Generated using the official AMS LATEX template v6.

On the attribution of weather events to climate change using empirically-fit

arXiv:2308.07560v3 [physics.ao-ph] 15 Dec 2024

extreme value distributions

Peter Sherman,a Peter Huybers,a,b and Eli Tziperman,a,b

a
School of Engineering and Applied Sciences, Harvard University, Cambridge, MA, USA
b
Department of Earth and Planetary Sciences, Harvard University, Cambridge, MA, USA

Corresponding author: Peter Sherman, [email protected]

1
ABSTRACT: Changes in extreme weather events are a potentially important aspect of anthro-
pogenic climate change (ACC), yet, are difficult to attribute to ACC because the record length
is often similar to, or shorter than, extreme-event return periods. This study is motivated by the
“World Weather Attribution Project” (WWA) and, specifically, their approach of fitting extreme
value distribution functions to local observations. The approach calculates the dependence of
distribution parameters on global mean surface temperature (GMST) and uses this dependence to
attribute extreme events to ACC. Applying this method to preindustrial climate simulations with no
time-varying greenhouse gas forcing, we still find a strong dependence of distribution parameters
on GMST. This dependence results from internal climate variability (e.g., ENSO) affecting both
extreme events and GMST. Therefore, dependence on GMST does not necessarily imply an effect
of ACC on extremes. We further consider for these three cases whether an extreme value, normal,
or log-normal distribution better represents the data; if a GMST-dependence of distribution pa-
rameters is justified using a likelihood ratio test; and if a meaningful attribution is possible given
uncertainties in GMST dependence. In one case we find that an attribution of Australia’s 2020–
2021 Bushfires to ACC is difficult due to the effects of internal variability. In another involving the
2019–2021 drought in Madagascar we find that the small number of available data points precludes
a meaningful attribution analysis. Overall, we find that the effects of internal climate variability on
GMST and the uncertain relationship between GMST and regional extremes may lead to inaccurate
attribution conclusions using the WWA approach.

2
1. Introduction

The Intergovernmental Panel on Climate Change (Eyring et al. 2021; IPCC 2023, 2021) found
that global mean surface temperature (GMST) between 2011–2020 was 1.09 ◦ C [0.95 ◦ C–1.20
◦ C] above preindustrial (1850–1900) levels. Although climate change is clearly manifested in
GMST warming, and there are reasons to expect extreme events to change with mean climate, it
is a complex task to attribute individual extreme weather events to anthropogenic climate change
(ACC). Extreme events—i.e., heatwaves, droughts, and heavy precipitation—are rare by definition,
so the observational record may be too short to detect anthropogenic signals. This issue is especially
pervasive in lower-income countries where data are sparse, and the historical record is short (Otto
et al. 2020). The situation is further problematic for extreme precipitation event attribution, which
is highly sensitive to the specific timescales and percentiles used to define extreme precipitation
events (Pendergrass 2018).
The field of attribution of extreme events has progressed significantly over the past decade and
encompasses several approaches (see example reviews and cited references within, Stott et al.
2010; Hegerl and Zwiers 2011; National Academies of Sciences 2016; Clarke et al. 2022; Philip
et al. 2020). The approaches that have been applied include the fraction of attributable risk Allen
(2003); Stott et al. (2004), using simulations from climate models, using constraints derived from
optimal fingerprinting techniques (Christidis et al. 2015), using climate models to validate observed
probability density functions (PDFs, Perkins et al. 2007; Lobeto et al. 2021), and the “Analogue-
Based Approach” (Lorenz 1969; Yiou et al. 2007; Vautard and Yiou 2009; Faranda et al. 2022).
Another methodology fits statistical distributions to the historical record (again, either observed or
from climate models) in what Stott et al. (2016) refer to as “Empirical-Based Approaches.” This
approach assesses how the return period of particular events has changed over the historical record,
with the purpose of detecting potential trends in the frequency of extreme events. These statistical
approaches have been used to study trends in extreme events such as flooding in Thailand in 2011
(Van Oldenborgh et al. 2012). Approaches using climate models allow for a better estimate of
the magnitude and direction of human-induced contributions to an extreme event than is possible
using observations alone.
The climate model approaches detailed in the above paragraph can also be performed with
Single Model Initial-condition Large Ensembles (SMILEs), composed of a climate model run

3
several times with slightly different initial conditions. Averaging over multiple climate model
ensemble members can reduce the signal from modes of internal variability, which are typically
uncorrelated between ensemble members, facilitating the identification of forced climate change
signals. This averaging generally improves the detection of trends in the magnitude or frequency of
an extreme event, which are difficult to quantify in a single simulation or in the observational record
where the signal from internal variability may be dominant (Yoshida et al. 2017; Yamaguchi et al.
2020). These simulations have been used in the context of tropical cyclone slowdown (Yamaguchi
et al. 2020), as well as drought events in South and East Africa (Lott et al. 2013; Pascale et al.
2020). The attribution literature is extensive, and for further detailed discussion, we refer the reader
to the above reviews and to recent studies (Stott et al. 2016; Knutson et al. 2017; Seneviratne et al.
2021).
One particular approach that is the focus of this paper and that has been used in recent years by
the WWA as a rapid response to extreme events attempts to fit the observed data to an extreme
value PDF (The World Weather Attribution project 2024). The fit allows for a dependence of the
distribution parameters on explanatory variables, typically the GMST (e.g., Ciavarella et al. 2021;
van Oldenborgh et al. 2021; Harrington et al. 2022; Philip et al. 2020). A GMST dependence is
then interpreted as a response of extreme events to ACC. This dependence is then used to attribute
extreme events by quantifying the effects of anthropogenic forcing on the probability of their
occurrence. Technically, this is achieved by writing the location parameter of the extreme event
distribution, µ, as a function of T , the (low-pass filtered, via a 4-year running average) GMST “as
a measure of anthropogenic climate change” (Harrington et al. 2022), such that µ = µ0 αT . The
GMST-dependence, α, is estimated via a fit to observations. The fitted PDF is then used to evaluate
the probability of extreme events at 1900, P µ0 αT 1900 relative to those at 2020, P µ0 αT 2020.
This allows for estimating how the probability of an extreme event changed between these times.
A change in the probability of an extreme event is inferred by the WWA approach if α is non-zero
and in proportion to the change in GMST.
Our focus is on the empirically-based approaches used by the WWA, but it should be noted that
WWA also requires evidence from climate model simulations before making any attribution state-
ments. WWA often uses multiple climate models and is careful to exclude models based on their
performance and return periods relative to observations (e.g., Pinto et al. 2023). In this study we use

