality) should be replaced by a reference to (Veridicality ).

73 The simple expression X1 D x1 shall be treated as a limiting case of such a conjunction.

74 In light of the discussion of Footnote 71, we should strictly speaking require that, where the
actual values of X1 : : : Xn represent states obtaining at t1 : : : tn the do.
/ function entails, for
each ti , the objective probability distribution over the variables in the model that would result
had interventions setting X1 D x1 & : : : &Xn D xn occurred by ti .

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 69

5.4 Analysing Causation in terms of Probabilistic Models

With the notion of an apt probabilistic model in hand, we’re in a posi-
tion to describe an analysis of (actual) causation that captures the notion of
de facto probability-raising more precisely. The following is a first-pass anal-
ysis that will be of heuristic value. By analysing causation in terms of de facto
probability-raising it helps with the problem, described in Section 5.2, of causa-
tion without probability-raising. But, as we’ll see, it can’t be our final analysis
because it doesn’t help with the converse difficulty of probability-raising
without causation.

PC

Where x and x0 (x ¤ x0 ) are possible values of the variable X and y is a

possible value of the variable Y, then X taking the value X D x rather than
X D x0 is a cause of Y taking the value Y D y iff X D x and Y D y are the
actual values of X and Y and, in some apt probabilistic model M , there’s
a path P from X to Y such that – where W is the set of variables that lie
off the path P and the actual values of those variables are W D w – M
entails that:

P.Y D yjdo.W D w  &X D x// > P.Y D yjdo.W D w  &X D x0 //

(IN)

Analysis PC is highly analogous to SEA.75 SEA instructs us to look for a

path P in an apt SEM such that holding fixed the variables that lie off P at
their actual values reveals a relation of de facto counterfactual dependence of
Y D y upon X D x. Analogously, PC instructs us to look for a path P in an
apt probabilistic model such that holding fixed the variables that lie off P at
their actual values reveals a relation of de facto probability-raising of Y D y
by X D x.
Analysis PC counts M D 1 (rather than M D 0) as a cause of E D 1 in
PP  . To see this, consider the path hM; T; Ei. There are two variables – D and
Y – that lie off this path. Moreover, holding D and Y fixed at their actual values

75 One disanalogy is that, while SEA builds in contrastivity on both the cause and the effect sides,
PC builds in contrastivity on the cause side only. Though it’s possible to incorporate contrastiv-
ity on the effect side too, this is a bit less natural in the probabilistic case where there’s often not
a determinate fact about what value the effect variable would have taken (but only a probability
distribution over alternative values) had the cause variable taken its contrast value.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 70 Philosophy of Science

D D 1 and Y D 0 reveals a relation of de facto probability-raising of E D 1 by

M D 1, as required by condition IN of PC:

P.E D 1jdo.D D 1&Y D 0&M D 1// > P.E D 1jdo.D D 1&Y D 0&M D 0//
(10)
This simply says that, given your decision but your failure to place your
material near the Geiger, the probability of the explosion is higher if I place
my material there than if I don’t. This is of course true in Probabilistic
Preemption.
Analysis PC is, however, in danger of giving the incorrect result that your
decision to place your material near the Geiger iff I don’t place mine there,
D D 1, is also a cause of the explosion, E D 1. To see this, first note that it’s
presumably the case that if both of us place our material near the Geiger then
the probability of the threshold being reached is higher than if I alone place
my material there. Next note that, since the case is probabilistic, we might rea-
sonably allow that your decision merely makes it very likely that you won’t
place your radioactive material near the Geiger if I do rather than deterministi-
cally preventing you from doing so. But now consider the path hD; Y; Ei from
D to E. There’s one variable – M – that lies off this path. But it could be that
D D 1 raises the probability of E D 1 even holding M fixed at its actual value
M D 1.76

P.E D 1jdo.M D 1&D D 1// > P.E D 1jdo.M D 1&D D 0// (11)

That is, holding fixed that I place my material near the Geiger, your decision
still raises the probability of the bomb’s exploding because there’s some chance
given your decision that you’ll place your material near the Geiger too. In this
case, PC counts your decision as a cause even though you don’t in fact place
your material near the Geiger.
To think about how to modify PC to give a more adequate analysis of causa-
tion, it helps to consider why we don’t regard your decision as a cause. The key
reason is presumably that your placing your material near the Geiger (some-
thing which you don’t in fact do) is an essential part of the causal process
via which your decision had the potential to bring about the explosion. Of

76 To secure this result we need to assume that the alternative to your making the decision that you
do is for you to make some decision (e.g. to simply walk away from the scene) that makes it
less likely than it in fact is that you will also place your material near the Geiger. Now we could
either simply stipulate this, or we could model the case with a multi-valued variable representing
the various possible decisions open to you. We could then make the explicitly contrastive point
that PC (incorrectly) treats your making the decision that you do rather than (e.g.) deciding to
walk away as a cause of the bomb’s explosion.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 71

course, it wouldn’t be very helpful to simply build into our analysis the require-
ment that ‘there be a complete causal process’ from putative cause to putative
effect without further characterisation of what a ‘causal process’ is and what
it is for such a process to be incomplete.77 And, as discussed in Section 1, it
seems impossible to give an account of causal processes without falling back on
counterfactuals (which, in the probabilistic case, might include counterfactuals
about probabilities).
Fortunately, we can sidestep these issues because it appears there’s a probabi-
listic symptom of the incompleteness of a causal process. Take the relationship
between your decision and the bomb’s explosion and note that, although the
probability of the bomb’s exploding may (even holding fixed my placing my
radioactive material near the Geiger) be higher if you take your decision than
if you don’t, the probability of the bomb’s exploding given that you take your
decision but you don’t place your radioactive material near the Geiger is (even
holding fixed my placing my material near the Geiger) no higher than if you’d
never taken your decision in the first place. That is, although Inequality 11
holds, it’s also the case that:

P.E D 1jdo.M D 1&D D 1&Y D 0//  P.E D 1jdo.M D 1&D D 0// (12)

There’s thus a sense in which the (de facto) probability-raising of E D 1 by

D D 1 is ‘non-robust’ in that it doesn’t hold up when we take into account the
actual value of a variable on the path from D to E.
By contrast, the de facto probability-raising of E D 1 by M D 1 revealed by
Inequality 10 is robust. That is, it continues to hold when we take into account
the actual value of the only intermediate variable, T, on the path hM; T; Ei along
which we discovered de facto probability-raising. That is:

P.E D 1jdo.D D 1&Y D 0&M D 1&T D 1//

> P.E D 1jdo.D D 1&Y D 0&M D 0// (13)

In words, holding fixed the fact that you make your decision but don’t place
your material near the Geiger, the probability of the bomb’s exploding is (not
only higher if I place my material there than if I don’t, but is) higher if I place
my material there and the threshold is reached than if I don’t place my material
there in the first place.

