CAUSAL COUNTERFACTUAL

THEORY FOR THE ATTRIBUTION

OF WEATHER AND CLIMATE-
RELATED EVENTS
BY A. HANNART, J. PEARL, F. E. L. OTTO, P. NAVEAU, AND M. GHIL

Causal counterfactual theory provides clear semantics and sound logic for causal reasoning
and may help foster research on, and clarify dissemination of, weather and climate-related
event attribution.

A
signif ica nt a nd grow ing par t of climate with the challenge of generating causal information
research studies the causal links between about episodes of extreme weather or unusual climate
climate forcings and observed responses. This conditions. This challenge arises from the needs for
part has been consolidated into a separate research public dissemination, litigation in a legal context,
topic known as detection and attribution (D&A). adaptation to climate change, or simply improvement
The D&A community has increasingly been faced of the science associated with these events (Stott et al.
2013). For clarity, we start by introducing a few no-
tations that will be used throughout this article: an
AFFILIATIONS: HANNART—IFAECI, CNRS/CONICET/UBA, event here is associated with a binary variable, say Y,
Buenos Aires, Argentina; PEARL—Computer Science Department, which is equal to 1 when the event occurs and to 0
University of California, Los Angeles, Los Angeles, California; when it does not, and we use the term event Y as an
OTTO —Environmental Change Institute, University of Oxford, abbreviation for the event defined by Y = 1. In any
Oxford, United Kingdom; NAVEAU —LSCE, CNRS/CEA, Gif-sur- event attribution study, the precise definition of the
Yvette, France; GHIL—École Normale Supérieure, Paris, France, event to be studied—that is, the choice of the vari-
and Department of Atmospheric and Oceanic Sciences, University
able Y—is crucial. Often, Y is defined ad hoc in the
of California, Los Angeles, Los Angeles, California
CORRESPONDING AUTHOR: Alexis Hannart, IFAECI, Ciudad
aftermath of an observed extreme situation based on
Universitaria, Pab. II, Piso 2, 1428 Buenos Aires, Argentina exceedance over a threshold u of a relevant climate
E-mail: [email protected] index Z, where both the index and the threshold are to
a large extent arbitrary. In the conventional approach,
The abstract for this article can be found in this issue, following the table
of contents.
which was introduced one decade ago by M. R. Allen
DOI:10.1175/BAMS-D-14-00034.1 and colleagues (Allen 2003; Stone and Allen 2005),
one evaluates the extent to which a given external
In final form 8 February 2015
©2016 American Meteorological Society climate forcing f F —where F encompasses, for
instance, solar irradiation, greenhouse gas (GHG)

AMERICAN METEOROLOGICAL SOCIETY JANUARY 2016 | 99

 emissions, ozone, or aerosol concentrations—has to be an object followed by another, where, if the first
changed the probability of occurrence of the event Y. object had not been, the second never had existed”
For this purpose, one compares the probability of oc- (Hume 2004, p. 48). Or, where X and Y are events:
currence of said event in an ensemble of model simu- Y is caused by X if and only if (iff) were X not to oc-
lations representing the observed climatic conditions, cur, then Y would not occur. Despite its dating back
which simulates the actual occurrence probability in to the eighteenth century, the above counterfactual
the real world, with the occurrence probability of the definition and the general approach to causality that
same event in a parallel ensemble of model simula- it implies is still relevant. Yet over the past decades,
tions, which represents an alternative world. The this definition has been further extended and refined
latter world is referred to as counterfactual, and it is within a probabilistic and graph–theoretical frame-
the one that might have occurred had forcing f been work, allowing for the counterfactual approach to
absent. To be precise, we introduce the binary variable be applied to actual datasets and to lead to reliable
Xf to indicate whether or not the forcing f is present. causal inference.
The probability p1 = P(Y = 1|Xf = 1) of the event oc- Overall, the current event attribution framework
curring in the real world, with f present, is referred obeys the spirit of counterfactual logic, and it is thus
to as factual, while p0 = P(Y = 1|Xf = 0) is referred to loosely connected to the above-mentioned corpus.
as counterfactual. Both terms will become clear in Yet it would be beneficial to tighten this connection
the light of what immediately follows. The so-called by adding several important concepts, definitions,
fraction of attributable risk (FAR) is then defined as and mathematical results of causal counterfactual
theory that, to the best of our knowledge, are lacking
. (1) in the current event attribution framework. Among
other lacking items, perhaps the most important one
The FAR is interpreted as the fraction of the likeli- regards the absence of definition for the word cause.
hood of an event that is attributable to the external Several recurrent controversial arguments in the
forcing f. Causal claims follow from the FAR and its realm of event attribution may possibly be related to
uncertainty, associated with model and sampling this lacking definition of causality: for instance, an
errors, resulting in statements such as “It is very argument often made (Trenberth 2012) is that any
likely that over half the risk of European summer single event has multiple causes, so one can never
temperature anomalies exceeding a threshold of assert that CO2 emissions, nor any other factors, have
1.6°C is attributable to human influence” (Stott et al. actually caused the event. Following this logic, single
2004, p. 612). events are thus inherently never causally attributable
This conventional framework and the FAR were at all. It is arguably difficult to clearly address this ob-
initially adapted from best practices in epidemiol- jection—or possibly many others—without a precise
ogy (Greenland and Rothman 1998), a field in which definition of causality in hand.
causal inference has always been of primary impor- The purpose of this paper is to propose a set of
tance. Best practices in epidemiology are themselves definitions and methodological extensions to the cur-
to some extent anchored in what can be referred to as rent event attribution framework that are rooted in
the standard theory of causality. Indeed, there exists a recent developments of causal counterfactual theory.
theoretical corpus of definitions, concepts, and meth- We start with a brief overview of the counterfactual
ods to define causality rigorously and to address the theory, emphasizing the most relevant concepts, and
issue of evidencing causal relationships empirically then proceed to illustrate the proposed extensions by
(e.g., Pearl 2000). The latter are readily accessible to revisiting the historical case study of the European
users and are progressively being implemented in a heatwave of 2003. Implications for causal claims are
growing number of fields. As a classic example taken finally discussed.
from epidemiology, statements of great importance
for public health, such as smoking causes lung can- A BRIEF OVERVIEW OF THE THEORY OF
cer, are often based on these shared definitions and CAUSALITY. We all deal with cause and effect
methods to investigate causality. The same is true of in our everyday life. Yet the notion of causality has
many causal studies that can be found in the fields of long been shrouded in controversy, and the field of
economics, social science, or artificial intelligence, to climate science is no exception in this respect. One
mention but a few domains of application. One point may argue that the main reason for this state of af-
of entry into the standard theory consists of the fol- fairs is the lack of clear semantics for causal claims;
lowing historical definition: “We may define a cause scientists and philosophers have indeed struggled to