4
a single model (CESM) to demonstrate some of our points but focus on assessment of observational
methodologies associated with fitting of extreme value distributions to observations. Our goal is
limited to evaluating the observational approach described by WWA and suggesting additional
tests to examine its robustness, as opposed to developing alternative attribution approaches.
Motivated by these pioneering efforts based on the fit to extreme event PDFs, we reexamine
three test cases analyzed by WWA: the Siberian heat wave analysis of Ciavarella et al. (2021,
hereafter CC2021), the analysis of high temperatures associated with the Australian Bushfire
by van Oldenborgh et al. (2021, hereafter OK2021), and the Madagascar drought analysis of
Harrington et al. (2022, hereafter HW2022). These cases were chosen as they represent a range of
applications, including the attribution of heatwaves at high and low latitudes (Siberia vs. Australia)
and of an extreme drought. Each of these studies leverages either a Generalized Extreme Value
(GEV) or Generalized Pareto Distribution (GPD) fit to assess potential contributions to extreme
events from ACC. WWA studies use both climate model simulations and observational analysis,
and we focus on the observational component. In each of the three cases we use the CESM Large
Ensemble (Kay et al. 2015) to put the observed instrumental record into perspective, and we explore
potential caveats concerning the methods examined and suggest how these limitations might be
addressed.
The issue exposed by our analysis is simple: internal variability modes such as ENSO affect
both local temperature extremes and GMST. In some cases, the effect on local extremes (say the
effects of ENSO on heatwaves in Australia) may be large, but the effect on GMST is small. This
might lead to a false conclusion that a small GMST signal leads to a comparatively large effect
on local extremes. To demonstrate this point, we analyze a preindustrial climate model run with
no anthropogenic forcing. We find, for example, that the fit of Australian heat extremes leads to
values of α that are substantially greater than zero in the absence of ACC because of the effects
of internal climate variability modes such as ENSO on both GMST and on local extremes. Our
findings show that α can be strongly affected by internal variability, biasing the attribution results
and complicating the WWA approach.
Our study addresses four questions: (1) Is a GMST dependence of the PDF parameters necessarily
an indication of the signal of ACC? (2) Is a GEV or GPD necessary to fit the available observations,
or are more standard normal or log-normal distributions sufficient? (3) Is the addition of GMST-

5
dependent distribution parameters, which form the basis for the attribution analysis, justified
statistically, and is the uncertainty range of these parameters, specifically α, sufficiently small to
allow for attribution? (4) What do SMILEs tell us about the amount of data needed to determine
the GMST dependence with confidence? WWA uses multiple attribution tools, including analysis
of simulations from multiple models and observations (Philip et al. 2020). We focus only on the
latter; specifically, on the fitting of extreme value distributions to observations. Our goals are
limited to evaluating this observational approach and suggesting additional tests to examine its
robustness, and do not include developing alternative attribution approaches.
The rest of the paper is structured with Section 2 describing the statistical models, observation
datasets, and methodologies. Section 3a shows that a GMST-dependence of distribution parameters
does not necessarily represent the effects of ACC using an analysis of a CESM preindustrial
simulation (i.e., without time-varying anthropogenic and natural forcings). We then further examine
the three attribution cases in the remainder of section 3. Section 4 provides a summary and
discussion of the results.

2. Methods and Data

In this section, we briefly describe the extreme value distributions used in the following sections (a
and b) and discuss uncertainty estimates and empirical cumulative distribution functions (CDFs)
that we will use later (c). We describe the deviance statistic test which we use to examine the
justification for adding GMST-dependent parameters (d), and the mean residual life plot used to
test the justification for using a GPD and its threshold value (e). The data used are described in
section (g).

a. The Generalized Extreme Value (GEV) distribution

Extreme value distributions are a family of PDFs that can be used to represent the statistical
behavior of block maxima or minima of a record. An example is an annual time series of the
maximum daily temperature at each year, the “block” being one year in this case. It can be shown
that the distribution of such maxima follows one of three classes of PDFs: Gumbel, Fréchet, or
Weibull (Coles et al. 2001), depending on the shape of the tail of the distribution of the events for
which block maxima are calculated. Rather than explicitly specifying one of these three classes in

6
fitting extreme data, one can combine these distributions into one functional form, represented by
the Generalized Extreme Value (GEV) distribution. Large quantities of data are typically needed
to accurately fit a GEV and calculate its parameters (e.g., Philip et al. 2020; Trevino et al. 2020).
The PDF of a variable x whose statistics are governed by the GEV is given by

1
GEV x, µ, σ, ξ = tx, µ, σ, ξ ξ1 e−tx,µ,σ,ξ , (1)
σ

where
−1ξ
x−µ

tx, µ, σ, ξ = 1 ξ , (2)
σ
under the assumption that ξ ≠ 0. If ξ = 0, tx, µ, σ, ξ is defined as,

x−µ
tx, µ, σ, ξ = e− σ . (3)

In some of the attribution studies we follow for this paper (specifically, CC2021 and OK2021),
the location parameter µ is assumed to vary linearly with the 4-year smoothed GMST T , as,

µ = µ0 αT . (4)

Estimation of the parameter α is the key to the attribution of extreme events. However, this
determination may be limited due to the short observational record (Zwiers et al. 2011; De Paola
et al. 2018). To estimate the mean µ0 , scale parameter σ, the GMST dependence α, and the
shape parameter ξ, we follow Coles et al. (2001) and WWA (Philip et al. 2020) and maximize the
log-likelihood of the data points xt using a GEV distribution,

m

lµ0 , σ, ξ, α = log GEV xt , µ0 , σ, ξ, α . (5)
t=1

We apply this approach, following CC2021 and OK2021, to observations and to data from the
CESM Large Ensemble (Kay et al. 2015).

7
b. Generalized Pareto Distribution (GPD)

In some applications, one is interested only in points in a time series that exceed a specified
threshold, whose statistics are represented by the Generalized Pareto Distribution (Coles et al.
2001). In order to study the extreme drought events in Madagascar using a record of precipitation,
HW2022 fit a Generalized Pareto Distribution (GPD) to the bottom 20th percentile of precipitation,
i.e., years associated with drought. This low tail was converted to a high tail by multiplying the
precipitation data by −1. Although this converts the lowest precipitation values to maxima, the
transformed data have a hard maximum at zero. We note that the validity of using a GPD, in
this case, is questionable because no such hard maximum is allowed by—or represented in the
functional form of—the GPD.
The PDF of the GPD is given by

− 1 1
1 x−u

ξ
GP Dx, u, σ, ξ = 1ξ . (6)
σ σ

HW2022 assumed that the threshold parameter, which they denote u (denoted µ by Philip et al.
2020) and scale parameter σ both depend exponentially on the 4-year running-averaged GMST, T
such that,

u = u0 expαT u0 ,
σ = σ0 expαT u0 . (7)

The exponential fit was assumed by HW2022 in order to ensure that the scaling factor expαT u0
applied to both u0 and σ0 is strictly positive and maintain a fixed ratio. We obtain the optimized
GPD parameters (u0 , σ, ξ, α) by again maximizing the log-likelihood (Eq. 4.10, Coles et al. 2001),

m
lu, σ, ξ = logGP Dxt , u, σ, ξ, (8)
t=1

using the trust-region constrained algorithm in the Scipy Python package (‘trust-constr’ method).
The GPD requires the bound σ > 0 and consistency relations of the form x ≥ u (for ξ ≥ 0),
and u ≤ x and x ≤ u − σξ (for ξ < 0). As the optimization iteratively searches for the optimal
parameter values, these constraints are occasionally violated, leading to complex numbers or NaNs