77 Though, as we saw in Section 3.4.3, an account needn’t be reductive to be illuminating, we

would surely like an account that tells us something about what a causal process is rather than
simply falling back on intuitions about this notion.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 72 Philosophy of Science

This suggests that we might try to revise PC so that the revised analysis
requires not merely that there’s de facto probability-raising of effect by cause,
but also that this probability-raising be suitably robust. This is the approach
taken by Fenton-Glynn (2017). The following is a slightly streamlined version
of that analysis.

PC1

Where x and x0 (x ¤ x0 ) are possible values of the variable X and y is a

possible value of the variable Y, then X taking the value X D x rather than
X D x0 is a cause of Y taking the value Y D y iff X D x and Y D y are the
actual values of X and Y and, in some apt probabilistic model M , there’s
a path P from X to Y such that – where W is the set of variables that lie
off P, where Z is the set of variables that lie intermediate between X and
Y on P, and where the actual values of the variables in W are W D w
– then for all subsets Z0 of Z (where Z0 D z  are the actual values of the
variables in Z0 ), M entails that:

P.Y D yjdo.W D w &X D x&Z0 D z  // > P.Y D yjdo.W D w &X D x 0 //

(IN1)

Analysis PC1 is complex and one can best get to grips with it by seeing it in
action. To start with, observe that it yields the correct results about Probabi-
listic Preemption. First, it correctly counts M D 1 as a cause of E D 1. To
see this, consider the path hM; T; Ei in PP  . There are two variables – D and
Y – that lie off this path, and one – T – that lies intermediate between M and
E upon it. The set of intermediate on-path variables is therefore just fTg which
has two subsets: fTg and ;. Inequality 10 shows that M D 1 raises the proba-
bility of E D 1 when we hold the off-path variables fixed at their actual values.
This de facto probability-raising is trivially robust (in the sense captured by
Inequality IN1 of PC1) relative to the empty subset of fTg. But it’s also robust
relative to the non-empty subset of fTg. This is indicated by the holding of Ine-
quality 13. The fact that the de facto probability-raising of E D 1 by M D 1 is
robust relative to both subsets of the set of intermediate on-path variables, fTg,
is precisely what is required for PC1 to diagnose M D 1 as a cause of E D 1.
Second, PC1 yields the correct result that D D 1 isn’t a cause of E D 1. To
see this note that hD; Y; T; Ei is the only path from D to E in PP  . The variable
M is the only one that lies off this path and its actual value is M D 1. Moreover,
although D D 1 may raise the probability of E D 1 when we hold fixed that
M D 1, as revealed by the holding of Inequality 11, this probability-raising

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 73

isn’t robust relative to the subset fYg of the intermediate on-path variables, as
is revealed by the holding of Inequality 12.
Probabilistic Preemption is an early preemption case. But it’s quite easy
to verify that PC1 yields the correct results in certain cases of late preemption
too. Consider Figure 3 from Section 4.6.2. This was a graph of the SEM that we
constructed for Late Preemption. But Figure 3 could just as well represent a
version of Late Preemption which differs from the original in that each of the
relations is merely probabilistic: specifically, Bob’s intention merely makes it
likely that Bob will shoot if the deer isn’t dead by 1:00pm, Alice’s shot merely
makes it likely that the deer will be dead by 1:00pm, Bob’s shooting would
merely make it likely that the deer would be dead by 1:01pm if Alice didn’t
shoot. In such a case, PC1 entails that Alice’s shot A D 1 was a cause of D1W01
= 1. To see this, consider the path hA; D1W00 ; D1W01 i. A D 1 raises the probability
of D1W01 D 1 when we hold the two off-path variables – BI and B – fixed at
their actual values BI D 1 and B D 0. That is:

P.D1W01 D 1jdo.BI D 1&B D 0&A D 1//

> P.D1W01 D 1jdo.BI D 1&B D 0&A D 0// (14)

That is, given that Bob intends to shoot if the deer isn’t dead by 1:00pm, but in
fact Bob doesn’t shoot, the probability that the deer is dead by 1:01pm is higher
if Alice shoots than if she doesn’t.
This de facto probability-raising is robust in the sense that it continues to hold
when we take into account the actual value of the (only) intermediate on-path
variable D1W00 , namely D1W00 D 1. That is:

P.D1W01 D 1jdo.BI D 1&B D 0&A D 1&D1W00 D 1//

> P.D1W01 D 1jdo.BI D 1&B D 0&A D 0// (15)

This says that, holding fixed that Bob intends to shoot if the deer isn’t dead by
1:00pm but in fact Bob doesn’t shoot, the probability that the deer is dead by
1:01pm is higher if Alice shoots and the deer is dead by 1:00pm than if simply
Alice doesn’t shoot in the first place.
Analysis PC1 also implies that Alice’s shooting was a cause of the deer’s
being dead by 1:00pm. To see this, consider the path hA; D1W00 i. Holding fixed
the off-path variables BI, B, and D1W01 at their actual values BI D 1, B D 0, and
D1W01 D 1, A D 1 raises the probability of D1W00 D 1. That is:

P.D1W00 D 1jdo.BI D 1&B D 0&D1W01 D 1&A D 1//

> P.D1W00 D 1jdo.BI D 1&B D 0&D1W01 D 1&A D 0// (16)

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 74 Philosophy of Science

Note that an intervention to ensure that D1W01 D 1 wouldn’t count as such if it

worked by ensuring (or even affecting the probability) that D1W00 D 1, which is
why we can be sure that Inequality 16 holds. Since there are no variables inter-
mediate between A and D1W00 on the path hA; D1W00 i it’s automatically true that
the de facto probability-raising revealed by Inequality 16 is suitably robust.78
Hence PC1 correctly counts A D 1 as a cause of D1W00 D 1.
Analysis PC1 also yields the correct result that Bob’s intention wasn’t a
cause of the deer’s being dead by 1:01pm (it implies that it wasn’t a cause of
the deer’s being dead by 1:00pm simply in virtue of the absence of a path from
BI to D1W00 79 ). To see this observe that there’s only one path – hBI; B; D1W01 i –
from BI to D1W01 but that BI D 1 doesn’t raise the probability of D1W01 when we
hold fixed the off-path variables A and D1W00 at their actual values, A D 1 and
D1W00 D 1. Specifically:

P.D1W01 D 1jdo.A D 1&D1W00 D 1&BI D 1// D 1

D P.D1W01 D 1jdo.A D 1&D1W00 D 1&BI D 0// (17)