100 | JANUARY 2016

define precisely when one event truly causes another random variables that is, of course, different from
and conversely when it does not. For instance, while that of causal dependence but proves instrumental
we all understand that barometers do not cause rain, in the formalization of causality. In the rainy episode
even such a simple fact cannot be easily translated example above, it is clear that the variables B and R are
into a precise formalization or a mathematical equa- dependent, which of course does not imply anything
tion. Beside this semantic difficulty, a fundamental about their causal relationship. If we now introduce
question is to determine what evidence is required to the variable W to denote whether or not a road near O
justify the causal claim that the falling barometer did is wet, then the rain R and the wet road W are clearly
not cause the rainy episode and how such evidence dependent and this is also the case of the barometer
may be extracted from observations. B and the wet road W. Once we know, however, that
Consider a naive observer O who knows nothing it has rained, we can deduce that the road is certainly
about either meteorology or barometers. By recording wet no matter the evolution of the barometer, so that
the movements of the barometer’s needle together W is independent of B conditionally on R. This im-
with the changes in weather during a few weeks, O portant property is called conditional independence
may be tempted to infer from the repeated observa-
tion of rainy episodes being preceded by a barometer P(W|B,R) = P(W|R); (2)
fall and of sunny ones being preceded by a rise that the
needle’s movement actually did cause the weather to this equation basically expresses that R screens off B
change—even without a clue with respect to (wrt) the from W. If we further complement our illustration
physical mechanism that may account for this causal by introducing L, which denotes whether or not a
relationship. However, O’s causal hypothesis will be low pressure meteorological system is present above
quickly ruined if she/he has a flash of inspiration to O, one can see by following a similar reasoning that
start experimenting with the barometer; forcing its P(R|B,L) = P(R|L) and P(W|R,L) = P(R|L), that is, that
needle up and down will soon convince O that acting L screens off B from R and that R screens off L from W.
on the barometer does not induce a weather change. Oriented graphs are a very useful tool to visualize
This simple example illustrates two aspects of causal- these considerations and can be considered as the
ity: first, that causal investigation relies crucially on second building block of causal theory (Pearl 2000).
observations, and second, that two different types of Skipping the rigorous definitions, a graph can be de-
observations may be used by the causal investigator scribed as a mapping of the conditional dependence
(experimental and natural, i.e., nonexperimental). relationships prevailing within a given joint prob-
While both of these aspects may seem obvious, the ability distribution P(Z1,Z2,...,Zn) under study (Pearl
difficulty starts with the implementation; given a 2000; Ihler et al. 2007). Each variable Zk is thus repre-
piece of data, experimental or not, what causal con- sented by a node, which is connected to one or more
clusions can be drawn from it? And what is the level of nodes by arrows; each arrow points from a parent to a
confidence associated with such causal conclusions? child. It is thus intuitive that graphs complement the
Over the past decades, a rigorous theory of causality purely probabilistic notion of dependence, which is
has emerged and been consolidated, with the purpose symmetric and noncausal, by introducing an asym-
of addressing these questions. Its main ideas and metry in the connections between variables, which
concepts are exposed next. is suited to encode causal relationships. The graph
associated with (Z1,Z2,...,Zn) may be understood as a
The mathematical basis of causal theory. The counter- visual representation of the following factorization:
factual definition of causality given by David Hume
and spelled out above—that is, Y is caused by X iff Y , (3)
would not have occurred were it not for X—can be
used to introduce this brief overview. For instance, where P k denotes the parents of variable Zk . The
let R be a rainy episode and B be a downward move graph representing causality in our illustrative wet
of the barometer’s needle; then, observing R while road example is shown in Fig. 1a and visually encodes
impeding B—that is, by holding the barometer’s the following factorization:
needle—provides counterfactual evidence that falling
barometers do not cause rain. Applying this approach P(B,R,W,L) = P(L)P(B|L)P(R|L)P(W|R). (4)
to data requires a few mathematical concepts from
the theory of probability and from graph theory. The Causal relationships among a set of variables
former entails the notion of dependence between can thus conveniently be represented by their joint

