Hunting Causes and Using Them: Is There No Bridge from Here to There?

Abstract

Causation is in trouble—at least as it is pictured in current theories in philosophy and in economics as well, where causation is also once again in fashion. In both disciplines the accounts of causality on offer are either modelled too closely on one or another favoured method for hunting causes or on assumptions about the uses to which causal knowledge can be put—generally for predicting the results of our efforts to change the world. The first kind of account supplies no reason to think that causal knowledge, as it is pictured, is of any use; the second supplies no reason to think our best methods will be reliable for establishing causal knowledge. So, if these accounts are all there is to be had, how do we get from method to use? Of what use is knowledge of causal laws that we work so hard to obtain?

Full text

Centre for the Philosophy of Natural and Social Science
Contingency and Dissent in Science
Technical Report 05/09
Hunting Causes and Using Them:
Is There No Bridge from Here to There?
Nancy Cartwright
(with Sophia Efstathiou)
Series Editor: Damien Fennell

1
The support of The Arts and Humanities Research Council (AHRC) is gratefully
acknowledged. The work was part of the programme of the AHRC Contingency and Dissent
in Science.
Published by the Contingency And Dissent in Science Project
Centre for Philosophy of Natural and Social Science
The London School of Economics and Political Science
Houghton Street
London WC2A 2AE
Copyright © Nancy Cartwright 2009
ISSN 1750-7952 (Print)
ISSN 1750-7960 (Online)
All rights reserved.
No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in
any form or by any means, without the prior permission in writing of the publisher, nor be
issued to the public or circulated in any form of binding or cover other than that in which it is
published.

2
Hunting Causes and Using Them:
Is There No Bridge from Here to There?
1
Nancy Cartwright
(with Sophia Efstathiou)
Editor’s Note
This paper addresses the $64,000 dollar question of practicable causal inference: how to analyse causality to
underpin the inference of causal claims that are useful in practice? Cartwright argues that the existing analyses
are either good at hunting causes in special cases or good for obtaining claims that are useful (e.g. predictions
of what will happen under intervention) for the specific contexts from which they were inferred. She argues
that an analysis of causality that bridges theses two ends, the hunting of causes and their use, remains to be
found.
Abstract
Causation is in trouble –at least as it is pictured in current theories in philosophy and in
economics as well, where causation is also once again in fashion. In both disciplines the
accounts of causality on offer are either modelled too closely on one or another favoured
method for hunting causes or on assumptions about the uses to which causal knowledge can
be put – generally for predicting the results of our efforts to change the world. The first kind
of account supplies no reason to think that causal knowledge, as it is pictured, is of any use;
the second supplies no reason to think our best methods will be reliable for establishing
causal knowledge. So, if these accounts are all there is to be had, how do we get from
method to use? Of what use is knowledge of causal laws that we work so hard to obtain?
1
This paper was presented at the Society for Philosophy of Science in Practice (SPSP) meeting in London in
the summer of 2007. It develops ideas presented in a more roundabout way in Cartwright (2006) and in
Cartwright (2007). Support for the research and completion of the paper was provided by the AHRC project
Contingency and Dissent in Science, by the University of California at San Diego Senate, by an Andrew White
Research Scholarship, by the Institute for Advanced Study at Durham University and by the Spencer
Foundation; we are grateful to all for their help.

3
1. Two Actions: Hunting Causes and Using Causes
Philosophic – and economic – accounts of causality are almost always rooted in ideas about
either how to HUNT causes or how to USE them; but not both simultaneously. Most are
almost directly read off either a favoured method of establishing causal claims – putting the
label ‘causal’ onto a claim – or a favoured use we expect to make of causal claims –
inferences to claims of practical use. Accounts based on methods for hunting causes include
the probabilistic theory of causality and its descendent Bayes-nets theories, accounts based
on experimental methods (like the Galilean experiments discussed below), most versions of
causal process theories and theories that rely on exchanges of conserved quantities or the like
and the usual counterfactual accounts. Those rooted in use include manipulation and
intervention accounts, those based in causal decision theory and counterfactual accounts that
allow realistic implementations.
Almost all these accounts are good ones, Cartwright has argued – good for specific purposes
in specific kinds of systems.
2
And, taken together, they cover both rules for entry into the
language of causality and for exit from causal language. The problem is that there is no
bridge from one to the other. What assures us that the knowledge we hunt at such great effort
and cost can be put to the uses we want to make of it? To be practicable a theory of
causation must simultaneously ground how we label features as ‘causes’ and the inferences
we make once the label is attached. So we need a theory of causation that gives us in one fell
swoop both methods for inferring causes and methods for using them.
For the purposes of this paper, trying to remain relatively neutral about just what causal laws
are since we will consider a variety of different accounts, we propose that one think of a
causal law as a law-like causal regularity: ‘C causes E’ constitutes a causal law if C regularly
causes E – and that’s no accident.
To illustrate our points about hunting and use we shall discuss four philosophical accounts of
causal laws besides Cartwright’s own
2
Nancy Cartwright (2007, p.11-24).

4
• James Woodward’s ‘level’ invariance account
3
• related accounts based on ‘Galilean’ experiments
4
• Lewis-style (‘miracle’-based) counterfactuals
5
• probabilistic theories of causality.
6
All these are geared towards hunting causes: they ensure that we are “correct” in putting
causal labels on. But they say little about what we can do with “causal” knowledge once we
have it. On the other side we shall describe Woodward’s ‘modularity’ assumption and the
kind of assumptions economists are wont to make about causality that guarantee its
usefulness.
7
We pick these for discussion because they allow us to illustrate our points fairly
simply and vividly.
2. Two Problems: Unstable Enablers and External Validity
We tend to assume that the knowledge we secure with our best methods for testing casual
laws carried out in the best circumstances is knowledge that we can use directly: in knowing
the causal law we know how to change effects by changing their causes and we can make
reliable predictions about the results of so doing. We seem to take some connection there as
given.
8
But why should one make this assumption? When if ever is it justified?
Philosophic accounts of causation based on hunting methodologies give no satisfactory
answer to these questions. The conditions our accounts need to secure causal knowledge are
not sufficient to secure the inferences that would put them to use. These accounts face two
problems.
1. Unstable enablers: changes in the enabling factors that support causal laws – a
problem commonly recognized in economics.
3
James Woodward (2003).
4
Cf. Ronald Giere (1979).
5
David Lewis (1973).
6
Esp. Patrick Suppes (1970).
7
Among philosophers we discuss Woodward rather than other accounts (eg. Peter Menzies and Huw Price
(1993) or Gasking (1955) or Georg Henrik von Wright (1971)) because his two separate conditions directly
illustrate the distinction between characterizations of causality rooted in hunting methodologies and those that
work to guarantee use.
8
Cf. Cartwright’s early paper (1979) or the spate of work at about the same time on causal decision theory (for
instance in Harper et al. (eds.) (1981)).