This might seem a bit of a ‘cheat’ because it relies on the fact that the connection
between the variables D1W00 and D1W01 is deterministic: given that the deer is
dead at 1:00pm it’s dead at 1:01pm with probability 1. But note that, even if this
link were merely probabilistic (perhaps there’s some probability of the deer’s
being resuscitated between 1:00pm and 1:01pm!), PC1 would still yield the
correct result. Specifically, in that case, although BI D 1 would de facto raise
the probability of D1W01 D 1 (Equality 17 would – assuming that, given Bob’s
intention, there’s some chance of his shooting even if the deer is dead at 1:00pm
– now be converted into an inequality, with the term on the LHS being greater
than the term on the RHS), this de facto probability-raising wouldn’t be robust
once we took account of the actual value – B D 0 – of the intermediate on-path
variable in the usual way. That is:

P.D1W01 D 1jdo.A D 1&D1W00 D 1&BI D 1&B D 0//

D P.D1W01 D 1jdo.A D 1&D1W00 D 1&BI D 0// (18)

78 Note that, since the causal process connecting Alice’s shot to the deer’s being dead at 1:00pm
is complete, interpolating variables along the path hA; D1W00 i wouldn’t affect the fact that this
probability-raising is robust. For instance, (holding the off-path variables fixed at their actual
values) the probability of the deer’s being dead at 1:00pm would be higher if Alice shot and her
speeding bullet passed mid-air between her and the deer than it would be if Alice hadn’t shot
in the first place. So interpolating a variable M representing the presence of the speeding bullet
along the path hA; D1W00 i wouldn’t affect the verdict of PC1.
79 Plausibly, the absence of such a path is no mere idiosyncrasy of our model, but rather will be
a feature of any model of the scenario satisfying (Veridicality*). That’s because the absence of
such a path reflects the assumption of the example that Bob’s intention is to give Alice enough
of an opportunity to kill the deer that he won’t himself shoot the deer dead by 1:00pm.
https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press
 Causation 75

That is (even holding the off-path variables fixed at their actual values) the
probability of the deer’s being dead by 1:01pm given Bob’s intention but also
the fact that Bob doesn’t shoot is no higher than it would be if simply Bob
hadn’t had the intention in the first place.

5.5 Overdetermination of Probabilities

It’s at least in principle possible that probabilistic scenarios analogous to Sym-
metric Overdetermination should arise. These are cases in which an effect
isn’t overdetermined but, as it were, its probability is overdetermined. The
following is an example.80

Overdetermination of Probabilities

Consider a specific one-second interval, i, and suppose the probability that

atom a will decay within i is (i) 0:1 if a doesn’t absorb a neutron at the
start of i; (ii) 0:8 if a absorbs neutron b alone or neutron c alone at the
start of i; and (iii) 0:8 if a absorbs both neutron b and neutron c at the
start of i. Suppose, moreover, that the chance of a absorbing both b and c
at the start of i is 1 and there’s probability 0 that a will absorb any other
particle during i. In fact, a absorbs b and c simultaneously at the start of i
and decays within i.

Suppose that, although absorbing neutrons brings about decay probabilisti-

cally, there’s no intermediate mechanism via which it brings about decays. That
is, absorbing a neutron results in a certain probability of decay and then either
the decay occurs or it doesn’t, with no further mechanism needed to bring about
the decay.
I don’t claim that this example is physically realistic, just that a case with
this probabilistic structure is conceivable and, insofar as we’d like our analysis
to apply across possible worlds, then our analysis ought to give a reasonable
treatment of cases like this. Moreover, symmetric overdetermination can be
considered the limiting case of overdetermination of probabilities where the
probability is overdetermined to be 1. And presumably we’d like our analysis
to be able to handle cases of symmetric overdetermination if we have ambitions
for it to hold in deterministic worlds.
Consider a model OP  of Overdetermination of Probabilities with a
do.
/ function that entails the probabilities specified in Overdetermination
of Probabilities, and with the additional features described as follows.

80 I stipulate that some of the probabilities in this example are 0 or 1 for simplicity. More complex
examples can be given in which all probabilities are between 0 and 1.
https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press
 76 Philosophy of Science

OP 

Variables: fBA; CA; ADg

Interpretation: BA D 1=0 if b is/isn’t absorbed by a at the beginning of i;

CA D 1=0 if c is/isn’t absorbed by a at the beginning of i; AD D 1=0 if a
does/doesn’t decay within i.

Figure 7: Overdetermination of Probabilities

Relative to OP  , PC1 counts neither BA D 1 nor CA D 1 as a cause of

AD D 1 (which is analogous to the way in which SEA doesn’t count overde-
terminers as causes). To see this in the case of BA D 1 (the reasoning is exactly
analogous for CA D 1) note that there’s only one path from BA D 1 to AD D 1:
namely hBA; ADi. There’s one variable lying off this path: namely CA. But
BA D 1 doesn’t raise the probability of AD D 1 when we hold fixed the
off-path variable at its actual value, CA D 1. Specifically:

P.AD D 1jdo.CA D 1&BA D 1// D 0:8 D P.AD D 1jdo.CA D 1 BA D 0// (19)

I suspect that, as in cases of symmetric overdetermination so in cases

of symmetric overdetermination of probabilities more generally, the major-
ity of people will have the opinion that both ‘overdeterminers’ are causes,
or at least that this is a preferable result to counting neither as causes.
This, then, is a limitation of PC1 (and it’s easily verified that it similarly
fails to count symmetric overdeterminers as causes in true overdetermina-
tion cases). One might therefore be tempted to consider a liberalisation of
PC1 that allows that we may sometimes vary features of the actual situ-
ation in seeking latent probability-raising of effect by the putative cause.
This would be analogous to the liberalisation of SEA considered in Section
4.6.4. For instance, were we allowed to suppose that CA D 0, then we’d
recover probability-raising of AD D 1 by BA D 1 in Overdetermination of
Probabilities.
https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press
 Causation 77