AMERICAN METEOROLOGICAL SOCIETY JANUARY 2016 | 101

FIG. 1. Graphs representing dependencies (a) among the four variables (R, B, W, L) used in our illustrative ex-
ample and (b) among forcings (X1, X2) and climate response Y. Dotted arrows represent dependency upon the
unobserved variable υ.

probability distribution, provided conditional relevant because other factors potentially affecting Y
dependence relationships are fully specified; such may not be controlled in the experimental setup. The
specification is conveniently encoded by using an ori- notion of intervention was hence introduced to de-
ented graph in which each arrow represents a causal scribe the situation where X is set by the experimenter
relationship. The existence of causal relationships at a chosen value x; it is denoted do(X = x). The notion
has various implications on the joint dependence of interventional probability then corresponds to the
structure; for example, independent causes become distribution of Y obtained in an experiment under the
dependent conditional upon their common effect, intervention do(X = x). It is denoted P[Y|do(X = x)] or
and dependent effects become independent condi- alternatively P(Yx), where Yx denotes the new random
tional upon their common cause. From the moment variable obtained for Y subject to the intervention
we have access to enough observations to infer the do(X = x). The set {P(Yx = y)|x,y = 0,1} obtained by
dependence structure, we are able to detect these collecting all the interventional probabilities of Y for
signatures and thereby provide evidence of causal every possible value of X is termed the causal effect of
relationships. Algorithms such as those described in X on Y. It is important to note that, in general,
Spirtes et al. (2000) and Shimizu et al. (2006) basically
follow this strategy and could perfectly be applied to P[Y|do(X = x) ≠ P(Y|X = x), (5)
the natural observations of R, B, and L collected by O.
An important limitation of using natural data which is why the notation do(X = x) is required.
is that several graphs can be compatible with the Indeed, P(R = 1|B = 1) reads in our example the
same joint distribution and hence with the same probability of rain knowing that the barometer is
observations; identifiability is an issue. For instance, decreasing in a nonexperimental context in which the
simultaneous changes in X and Y are compatible barometer evolution is left unconstrained, whereas
with both the causal relationships X → Y and Y → X P[R = 1|do(B = 1)] reads the probability of rain forc-
whenever only these two variables are observed (e.g., ing the barometer to decrease in an experimental
when observing R and B but not L). The experimental context in which the barometer is manipulated. The
approach is thus required for disambiguation of the two probabilities are obviously distinct, and it is their
causal relationship between X and Y. Several out- difference that allows for disambiguation, as it reveals
comes Y are thereby experimentally collected for each the absence of a causal link between B and R.
tested value of X. The value of X is thus chosen by the Nonetheless, confusion is still possible because
experimenter, and treating it as a random variable is P[Y|do(X = x)], and P(Y|X = x) may also sometimes
no longer relevant in this experimental context. How- be equal. This is the case when X satisfies a property
ever, a probabilistic treatment of the response Y is still called exogeneity wrt Y. Without going into details,

102 | JANUARY 2016

a sufficient condition for X to be exogenous wrt any to discuss which of the three probabilities are most
variable is to be a top node of a causal graph. In the relevant for causal attribution, in which context, and
present context, radiative forcings under causal scru- how they should be interpreted.
tiny are actually modeled in a physical setting, such On the one hand, PN closely matches the reason-
as a general circulation model (GCM), as prescribed ing used in lawsuits, where legal responsibility is
conditions that are external to the climate system; they understood counterfactually, that is, in the sense of
are thus exogenous by construction. Provided D&A necessary causation. In such a context, PN equals
keeps on focusing on causal relationships between the probability that the damage Y suffered by the
variables that are exogenous, the otherwise critical plaintiff would not have occurred were it not for the
distinction between conditional and interventional defendant’s action X, and the latter is declared guilty
probability is therefore not of utmost importance here whenever it can be proven that PN is high enough;
because both quantities are actually the same. the threshold is explicitly set to 1/2 in a civil case
(preponderance of the evidence) and to an unspeci-
Necessity, sufficiency, and probabilities of causation. To fied value that is supposedly very close to one in a
assess how likely it is that one event was the cause criminal case (beyond reasonable doubt). Assume for
of another, the probability PN of necessary causal- instance that individual A fires a gun (X) in a seem-
ity is defined, in agreement with the counterfactual ingly deserted but public place. Unluckily, individual
principle, as the probability that the event Y would B, who happens to be standing 1 km away, is hit and
not have occurred in the absence of the event X given injured (Y). Legally speaking, A is the obvious culprit
that both events Y and X did in fact occur. The prob- for the injury of B and will likely be convicted in case
ability PN thus quantifies how likely it is that X has of a trial because PN is very close to unity here; B
caused Y in a necessary causation sense; here X is a would be safe and sound had it not been for A shoot-
necessary cause of Y means that X is required for Y ing. Nevertheless, the probability of the bullet hitting
to occur but that other factors might be required as someone from such a long distance is very low, the
well. In other words, it means that Y would not occur lightest wind gust could possibly have deviated its
were it not for X. Sufficient causation, on the other trajectory and saved B. The probability of sufficient
hand, as in X is a sufficient cause of Y, means that X causation PS is thus close to zero here, but this is not
always triggers Y but that Y may also occur for other important in a legal context, in which it is only PN
reasons without requiring X. The probability PS of that matters, while PS does not.
sufficient causation is defined to be the probability In contrast, consider the case of a policymaker who
that Y would have occurred in the presence of X, given aims at reducing the number of casualties from acci-
that Y and X did not occur. Note that PN and PS are dental shootings (Y) through a policy (X). An abrupt
thus simultaneously interventional and conditional policy prohibiting gun sales altogether will clearly be
probabilities. To complete the probabilistic setting, sufficient but arguably not necessary since a smoother
PNS is the probability of necessary and sufficient policy based on tightly regulated sales may achieve a
causation. It is defined as the probability that Y would similar result. In parallel, improving the dissemina-
have occurred in the presence of X and that Y would tion of safety information to gun owners is arguably
not have occurred in the absence of X. These three necessary but will likely not be sufficient. In any case,
definitions are formally expressed as follows (Pearl it is a high PS that guarantees that the desired objec-
2000, p. 286): tive Y will be met by the policy X, not a high PN; PS
therefore tends to be more important than PN in the
PN =def P(Y0 = 0|Y = 1, X = 1), context of elaborating and assessing policies.
PS =def P(Y1 = 1|Y = 0, X = 0), (6) Even though all three probabilities relate to coun-
terfactual worlds, it is worthwhile underlining that
PNS =def P(Y0 = 0, Y1 = 1). these quantities are not nebulous metaphysical no-
tions: the definitions are precise and unambiguously
The three probabilities PN, PS, and PNS are of implementable, as long as a fully specified probabi-
utmost importance because they provide a complete listic model of the world is postulated. That being
characterization of the causal relationship between X said, it is still a difficult task to derive them under
and Y as well as of the associated uncertainties. Their general assumptions and one that remains an active
estimation can thus be viewed as the ultimate purpose and challenging research topic in causal theory at
of a causal attribution study. Before addressing the present. Important results were obtained, however,
issue of deriving them in practice, it is enlightening by introducing some additional assumptions. For