8
(not a number) in the likelihood. The GPD optimization is notably more prone to such failures
and sensitive to the initial guesses of α, u0 , ξ, σ than the GEV optimization. In order to avoid such
search failures, we replace complex and NaN terms in the log-likelihood sum with large penalty
terms whenever they occur due to constraints being violated. In addition, if the optimization fails,
we re-run the optimization again with new randomly selected initial guesses until convergence to a
solution that maximizes log-likelihood and satisfies the constraints is found. We note that related
solutions have been used to deal with GEV/GPD convergence issues by Robin and Ribes (2020a).

c. Uncertainty estimates and empirical cumulative distribution functions

We follow WWA studies and use non-parametric bootstrapping to estimate uncertainty ranges
for the optimized parameters. In particular, we use the uncertainty in α as one way to evaluate
the validity of the attribution results. We also compare the appropriately fitted distributions (GEV
and normal distributions for extreme temperatures or GPD and log-normal for droughts) to the
empirical cumulative distribution functions (CDFs) of the observations to assess the quality of the
fit. The empirical CDF is calculated by sorting extreme values (either temperature or precipitation)
from smallest to largest, calculating the cumulative sum, and normalizing to a maximum CDF
value of one. We calculate error bars in the empirical CDFs based on bootstrap resampling with
replacement of the observations. The empirical CDF is calculated for each sampling, and we then
calculate the 90th and 95th percentiles from the resulting distribution at each value of the variable
whose CDF is estimated. In order to estimate uncertainty in the parameters estimated from our
GEV or GPD maximum likelihood fit, we similarly use bootstrap resampling. We sample with
replacement the GMST and corresponding extreme temperature record 5,000 times and optimize
the parameters for each resampling. We then estimate the 90th and 95th percentiles from 5,000
bootstrap samples. For the Madagascar precipitation case, we perform 5,000 bootstrap resamples
when analyzing the observations, but 1,000 bootstrap resamples when analyzing the CESM Large
Ensemble due to the slow convergence and expensive computations using the GPD log-likelihood
maximization.

9
d. Deviance Statistics (Likelihood Ratio Test)

In general, when fitting a model to data, if the number of fit parameters is increased, one
expects a better fit because of the greater flexibility of the fitted model. Specifically, in the case
considered here, adding a parameter α to include the GMST dependence of the location/threshold
parameter will increase the maximum log-likelihood of the data. The salient question is whether
the improvement in the fit exceeds that expected simply on account of the larger number of model
parameters. To evaluate the improvement of the fit, Coles et al. (2001, theorem 2.7) recommend the
use of a deviance statistic, also known as a likelihood ratio test. This test uses a measure, D, equal
to twice the difference of the sum of log-likelihoods of the model with k additional parameters
minus that of the simpler model. Because more parameters lead to a better fit and thus higher log-
likelihood, D is non-negative. D is then compared to a χ2k distribution to assess if the k additional
parameters significantly improve the likelihood. For example, with k = 1, the 95th percentile of χ21
is equal to 3.8. If D is larger than this threshold, we conclude that adding an additional parameter
significantly improves the fit. It can then be inferred that the added parameter—in our specific
case, the GMST dependence, α—meaningfully explains features of the data. On the other hand,
if the deviance statistic is insignificant, one concludes that the addition of another parameter to
the statistical fit is not justified by the data and, therefore, that this parameter cannot be used to
draw conclusions about the data. Specifically, if one finds that the addition of α is not justified, the
implication is that the estimated value should not be used to calculate the effects of climate change
on the extreme events under examination. We also estimate the distribution and uncertainty of D
through bootstrap resampling of the data 5,000 times.

e. Mean residual life plot

In the process of fitting a GPD, a minimum threshold is chosen to select the data that represent
extreme values that should be fit. The GPD fit can be sensitive to the selection of this threshold,
where too low of a threshold could violate the asymptotic nature of the GPD, and too high of a
threshold would provide only a few relevant data points and lead to an unstable fit. In order to
determine an appropriate threshold, one leverages what is known as the mean residual life (MRL)
plot. The MRL uses the GPD-based expectation value of the data (Eq. 4.8, Coles et al. 2001). For
a given threshold, u0 , the expectation value (mean) of the data X that exceed the threshold can be

10
written for ξ < 1 as,
σu0
EX − u0 |X > u0 = .
1−ξ
It can be shown that for all thresholds u > u0 , this expectation value is (Eq. 4.9, Coles et al. 2001),

σu0 uξ
EX − u|X > u = . (9)
1−ξ

That is, for u > u0 , EX − u|X > u is a linear function of u. This expected linearity of Equation 9
with u—for values of u for which the GPD fit is appropriate—can be used to test the validity of the
chosen threshold used in the GPD analysis via a plot which we use in our results section known as
the MRL plot.

f. Hypothesis testing

Although WWA does not provide a detailed discussion of the hypothesis that they assume, we
interpret that they adopt a null hypothesis of α equal to zero implying no ACC influence on extremes
and an alternative hypothesis that, if α is not equal to zero, ACC does influence extremes. This
inference is based on the quote, “If the one- or two-sided 95% confidence interval [for the change
in probability, which has the same effect as α] excludes zero, the change is statistically significant”
(see Section 4.3.1 of Philip et al. 2020). We assess whether rejection of a null hypothesis of α
equal to zero permits accepting an alternate of ACC influencing extreme events. In other words,
we assess if a nonzero α necessarily means a role for ACC. We note that another null hypothesis
for purposes of heat could be that the location of PDF of heat extremes is shifted by an amount
equaling the GMST warming, as would be represented by α = 1. For purposes of consistency with
the null implied by WWA, however, we only consider a null of α equal to zero.
WWA uses two variants of their analysis approach, one in which the extreme data in question is
incorporated into the estimate of the PDFs, which we follow here, and one in which it is excluded
(Philip et al. 2020). The act of choosing to study a class of events immediately following the
occurrence of a particularly extreme example is known to induce a bias related to whether it is
included or excluded (Miralles and Davison 2023, see). We choose to include the extreme value
in question, which tends to make the parametric dependence on GMST larger and tends to make
it easier to reject the null of no change in favor of the alternate hypothesis. Within the context of

11
the WWA hypothesis test, this choice to include the extreme in question makes it more likely to
conclude that AAC significantly influences extreme events.

g. Data

We briefly review the observations and our processing used in each of the three case studies re-
considered here. In addition, we describe the model Large Ensemble and preindustrial simulations
that we analyze.