5
2. External validity: the problem of generalizing from a particular setting or population
to the one of interest.
2.1 Unstable Enablers
Economists have worried about the stability of causal laws and their use for policy prediction
since at least the time of John Stuart Mill.
9
We go to the trouble of setting economic policy
on the assumption that the factors we propose to use as policy levers can change and that our
casual laws will allow us to predict what will follow on from that. But features in the
economy that permit change in the ways we like and can predict allow for change in ways
that we don’t like and worse, often in ways that we can’t predict. Our very attempts to use
our causal laws may undermine the laws themselves. Instabilities in the causal laws we
establish may not only result when we interfere to set policy, they may arise naturally as
well, and at times and places we often have no way of knowing. What use then are causal
laws for prediction?
Mill was pessimistic. He argued that economics cannot be an inductive science because the
background arrangements of causes fluctuate and do so erratically, both naturally and as a
result of deliberate actions by us. Any effect consequent on a particular cause depends, he
argued, on a large background of other causal factors simultaneously at work as well, factors
we are rarely able to identify. The effects consequent on a cause on one occasion, or over a
period of time, cannot be relied on to occur on other occasions since this myriad of other
causes is likely to change and in ways we cannot predict. So, concluded Mill, the sea of other
causal factors that enable a particular factor to bring about a particular effect are rarely in
place long enough to allow us to make new predictions based on past regular associations
between the two factors.
Besides variations in the arrangements of causes themselves Mill was also concerned about a
second kind of possibly unstable enabler: changes in the underpinning structures that give
rise to causal laws in the first place. It is a common experience of everyday life that what
9
John Stuart Mill (1836 [1967]).

6
causal laws hold in a situation depends on the underlying structure that gives rise to them.
Pressing a lever in a toaster causes the bread to drop into the toaster and brown, pressing a
lever on the floor of the driver’s side of a Rover causes the car to accelerate. Putting a pound
coin into a vending machine in the UK causes a bag of crisps to drop into the tray, putting a
pound coin into a similar looking machine in the USA just gums up the works. For most of
these man-made devices the usual ways of manipulating the cause to achieve the predicted
effect will not gum up the works. They are typically well-shielded to prevent just that
happening.
The structural features that give rise to causal laws at work in the economy may be far more
porous however. Like Mill, the early econometricians worried about just that. More recently
Chicago School economists like Robert Lucas argue more strongly: not only can active
policy intervention change the underlying structural arrangements in ways that undermine
the very causal laws that are being relied on to predict the outcomes of those interventions; it
is very likely this will happen. If interventions are expected, so argues the Chicago School,
agents will change their behaviours and in just such a way as to undermine the established
causal relations that predictions are based on.
10
On the other hand, despite being worrisome, causal laws can be unstable in predictable, even
happy ways. Take an example from Mill’s On the Subjugation of Women.
11
In this essay Mill
notes that many (like August Comte, one of his chief adversaries on this issue) maintain that
there is a well-established (very law-like) causal regularity between being a woman and
being inept at leadership, reasoning and imagination. Happily this causal law is not stable
under changes in the underlying enabling structure that supports it. Change the social
structure so that the upbringing, roles, education and experiences of women are more like
those of 19th century British middle-class men and the causal laws relating sex to leadership,
reasoning and imagination shift dramatically. The enabling factors that hold the regularity
fixed change in more or less expected ways, and luckily so.
10
Robert Lucas (1981), and Robert Lucas (1988).
11
John Stuart Mill (1869 [1997]).

7
To see how the problem of unstable enablers plays out in a current philosophical account of
causation, let us turn to James Woodward.
12
Woodward claims that two requirements must
be satisfied by a relation before we can call it a causal law. The first requirement is level
invariance (modularity comes later):
• Level invariance: a relationship between putative causes and an effect is level
invariant if the relationship stays fixed as any putative cause in the relationship varies
‘by intervention’.
‘Intervention’ is hard to define properly. It is something like a ‘miracle’ in the Lewis account
of counterfactuals: a change in the cause at the last stage while all other causes from the set
of causal laws operating in the situation remain the same, except for changes induced by the
change in the targeted factor (sometimes called by philosophers, changes ‘causally
downstream’ from the targeted change).
When it comes to causality Woodward focuses on systems of linear deterministic causal laws
that look like this:
Linear deterministic causal law
z c= ax + by
The symbol ‘c=’ means that the left and right-hand side are equal and that the factors on the
right are a complete set of causes of those on the left. (Reference to the population and
circumstances is repressed as is usual in presentation.) Applied to z c= ax + by, level
invariance dictates that the functional relationship z = ax + by must hold as x varies
13
while
y and all the true causes in causal laws at work in the situation (except those causally
downstream from the change in x) stay fixed; and the functional relation also holds as y
varies and x and other causes stay fixed.
12
Cf. James Woodward (2003) and Sandra Mitchell (2003) who stresses the need for invariance without a
detour through causality.
13
The range of permitted variation of variable must be specified as well. Mention of the relativization to these
ranges is suppressed as well.