AMERICAN METEOROLOGICAL SOCIETY JANUARY 2016 | 103

 instance, under the assumption of monotonicity, the perspective of the plaintiff—or the judge or insurance
following exact expressions hold: contract holder—and the ex ante perspective of the
planner—or the policymaker or campaigner. In the
first case, the question of who is to blame for the event
that occurred—with potentially many implications of
its answer—is central. The problem of climatic event
(7) attribution can thus be compared to a lawsuit and
actually does already appear in courts (Adam 2011);
we may primarily seek to determine responsibilities
for the event and its aftermaths, where responsibility
where variable Y is said to be monotonic wrt variable is understood in a legal sense, that is, in a necessary
X iff for any realization ω, in the probability space causation sense. Event attribution thus requires the
Ω, Yx(ω) is a monotonic function of x. Furthermore, adversarial debate typical of a lawsuit in order to
when assuming exogeneity of X wrt Y in addition to cautiously balance incriminating versus exonerating
monotonicity, the expressions given in Eq. (7) sim- evidence, that is, to evaluate the main cause under
plify because interventional and conditional prob- scrutiny, for example, anthropogenic forcings, as well
abilities are then equal, that is, px = P(Yx = 1) for as each and every possible alternative explanations,
x {0.1}, and thus for example, natural forcings or internal variability of
the climate system, which may have led to the same
outcome. If the resulting PN is high enough, then
human responsibility is established and a ruling may
in theory follow, as it does in litigation cases. In any
(8) case, as in the imprudent shooter example, PS does
not matter here, only PN does.
By contrast, the planner is looking forward and
may ask instead the general type of question what
Note that, under such conditions and provided p1 ≥ p0, should be done today wrt events that may occur in
PN matches with the FAR; we elaborate on this co- the future? For instance, in the context of mitigation,
incidence further in this article. Another important two causal questions are at stake: on the one hand,
result of causal theory that is linked to Eq. (8) is what is the, expectedly beneficial, effect of limiting
that under exogeneity and releasing the assumption CO2 emissions? And, on the other hand, what is the,
of monotonicity, the probabilities of causation are expectedly costly, effect of not limiting them? The
then no longer identifiable, but the three quantities first question seeks a causal guarantee that remov-
1 – p0/p1, 1 – (1 – p1)/(1 – p0), and p1 – p0 provide lower ing the forcing will make the event less frequent,
bounds respectively for PN, PS, and PNS. Figure 2 and the concern is thus predicated on necessary
shows a plot of the expressions given in Eq. (8); it causality. Conversely, the second question seeks a
can be seen that PN is more sensitive to p0 than to causal guarantee that maintaining the forcing will
p1 and conversely that PS is more sensitive to p1 than maintain the event frequency, and the concern is thus
to p0. Necessary causation is enhanced further by an predicated on sufficient causality. Therefore, PS is
event being rare in the counterfactual world, whereas the appropriate focus for the planner when assessing
sufficient causation is enhanced further by its being the future costs that inaction will imply, but PN is at
frequent in the real one. This being said, PN and PS stake when assessing the future benefits of enforcing
are clearly not independent and coincide under two strong mitigation actions. Policy elaboration requires
situations: (i) when p0 + p1 = 1 (e.g., in a deterministic both sides of this assessment; thus, both PN and PS
context where p1 = 1 and p0 = 0, then both PN and are of interest here. To summarize, depending on the
PS = 1), and (ii) when p0 = p1 (e.g., where the counter- context, PN, PS, or both may be relevant and can help
factual and real worlds’ responses are identical, then answer different causal questions.
both PN and PS = 0).
Methodological proposal. Our methodological pro-
CAUSAL ATTRIBUTION OF CLIMATE- posal for the attribution of weather and climate-related
RELATED EVENTS. Choosing to focus on PN or events is rather straightforward, and it is derived from
PS is a matter of point of view. To illustrate this issue, previous considerations. It consists of deriving the
we can consider two typical perspectives: the ex post probabilities of necessary and of sufficient causality,