(i) Siberian Heatwave. Following CC2021, we rely on two sets of observations: (1) daily maxi-
mum temperature extremes for each year from station data in Verkhoyansk as discussed in CC2021;
and (2) 4-year smoothed GMST anomalies from GISTEMP (Lenssen et al. 2019). CC2021 also
performed a GEV analysis on temperature anomalies over a region in Siberia, which we do not fo-
cus on here in order to study specifically the extreme values detected at the Verkhoyansk station. At
the time of the CC2021 study, the daily maximum temperature extreme in 2020 at the Verkhoyansk
station was recorded as 38◦ C. This temperature extreme currently appears as a missing data point
on the National Climatic Data Center website (https://www.ncei.noaa.gov/cdo-web/search, station
RSM00024266, VERHOJANSK, data downloaded Aug 2023) but we still use this unconfirmed
value in our analysis.
We also study GMST anomalies and extreme temperatures over the Siberian region defined in
CC2021 within the CESM Large Ensemble (Kay et al. 2015) over the period 1926–2019. The
CESM Large Ensemble contains 40 simulations of the climate over the period 1920–2100 under
historical and RCP 8.5 external forcing, of which we use the 35 that were run on the NCAR
supercomputer and not those runs done on the University of Toronto supercomputer that contain a
global mean temperature bias National Center for Atmospheric Research (2024). These simulations
differ only in small perturbations to the initial conditions, which lead to different temporal sequences
of modes of internal variability. We use model output only up to the year 2020, and, as a result,
there is little sensitivity to the emissions scenario used (DAMIP, Gillett et al. 2016, uses RCP 4.5)
because emissions scenarios only significantly diverge later.
In addition, we study GMST anomalies and extreme temperatures within a 1,800-year CESM
preindustrial simulation, yielding 18 segments of 94 years, chosen for consistency with the length
of the observed record. This preindustrial simulation uses fixed greenhouse gas concentrations,

12
volcanic activity, anthropogenic aerosols, land use, and solar forcing all set to their respective
levels in the year 1850. Most importantly in the context of this study is the lack of time-varying
anthropogenic forcing in this simulation.

(ii) Australian Bushfire Heatwave. Following OK2021, we again use two sets of observations for
the years 1920–2019: (1) annual (July–June) maxima of 7-day moving average daily-maximum
surface temperature data from the Australian Water Availability Project (AWAP) over the South-
eastern region in Australia defined in OK2021; and (2) 4-year smoothed GMST anomalies from
GISTEMP. We apply a similar analysis to the CESM Large Ensemble and to the long CESM
preindustrial run as for the Siberian Heatwave case except, in this case, using 18 segments of 98
years for consistency with the observations.

(iii) Madagascar Drought. Following HW2022, we use two datasets. The first data are 2-year
means of precipitation from ERA5 (Hersbach et al. 2020) over the region of Madagascar that was
defined in HW2022. HW2022 explained that they used a “24-month running mean rainfall data
from July to June.” Our approach is to calculate total annual precipitation and then calculate a
running 2-year average of this annual data resulting in a new smoothed annual time series. For 1951–
2020 the bottom 20% precipitation 2-year periods are selected for the analysis, corresponding to
the following 14 two-year periods: 1956–1957, 1957–1958, 1958–1959, 1959–1960, 1962–1963,
1990–1991, 1991–1992, 1992–1993, 2008–2009, 2009–2010, 2015–2016, 2016–2017, 2017–
2018, 2019–2020. The second data are 4-year smoothed GMST anomalies from GISTEMP. We
note that the analysis following HW2022 relies on a fit to only 14 data points. We discuss the issues
that are involved in fitting extreme distribution functions to such a small number of data points
below. We also study the equivalent variables within the CESM Large Ensemble over the period
1950–2019, consistent with the period of observations. We again apply a similar analysis to the
CESM Large Ensemble and to the long CESM preindustrial run, in this case using 18 segments of
70 years.
We note that in each of these three case studies, WWA leverages many datasets, including ERA5
reanalysis, station observations, and the multimodel CMIP5 ensemble. We use a subset of the data
that WWA study but also the CESM Large Ensemble data in a consistent manner across the three
example cases that we evaluate.

13
3. Results

We first describe our analyses of how extreme events depend on GMST in the context of
the long CESM pre-industrial simulation in subsection 3a. The simulation contains no time-
varying anthropogenic forcing, making it useful for evaluating a null-hypotheses associated with
no anthropogenic influence. We then consider each of the three case studies in subsections 3b,
3c, and 3d. For each case we assess how GEV or GPD distributions fit the observations relative
to normal or log-normal distributions. We then evaluate evidence for the inclusion of a GMST
dependence within distributions using the deviance statistic and estimate the uncertainty level of the
parameter α used for the attribution. Finally, we examine the amount of data needed to adequately
constrain GMST dependence in a GEV or GPD fit.

a. Extremes analysis of a preindustrial simulation

WWA examines the dependence of the distribution parameters of extreme events on the 4-
year running mean GMST. Commonly a linear dependence of a location parameter on GMST, as
represented by the parameter α, is used to estimate the PDF of extreme events in, for example, 1900
versus 2020. Changes in the PDF between these epochs are interpreted as owing to the effects of
anthropogenic climate change, implying that α indicates the extent to which anthropogenic climate
change affects extremes. We show here that it is possible, however, for internal climate variability,
such as ENSO, to produce covariance between GMST and extremes that is expressed as a non-zero
α even without anthropogenic influence. This covariance is demonstrated most clearly in this
section by using a pre-industrial climate model run, but we also elaborate upon this basic result
using the CESM Large Ensemble. For attribution purposes this results suggests that the value of
α calculated in the presence of a GHG increase can be affected by internal variability, biasing the
attribution results. It should be noted that in a realistic scenario α is affected by both anthropogenic
climate change and internal variability. The scenario of internal variability alone analyzed here is
artificial, of course, and does not capture that, in the current climate, GMST is a reasonable proxy
for anthropogenic climate change. Analyzing simulations that only include internal variability
nevertheless provides insight into the effects of internal variability in more complete scenarios.
To examine this potential role of internal variability, we start with the case of the hot weather
associated with the bushfires in Southeastern Australia and perform a GEV analysis using a 1,800-

14
year CESM preindustrial run. This run holds anthropogenic CO2 , other anthropogenic greenhouse
emissions, volcanic (natural) emissions, land use changes, and anthropogenic aerosols at values
estimated for the year 1850. We divide the run into 18 data segments of length 94 years, equal to the
length of the observed record and analogous to having 18 simulations of historical climate without
ACC. The distribution of α values for the GEV preindustrial analysis is shown by the green bars
in Fig. 1a. We find a large and positive α calculated from essentially all model output segments
centered around a median value of 4.9 ◦ C per ◦ C, or that extreme summer temperatures increase by
4.9 ◦ C for every 1 ◦ C GMST warming. Note that, following WWA, we are using four-year-running
average GMST values for this computation.
It is difficult to imagine a physical causal mechanism that would lead to such an enormous
amplification of the GMST signal in Australian heatwaves. A more plausible interpretation is
that internal variability — for example, ENSO — could increase both extreme temperatures over
Australia and, to a smaller degree, GMST, generating covariance between extreme events and
GMST. In this scenario GMST is not an indicator of forcing of local extremes; rather, both
local extremes and GMST are influenced by a third factor. It is known that several modes of
internal variability affect extreme temperatures in Australia, including ENSO, the Indian Ocean
Dipole, and the Southern Annular Mode (Hendon et al. 2007; Ummenhofer et al. 2009; King et al.
2020). Min et al. (2013) find that ENSO and the Indian Ocean Dipole can lead to warming of
extreme temperatures in Southeastern Australia by around 1◦ C, indicating that internally generated
variability can have an effect that is similar in magnitude to ACC.
WWA recognizes that α may be influenced by internal variability and, as noted, use a 4-year
running mean GMST to filter out the effects of short-term internal variability. This approach does
not fully remove the ENSO signal, however, because much of the variability is at longer periods.
Spectral analysis of the annual-average Bivariate ENSO time series (Smith and Sardeshmukh 2000)
between 1870–2022, for example, indicates that 63% of the variance in the time series resides at
frequencies below 1/(4 years). There is also the potential for nonlinear interactions between ENSO
and GMST and the presence of slower modes of variability such as the Pacific Decadal Oscillation
(PDO).
A similar analysis of Australian heat extremes made using the CESM Large Ensemble, which
includes anthropogenic forcing, gives a distribution of α that is closer to zero (blue bars in Fig. 1)