5.6 Type Causation, Interventions, and Epistemology

The focus of this section has been on actual causation: the sort of token causal
relation that’s usually at issue when scientists, lawyers, and pretty much every-
one else talks about causation between particulars rather than types. But, just as
in the deterministic case, we might sometimes have use for a notion of causa-
tion between variables themselves (as discussed in Section 4.6.5) and not just
between variable values.
In fact, the definitions of relations of causation between variables given by
Woodward (2003) and discussed in Section 4.6.5 carry across neatly to the
probabilistic case. As in that case, X can be defined as as a direct cause of
Y in a model iff X is a parent of Y in the model. The only difference from
the deterministic case is that the ‘model’ in question is now understood as a
probabilistic model rather than a SEM.
Woodward (2003, 51, 57) himself provides probabilistic versions of his def-
initions of total cause and contributing cause. Specifically, X is defined as a
(probabilistic) total cause of Y iff there’s a pair of values x, x0 (x ¤ x0 ) of X
such that, had (due to an intervention) X taken X D x, then the probability dis-
tribution over the possible values of Y would have been different than had (due
to an intervention) X taken X D x0 . On the other hand, X is defined as a (prob-
abilistic) contributing cause of Y relative to a model iff there’s a path P from
X to Y in the model, a pair of values x, x0 (x ¤ x0 ) of X, and an assignment of
values to all off-path variables such that, had (due to an intervention) X taken
X D x and had (due to interventions) the off-path variables taken the assigned
values, then the probability distribution over the possible values of Y would
have been different than had (due to an intervention) X taken X D x0 and had
(due to interventions) the off-path variables taken the assigned values.
With the concept of a probabilistic model together with these definitions of
(probabilistic) causation between variables in hand, the notion of an ‘interven-
tion’ in the probabilistic case can be formally defined by strict analogy to the
definition for the deterministic case.81 Specifically, an intervention on X with
respect to Y consists in X’s taking some value X D x as a result of an interven-
tion variable IX;Y taking one of its ‘switching’ values (‘switching’ is defined in
the same way as in Section 4.6.6, but with ‘depends’ now replaced by ‘proba-
bilistically depends’), with IX;Y counting as an intervention variable for X with
respect to Y iff (a) IX;Y is a contributing cause of X, (b) IX;Y acts as a switch for
X, (c) there’s no directed path from IX;Y to Y that doesn’t go via X; (d) there’s no
variable that’s an ancestor of IX;Y that lies on a directed path to Y that doesn’t
run via IX;Y .

81 Though, as noted in Footnote 71, in the probabilistic case we might also have use for the notion
of a probablistic intervention.
https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press
 78 Philosophy of Science

Just as in the deterministic case, the various definitions of causal relations

between variables don’t necessarily capture type-causal notions. For example,
according to these definitions, M counts as a direct, total, and contributing
cause of Y in our model PP  of Probabilistic Preemption. Yet these variables
represent whether or not certain token events occur. But, just as in the deter-
ministic case, it’s of course possible to use variables to represent things that
are non-particular. For instance, we could construct a type-level probabilistic
model of the causes of heart disease with variables representing such things as
average number of cigarettes smoked, minutes of exercise taken, saturated fat
consumed, etc. per day. Such a model would be ‘type-level’ in the sense that
it can be taken to model the causes of heart disease across a population rather
than for any specific individual.
An interesting thing to note about Woodward’s definitions of causation
between variables is that they don’t focus on probability-raising, but rather sim-
ply require that interventions on the putative cause variable X make a difference
to the probability distribution over the values of the putative effect variable Y
(when other variables are held fixed). Part of the reason for this is that it simply
doesn’t make sense to talk about X raising the probability of Y where X and Y
are variables. What we can sensibly say is that the value of X affects the proba-
bility distribution over Y. Still, we might ask, when it comes to actual causation
(which has been the principal focus of this section) why focus (as we have) on
(de facto) probability-raising rather than probability-lowering, given that both
notions are well-defined when we’re concerned with the relations between spe-
cific variable values? This is an important question and one that will be taken
up in the next subsection.
Before turning to that however it is worth observing that, when it comes to
probabilistic models, the interpretation of the probabilities figuring in a model
of a particular (token) system is presumably different from the interpretation
of the probabilities figuring in a ‘type-level’ model. In giving a probabilistic
model of a particular system (an individual person or, for instance, the Gei-
ger counter set-up described in Probabilistic Preemption), we’re assuming
that it makes sense to assign probabilities to particulars: these are sometimes
known as ‘single-case’ probabilities. Such probabilities can’t be straightfor-
wardly taken to be frequencies/statistical probabilities. That’s because there’s
only one of the token system in question, so the ‘frequencies’ (i.e. statistics)
for it are trivial. For instance, the bomb either explodes or it doesn’t, so the
frequency with which it does is either 1/1 (=1) or 0/1 (=0).
Of course, one might take population-level frequencies to imply single-case
probabilities. For instance if, on average, Jo Doe smokes 10 cigarettes, exer-
cises for 30 minutes, and consumes 20g of saturated fat per day, then one

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 79

might take her individual probability of suffering heart disease to equal the
population-level frequency with which people who on average smoke 10-a-
day, exercise for 30 minutes, and consume 20g of saturated fat develop heart
disease. Thus a type-level model might be used to calculate the risk of heart
disease in a particular person. But this is far from an innocuous move, since it
gives rise to the notorious reference-class problem.82
While this isn’t the place to enter into extended discussion of the issue, it’s
worth pointing out that there are alternative ways of understanding single-case
probabilities, including ‘propensity’ interpretations (e.g. Popper 1959; Hack-
ing 1965; Mellor 1971) and ‘best system’ interpretations (e.g. Lewis 1994).
When it comes to ‘type-level’ models, on the other hand, a frequency inter-
pretation is tempting, but it’s not the only possibility. For instance, the ‘best
system’ approach provides an interpretation of type-level as well as single-case
probabilities.
When we’re dealing with models that represent type-level phenomena and
frequency/statistical information about them, there are quite sophisticated tech-
niques for causal discovery (see, e.g., Spirtes et al. 2000). In particular, we can
often discover a lot about the causal structure (e.g. direct, total, and contribut-
ing causal relations among variables) simply by examining the conditional
and unconditional correlations between variables, at least if we’re prepared to
make certain general assumptions about the relationship between causation and
probability. Such general assumptions include that events that don’t cause one
another are probabilistically independent conditional upon any common causes
they might have, that if A causes C (only) by causing B then A and C are proba-
bilistically independent conditional upon B,83 and that independent causes of a
single effect become probabilistically dependent when that effect is conditioned
upon.84 We can get even further if we have statistics available from randomised
control trials (RCTs). That’s because RCTs act like type-level ‘interventions’:
in an RCT, values of some variable X are assigned randomly to members of
a population thus, within that population, breaking the usual dependence of X

82 Venn (1866) was the first to systematically describe this problem. Hájek (2007) provides a
contemporary discussion.
83 These two assumptions are captured by the so-called Causal Markov Condition (Spirtes et al.
2000, 11).
84 To illustrate: patients with MS are often misdiagnosed as having lupus and vice versa because
many of the symptoms are the same. Loss of coordination, for example, is an effect of both
conditions. So consider the class of patients presenting with coordination loss. We would expect
that the incidence of MS is lower in that subset of such patients who have lupus than it is among
that subset of such patients who don’t have lupus. This explains why, for instance, a doctor
presented with such a patient who rules out lupus becomes more confident that MS is the correct
diagnosis and, conversely, why a doctor who discovers that such a patient has lupus wouldn’t
usually investigate whether the patient additionally has MS.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 80 Philosophy of Science

upon its parents (i.e. acting as a ‘switch’). This allows us to determine the causal
influence of X upon a target variable Y while being reasonably confident that
any association between X and Y isn’t due to common causes or indeed due to
Y’s being a cause of X.
It’s unquestionable that our discovery – by such methods – of type-level
causal relations and statistical information about their strength informs our
judgements about causal relations and probabilities at the token level. But the
relationship between the type- and the token level, as ever, isn’t straightfor-
ward. For instance, statistical associations at the type level needn’t be reflected
by genuine indeterminism at the token level (still less need they match the
single-case probabilities): probabilities at the type level may reflect unmod-
elled causes, measurement errors and the like. Nor can actual causal relations
simply be ‘read off’ from those at the type level. For instance, driving with a
damaged tyre might be a type-level cause of car crashes (because of the risk of
blow out) but, in a particular case where a driver drives with a damaged tyre
and crashes, the former needn’t be an actual cause of the latter (suppose the
driver fell asleep at the wheel and there was no blow-out). However, delving
deeper into the complexities surrounding the relation between type and token
levels is a task that must await another occasion.