8
Woodward supports his claims arguing by contradiction: he considers various examples in
which level invariance is violated by relations known to be spurious.
14
The canonical
example is the relationship between a storm and the reading of a barometer. Storms are
caused by fluctuations in the relative pressure of atmospheric fluids so we can relate the
presence of a storm to a function of atmospheric pressure by writing, say,
(1) storm = f (pressure).
The level of a well-functioning barometer, an instrument designed to measure pressure,
should also depend on atmospheric pressure, according to a relationship like
(2) barometer level = g (pressure).
One can mathematically solve (2) for pressure in terms of the inverse function g’ and
substitute into (1) to relate the presence of a storm to the reading of a barometer:
(3) storm = f (g'(barometer level)).
The question is which of (1), (2) or (3) describes a genuine causal law.
The true causal laws at work in the situation contain three variables: pressure, storm,
barometer level. We expect that equation (1) would continue to obtain if pressure were
varied ‘by miracle’ with barometer level staying fixed. So it should pass the level invariance
test. So too does (2) since it would not change if pressure were changed by a miracle while
the presence or absence of the storm stays fixed. Equation (3) on the other hand, while it may
be useful for calculating the likelihood of a storm given that the instrument is functioning
correctly, does not express a causal law and this shows up in the level invariance test: if the
pressure stays fixed while the barometer level changes ‘by miracle’ (say it just explodes), the
relationship between the storm and the reading of the barometer changes. It is well-known
that you can’t bring on the storm by breaking the glass!
Cartwright has elsewhere supported Woodward’s emphasis on level invariance by showing it
that is a sufficient condition for causality.
15
She does so by proving a kind of representation
theorem for level invariance vis-à-vis causal laws. The theorem uses some fairly
14
Cf. James Woodward (2003).
15
See Cartwright (2003).

9
straightforward axioms that a set of causal laws should satisfy – like asymmetry, irreflexivity
and the assumption that any functional relations that hold are generated by genuine causal
relations. These are all axioms that are presupposed in most discussions of causality and in
particular in the examples that Woodward employs. The theorem then shows that any
functional relation generated by a set of causal laws will be one of those causal laws if and
only if it is level invariant.
16
So given a set of ‘true’ causal laws like (1) and (2) no spurious
relation derived from them will be level invariant.
17
Despite the fact that level invariance ensures that a functional relationship is a causal law
given the axioms laid down for causal laws, Woodward thinks this is not enough. He adds a
second condition, modularity, which we shall discuss below. But Cartwright, to the contrary,
takes the theorem to prove that level invariance is sufficient. In that case level invariance is a
good – indeed sure – way to HUNT causes, to put a causal label onto a relationship. If a
relationship satisfies this condition we have conclusive reason to think it is causal. But does
level invariance get around Mill and Lucas’s problem?
Unfortunately it does not. (And we will later argue that modularity doesn’t really do better
either.) Looking for level invariant laws presupposes that factors in causal laws are not
erratically interacting with the causal structures that give rise to them.
18
Consider again Mill’s original example of the subjugation of women. The relationship
between sex and leadership could well have been constant across a very great many
16
Cartwright takes these axioms to be fairly innocuous and to be true of causal laws even if her singular-
causings account of causal laws is mistaken. Cartwright notes that there has been some objection that the
axioms are not so innocuous because a transitivity axiom is included. But the transitivity axiom assumes only
that if x appears as a cause of y in a linear deterministic causal system, we still have a causal law for y if we
substitute for x the right-hand side of any causal law that has x as effect. This is necessary unless we are willing
to assume that causation in nature is not continuous in time, so that there is a notion of direct causal law (the
‘last’ law in operation before the effect is produced) that is not representation relative and that it is this notion
of direct causal law that we are trying to characterize.
17
Cartwright (2003).
18
It also supposes a relatively clear separation between the structure (what Cartwright calls the ‘nomological
machine’) that gives rise to a set of causal laws and the laws themselves. Cartwright has defended this division
in Cartwright (1989) and in Cartwright (1999). Cartwright does not think it is universally applicable though.
But she does think it can be made practically everywhere where there are causal laws at work that can be
represented in the usual triangular array of equations, whether those equations represent deterministic or
probabilistic causality and whether they are linear or not. For more discussion see Efstathiou (2009a).

10
variations in other causes affecting leadership over the centuries, constant enough to count it
as level invariant. It was throughout a true causal law: being a woman caused one to be weak
at leadership. But if we transform the structure of society changing the role and experiences
of women from birth onwards, the relationship may well no longer satisfy Woodward’s
condition of level invariance. So the well-established causal law would be no guide for
predicting the effect of sex on leadership skills if the structure is changed. The invariance of
a relationship under changes in variables in the other causal laws at work does not ensure
predictability unless underlying enabling structures stay fixed. But that we cannot expect
structural causal enablers to remain stable is just the Lucas critique!
2.2 External validity
A second problem many current theories of causality have to grapple with is external
validity. This is a well-known problem in methodology. Many of the methods that allow us
to establish causal-law claims most securely can be applied in only very narrow settings. We
establish results very securely in a particular experimental setting or a particular test
population. But the method itself provides no basis for extending the results to a population
or setting different from that in the test. Theories of causality that are too closely tied to these
kinds of method will suffer from the same problem; the very way causal laws are
characterized makes causal laws very narrow in their scope and thus very limited in their
predictive power.
Consider a standard method for hunting causes: Galilean experiments. The goal in a Galilean
experiment is dual. First we wish to eliminate all confounders, then to establish a law-like
regularity between cause and effect with no confounders to interfere. We can eliminate
confounders by physically isolating an experimental system from background interference
and/or by making various idealizing assumptions. The problem is that establishing the
presence of a causal relation in absence of confounders does not ensure that the same causal
relation persists once confounders are present.