15
than in the unforced simulations (green bars). This indicates that both ACC and internal variability
influence regional extremes and GMST, such that estimates of α will reflect multiple processes. In
simulations with time-varying anthropogenic forcing, however, there is no way to disentangle the
effect of internal variability on α using the WWA approach as described in published literature.
We conclude that the magnitude of α is a biased indicator of anthropogenic influence. Including
indices of internal variability as covariates may help to disentangle their respective contributions,
though the projection of internal variability onto GMST can still confound the results.
These results highlight a potential major difficulty with analyses based on a fit of distribution
parameters to GMST in a more realistic scenario that includes both internal variability and ACC:
the calculated α, and therefore the shift in the probability of extreme events, may be due to
internal variability rather than only ACC. When analyzing the observational record in section c,
we find α ≈ 2 ◦ C per ◦ C. This suggests a two-degree warming of extreme events for a one-degree
warming of the GMST. Based on our results from the CESM preindustrial run with a time-invariant
anthropogenic forcing, however, we suspect that the estimate obtained for α from the observed
record is biased high by internal variability.
Analysis of the CESM preindustrial run for the case of Siberian heatwaves (green bars in Fig. 1b)
shows that the value for α deduced from a record similar in length to that of the observations ranges
between 1 to 4, although there is again no time-varying anthropogenic forcing in the corresponding
model run. This large and variable response implies that a non-zero α does not necessarily imply
an effect of ACC. This analysis shows that α equal to zero is an unsuitable null hypothesis for the
purposes of evaluating the presence of an anthropogenic influence on extreme events.
For the case of Madagascar Drought (Fig. 1c) we find a large spread in α for both the preindustrial
run as well as the CESM Large Ensemble, with both centered near zero. The fact that the α
distribution from the preindustrial run is centered around zero suggests that a null hypothesis of
α equal to zero may be appropriate, but the fact that α from the Large Ensemble is also centered
around zero indicates that the there is little potential for this test to discern an ACC influence on
extremes. That is, low extremes of precipitation show no systematic increase or decrease related to
GMST in either the forced or unforced runs such that this test has low statistical power. We further
consider this case below using observations.

16
 The analysis in this paper shows that internal variability can lead to a large α. We conclude
that in a more typical scenario analyzed using the WWA methodology and applied to observations
under varying anthropogenic GHG and aerosols, the value of α may still be affected by internal
variability and, therefore, may bias the estimates of the effect of ACC on the return time of extreme
events. One could justify the WWA use of the 4-year averaged GMST as being a proxy for known
greenhouse gas concentrations. If CO2 were indeed used, the analysis of the preindustrial run
would, of course, show no dependence, eliminating this problem, though then introducing issues
associated with lags between GHG concentrations and warming (Proistosescu and Huybers 2017).
Because WWA studies use GMST, we follow that methodology here.
The result that internal variability may strongly affect α when the 4-year average GMST is used to
link extreme events to ACC suggests that it might be useful to reevaluate some of the WWA results
based on the fit to extreme distributions in at least the three studies that we review. It is possible,
of course, that a strong anthropogenic signal in a few decades could significantly affect extreme
events and overwhelm the effects of internal variability in cases such as the Siberian Heatwave.
The attribution of extreme temperatures may be more difficult in cases like the Australian example,
where the effects of internal variability are large, as indicated by the large α detected in the
CESM preindustrial simulation. In such cases, it would be important to correctly identify all
relevant covariates — such as the possible effect of ENSO in the Australian case — and note that
misleading results for α may be obtained if one relies only on GMST.

b. Siberian Heatwave

We now explore other facets of the attribution of the Siberian Heatwave in the summer of 2020
to anthropogenic change, notwithstanding the foregoing finding that the interpretation of α is
complicated by the presence of internal variability. We focus specifically on the CC2021 analysis
of the June daily maximum temperatures because it allows us to address the question of whether
the data justify the use of non-stationary (GMST-dependent) extreme value distributions.
While it is understood based on theoretical considerations that a GEV is the appropriate rep-
resentation of block-maxima of data (Coles et al. 2001), it is not obvious for a given problem
what minimum block size justifies using a GEV, and specifically if a year of daily maximum data
suffices. This is especially an issue given that the temperature is auto-correlated and, therefore,

17
 (a) Australia
50 med( CTRL) = 4.9 CESM-CTRL
CESM-LE
med( LE) = 0.44

0
2 0 2 4 6 8 10
(b) Siberia
50 med( CTRL) = 1.1
med( LE) = 2.5
PDF

0
0 1 2 3 4 5
(c) Madagascar
10 med( CTRL) = 0.023
med( LE) = 0.095

0
0.4 0.2 0.0 0.2 0.4

Fig. 1. Effect of internal climate variability on α distribution. PDFs of α for (a) The Australian 7-day
moving mean of the maximum temperature related to the analysis of the 2019–2020 bushfires. (b) The Siberian
maximum daily temperatures that were used in the analysis of the 2020 heatwave. (c) The Madagascar 2-year
precipitation used in the analysis of the drought of 2019–2021. The analyses are based on the 1,800-year CESM
preindustrial run with 94-, 98-, and 70-year segments, respectively (green), and the 35 Large Ensemble members
(blue). The lengths of these segments correspond to the extent of the observed records used in CC2021, OK2021,
and HW2022. The median values for the CESM preindustrial and Large Ensemble for each of the cases are
indicated in the boxes. The median values are Australia: αCT RL = 4.9, αLE = 0.4; Siberia: αCT RL = 1.1,
αLE = 2.5; Madagascar: αCT RL = −0.02, αLE = 0.09.