104 | JANUARY 2016

PNf and PSf associated with
the causal relationship be-
tween each forcing f F ,
and an event Y of interest.
As outlined in the introduc-
tion, the choice of Y is based
on a climate variable Z and a
threshold u; this choice de-
pends on the causal focus of
the study and is otherwise
rather arbitrary. Once Y has
been duly defined, the caus-
al chain to be investigated is
actually quite simple, not-
withstanding the complex-
ity of the climate system. It
can be represented by the
single, standard graph of
Fig. 1b, independently of
the specificities of the event
Y under scrutiny. A set of
binary variables {Xf : f F }
that represent the external
forcings occupy the top
nodes in this graph and
are thus exogenous. The
event variable Y has parents
P = {Xf : f F }, and it is FIG. 2. Contour plots of (top left) PN, (top right) PS, (bottom left) PNS, and
also influenced by internal (bottom right) PN – PS as functions of the counterfactual probability p 0 (hori-
climate variability υ that zontal axis) and of the factual probability p1 (vertical axis).
is treated here as random
terms (Ghil et al. 2008).
Next, we can apply Eq. (8) because all the forc- allowing for a direct estimation of p1 and p 0. Most
ings are exogenous, and one may also assume that unfortunately, in the climate sciences, no such sample
the event Y is monotonous wrt the forcing. Indeed, of Earth-like climate systems is accessible to natural
assuming that the latter does not hold would imply observation and even less so to experimental testing.
that despite the event being more frequent in the The paleoclimatic record may in theory palliate this
factual world than in the counterfactual one (i.e., difficulty by considering several remote episodes
p1 > p0), there exists some realizations ω Ω, such of Earth’s climatic history as a sample (National
that Y0(ω) = 1 and Y1(ω) = 0. That is, one can find Research Council 1995). An important limitation of
some conditions under which the event does occur this approach, however, is the limited size and high
when the forcing is turned off but no longer occurs uncertainty of the indirect paleoclimatic estimates
only by turning it on—other conditions being held of both the response Y and the forcings Xf over the
unchanged. Such conditions are arguably not realis- distant past. Furthermore, such nonexperimental
tic physically for a broad class of events and for the analysis is inherently restricted to forcings that can be
forcings usually considered in D&A. We thus derive traced to paleoclimatic perturbations that did occur
PN = 1 – p0/p1 and PS = 1 – (1 – p1)/(1 – p0) for each and for which exogeneity is guaranteed. With such
forcing f and omit hereinafter, for simplicity, the index strong limitations on the natural observation side and
f. Hence, the challenge is now to estimate the causal with in situ experimentation inaccessible, we are left
effects {p0, p1}. In many fields, experimental and/or with the only remaining alternative: so-called in silico
natural observations of a response Y—say, in epidemi- experimentation. This option is rendered plausible by
ology, a disease—and of a factor X—say, a bad habit or the increasing realism of climate system models that
a treatment—are available for a sample of individuals, were developed partly for this purpose. Estimates of