11
Suppose we establish via a perfect Galilean experiment that C causes E. Call the set of
confounding factors – the ones we work so hard to eliminate in the experiment – ‘N’ (for
‘nasties’). Then what we have actually established in the experiment is
C&¬N c E
from which we cannot infer
C&N c E.
But that is the kind of result we need for reliable prediction if we want to use C to control E
in real life settings. Inferring that a cause will have the same effect whether or not
confounders are present is a logical fallacy.
Any characterization of causal laws that reads them off from the results of a Galilean
experiment will have even worse troubles. In that case causal laws will end up by definition
to hold only in situations where confounders are absent.
External validity is also a problem for David Lewis’s counterfactual account of causation.
Lewis counterfactuals establish a causal link by changing the cause at the last instant by a
miracle-like intervention that removes the cause while leaving everything else the same. If
the effect does not obtain once the cause – and only the cause – is absent then a causal
relationship is established.
Although Lewis offers his account as an account of singular causal relations, it is also the
basis for related accounts of causal laws. But as an account of causal laws it suffers from the
problem of external validity. Any inferential power we get from Lewis counterfactuals
pertains to a particular setting – whatever setting is under consideration, the setting in which
the miracle will intervene to change the putative cause. Even though the setting is not an
idealized one, if we follow Lewis we establish a law-like causal regularity for only one kind
of setting. Unless further assumptions are made there are no guarantees that the causal law
established is resilient to a different background arrangement of confounders. So Lewis
counterfactuals can lay down causal laws for settings where confounders occur, but only for
the specific arrangement of them under consideration. It is no help for any different
arrangement.

12
What we know from Lewis counterfactuals is that for some specific N
C&N cE.
This does not imply any results for some different N'. In particular it is a logical fallacy to
conclude that
C&N' c E.
The same cause need not have the same effect under a new arrangement of confounders.
External validity is also a problem for an account more akin to Cartwright’s own: the
probabilistic theory of causality developed by Patrick Suppes. This account is almost read
off statistical methods for inferring causal laws from observational (as opposed to
experimental) data: stratify on all possible confounders, then look for correlations between
the putative cause and effect within each stratum. One fairly good attempt at formulating the
probabilistic theory (for yes-no variables) says that C is a cause of E in a particular
arrangement of confounders, say K if and only if the cause increases the probability of the
effect in that arrangement: P(E/C&K) > P(E/¬C&K).
19
This of course makes it explicit that
casual laws are situation relative.
20
C causes E simpliciter if it does so in all arrangements of
confounders.
But once more, the mere hap that under a particular arrangement of confounders the cause
increases the probability of the effect gives little information about what happens in a
different arrangement of confounders. As before
P(E/C&K) > P(E/¬C&K)
does not imply
P(E/C&K') > P(E/¬C&K').
We have again no logical grounds for inferences in settings different from the ones our laws
are derived in. And again as before, if we use the probabilistic theory to characterize what a
19
The formulation given here still isn’t quite right because K must not hold fixed any causal intermediaries by
which C causes E on a given occasion. Cartwright’s own best attempt relies on reference to singular causings
even in the formulation of the probabilistic theory. See Cartwright (1989), sections 2 and 3 and Cartwright
(2007) for a fuller discussion.
20
As usual the population relativity is buried in the probability measure. The laws hold for any population for
which the assumed probability measure holds.

13
causal law is the problem is even worse since by definition the same law cannot hold in
different settings. One may of course look to the non-relativized law: C causes E simpliciter.
In this case causal laws become more useful but incredibly hard to hunt – we then need
reason to suppose that the increase in probability will hold across all arrangements of
confounders. That we certainly do not get just by looking at what happens in any one or
handful of such arrangements.
So, here we see three familiar accounts of causality – the probabilistic theory, the
counterfactual account, and accounts read off from the theory of the Galilean experiment. All
three are rooted in methods for hunting causes. And all three are very good at what they do –
putting a causal label on a relationship only where it seems correct to do so. That’s not
surprising since the methods they are based on are among our surest methods for casual
inference. But all three face the trouble of external validity once it comes to USING causes.
Galilean experiments establish conditions sufficient for attaching a causal label on a feature.
If changes in C and C alone are followed by E in idealized settings with no confounders this
is sufficient for C to be a cause of E. But notice, the changes must be in C alone and the
setting must be ideal. Similarly Lewis counterfactuals are supposed to give conditions
sufficient for attaching a causal label. If changes in C and C alone are followed by E given a
specific arrangement of confounders this is sufficient for C to cause E – but only in cases
where C alone is changed and where the confounders are arranged in just the same way.
Probabilistic causality tells us to attach a causal label if changes in C and C alone are
followed by an increase in the probability of the effect given a specific arrangement of
confounders. This is again sufficient for C to cause E but only when C alone is changed in
just the specified arrangement of confounders.
How then can we use causal knowledge obtained in a Galilean experiment for predictions
about what will happen when confounders are in place or when factors other than C change
as well? How can we infer something from one “nasty” setting about what effects follow the
cause in another one using Lewis counterfactuals? How can we use probabilistic causality to
infer from one case where the presence of the cause makes the effect more probable

14
something about other cases where the same cause is known to be present but other causes
have changed?
These are well known problems in the methodology of causal inference. But methodologists
have a way to deal with them that philosophers cannot adopt. Methodologists often seem to
suppose that there is something, they know not what exactly it is – a causal law – that can be
established in a variety of different ways, then used for prediction in a variety of ways and
situations different yet again. Causal laws are something like the charge of an electron or
valency of oxygen in this respect.
We philosophers, however, aim to say what a causal law is, or at least to give some
significant characterizing features of it, and to do so in a way that makes sense both of the
ways we hunt causes and the ways we use them. But many of our attempts are too
operationalistic so not surprisingly they suffer from the standard troubles of operationalism.
If a concept is defined too closely to one or another way we test for it, we cannot account for
why other test methods are reasonable and we cannot underwrite the usual inferences we
make with the concept. Conversely, if a concept is defined too closely to one or another of
the inferences we make using it, nothing about the concept supports other inferences nor
grounds any methods for testing if the concept applies.
Current theoretical accounts of causation based on hunting methodologies are sufficient for
identifying the cause of an effect in a particular situation; they do not give us methods for
navigating to new settings. They only tell us where we are, not how we can get to where we
want to go.
3. Unstable Enablers and External Validity: The Same Problem from Different
Perspectives
Although the two problems of unstable enablers and external validity have a different source,
when it comes to the structure of inference there is a sense in which they are equivalent. This