18
more data are needed per averaging block to justify the use of a GEV. It is, therefore, worthwhile
to check how well other standard distributions perform in fitting the data.
A cumulative distribution estimate of the highest June daily-maximum temperature observed
each year between 1926–2020 at the Verkhoyansk station is shown in Fig. 2a along with a 95%
confidence interval of the CDF obtained by bootstrapping. In the first step of our analysis, we
consider cumulative distribution function fits to the observed distribution, assuming the data follows
either a GEV or normal distribution. The fitted distributions are specified here to be stationary (by
setting α = 0 and µ = µ0 in Eqns. 1–4). Fits are obtained by maximizing the likelihood of the data,
and the results are shown in Fig. 2a. Both distributions generally fit the observations well, with
the normal fit completely within the 95% range of the observed distribution, and with the GEV
fit being good except at the highest temperature, where it slightly underestimates the probability
of obtaining the highest temperatures. There is only one data point above 35 °C, so the empirical
CDF is poorly constrained there in any case. Interestingly, the normal distribution and GEV CDFs
look similar and lie within the confidence intervals defined for the empirical CDF, despite the
normal distribution having fewer parameters than the GEV. It is worth noting that a GEV and a
normal distribution may assign very different probabilities to extreme events, even when the two
distributions provide an adequate fit to the observed CDF. These modeling decisions are therefore
of critical importance for attribution studies, which usually focus as much or more on estimation
of the exceedance probability and return period of an event — and, consequently, on the changing
likelihood and fraction of attributable risk of an event — as they do on the change in intensity. It
is therefore useful to consider alternate distributional forms and, in many ways, the normal is the
simplest and often a useful default.
The quality of fits indicates that an extreme value distribution is not empirically a better choice
than a normal distribution for fitting the data. Similarly, a non-stationary distribution with α ≠ 0
may not be necessary given that the stationary distributions seem to provide a satisfactory fit to
the empirical cumulative distributions, and we further examine this next. CDFs are commonly
used to differentiate between distribution functions, and the results of a Kolmogorov-Smirnov
test (Supplementary Fig. SI-1 for normal, GEV and empirical distributions) also suggests that the
normal distribution is consistent with the observed data. An application of Occam’s razor principle
suggests that the simpler normal distribution should be used rather than a GEV if there is not a

19
significant improvement in the distributional fit with the GEV. This is also discussed by Philip et al.
(2020) who mention explicitly that the extreme value distributions may not outperform a simpler
(normal) distribution in some cases. Otto et al. (2018) opted to use a normal distribution in lieu of
a GPD for attribution analysis of a cold spell in Peru because the normal “uses all data, rather than
the GPD, which samples only the cold tail”. CC2021 found that a normal distribution fits the data
well but proceeded to use a GEV for the bulk of the analysis. These results question the need for
extreme value distribution functions when it comes to attribution analysis. One might want to use
proper selection criteria (the Akaike Information Criterion, AIC, or deviance statistic or similar)
to rigorously decide which of the models performs best.
We now proceed to examine the role of GMST in the fit to a GEV distribution (eq. 4) in this
particular application, following CC2021. A linear regression of GMST anomalies against daily
max temperatures gives a slope of 1.1 ◦ C per ◦ C (Fig. 3a), where this value indicates that the
annual maximum daily temperatures at Verkhoyansk increase slightly more quickly than GMST.
The squared Pearson correlation, however, is only r2 = 0.01, with a p-value of 0.2, indicating that
GMST is only a weakly significant indicator of trends in Siberian temperature maxima.
Fig. 3b shows the maximum likelihood fit of a nonstationary GEV, including α, for multiple
bootstrapping samples as specified in the Methods section. The values of α estimated from the
observations is 1.8 ◦ C per ◦ C, larger than the slope of 1.1 obtained from the linear regression. This
discrepancy between α and the slope suggests that the conclusions from an attribution analysis can
be highly sensitive to the choice of the fitted distribution, again underscoring the importance of
model selection. The [2.5%, 97.5%] confidence interval is [−1.2, 4.8], while the 5% value is −0.7
(Fig. 3b; red plots). Both indicate that α is not significantly different from zero. It is noteworthy that
the red distribution in Fig. 3b is bimodal, presumably because some of the bootstrapped samples
included the 2020 event and some did not. This suggests that the fitted distribution is very sensitive
to the inclusion or exclusion of the event (again consistent with Miralles and Davison (2023));
that is, α would be significantly non-zero if the event is included, but is not significant when it
is omitted. While CC2021 did not provide an uncertainty range for α, they note that the change
in intensity of Siberian heatwaves at a fixed probability is estimated from their inferred PDF as
1.04 ◦ C (with a 95% confidence interval from 0.35 to 3.4). α is directly related to the change in
intensity: the change in intensity is k ∗ α, where k is the observed GMST change between the two

20
 (a) (b)
1.0
0.8

Deviance Statistic
0.6 101
CDF

0.4
0.2 Empirical CDF
GEV fit without GMST
Normal fit
95th %ile 100
0.0 90th %ile

25 30 35 ObsLE-5 LE-15 LE-25 LE-35

Siberian Daily Max T ( C) Ensemble Member #

Fig. 2. Siberian Heatwave: (a) Examining the fit of a GEV and normal distribution to Verkhoyansk
station data. The red curve indicates the empirical CDF of the highest daily maximum temperature for each
year for the Verkhoyansk station data. The purple and blue lines denote fits by a GEV (with a constant location
parameter) and by a normal distribution, respectively. The light and dark green shadings correspond to 90th and
95th percentile confidence intervals (section 2c). (b) Justification for using a GMST-dependent mean. Box
plots of the deviance statistic between a GEV distribution without a GMST-dependent mean relative to a GEV
distribution with a GMST-dependent mean using the Siberian heatwave data with 5,000 bootstrap resamples. The
edges of the boxes correspond to the first and third quartiles. The deviance statistic for the station observations
is shown in green as the leftmost box. The deviance statistic was also computed for each of the 35 members
from the CESM Large Ensemble (blue). The 95th percentile significance level from a χ2k=1 distribution with one
degree of freedom is shown by the horizontal red line, corresponding to the addition of one degree of freedom
by including α in the fitted model.

periods of interest, usually 1.2◦ C. Assuming a GMST change of 1.2◦ C, we infer the value of α
in CC2021 is 1.3 with a 95% confidence interval from 0.42 to 4.1. The central estimate of α is
similar to our results but entails a smaller uncertainty range, likely due to our decision to include
the 38◦ C event in our bootstrap resampling, but which was not done in CC2021. The bootstrapped
distribution for all GEV parameters for the observations is given in Supplementary Fig. SI-2.
The deviance statistics, D, (section 2d) indicate that the addition of a model parameter, α,
representing the GMST dependence, is justified in the case of the Verkhoyansk observations,

21
((Fig. 2b, leftmost bar is above the threshold indicated by the red horizontal line). Yet the deviance
statistic uncertainty range includes a large range that is below the significance line, which seems
appropriate given that the uncertainty range for α includes negative values.

(a)
37.5 slope = 1.1
p2= 0.2
Siberian Daily Max T ( C)

35.0 r = 0.018

32.5
30.0
27.5
25.0
22.5 Linear Regression
GEV Linear Fit

0.2 0.0 0.2 0.4 0.6 0.8

4-yr Running GMST anomaly ( C)
25 (b) Observations
OBS = 1.8 CESM-LE
CESM LE = 2.6
20
15
PDF

10
5
0
4 2 0 2 4

Fig. 3. Siberian Heatwave: uncertainty in α. (a) Scatter plot of the daily max temperature at the Verkhoyansk
station against GMST. The red line indicates a linear regression fit to the data, and the blue line indicates a linear
fit using the location parameter and α derived from the GEV maximum likelihood estimate. (b; red) PDF of the
GEV-fit α distributions from the station data, using 5,000 bootstrap resamplings of the observed time series. The
α value from the observations (no resampling) is shown by the vertical red line. (b; blue) Shown is a histogram
of the GEV-fit α PDFs derived from the CESM Large Ensemble, where the α values are calculated from each
ensemble member with no resampling. The average α value across the Large Ensemble is indicated with the
blue line. There are 94 annual data points for each ensemble member.