5.7 Probability-Raising vs Probability-Lowering

In the previous subsection, we deferred an important question. This is the ques-
tion of why, when we are analysing actual causation, we focus on (de facto)
probability-raising rather than probability-lowering? To address this question,
it is helpful to consider an argument given by Mellor (1995, 58–66, ch. 7) who
(as we saw in Section 3.4.1) claims that causation has several important ‘conno-
tations’, including that effects are evidence for their causes, that causes explain
their effects, and that causes are means for bringing about their effects as ends.
Mellor says:

Any apparent cause that failed to raise its effects’ chances would not only be
neither evidence for them nor explain them, it would not be a means to them if
they were ends. And nothing that satisfied none of these three connotations
would be a cause: the whole point of calling it a cause would be lost. In
particular, to call something a cause that provides no way of bringing about
its effects seems to me an obvious contradiction in terms. (Mellor 1995, 88)

Anyone who wishes to allow that there can be probability-lowering causes –

as my placing my U-232 near the Geiger appeared to be in Probabilistic Pre-
emption – has either to reject the notion that causation has the connotations

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 81

that Mellor describes, or has to reject the idea that probability-raising is neces-
sary to satisfy these connotations. Doing the former seems rather unpalatable,
but the advocate of the de facto probability-raising approach can explain why
doing the latter is less so.
Continuing to focus on the example of Probabilistic Preemption, note that,
given that you didn’t actually place your Th-228 near the Geiger in Probabi-
listic Preemption, my placing my U-232 there clearly explains why the bomb
exploded. Also fairly clear is that my placing my U-232 near the Geiger was a
means to the bomb’s exploding. Of course, there’s a sense in which my failure
to place my U-232 there would also have been a means (indeed a more effective
one) to the bomb’s exploding. Yet, in the scenario, I didn’t know this since I
didn’t know of your presence. Nevertheless, although it’s plausibly a platitude
about causation that causes are means to their effects, it’s surely not a plati-
tude that for anything to count as a cause it must be the best means to its effect
or even the best available in the circumstances. For example, if someone who
works in asbestos removal also smokes, surely her smoking has the potential
to cause her to suffer lung cancer even if she could have more surely brought
about the latter by removing her protective mask at work.85
Is my placing my U-232 evidence that the bomb exploded? It depends.
What is evidence for what presumably depends upon the epistemic state of
the explainee. For instance, it would be odd to say that the causes of the explo-
sition count as evidence of its occurrence for someone who actually witnessed
the explosion. Likewise, for someone who hasn’t witnessed the explosion but
who knows of your intention to place your Th-228 near to the Geiger if I don’t
place my U-232 there, the fact that I do place my U-232 there is plausibly
evidence against the explosion. Still, for someone who doesn’t know of your
intention, or who knows of your intention but also knows that in fact you don’t
place your Th-228 near to the Geiger, my action is presumably evidence that
the bomb explodes. I take it that this is sufficient for the relevant connotation
of causation to be satisfied.
Interestingly, then, if the de facto probability-raising account is correct, then
it might be true to say that c caused e (since c de facto raised the probability of

85 Of course there’s a disanalogy between this case and Probabilistic Preemption in that smoking
and not wearing a protective mask to work are not mutually exclusive, but my placing my U-
232 near the Geiger and my not doing so are. So perhaps one could argue that, where c is
a member of a set S of mutually exclusive (and jointly exhaustive?) possibilities, c is only a
cause of e if c is the best means to e in S? Yet I suspect that, in doing so, one leaves the realm
of straightforward ‘connotations’ or platitudes about causation for the realm of philosophical
theorising. Moreover, such an argument seems to simply beg the question against the advocate
of the de facto probability-raising account.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 82 Philosophy of Science

e) even though there’s a sense in which e occurred despite c (since c straight-

forwardly lowered the probability of e).86 But arguably this is just as it should
be: in Probabilistic Preemption we appear to be tempted to regard my action
as a cause, even though there’s also a sense in which the explosion occurred
‘despite’ my action. Focusing on the role that my action played in preventing
your more efficacious action can incline us to adopt the ‘despite’ locution. The
advocate of the de facto probability-raising approach will therefore argue that
there’s no contradiction between asserting ‘c caused e’ and that ‘e occurred
despite c’, though perhaps – because the former emphasises c’s role in bring-
ing about e, while the latter emphasises c’s (unmanifested) potential to prevent
e – it’s infelicitous to utter them in the same breath.
This leaves the question of why our concept of actual causation should track
de facto probability-raising rather than overall probability-raising? Presumably
the advocate of the de facto probability-raising account will seek to answer this
in terms of the connotations of (actual) causation: for instance, the fact that my
placing my U-232 near the Geiger would serve as evidence and explanation for
the explosion for an agent who knew enough about the circumstances to know
that you didn’t place your Th-228 there, and that I successfully performed my
action as a means to the explosion. In contrast, consider the following example
due to Cartwright (1979, 425):

I consider eradicating the poison oak at the bottom of my garden by spraying

it with defoliant. The can of defoliant claims that the spray is 90 per cent
effective; that is, the probability of a plant’s dying given that it is sprayed is
.9, and that the probability of its surviving is .1.

Spraying the plant lowers the probability of the plant’s survival. But if the plant
survives then, as Cartwright points out, the spraying can’t explain its survival.
Likewise the spraying is neither evidence for the survival nor is it in any sense
a means to bringing about the survival. Plausibly this is because the spraying
doesn’t even bear a de facto probability-raising relation to the survival, and for
this reason isn’t an actual cause.