15
suggests a deeper epistemological question that needs answering and also that there might
possibly be a common answer to them both.
We worry about “unstable enablers” when we study a particular case – e.g. a set of agents,
the direction of a particular effect, a particular variable – and we care to keep studying this
case in midst of a – possibly – changing background structure. The problem of “unstable
enablers” becomes visible after we have identified a causal relationship (e.g. pushing down
the lever producing toast, putting the pound coin in the slot producing a packet of crisps, the
mass of the sun causing the earth to orbit around it, stepping on the throttle causing the car to
accelerate) within some perceived structure; some possibly actual but for our purposes
conceptualized space (like a toaster, the vending machine, the planetary system or a Rover
500 automobile). We take the case we are interested in as momentarily fixed against this
underlying structure and worry about how the changing structure may affect what we care
about. (e.g. What would happen if the toaster were wet? If the machine ran out of crisps?,
etc.) The problem of “unstable enablers” speaks of shifts of known and unknown structural
factors which seem to happen in time and affect the case we fixed our interest on.
Compare this with the problem of “external validity”. External validity becomes an issue
when we shift our interest from a case already studied to one, hopefully similar, but other
than the one studied. We think of the problem as arising when we export causal knowledge
gained from a particular case to some new situation of interest. (e.g. Does pressing down the
lever work in any toaster? Is the collision impact on the test dummy the impact suffered by
the driver of a Rover 500?, etc.) The problem of “external” validity speaks of changes as
arising when we move in space, from places “interior” to those “exterior” to our case study.
So we might say that the problem becomes relevant when we think of causal knowledge as
shifted across space rather than when things change with time.
Let’s take a step of abstraction. Using “space” and “time” variables, Σ and T, we could re-
describe the two problems as follows:
1. Unstable Enablers:
Hypothesis & Structure (Σ1, T1) => Conclusion

16
but Hypothesis & Structure' (Σ1, T2) =>New Conclusion
2. External Validity:
Hypothesis & Setting (Σ1, T1) => Conclusion
but Hypothesis & Setting' (Σ2, T1) => New Conclusion
The differences in the effect depend on change that occurs across a perceived temporal or a
spatial setting:
1. Unstable Enablers: ∆Ε= φ (∆T)
2. External Validity: ∆Ε= ψ (∆Σ)
The “space” and “time” variables, Σ and T, here do not refer to physical space-time, but
rather to dimensions of our experience (or better yet, our talk) of things causally interacting
in physical space-time. But the way we talk about these problems is contingent. If we think
of these dimensions as interchangeable, talking about here and there becomes the same as
talking about now and then. When we take the instability of causal enablers as arising in
time, before and after we modify policy say, we get what we have termed the problem of
“unstable enablers”; when we take the instabilities of causal factors to occur across space –
from the test population to a target population say – we get the problem of external validity.
The two problems are equivalent in this sense (Figure 1 depicts how the distinction is
made).
Figure 1: What changes from one labelled problem to another is what holds our
interest not the logical structure of the situation.
(1) Unstable Enablers: (2) External Validity:
Interest (i) is fixed on a case;
worry about how background
affects the case studied.
Interest (i) is shifted to a case “outside” the
case studied; worry about whether conclusions
about the case studied apply here
Background
case
i
case
i
Background

17
So what? There is no denying that unstable enablers and external validity are different
methodological problems for us. Well, two things are made visible here. First, that we need
to address both problems when drawing up our methodology at the same time. The problem
of external validity it would seem is always present no matter (or on top of) whether we take
the “underlying” causal structure to be changing or stable. Conversely, why should we think
that the underlying structures are stable during the time it takes us to export our causal
knowledge? Looking at the two problems from the perspective of changes in ‘space’ and
‘time’ brings home how likely it is that we face both problems much of the time.
Besides noting that the two problems are probably happening at the same time [for which
one can argue without abstractions] our point is to note that the problems share a form and
may also share a solution on a more abstract theoretical plane than either policy or
methodology.
21
4. Cartwright’s Causal Laws and Capacities
Cartwright’s account of causal laws takes singular causation as primary and builds laws from
there: C causes E in Ф if and only if some C’s regularly cause E’s in Ф (in the ‘long run’).
Cartwright then argues that where probabilities apply, the probabilistic theory of causality as
formulated in section II.2 provides necessary and sufficient conditions for causal laws. This
is a good study to consider because it has all the problems we raise, writ large. As a version
of the probabilistic theory, this account of causal laws is entirely rooted in a hunting
methodology, a very reliable methodology if ideally carried through, but concomitantly very
narrow in the range of the claims established and hence of extremely limited use. This is
reflected in the fact that causal laws, as Cartwright sees them, are always relative; in her
account they are relative both to a particular arrangement of confounders and to the
nomological machine – the underlying structure – that gives rise to the causal regularity.
22
21
For further discussion see Efstathiou (2009b).
22
Cf Cartwright (1989). Also note as in section 2.2 that we can drop the relativization to the arrangement of
confounders by relativizing instead to a population. Clearly C will cause some E’s in a population if it is

18
This means Cartwright’s causal laws are fragile: they are open to both the problems of
external validity and of unstable enablers.
What then about use? Cartwright gets this by an altogether different route, not from causal
laws at all but from capacities. Capacities are powers that agents possess to contribute in a
fixed way to what results whenever they are present. For example, because of their nuclear
structure permanent magnets have the capacity to attract metallic objects. This is a capacity
we measure in various experiments but the experiments themselves are not enough to tell us
that there is a capacity to be measured in the first place. The claim that magnets have such a
capacity is grounded in a large extended and complicated network of theory, experiment and
successful prediction.
Because the capacity will produce its contribution whenever it is present (or is properly
triggered)
23
knowledge of capacities can be extremely useful. But beware. For the magnet,
‘attraction’ is the contribution, not the actual motion (or not) that occurs when the magnet
operates. The attraction is always there even if the metallic object never moves. This is a
piece of information we can use reliably across a huge variety of situations. But the
predictions it gives rise to may not be as helpful as we wish. The metallic earring is stuck
between the floorboards. Shall we buy a magnet to get it out? We can reliably predict that
the magnet will attract the earring but we need a whole lot more information to predict that
the earring will move. The most we can definitely predict with the kind of knowledge we
usually have in these situations is that the magnet may very well pick up the earring that fell
between the floorboards.
So, even with capacities, predictive power is weak. But that is not the point here. What
matters for our worries about causal laws is that causal laws and capacities are entirely
distinct. This is so even if one has a very different account of causal laws from Cartwright’s.
None of the accounts of causal laws currently discussed in philosophy look anything like an
account of capacities; nor do our standard methods for testing causal laws serve well for
guaranteed to cause some E’s in some subpopulation of that population (i.e. a subpopulation that is
homogeneous with respect to confounding factors).
23
Except for chancy capacities, which produce their contributions only spasmodically.