22
 To provide context for the interpretation of the Verkhoyansk observational results, we also
examine output from the CESM Large Ensemble, which includes time-varying anthropogenic
forcing, at the grid point containing the Verkhoyansk station. We find that all ensemble members
within the CESM Large Ensemble indicate a positive value of α, with a mean value of 2.60 ◦ C
per ◦ C (Fig. 3b; blue plots). Similarly, 33 of 35 ensemble members give a deviance statistic that
supports the use of a non-stationary component (that is, a GMST-dependence) in the GEV statistic
(Fig. 2b).
The fact that the LE results indicate a stronger link between Siberian extremes and GMST (that
is, a larger α) is consistent with the LE, on average, simulating an Arctic-wide summer warming
rate of 1.8 ◦ C per ◦ C warming in GMST, whereas an observational analysis only indicates a rate
of 1.3 ◦ C Arctic JJA warming per ◦ C in GMST (using GISTEMP for both; Fig. 4a,b). This greater
amplification of Arctic warming may be related to the fact that CESM1 has a relatively high climate
sensitivity of 4.1 K (Gettelman et al. 2019). The WWA approach leverages a range of climate
models so that results are not dependent entirely on one GCM. That said, one expects heatwaves to
be strongly influenced by the JJA climatology, and Fig. 4c shows that different ensemble members,
corresponding to different realizations of internal variability, show very different dependencies of
the Arctic JJA climatology on GMST. It is difficult, therefore, to estimate from the single realization
corresponding to the observed record to what degree the Arctic JJA climatology is affected by ACC
versus internal variability, adding to the uncertainty in attributing Siberian heatwaves. To address
that, the WWA protocol includes a step where the same analysis is applied to climate models, and
results are reported from both sources.
Unsurprisingly, if data from multiple ensemble members are analyzed simultaneously as a single
dataset, which mimics the availability of more data, the distribution of the GEV parameters narrows,
generally converging around the mean value of the parameters across individual ensemble member
fits (Fig. 5). The 95% confidence range for α (red asterisks in panel a) decreases from a large range
of [1,4] (◦ C per ◦ C) with one ensemble member to a smaller [2,3] range with 10 ensemble members.
Similar behavior is seen for the other GEV parameters. This implies that with a sufficiently long
record, one can reduce the uncertainty in the GEV parameters, as expected, but that might require
much more data than is expected to be available in the historical record anytime soon.

23
 The overall result for the GMST dependence for the Siberian Heatwave analysis, which is the
basis of the attribution to ACC, is largely consistent with the results found from CC2021 but with
three important caveats. First, we found in the previous subsection that internal variability can lead
to GMST dependence that biases the value of α and hence the attribution results (Fig. 1a). Second,
whereas a GEV fit a priori make sense to fit to block maxima of daily temperature extremes, a
normal distribution with α = 0 does not appear to be ruled out by the observations (Fig. 2a). This
may be a result of both the relatively short observational record and the autocorrelated nature of
the temperature extremes. Still, a non-zero α is found to meaningfully improve the statistical fit to
the CESM Large Ensemble Siberian extremes based on deviance statistics. Finally, the uncertainty
range for α, which was not explicitly evaluated in the original WWA analysis, includes negative
values (Fig. 3b).

c. Australian Bushfire Heatwave

We now consider the analysis of possible connections between the hot weather that may have
contributed to the Australian Bushfires of the summer of 2020–2021 and ACC, using both ob-
servations and the CESM Large Ensemble, following the observational analysis of OK2021. We
again find that a normal distribution and a GEV distribution (both with a constant mean/location
parameter) fit the 7-day moving mean of daily temperature maxima in the observations without a
GMST dependence well (Fig. 6a). In this case, a normal distribution fit to the empirical distribu-
tion has a mean absolute error (MAE) of 0.01, equal to the MAE for the GEV fit. Supplementary
Fig. SI-3 shows results for normal, GEV, and empirical distributions that again suggest that the
normal distribution aligns well with the observed data. The alignment may reflect that the small
number of available data do not constrain the PDF very well. Given that the normal distribution
has one less parameter, it seems to offer a simpler approach to representing the data. The deviance
statistic for the observations does suggest a significant GMST component (leftmost bar in Fig. 6b).
However, given the possibility that this component is driven by internal variability rather than by
anthropogenic change (section 3a, Fig. 1a), one must conclude again that even if α significantly
differs from zero, this might reflect a bias introduced by the effects of internal variability. This
underscores the need to include covariates representing relevant internal variability based on our
physical understanding. The deviance statistic test for the CESM Large Ensemble shows that in

24
 (a) Obs (b) CESM-LE
slope = 1.3 slope = 1.8
2 p = 0.001 2 p = 0.001

Arctic Annual T ( C)

Arctic Annual T ( C)
r2 = 0.47 r2 = 0.61

1 1
0 0
1 1
2 2
0.0 0.5 1.0 0.0 0.5 1.0
GMST4year ( C) GMST4year ( C)
(c) Slope density plots in Obs and CESM-LE
100 slopeCESM LE = 1.8
slopeOBS = 1.3
80
60
PDF

40
20
0
0.75 1.00 1.25 1.50 1.75 2.00 2.25 2.50 2.75
Slope
Fig. 4. Internal variability and the attribution of Siberian heatwaves. (a) A scatter plot of the annual
average JJA Arctic (north of 70◦ N) surface temperatures against the 4-year running mean of GISTEMP annual
GMST. (b) Same, for the 35 members of the CESM Large Ensemble over the period 1926–2019. (c) PDF of
the slopes of annual average JJA Arctic surface temperatures against the 4-year running mean of annual GMST
within each member of the CESM Large Ensemble with 5,000 bootstrap resamples (red). The same distribution
is indicated for GISTEMP observations in blue. The vertical red line indicates the average slope across the
CESM Large Ensemble, and the vertical blue line indicates the slope determined from observations.

only5 of the 35 ensemble members is the extra complexity of adding GMST dependence to the
distribution justified (Fig. 6b). This raises the question of whether the influence of GMST, and
therefore of ACC in the WWA formulation, is a necessary part of the statistical description of
extreme temperatures in Australia as represented in the CESM Large Ensemble. It should be noted
that our results relate only to one climate model that, of course, does not necessarily represent all
the relevant physical processes at work in the real world. The use of multiple climate models in
the WWA analysis helps ensure a wider range of process representations are included.

25
 (a) P( ) (b) P( )
4 -0.1 10
1.5
3.5
8

3 -0.2
1
6
2.5
4
2 0.5 -0.3

2
1.5

1 0 -0.4 0
10 20 30 10 20 30

(c) P( ) P( 0 )
(d)
2.1 3.5
7 14.4
2 3
6 14.2
2.5
1.9 5
14
4 2
0

1.8
13.8
3 1.5
1.7 13.6
2 1
1.6 1 13.4 0.5

1.5 0 13.2 0
10 20 30 10 20 30
# of ensemble members # of ensemble members

Fig. 5. Siberian Heatwave: uncertainty in GEV parameters as a function of the number of ensemble
members used. Probability density function contour plots for each of the GEV parameters as a function of the
number of ensemble members used from the CESM Large Ensemble and the parameter value. One ensemble
member is equivalent in length to the historical record. Panel (a) represents the PDF for α, (b) ξ, (c) σ and (d)
µ0 . The PDFs were derived from the highest daily maximum Siberian temperature in June, as in CC2021, using
the CESM Large Ensemble. Contours are drawn every 0.3 for (a), 2 for (b), 1.5 for (c), and 0.7 for (d), all starting
at 0.05. We denote the 95% confidence range for each number of ensemble members by the red asterisks.