AMERICAN METEOROLOGICAL SOCIETY JANUARY 2016 | 105

 the causal effects {p 0, p1} can be obtained from an corresponding two ensembles by fitting a general-
ensemble of numerical experiments consisting of r1 ized Pareto distribution to each one (cf. Fig.
and r0 runs under factual and counterfactual condi- 3a). The inference procedure yields two ranges
tions, respectively, wrt one or more forcings f. An of values for the return periods: 350 ≤ T0 ≤ 2500 and
obvious estimation strategy is to use the empirical 100 ≤ T1 ≤ 1000. For the sake of clarity, we choose
^
frequencies px = for x {0,1}, where Yx(k) to concentrate here on two values that are arbi-
is the event occurrence in the kth run of the factual trarily chosen within these ranges: T0 = 1250 yr and
or counterfactual experiment. This option presents T1 = 125 yr, implying p0 = 0.0008 and p1 = 0.008.
^
a major shortcoming since px, as well as PN and PS, • These values of p 0 and p1 yield PN = 0.9 and
are affected by high sampling uncertainty. In practice, PS = 0.0072, by applying Eq. (8).
because of restrictions on computer resources, rx is
typically in the range of 10–100, while asymptotic It follows that CO2 emissions are very likely to
convergence requires rx to be large compared to the be a necessary cause, but are virtually certainly not
return period Tx 1/px of the event; the latter is a sufficient cause, of the summer of 2003 heatwave.
clearly out of reach for the rare events usually at stake. This statement highlights a distinctive feature of
Another serious difficulty is that climate models, unusual events: several necessary causes may often
including the most detailed GCMs, are simplified be supported by the data but rarely a sufficient one.
representations of reality that are affected by both To further illustrate this point, we plot PN, PS, and
numerical and physical modeling errors. Thus, the PNS as a function of the threshold u in Fig. 3b. It is
real causal effects may differ from the model causal clear from this figure that the causal evidence shifts
effects. While both these difficulties are serious, they from necessary and not sufficient when u is large
can be addressed by introducing additional assump- (unusual event) to sufficient and not necessary when
tions on the distribution of the climate variable Z and u is small (usual event). This shift occurs because, in
by treating model error as an additional random term the latter case, it is the nonoccurrence of event Y that
influencing the response variable Y. Discussing such becomes an unusual event. But this rare “nonevent”
approaches is beyond the scope of this paper. The tends to be less unusual in the counterfactual world
probabilities PN and PS are then derived from the than in the factual one, which implies necessity for
^ ^
estimates p1 and p0 so obtained. the “nonevent” and thus sufficiency for the event by
Causal claims are eventually formulated from the definitions of PN and PS, respectively, in Eq. (6).
these probabilities and translated into words based In any case, a low threshold conversely yields
on standardized uncertainty wording, such as the PN 0 and PS 1; it follows that anthropogenic CO2
one used in IPCC (2013). Summarizing, the general emissions are virtually certainly a sufficient cause,
methodological approach proposed herewith consists and are virtually certainly not a necessary cause, of
of the following: the fact that the summer of 2003 was not unusually
cold. Therefore, this symmetrically illustrates that
• Define a response variable of interest Y based on the occurrence of a usual event—or equivalently, the
a climate index Z and threshold u. nonoccurrence of a rare event—is thus often prone
• Infer the causal effects associated with Y, based on to have a sufficient cause but rarely necessary ones.
in silico experimentation. The above analysis defines the occurrence of the
• Derive PN and PS for each forcing and formulate 2003 European heatwave wrt to the particular year
associated causal claims by using, for instance, the when it occurred. Such a definition of the event inher-
IPCC (2013) uncertainty terminology. ently considers that the particular year of occurrence
(2003) is a relevant feature thereof and consequently
2003 European heatwave. We illustrate our approach builds this feature into the causal analysis. This ap-
by revisiting one of the first counterfactual event attri- proach is particularly relevant in the context, say, of
bution studies (Stott et al. 2004), which focused on the an insurance contract, which may often apply only
European heatwave of the summer of 2003. Applying to a single specified year. But a broader perspective
our notation and the above three steps to this study, focusing on longer time scales is arguably more rel-
evant in other contexts, such as elaborating adapta-
• Z is the mean summer temperature anomaly over tion and mitigation policy, which has no reason to
Europe, and u is set at 1.6°C. grant any particular importance to the year 2003. In
• The factual and counterfactual probability den- such a context, one would release the year 2003 as
sity functions (PDFs) of Z are obtained from the an event feature and focus instead on the fact that a

106 | JANUARY 2016

severe European heatwave did occur. The meaningful
temporal feature retained here would be occurrence
during the industrial period instead of occurrence
during year 2003. It is straightforward to translate
this approach into our proposed framework by going
through the same three steps again. In what follows,
for clarity, we denote with an asterisk the new vari-
ables Y*, Z*, and u*:

• The variable Z* is defined as the number of occur-

rences of European heatwaves over a time period of
length τ ending in 2003, where in any given year a
heatwave occurrence is defined as above by Z ≥ u,
and the threshold u* is set to 1. The event Y* thus
occurs if at least one heatwave took place in Europe
during the time interval 2004 – τ ≤ t ≤ 2003.
• Deriving the new causal effects {p0*, p*}
1
is straight-
forward, subject to assuming stationarity wrt time
(see discussion immediately below) based on the
previous causal effects {p0*, p*}:
1

p*x = P(Z*x ≥ 1) = 1 – (1 – px)τ. (9)

For τ = 1, this equation reduces to p*x = px since

Y* = Y in this case. When τ is large compared to
the return period of event Y (i.e., τ large compared
to 1/px), it implies p*x 1; this is also unsurprising
because in either the factual or the counterfactual
world, the occurrence of a heatwave, no matter
how rare in any given year, is certain over a suf-
ficiently long period.
• Plotting in Fig. 3c PN* and PS* as a function of τ,
based on Eq. (9), we see that the causal evidence FIG. 3. Causal inference for the 2003 European heat-
wave. (a) Counterfactual and factual PDFs of the tem-
shifts from necessary and not sufficient in the
perature anomaly index, using a generalized Pareto
limiting case τ = 1 (since Y* = Y) to sufficient and distribution fit after Stott et al. (2004); (b) probabilities
not necessary when τ gets asymptotically large. For PN, PS, and PNS as a function of the threshold u; and
τ = 200 yr—that is, the industrial period, which (c) PN, PS, and PNS as a function of the length of the
matches approximately the instrumental record observation period τ.
length—we find from Eq. (9) that p 0 = 0.14 and
p1 = 0.80 and next that PN* PS* 0.8. under scrutiny, or not, has crucial implications for
the associated level of causal evidence. Replacing the
It follows that anthropogenic CO2 emissions are feature year of occurrence by the feature occurrence
likely to be both a necessary cause and a sufficient during the industrial period may be more relevant to
one for a 2003-like heatwave to have occurred at the analysis in many situations and yield more power-
least once over the industrial period. In summary, ful causal evidence.
sufficient causality does not apply to the event oc- This being said, the stationarity hypothesis under-
currence on the particular year when it did occur, lying Eq. (9) is unrealistic because mean temperature
but it does for such an event to have occurred at least did change over the period considered and so did
once over the entire period. Evidence of necessary extremes. This convenient assumption was made here
causality, on the other hand, is strong in both cases. for the sake of illustrating in a simple and qualitative
This illustrative example thus shows that whether one way the effect on PN and PS of defining the event
considers something as fortuitous as its particular occurrence on a longer period of length τ. While a
year of occurrence to be a relevant feature of the event realistic nonstationary treatment of this case study