19
establishing capacities. Capacities can be of use for predicting what happens when we set
policy or build new technological devices. That, however, does not salvage causal laws. In
introducing capacities Cartwright never meant to undermine causal laws. But focussing on
the distinction between capacities and causal laws points up the problem in bold relief: we
are very good at finding out about causal laws; but once we have done so, of what possible
use are they?
5. Solutions?
There are a number of strategies one might adopt to deal with worries about the usefulness of
causal laws. None really work.
a. Chuck Causal Laws. The first, obvious solution to these problems is to chuck causal laws.
Go for what you need. In Cartwright’s case this is capacities. It is not a causal law that is in
operation whenever the magnet succeeds in picking up the earring. It is instead the capacity.
But Cartwright’s account is metaphysically heavy. It postulates powers and in exactly the
way Hume despised. Capacities require a three-fold distinction between 1) the presence of
the capacity (e.g. whenever the magnet is present so too is the capacity to attract metallic
objects); 2) the exercise of the capacity (the attracting of the earring) and 3) the actual result
(the movement – or not – of the earring). Hume allows at most two of these but all three are
necessary to do the job.
If we follow Sandra Mitchell’s attack on ‘laws’, we can also chuck causality.
24
Mitchell
points out that for use we do not need laws; ipso facto we do not need causal laws. Any truth
can be useful so long as it is true where you propose to use it for prediction.
Classic instrumentalism can also be seen as a way of chucking causality and going directly
for what we need, much along the lines of the proposal we take from Mitchell. Science need
not establish laws; what’s needed are instruments that give us correct (enough) results for the
predictions we want to make.
24
Mitchell (2003) stresses the need for invariance without a detour through causality.

20
As liberating as it sounds, chucking causality doesn’t get rid of our troubles. First of all it
does not salvage causal laws. Are these just a wasted effort after all? Second, we still need to
be told both how to establish and how to use whatever substitute notions are proposed.
Which claims will be true in a particular setting, how should we use them and how do we
establish that they will be true? Chucking causality gives no bridge from language entry to
exit; from method to use. It just promises you can build another bridge, further down the
road.
b. Dubbing. A second solution is to go for what you need and dub it causality. This approach
is typical in econometrics. What makes for causality in econometrics is “structure”,
represented in structural equations. So, what’s structure? Structural equations are more or
less the ones that can be relied on for the predictions we want. But again this builds no
bridge from language entry to exit; from method to use. Econometrics is very good at telling
how to estimate parameters in structural equations – this looks much like language entry. But
there’s no good theory of what structure is that fits with both some good theory of ‘model
adequacy’ – what makes it okay to assert a set of ‘structural’ equations in the face of data –
and simultaneously with the assumption that the model can be used for prediction under
intervention or in new situations.
To reinforce this point, let us turn to an account of causality offered by macroeconomist and
methodologist Kevin Hoover that seems geared very much to use, unlike the philosophical
accounts we have looked at so far and more like familiar manipulation and intervention
accounts in philosophy.
25
(We discuss Hoover rather than the philosophers in part as a way
of introducing his work to readers outside the philosophy of economics and partly because
we can illustrate our point so cleanly with it.) Hoover defines causality directly in terms of
the effects that can be achieved by manipulation, real manipulation.
25
See Hoover (2001). It should be noted that the description given here of Hoover’s account is not one he is
happy with. Cartwright claims it is what his definitions say and takes the kind of causal relation described by
the definitions as a very important one different from more ‘mechanical’ kinds of causal relations. He maintains
that he intends his account to cover the more conventional notion of ‘mechanical causation’ and that various
caveats he offers allow his definitions to do so. For further discussion, see Cartwright (2007), Chapter 14.

21
Hoover: C causes E iff anything we can do to fix C partially fixes E but not the
reverse.
26
Although this definition secures a connection between causation and manipulation the kinds
of relations it calls ‘causal’ would not count as causal in everybody’s books – like
probabilistic theories of causation, causal process theories or Lewis-style counterfactual
accounts. Figure 2 provides an example of a simple mechanism to illustrate, where the u’s
are ‘policy levers’ – quantities we can manipulate, and the solid lines with arrows depict
pure ‘mechanical’ causation, like pushing on a lever at one end to trip a switch at the other.
The dotted line depicts ‘Hoover’ causation.
Figure 2
u
x
x u
z
y z
If Hoover
27
is right about what causation is, we can see why causes by their very nature can
provide predictions about strategies for manipulating the world. But many will not want to
allow that Hoover’s is an account of causation at all. More important, the account is entirely
rooted in use and does not does not supply any bridge to get there from method. Knowing
Hoover-causal facts will certainly be helpful in predicting the effects of policy interventions.
But how does one learn these facts in the first place? Indeed many standard methods for
causal inference will yield wrong verdicts vis-à-vis Hoover causation (e.g. methods that look
for causal processes and physical connections), many will often not be applicable (e.g.
Galilean experiments and Lewis counterfactual investigation in cases where the ‘miracle’-
like intervention is not among those we can do), and many will give ‘no’ answers to causal
26
Hoover causation thus is closely associated with the kind of ‘implementation neutral’ counterfactual that
Daniel Hausman proposes for investigating casual claims, but with the range of implementations restricted to
implementations we are able to bring about. See Cartwright (2007), Chapter 16.
27
Or better, with footnote 25 in mind: ‘Hoover as described here’.