6 Conclusion
In this Element, we’ve examined three broad traditions in the philosophy of
causation: the regularity, counterfactual, and probabilistic approaches. Our
discussion showed how contemporary thinking about causation has been influ-
enced by a long history of intellectual thought on the subject, but we’ve also

86 Cf. Hitchcock (2001b, esp. 382).

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Causation 83

ventured to the cutting edge of contemporary research: for instance, by exam-

ining structural equation and probabilistic modelling approaches. The latter
approaches constitute a rich – and increasingly interdisciplinary – vein of cur-
rent investigation. In addition to examining particular approaches, we’ve inves-
tigated issues that are sometimes informed by, but to some extent transcend,
particular theories of causation: for instance, questions about the relationship
between type and token causation, about causation involving absences, and
about what kinds of, and how many, relata the causal relation should be taken
to have.
The discussion has, of necessity, been selective. For reasons discussed in the
Introduction, we haven’t examined the rich literatures on causal processes and
causal mechanisms. There are also some general questions about causation that
we haven’t tackled, such as whether causation comes in degrees (Halpern and
Hitchcock 2015, Bernstein 2017, Sartorio 2020) and whether Hart and Hon-
oré (1959) and Lewis (1973a) are indeed correct that our tendency to pick out
abnormal factors as causes is something to be explained by a pragmatics of
causal discourse, as opposed to a more recent suggestion that it’s to be baked
into the semantics of causal claims.87 Yet by introducing the reader to core the-
ories, concepts, and tools present in the contemporary debate it is hoped, not
merely that the reader’s enthusiasm for further investigation has been piqued,
but also that they have been equipped to get the most out of such investigation,
and indeed to participate in the debate themselves.
Of the approaches to causation that we’ve examined, some appear more
promising than others. For my part, I find approaches that appeal to de facto
dependence/probability-raising to be more promising than some of the earlier
accounts in the counterfactual and probabilistic traditions. Most investigators
these days seem to regard these traditions as more fruitful than the regularity
tradition, though I suspect that ways of sophisticating the regularity approach
(e.g. Strevens 2007, Baumgartner 2013) have been unfairly neglected. Yet even
among the approaches that I’ve suggested are most promising, the specific
accounts that have been developed – while elegantly dealing with some of the
problems that afflicted earlier accounts – are not immune to difficulties.
Of course the hope is that further progress can be made – these research
programmes are very much alive and kicking – but if anyone has come up
with an entirely problem- and controversy-free account of any philosophically
interesting notion, I’ve yet to see it! Part of the difficulty is that the various
aspects of our pre-theoretical thinking about notions such as causation probably

87 For debate over this latter point, see e.g. (McGrath 2005), (Halpern and Hitchcock 2015) and
(Blanchard and Schaffer 2015).

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 84 Philosophy of Science

don’t admit of a complete and elegant regimentation. Yet even if no single,

unified account is possible, it’s to be hoped that we’ll find that we have a range
of accounts available each of which captures an important aspect of our causal
thinking and whatever it might be ‘in the world’ that that thinking tracks.
In any case, not all progress in deepening and sharpening our understanding
of a phenomenon like causation consists in constructing analyses of it. Much
progress consists in getting a clearer picture of the terrain by, for instance,
gaining clarity on the nature and variety of difficult cases (e.g. the varieties
of preemption, the different ways causation can come apart from probability-
raising, the existence of apparent counterexamples to causal transitivity, etc.)
and in developing tools and concepts (e.g. sophisticated counterfactual seman-
tics, the notion of an ‘intervention’, the concept of an ENF counterfactual, the
formal techniques of causal modelling and graph theory) for thinking about
these and other cases. At the very least, this increased understanding of the
nature of causation can help us avoid making basic errors in our thinking about
it and, given the centrality of causation to the legal and scientific worlds, the
importance of this can’t be overstated.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 References
Albert, D. Z. (2000). Time and Chance. Cambridge, MA: Harvard University
Press.
Anscombe, G. E. M. (1971). Causality and Determination. Cambridge:
Cambridge University Press.
Armstrong, D. M. (1983). What Is a Law of Nature? Cambridge: Cambridge
University Press.
Baumgartner, M. (2013). A regularity theoretic approach to actual causation.
Erkenntnis 78, 85–109.
Beebee, H. (2004). Causing and nothingness. In J. Collins, N. Hall, and
L.A. Paul (eds): Causation and Counterfactuals, pp. 291–308. Cambridge,
MA: MIT Press.
Bernstein, S. (2017). Causal proportions and moral responsibility. In
D. Shoemaker (ed.), Oxford Studies in Agency and Responsibility: vol. 4,
pp. 165–82. Oxford: Oxford University Press.
Bigaj, T. (2012). Causation without influence. Erkenntnis 76, 1–22.
Blanchard, T. and J. Schaffer (2015). Cause without default. In H. Beebee,
C. Hitchcock, and H. Price (Eds.), Making a Difference. Oxford: Oxford
University Press.
Callender, C. and J. Cohen (2010). Special sciences, conspiracy and the better
best system account of lawhood. Erkenntnis 73, 427–47.
Carlson, G. and F. Pelletier (Eds.) (1995). The Generic Book, Chicago, IL.
University of Chicago Press.
Carroll, J. (1991). Property-level causation? Philosophical Studies 63,
245–70.
Cartwright, N. (1979). Causal laws and effective strategies. Noûs 13, 419–37.
Cartwright, N. (2002). Against modularity, the causal markov condition, and
any link between the two: Comments on Hausman and Woodward. British
Journal for the Philosophy of Science 53, 411–53.
Cartwright, N. (2004). Causation: One word, many things. Philosophy of
Science 71, 805–19.
Cat, J. (2017). The unity of science, In E. Zalta (ed.): Stanford Encyclopedia
of Philosophy, https://plato.stanford.edu/archives/fall2017/entries/scientific-
unity/
Choi, S. (2002). Causation and gerrymandered world lines: A critique of
Salmon. Philosophy of Science 69, 105–17.
Collins, J. (2000). Preemptive prevention. Journal of Philosophy 97, 223–34.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 86 References

Craver, C. and J. Tabery (2019). Mechanisms in science. In E. Zalta (ed.):