AMERICAN METEOROLOGICAL SOCIETY JANUARY 2016 | 107

 is beyond our scope, it is important to underline a simple methodology for its application to D&A
that including assumptions of nonstationarity into studies. We hope that this methodological frame-
a causal inference study presents no particular dif- work—along with the more precise vocabulary it
ficulties in general. For instance, in the present case relies on—will help clarify discussions between D&A
study, this may be done merely by using the more experts as well as communication to wider audiences.
general expression We have shown, with simple examples, that it
is important to distinguish between necessary and
(10) sufficient causality. Such a distinction is, at present,
lacking in the conventional event attribution frame-
work. Any time a causal statement is being made
in place of Eq. (9) in order to determine the causal about a weather or climate-related event, part of the
effects {p0*, p*}.
1
In Eq. (10), px,t denotes the probability audience understands it in a necessary causation
of occurrence of a heatwave in year t and is thereby sense, while another part understands it in a suf-
allowed to change over time. In practice, (px,t)t=1 τ
may ficient causation sense, which can give rise to many
be estimated based on an ad hoc statistical model potential misunderstandings. Introducing the clear
accounting for nonstationarity. For instance, a com- distinction may thus clarify discussions. Specifically,
monplace choice for the latter is to specify the PDF it may for instance help address the claim recalled in
of the index Z in year t conditionally on a covariate the “Background and rationale” section, according to
that changes in time (e.g., mean temperature) and/ which single events are never attributable since they
or an explicit parametric dependence to time t (e.g., are multicaused. In light of what precedes, this claim
a linear trend). Note that Eq. (10) would clearly be intrinsically postulates that a cause qualifies as such
required for the estimation of p*1 because the factual only if it is both necessary and sufficient. The latter is
world has undeniably changed. Yet Eq. (9) may still be arguably far too restrictive an approach of causation.
considered acceptable for the estimation of p0* since Our revisiting the well-known case study of the
the counterfactual world would arguably have suf- European heatwave of 2003 should clarify an apparent
fered limited changes. Accordingly, one may expect paradox in the interpretation of such studies. Even in
that when moving to a nonstationary treatment (i) p0* the few such cases where evidence supporting neces-
would only be marginally affected and (ii) p*1 would sary causation is strong, assertive causal statements
potentially be substantially affected. More precisely, appear to have been shied away from, possibly by the
one would expect p*1 to have a lower value because perception that sufficiency was lacking. A statement
px,t is expected to be lower than its value in the year such as “CO2 emissions have not caused the particular
2003 for any year t preceding it. Therefore, based on event Y: they have only caused the probability of occur-
the above considerations and on Fig. 2, accounting rence of Y-like events to increase” may actually often
for nonstationarity would expectedly translate here be too conservative and even wrong; as in the above
into a slight decrease in PN, a potentially pronounced example, it may indeed be the case that CO2 emissions
decrease in PS, and a lower level of causal evidence did cause event Y, although in a restrictively necessary
overall—as compared to the values given above for causation sense. Further, by defining the event to mean
illustration. not just occurrence in a particular year but during the
In any case, each of the different perspectives taken entire industrial era, it may be possible to establish
above addresses a causal question about the 2003 that event Y was in fact caused by increased CO2 emis-
heatwave that is different and may be of interest for sions—this time wrt both necessity and sufficiency.
distinct purposes. But while the questions only differ Our proposed methodology, like the conventional
slightly, the answers vary greatly. The answer to such one, relies on in silico experimentation to derive both
an open question as have CO2 emissions caused the the factual and counterfactual probabilities p1 and
2003 European heatwave is thus dramatically affected p0 , respectively; use the two to obtain the quantity
by (i) how one defines the event 2003 European heat- 1 – p0/p1 and then translate it into a causal statement.
wave and (ii) whether causality is understood in a nec- Our extended framework, however, has important
essary or sufficient sense. Precise causal answers about distinctive features. First, we have shown that 1 – p0/p1
climate events thus require precise causal questions. is associated only with the first facet of causality,
that of necessity, and we have introduced its second
CONCLUDING REMARKS. We have provided facet, that of sufficiency, which is associated with the
an introduction to causal theory, as used in causal symmetric quantity 1 – (1 – p1)/(1 – p0). Both have
studies across several disciplines, and proposed been shown to be relevant depending on the context.