22
relations where a more nuanced verdict would be of far more help (e.g. where E changes
under some manipulations of C but not others, C will not Hoover-cause E).
c. Add-ons. A third way is to be more demanding. Require conditions for both method and
use. Woodward takes this route. Besides level invariance he requires ‘modularity’ for a
relationship to pass as a causal law.
Modularity: there must be at least one way to ‘intervene’ to change any genuine
cause.
28
The effect of this requirement is that each variable appearing in a causal law operating in a
situation can be changed without changing anything else except the effects of changing that
variable. This again is a ‘miracle’-like change. What justifies this as a condition on
causality? Woodward is clear:
29
this addition allows us to use the relation in question for
predictions about manipulations. That, we take it, is why Woodward calls his account of
causality indifferently an ‘invariance’ account and a ‘manipulability’ account.
As an answer to our worries it is not satisfactory however. First, it underwrites the use of
causal laws for predicting the outcomes of manipulations not for the manipulations we might
be envisaging, but only for very special kinds of miracle-like interventions that change the
cause and nothing else. These are the kinds of manipulations that are demanded in a Galilean
experiment or in Lewis-style counterfactuals. That may well be how they come to play such
a special role in Woodward’s account. They are good for a very special way of testing for
causal laws. But they are no good for showing why knowledge of causal laws is useful for
real policy predictions.
28
This is not exactly how Woodward defines modularity but it is how he uses the notion sometimes and
especially to do just the job discussed here. See Woodward’s definition of modularity in Woodward
(2003,p.329)
29
Actually, he gives the same reason – causes must be usable to manipulate their effects – for both level
invariance and for modularity. We cite it only for modularity because level invariance does not provide
manipulability unless modularity is added and Cartwright at any rate has an alternative defense of level
invariance.

23
In this respect econometricians David Hendry and Robert Engle do better.
30
First they
require that causes be exogenous.
31
This is a technical notion that has to do with efficient
estimation from data. That facilitates language entry. What about language exit? They
require that as well: before they will call a relation ‘causal’ they demand that it not only be
exogenous but also ‘superexogenous’, which is a relative notion. A relation is
superexogenous relative to an envisaged policy change just in case it will remain true under
that change.
Accounts like Woodward’s and Hendry and Engle’s are what we call ‘add-on’ accounts.
They do not provide an account, a theory, of causality rich enough to justify our usual rules
for both language entry and language exit for causal laws. In particular they offer nothing
that shows why our standard methods for inferring causal laws leads us to claims that can be
used in the ways proposed. They just refuse to call something a causal law unless it can both
be admitted by their favourite test for causality AND can be relied on for the kinds of
predictions they describe.
This is no bridge at all from method to use. It is mere mereology. And mereology will not
save causal laws. Without the assurance of a bridge from standard method to use, the
assumption that the claim can be used in the ways described needs to be established on its
own, independently. But then why bother with the first half to begin with? What’s the point
of all the time, money, thought and effort that goes into putting the causal label on through
careful use of our best methods if the tests that do all the work in justifying our prediction
are altogether different ones? If mereology is the only alternative left, it seems preferable to
give up on causality altogether and instead adopt one of the first two strategies: give up on
causal laws and establish what we need. Then call it causality if you like.
30
See Hendry (2001) and Hendry (2004) and Engle et al (1983).
31
Also, in line with the probabilistic theory of causality there is in general the assumption that causes and
effects are probabilistically dependent.

24
6. Aside: Woodward versus Cartwright
It might be useful before concluding to juxtapose Cartwright’s capacities and Woodward’s
modularity demand. Woodward’s conditions on causality concern whole equations:
z = ax +by
If an equation is level invariant under miracle-like interventions on right-hand-side variables,
it must be a causal law under Cartwright’s axioms. Modularity demands that we call it a
causal law only if there is a miracle-like intervention for each right-hand-side variable.
Cartwright’s capacities are causal tendencies associated with individual features. The
strength of capacities is measured in Galilean experiments. In an equation like the one above,
if the quantities designated by x and y both have capacities with respect to z, then ‘a’ is the
strength of x’s capacity to produce z and ‘b’ is the strength of y’s capacity to produce z. Note
that there is no assumption that the equation is invariant. What equation holds is relative to a
particular setting, a particular arrangement of causes. The coefficients however are invariant.
The hypothesis that x and y have capacities with respect to z guarantees that they make the
same contribution (viz ax and by) whenever they are present. So they can be used to build
new equations to describe situations where they are present with other causes and without
each other.
Woodward’s mechanisms and Cartwright’s capacities are both difficult to establish. Beyond
that though they have complementary limitations and advantages. For Woodward, the whole
causal equation is invariant. This is good for prediction but bad for scope; we are restricted
to predictions in new situations that have exactly the same set of causes operating. Further,
Woodward causal laws describe what happens under miracle-like interventions. This is again
bad for scope. We cannot easily perform miracles so our causal laws won’t help much with
predictions about real life changes. Still, for Woodward there is at least a way to use a causal
law for prediction and maybe we’d better just find that way.
Capacities on the contrary are good for scope. Once established capacities can be carried to
new settings (addressing worries of external validity) and even, for ‘fundamental’ capacities,