Stanford Encyclopedia of Philosophy, https://plato.stanford.edu/archives/
sum2019/entries/science-mechanisms/
Davidson, D. (1967). Causal relations. Journal of Philosophy 64, 691–703.
Dowe, P. (1992). Wesley Salmon’s process theory of causality and the con-
served quantity theory. Philosophy of Science 59, 195–216.
Dowe, P. (2000). Physical Causation. Cambridge: Cambridge University Press.
Dunn, J. (2011). Fried eggs, thermodynamics, and the special sciences. The
British Journal for the Philosophy of Science 62(1), 71–98.
Eagle, A. (2015). Generic causation. Unpublished Manuscript;
http://philpapers.org/rec/EAGGC.
Earman, J. (1986). A Primer on Determinism. Dordrecht: Reidel.
Edgington, D. (1997). Mellor on chance and causation. British Journal for the
Philosophy of Science 48, 411–33.
Eells, E. (1991). Probabilistic Causality. Cambridge: Cambridge University
Press.
Elga, A. (2001). Statistical mechanics and the asymmetry of counterfactual
dependence. Philosophy of Science 68 (Supplement), S313–S324.
Emery, N. (2015). Chance, possibility, and explanation. British Journal for the
Philosophy of Science 66, 95–120.
Evans, R. (2013). Altered Pasts: Counterfactuals in History. Waltham, MA:
Brandeis University Press.
Fenton-Glynn, L. (2017). A proposed probabilistic extension of the Halpern
and Pearl definition of ‘actual cause’. British Journal for the Philosophy of
Science 68, 1061–124.
Feynman, R. (1965). The Character of Physical Law. London: BBC.
Glennan, S. (1996). Mechanisms and the nature of causation. Erkenntnis 44,
49–71.
Glennan, S. (2017). The New Mechanical Philosophy. Oxford: Oxford Univer-
sity Press.
Glymour, C., D. Danks, B. Glymour et al. (2010). Actual causation: A stone
soup essay. Synthese 175, 169–92.
Glynn, L. (2013). Of miracles and interventions. Erkenntnis 78, 43–64.
Goldszmidt, M. and J. Pearl (1992). Rank-based systems: A simple
approach to belief revision, belief update, and reasoning about evidence
and actions. In Proceedings of the Third International Conference on
Knowledge Representation and Reasoning, San Mateo, CA, pp. 661–72.
Morgan Kaufmann.
Hacking, I. (1965). The Logic of Statistical Inference. Cambridge: Cambridge
University Press.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 References 87

Hájek, A. (2007). The reference class problem is your problem too.

Synthese 156, 563–85.
Hall, N. (2000). Causation and the price of transitivity. Journal of
Philosophy 97, 198–222.
Hall, N. (2004). Two concepts of causation. In Causation and Counterfactuals,
pp. 225–76. Cambridge, MA: MIT Press.
Halpern, J. Y. (2015). Appropriate causal models and stability of causa-
tion. Unpublished manuscript, www.cs.cornell.edu/home/halpern/papers/
causalmodeling.pdf.
Halpern, J. Y. (2016). Actual Causality. Cambridge, MA: MIT Press.
Halpern, J. Y. and C. Hitchcock (2010). Actual causation and the art of mod-
eling. In R. Dechter, H. Geffner, and J. Y. Halpern (Eds.), Heuristics,
Probability and Causality: A Tribute to Judea Pearl, pp. 383–406. London:
College Publications.
Halpern, J. Y. and C. Hitchcock (2015). Graded causation and defaults. British
Journal for the Philosophy of Science 66, 413–57.
Halpern, J. Y. and J. Pearl (2005). Causes and explanations: A structural-model
approach. Part I: Causes. British Journal for the Philosophy of Science 56,
843–87.
Hart, H. L. A. and T. Honoré (1959). Causation in the Law. Oxford: Clarendon
Press.
Hausman, D. and J. Woodward (1999). Independence, invariance and the causal
markov condition. British Journal for the Philosophy of Science 50, 521–83.
Hesslow, G. (1976). Two notes on the probabilistic approach to causality.
Philosophy of Science 43, 290–2.
Hitchcock, C. (1995). Salmon on explanatory relevance. Philosophy of
Science 62, 304–20.
Hitchcock, C. (1996). Farewell to binary causation. Canadian Journal of
Philosophy 26, 267–82.
Hitchcock, C. (2001a). The intransitivity of causation revealed in equations and
graphs. Journal of Philosophy 98, 194–202.
Hitchcock, C. (2001b). A tale of two effects. Philosophical Review 110,
361–96.
Hitchcock, C. (2004). Do all and only causes raise the probabilities of effects?
In J. Collins, N. Hall, and L. Paul (eds.), Causation and Counterfactuals, pp.
403–17. Cambridge, MA: MIT Press.
Hitchcock, C. (2007). Prevention, preemption, and the principle of sufficient
reason. Philosophical Review 116, 495–532.
Hitchcock, C. (2009). Problems for the conserved quantity theory: Counterex-
amples, circularity, and redundancy. Monist 92, 72–93.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 88 References

Hitchcock, C. and J. Knobe (2009). Cause and norm. Journal of

Philosophy 106, 587–612.
Hoefer, C. (2007). The third way on objective probability: A sceptic’s guide to
objective chance. Mind 116, 549–96.
Hume, D. (1739). A Treatise of Human Nature. London: J. Noon.
Hume, D. (1748). An Enquiry Concerning Human Understanding. London:
A. Millar.
Ismael, J. (2009). Probability in deterministic physics. Journal of
Philosophy 106, 89–108.
Kim, J. (1973a). Causation, nomic subsumption, and the concept of event.
Journal of Philosophy 70, 217–36.
Kim, J. (1973b). Causes and counterfactuals. Journal of Philosophy 70,
570–72.
Kment, B. (2006). Counterfactuals and explanation. Mind 155, 261–10.
Kroedel, T. (2008). Mental causation as multiple causation. Philosophical
Studies 139, 125–43.
Kvart, I. (2004). Causation: Probabilistic and counterfactual analyses. In Cau-
sation and Counterfactuals, pp. 359–86. Cambridge, MA: MIT Press.
Lewis, D. (1973a). Causation. Journal of Philosophy 70, 556–67.
Lewis, D. (1973b). Counterfactuals. Oxford: Oxford University Press.
Lewis, D. (1979). Counterfactual dependence and time’s arrow. Noûs 13,
455–76.
Lewis, D. (1986a). Events. In D. Lewis (Ed.), Philosophical Papers, vol. 2, pp.
241–269. Oxford: Oxford University Press.
Lewis, D. (1986b). Philosophical Papers, vol. 2. Oxford: Oxford University
Press.
Lewis, D. (1994). Humean supervenience debugged. Mind 103(412), 473–90.
Lewis, D. (2004a). Causation as influence. In J. Collins, N. Hall, and L. Paul
(eds.), Causation and Counterfactuals, pp. 75–106. Cambridge, MA: MIT
Press.
Lewis, D. (2004b). Void and object. In J. Collins, N. Hall, and L. Paul (eds.),
Causation and Counterfactuals, pp. 277–290. Cambridge, MA: MIT Press.
List, C. and M. Pivato (2015). Emergent chance. Philosophical Review 124,
119–152.
Loewer, B. (2001). Determinism and chance. Studies in History and Philosophy
of Science Part B: Studies in History and Philosophy of Modern Physics 32,
609–20.
Machamer, P., L. Darden, and C. Craver (2000). Thinking about mechanisms.
Philosophy of Science 67, 1–25.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 References 89