108 | JANUARY 2016

Second, the interpretation given to 1 – p0/p1 differs counterfactual model simulations. This carries an
under both frameworks, which has deep implica- immediate implication wrt the design of standard-
tions for the formulation of causal statements and ized Coupled Model Intercomparison Project (CMIP)
the treatment of uncertainty. The quantity 1 – p0/p1 experiments that specifically address D&A purposes.
was coined as the fraction of attributable risk upon The present analysis suggests moving toward a fully
being introduced in event attribution, and similarly counterfactual design in the future—that is, all forc-
in other applied fields, terms like excess risk ratio, ings except f being on—instead of the mostly factual
attributable fraction, or attributable proportion are one prevailing at present—that is, forcing f only being
also used to name the same quantity. The FAR, as on. Generalizing this design would be a significant
well as these similar terms, is used to communicate step forward in attribution studies of weather and
the idea—particularly relevant in epidemiology from climate-related events.
which it originates—that the exposition to a given
risk factor X translates into an increase of, say, the ACKNOWLEDGMENTS. Part of this work was
frequency of a given disease Y. In this terminology, supported by the French Agence Nationale de la Recher-
the quantity 1 – p0/p1 is a frequency increase index; che Grants DADA (AH, PN and MG), MCSim (PN),
it corresponds to a statistical monitoring approach, and MOPERA (PN) and U.S. National Science Foun-
which is more descriptive than structural, in the sense dation Grants DMS-1049253 and OCE-1243175 (MG).
that it does not embed any precisely defined causal
meaning. For this reason, Pearl (2000) has argued that
the term attributable risk is a misnomer; because such REFERENCES
a precise causal meaning is lacking, the associated Adam, D., 2011: Climate change in court. Nat. Climate
statement can only address the increase in frequency. Change, 1, 127–130, doi:10.1038/nclimate1131.
Accordingly, uncertainty analysis conducted on the Allen, M. R., 2003: Liability for climate change. Na-
FAR by deriving its probability distribution cannot ture, 421, 891–892, doi:10.1038/421891a.
be easily translated into uncertainty on the causal Ghil, M., M. D. Chekroun, and E. Simonnet, 2008:
link at stake; instead, the focus on the frequency Climate dynamics and fluid mechanics: Natural
increase and its uncertainty yields statements like variability and related uncertainties. Physica D,
“there is a 90% confidence level that CO2 emissions 237, 2111–2126, doi:10.1016/j.physd.2008.03.036.
have increased the frequency of occurrence of Y-like Greenland, S., and K. J. Rothman, 1998: Measures of
events by a factor at least two.” effect and measures of association. Modern Epidemi-
In causal theory, the probability of necessary ology, 2nd ed. K. J. Rothman and S. Greenland, Eds.,
causation PN formally embeds the notion of causal Lippincott-Raven, 47–66.
attribution in its definition, given by Eq. (6). While PN Hume , D., 2004: An Enquiry Concerning Human
is not easily computable in general, it coincides with Understanding. Dover, 226 pp.
1 – p0/p1 under exogeneity and monotonicity. These Ihler, A. T., S. Kirshner, M. Ghil, A. W. Robertson,
two rather restrictive conditions are fortunately met and P. Smyth, 2007: Graphical models for statistical
in the context of D&A, thus the quantity 1 – p0/p1 usu- inference and data assimilation. Physica D, 230,
ally referred as FAR now has a precise causal meaning, 72–87, doi:10.1016/j.physd.2006.08.023.
instead of being merely an index of frequency increase. IPCC, 2013: Summary for policymakers. Climate
This shift in interpretation affects the associated causal Change 2013: The Physical Science Basis, T. F. Stocker
claim, which can now address more directly the actual et al., Eds., Cambridge University Press, 1–29.
causal link. Moreover, this shift has an immediate National Research Council, 1995: Natural Climate
implication in terms of assessing the uncertainty of Variability on Decade-to-Century Time Scales. Na-
the claim: the latter is indeed already quantified be- tional Academy Press, 630 pp.
cause PN is a probability, which inherently measures Pearl, J., 2000: Causality: Models, Reasoning, and Infer-
uncertainty. Therefore, based on the same supporting ence. Cambridge University Press, 384 pp.
data, the new interpretation translates into “CO2 emis- Shimizu, S., P. Hoyer, A. Hyvarinen, and A. Kerminen,
sions are likely to have caused event Y in a necessary 2006: A linear, non-Gaussian acyclic model for
causation sense,” a claim that is more direct, assertive, causal discovery. J. Mach. Learn. Res., 7, 2003–2030.
and clear from a causal attribution standpoint than Spirtes P., C. Glymour, and R. Scheines, 2000: Causa-
the previous one. tion, Prediction, and Search. 2nd ed. MIT Press, 543 pp.
Finally, at a more practical level, attribution stud- Stone, D. A., and M. R. Allen, 2005: The end-to-end
ies applying causal theory require the availability of attribution problem: From emissions to impacts.

AMERICAN METEOROLOGICAL SOCIETY JANUARY 2016 | 109

 Climatic Change, 71, 303–318, doi:10.1007/s10584 Society: Research, Modelling and Prediction Priorities,
-005-6778-2. G. R. Asrar and J. W. Hurrell, Eds., Springer, 307–337,
Stott, P. A., D. A. Stone, and M. R. Allen, 2004: Human doi:10.1007/978-94-007-6692-1_12.
contribution to the European heatwave of 2003. Trenberth, K. E., 2012: Framing the way to relate cli-
Nature, 432, 610 – 614, doi:10.1038/nature03089. mate extremes to climate change. Climatic Change,
—, and Coauthors, 2013: Attribution of weather and 115, 283–290, doi:10.1007/s10584-012-0441-5.
climate-related events. Climate Science for Serving

NEW FROM AMS BOOKS!

“ A thoughtful analysis of actions that

we need to take to reduce the impacts
of extreme weather…a must-read
for everyone with an interest in the
weather and climate.”
— F R AN K LIN W. NU T TE R ,
President, Reinsurance Association of America

Living on the Real World:

How Thinking and Acting Like
Meteorologists Will Help Save the Planet
WILLIAM H. HOOKE
Meteorologists sift through a deluge of information to make predictions every day.
Instead of being overwhelmed by the data and possibilities, they focus on small
bits of information while using frequent collaboration to make decisions.
With climate change a reality, William H. Hooke suggests we look to the
way meteorologists operate as a model for how we can solve the
twenty-first century’s most urgent environmental problems.
© 2014, PAPERBACK 978-1-935704-56-0
LIST $30 MEMBER $22
www.ametsoc.org/amsbookstore

110 | JANUARY 2016