25
across different underlying structures (dealing with unstable enablers).
32
For example, now
that it has been established that magnets have the capacity to attract metallic objects, the
attraction may be confidently relied on in new settings. But capacities are not as good as we
might hope for prediction. What is guaranteed with a capacity is that it will produce a fixed
contribution. That’s the bit that Hume would not like. What actually happens is far harder to
predict since it depends on what other causes are operating and what all their contributions
together add up to.
There are nice cases of course, in the sciences, where we know a set of capacities, each with
its fixed contribution, and we also know a rule of composition for how the contributions
‘add’.
33
Forces in mechanics and their rule of vector addition is the paradigm. And it is this
paradigm that Mill turns to when he defends the role of stable tendencies (from which
Cartwright’s capacities are copied) in the economy. But as Anna Alexandrova and Julian
Reiss argue, both from their studies of economic models,
34
for most cases, even cases where
we are strongly inclined to ascribe capacities, Mill’s hope for a rule of composition is daft.
Consider for example the case of the subjugation of women. We might well admit that
women do have the capacity for leadership and intelligence. What will result when this
capacity operates in a variety of real-life settings? Is it really reasonable to assume that in
each setting, a set, albeit possibly a large one, of other causes each with its own capacity is at
work and that the outcome of all acting together can be calculated by a fixed rule of
composition? Even in the case of the magnet this picture seems suspicious. To be sure, there
are cases where all the causes affecting the motion of a metallic object can be represented
neatly as vector forces, the magnetic force among, and the resultant motion calculated via
vector addition and the rule that the acceleration of the metallic object equals the resultant
force divided by the object’s mass. But it is a huge leap of faith to suppose that the dust and
spider webs between the floorboards can be regimented into this neat picture. The best
32
A good many capacities are derivative however. These too will depend on the underlying structure, or
‘nomological machine’, that gives rise to them.
33
Indeed, in most cases it is just because we know both a rule of composition and the contribution of a full set
of causes towards the effect that we can make sense of the idea of a contribution from any one of them.
34
Cf. Anna Alexandrova (2006) and Julian Reiss (2007).

26
Cartwright would be prepared to bet is true is that the magnet could well lift the earring. And
this remains a weak prediction!
7. Two Actions, Two Shores, Two Problems
We have distinguished between
1. two actions: hunting causes and using causes
taking place on
2. two shores: methodology and policy
that cannot be bridged because of
3. two problems: unstable enablers and external validity.
Together these paint a sad picture for causal laws. On current accounts it is either easy to
explain why some one or another of our best methods for hunting causal laws should be
reliable or it is easy to account for why causal laws can be used for predictions when we
propose to change the world. But no account on offer does both at once very well. There are
two obvious conclusions:
Conclusion 1: The state of philosophy reflects the state of nature. Causal laws are not worth
the paper they are written on. They can be found alright – and it is clear why our best
methods for finding them work so well: they just are ‘that which results from this method’.
Or they can be used just as we want. But our elaborate methods for testing are neither
necessary nor sufficient for claims that give true conclusions about policy manipulations. So
we might as well chuck all those elaborate tests.
Conclusion 2: There is a lot of work left for philosophy to do: to find good, rich theories of
causality that support method and use in one fell swoop.
We can hope for Conclusion 2, but as always the proof of the pudding will be when we have
the theory on our plate and have cut it open to find inside not just one but the two sixpences
of hunting and use together.

27
References
Alexandrova, Anna (2006), “Connecting Economic Models to the Real World”, Philosophy
of the Social Sciences, 36: 173-192.
Cartwright, Nancy (2007), Hunting Causes and Using Them, Cambridge University Press.
Cartwright, Nancy (2006), “Where is the Theory in our ‘Theories’ of Causality?”, Journal of
Philosophy, Vol. CIII, no. 2. 2006, 55-66.
Cartwright, Nancy (2003), “Two Theorems on Invariance and Causality”, Philosophy of
Science 70, 203-224; reprinted in Cartwright (2007) Hunting Causes and Using Them,
Cambridge University Press.
Cartwright, Nancy (1999), The Dappled World, Cambridge University Press.
Cartwright, Nancy (1989), Nature’s Capacities and their Measurement, Oxford University
Press.
Cartwright, Nancy (1979) “Causal Laws and Effective Strategies”, Nous, 13, 4, 419-37 (also
published in Cartwright (1983) How the Laws of Physics Lie, Oxford University Press).
Efstathiou, Sophia (2009a) “Nomological machines in science practice”, unpublished
manuscript.
Efstathiou, Sophia (2009b) “From Methods to Use: Is there a Here and There?”, unpublished
manuscript.
Engle, R. , Hendry, D. and Richard, J. F. (1983), “Exogeneity”, Econometrica, 51, 277-304.
Gasking, Douglas (1955), “Causation and Recipes”, Mind 64, 479-87.

28
Giere Ronald (1979), Understanding Scientific Reasoning. New York: Holt, Rinehart &
Winston.
Harper, W. L., Stalnaker, R. and Pearse, G. (eds.) (1981), Ifs: Conditionals, Belief, Decision,
Chance and Time, Dordrecht: Reidel.
Hendry, David (2001) Causality in Macroeconomics, Cambridge University Press.
Hendry, David (2004) ‘Causality and Exogeneity in Non-stationary Economic Time-Series’,
Causality: Metaphysics and Methods Technical Report CTR 18-04, CPNSS, London School
of Economics.
Hoover, Kevin (2001), Causality in Macroeconomics, Cambridge University Press,
Cambridge.
Lewis, David (1973), “Causation”, Journal of Philosophy, 70, 556-67.
Lucas, Robert (1981), “Economic Policy Evaluation: A Critique”, in Studies in Business
Cycle Theory, Basil Blackwell.
Lucas, Robert (1988), “On the Mechanics of Economic Development”, Journal of Monetary
Economics, January 1988, 22, 3-32.
Menzies, Peter and Huw Price (1993), “Causation as a secondary quality”, British Journal
for the Philosophy of Science 44, 187—203.
Mill, John Stuart (1869 [1997]), The Subjection of Women, Dover Publications, 1997.

29
Mill, John Stuart (1836 [1967]) “On the Definition of Political Economy and on the Method
of Philosophical Investigation in that Science”, reprinted in Collected Works of John Stuart
Mill, vol. IV, Toronto: University of Toronto Press
Mitchell, Sandra (2003), Biological Complexity and Integrative Pluralism, Cambridge
University Press.
Reiss, Julian (2007) Error in Economics: The Methodology of Evidence-Based Economics,
London: Routledge.
Suppes, Patrick (1970), A Probabilistic Theory of Causality, Amsterdam: North Holland.
Von Wright, Georg Henrik (1971), Explanation and Understanding, Cornell University
Press.
Woodward, James (2003), Making Things Happen, Oxford University Press.