Mackie, J. (1965). Causes and conditions. American Philosophical Quarterly 2,

245–64.
Mackie, J. (1980). The Cement of the Universe. Oxford: Oxford University
Press.
Malament, D. (2008). Norton’s slippery slope. Philosophy of Science 75,
799–816.
McDermott, M. (2002). Causation: Influence versus sufficiency. Journal of
Philosophy 99, 84–101.
McGrath, S. (2005). Causation by omission: A dilemma. Philosophical
Studies 123, 125–48.
Meek, C. and C. Glymour (1994). Conditioning and intervening. British
Journal for the Philosophy of Science 45, 1001–21.
Mellor, D. H. (1971). The Matter of Chance. Cambridge: Cambridge University
Press.
Mellor, H. (1995). The Facts of Causation. London: Routledge.
Menzies, P. (1989). Probabilistic causation and causal processes: A critique of
Lewis. Philosophy of Science 56, 642–63.
Menzies, P. (1996). Probabilistic causation and the pre-emption problem.
Mind 105, 85–117.
Mill, J. (1843). A System of Logic, Ratiocinative and Inductive, Vol. 1. London:
John W. Parker.
Moore, M. (2009). Causation and Responsibility: An Essay in Laws, Morals,
and Metaphysics. Oxford: Oxford University Press.
Northcott, R. (2008). Causation and contrast classes. Philosophical
Studies 139, 111–23.
Norton, J. (2008). The dome: An unexpectedly simple failure of determinism.
Philosophy of Science 75, 786–98.
O’Connor, T. (2002). Persons and Causes: The Metaphysics of Free Will.
Oxford: Oxford University Press.
Paul, L. A. and N. Hall (2013). Causation: A User’s Guide. Oxford: Oxford
University Press.
Pearl, J. (2009). Causality: Models, Reasoning, and Inference (2nd ed.).
Cambridge: Cambridge University Press.
Popper, K. (1959). The propensity interpretation of probability. British Journal
for the Philosophy of Science 10, 25–42.
Read, R. (ed.) (2014). The New Hume Debate: Revised Edition, Abingdon:
Routledge.
Reichenbach, H. (1971). The Direction of Time. Mineola, NY: Dover Publica-
tions. M. Reichenbach (ed.). First published in 1956.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 90 References

Reutlinger, A., Schurz, G., Hütteman, A. and Jaag, S. (2019). Ceteris

paribus laws, In E. Zalta (ed.): Stanford Encyclopedia of Philosophy,
https://plato.stanford.edu/archives/win2019/entries/ceteris-paribus/
Rosen, D. (1978). In defense of a probabilistic theory of causality. Philosophy
of Science 45, 604–13.
Salmon, W. (1994). Causality without counterfactuals. Philosophy of
Science 61, 297–312.
Salmon, W. (1997). Causality and explanation: A reply to two critiques.
Philosophy of Science 64, 461–77.
Salmon, W. C. (1980). Probabilistic causality. Pacific Philosophical Quar-
terly 61, 50–74.
Sartorio, C. (2005). Causes as difference-makers. Philosophical Studies 123,
71–96.
Sartorio, C. (2020). More of a cause? Journal of Applied Philosophy 37,
346–63.
Schaffer, J. (2001). Causes as probability raisers of processes. Journal of
Philosophy 98, 75–92.
Schaffer, J. (2004). Trumping preemption. In J. Collins, N. Hall, and L. A. Paul
(eds.), Causation and Counterfactuals, pp. 59–73. Cambridge, MA: MIT
Press.
Schaffer, J. (2005). Contrastive causation. Philosophical Review 114,
327–58.
Schaffer, J. (2013). Causal contextualisms. In M. Blaauw (ed.), Contrastivism
in Philosophy, pp. 35–63. New York: Routledge.
Spirtes, P., C. Glymour, and R. Scheines (2000). Causation, Prediction, and
Search (2nd ed.). Cambridge, MA: MIT Press.
Strevens, M. (2007). Mackie remixed. In J. Campbell, M. O’Rourke, and H. Sil-
verstein (eds.) Causation and Explanation, pp. 93–118. Cambridge, MA:
MIT Press.
Suppes, P. (1970). A Probabilistic Theory of Causality, Acta Philosophica
Fennica, vol. 24. Amsterdam: North-Holland.
Tucker, A. (1999). Historiographical counterfactuals and historical contin-
gency. History and Theory 38, 264–76.
Venn, J. (1866). The Logic of Chance. London and Cambridge: Macmillan
and Co.
Werndl, C. (2011). On choosing between deterministic and indetermin-
istic models: Underdetermination and indirect evidence. Synthese 190,
2243–65.
Weslake, B. (2020). A partial theory of actual causation. Forthcoming in British
Journal for the Philosophy of Science.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 References 91

Williamson, T. (2000). Knowledge and its Limits. Oxford: Oxford University

Press.
Woodward, J. (2003). Making Things Happen: A Theory of Causal Explana-
tion. Oxford: Oxford University Press.
Yablo, S. (2002). De facto dependence. Journal of Philosophy 99,
130–148.
Yablo, S. (2004). Advertisement for a sketch of an outline of a prototheory
of causation. In J. Collins, N. Hall, and L.A. Paul (eds): Causation and
Counterfactuals, pp. 119–137. Cambridge, MA: MIT Press.
Yang, H., Z. Rivera, S. Jube et al. (2010). Programmed necrosis induced
by asbestos in human mesothelial cells causes high-mobility group box
1 protein release and resultant inflammation. Proceedings of the National
Academy of the Sciences 107, 12611–16.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press
 Philosophy of Science

Jacob Stegenga
University of Cambridge
Jacob Stegenga is Reader in the Department of History and Philosophy of Science at the
University of Cambridge. He has published widely on fundamental topics in reasoning
and rationality and philosophical problems in medicine and biology. Prior to joining
Cambridge, he taught in the United States and Canada, and he received his PhD from
the University of California San par Diego.

About the Series

This series of Elements in Philosophy of Science provides an extensive overview of the
themes, topics and debates which constitute the philosophy of science. Distinguished
specialists provide an up-to-date summary of the results of current research on their
topics, as well as offering their own on those topics and drawing original conclusions.

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press

 Philosophy of Science

Elements in the Series

Scientific Knowledge and the Deep Past: History Matters
Adrian Currie
Philosophy of Probability and Statistical Modelling
Mauricio Suárez
Relativism in the Philosophy of Science
Martin Kusch
Unity of Science
Tuomas E. Tahko

Big Data
Wolfgang Pietsch
Objectivity in Science
Stephen John

Causation
Luke Fenton-Glynn

A full series listing is available at: www.cambridge.org/EPSC

https://doi.org/10.1017/9781108588300 Published online by Cambridge University Press