The regression analysis of Australian extreme temperatures and GMST, (Fig. 7a) as well as the
range of α values estimated from observations (Fig. 7b; red) both show a GMST dependence
near 2 ◦ C change in extreme temperatures per degree increase in GMST. As mentioned above, a

26
 (a) (b)
1.0
101
0.8

Deviance Statistic
0.6
CDF

0.4 100

Empirical CDF
0.2 GEV fit without GMST
Normal fit
95th %ile
0.0 90th %ile 10 1

30 32 34 36 Obs LE-5 LE-15 LE-25 LE-35

TX7x ( C) Ensemble Member #

Fig. 6. Australian Bushfire Heatwave: (a) Testing the justification for a GEV. The red curve indicates the
empirical CDF of the highest annual 7-day running mean of daily maximum temperature over the area defined in
van Oldenborgh et al. (2021) for observations. The purple and blue lines denote GEV (with a constant mean), and
normal distribution fits, respectively. The light and dark green shadings correspond to 90th and 95th percentile
confidence intervals. (b) Justification for using a GMST-dependent mean. Box plots of the deviance statistic
between a GEV without a GMST component and a GEV with a GMST component using the Australia Bushfire
heatwave example with 5,000 bootstrap resamples. The edges of the boxes correspond to the first and third
quartiles. The deviance statistic was computed for each of the 35 members from the CESM Large Ensemble
(blue) and compared with the 95th percentile significance level from a χ2 distribution with one degree of freedom
(red line). A box plot for the deviance statistic for observations is shown in green.

sensible null hypothesis for heatwaves in a warming climate is that the extreme values shift with the
mean warming (Tziperman 2022). The GMST dependence calculated here suggests that Australian
heatwaves increase at nearly twice the rate of the GMST. Such a strong response requires a physical
mechanism. Some candidates are (1) the fact that land heats up more quickly than the ocean, so
GMST always tends to lag behind land surface temperatures; (2) soil-drying can lead to further
increases in warming over land as less water is available for evaporative cooling, etc. It is not clear
that these can lead to a doubling of the GMST effect but further exploration of this amplification is
beyond the scope. The alternative is that this is a result of internal variability — possibly, ENSO
— affecting both Australian heat extremes and the GMST, consistent with Fig. 1a. Again, this

27
suggests that using this value of α for attribution may lead to biased results. Finally, Fig. 7b (blue)
shows a large range of GMST-dependencies for different ensemble members in the CESM Large
Ensemble, again highlighting that a large component of this dependence most likely represents
internal variability rather than anthropogenic change. The distribution of all GEV parameters for
the observations is given in Supplementary Fig. SI-4. How the GEV parameters scale with the
number of CESM Large Ensemble members used is indicated in Supplementary Fig. SI-5.

d. Madagascar Drought

Finally, we consider the attribution analysis for the 2019–2021 drought in Madagascar, following
the lead of HW2022. This case differs from the two examined above in that it analyzes precipitation
extremes that are positive by definition. HW2022, therefore, used a different extreme value
distribution function, the GPD, which is meant to represent the high tail of the distribution of
extreme events over a specified threshold (Coles et al. 2001). As discussed in Section 2, HW2022
analyzed the driest 20% of the 2-year averages of precipitation data from 1951–2020 multiplied by
minus one to turn minima (droughts) into maxima. We follow their example, with the important
caveat made in Section 2b that the validity of using a GPD this way for precipitation data is not
guaranteed, given that there is a hard maximum at zero precipitation that seems inconsistent with
the definition of the GPD. We also note that taking the lowest 20% of 2-year precipitation data
points amounts to only 14 data points, likely too low a number for a reliable attribution.
Fig. 8d shows the MRL plot (section 2e) for the Madagascar Drought case. As a reminder, this
plot needs to scale linearly within the range of the chosen threshold value (section 2e) in order for
the GPD analysis to be justified and self-consistent. For small and large values of precipitation,
the MRL curve appears to curve nonlinearly, indicating that a GPD fit is likely invalid using those
thresholds. For precipitation thresholds ranging from roughly 1.6 to 2 mm/day, the MRL curve
plateaus. The orange line shows the expected MRL line, (σ uξ1 − ξ); that slope does not align
with the MRL curve around the value of 1.9 corresponding to the 20th percentile precipitation
threshold chosen by Harrington et al. (2022). In general, we suggest that this MRL analysis should
be helpful in examining the self-consistency of applying a GPD to precipitation data, though the
small number of data points appears an obstacle for a reliable MRL analysis in this particular case.

28
 (a)
Linear Regression
slope = 1.9 GEV Linear Fit
p2= 0.001
TX7x ( C) 36 r = 0.15

34

32

30
0.2 0.0 0.2 0.4 0.6 0.8
4-yr Running GMST anomaly ( C)
30 (b) Observations
OBS = 1.7 CESM-LE
25 CESM LE = 0.48

20
15
PDF

10
5
0
3 2 1 0 1 2 3 4 5

Fig. 7. Australian Bushfire Heatwave: distribution of α values. (a) Scatter plot of the highest annual 7-day
running mean of daily maximum temperatures against GMST (following van Oldenborgh et al. 2021). The red
line indicates a linear regression fit to the data, and the blue line indicates a linear fit using the location parameter
and α derived from the GEV maximum likelihood estimate. (b; red) PDF of the GEV-fit α distributions from
observations, using bootstrap resampling. The α value from the observations without resampling is shown by
the red line (at 1.7) (b; blue) Effect of the internal variability and number of data from a model large ensemble
on the uncertainty in α for the Australian Bushfire heatwave example. Shown is a histogram of the GEV-fit α
PDFs derived from the α value from each ensemble member with no resampling. The average α value across
the Large Ensemble is indicated with the blue line.

Specifically, it is not obvious if the required linear range exists to a degree that justifies using a
GPD distribution and allows for selecting a threshold.

29
 We find again that an extreme value distribution is not required by the data distribution. The
GPD fit to the bottom 20th percentile precipitation data is no better than a log-normal distribution
fitted to all of the precipitation data (Fig. 8a). Furthermore, both the log-normal and GPD fit
indicate no need of a GMST dependence, in keeping with the small value of α found by HW2022.
A log-normal distribution seems a reasonable null hypothesis for the positive precipitation data.
In fact, the empirical CDF is practically unconstrained at its lower tail because there are no data
points below a value of 1.25 mm/day. Supplementary Fig. SI-6 for GPD, log-normal, and empirical
distributions also suggests the log-normal distribution looks reasonable, albeit acknowledging the
highly limited sample size in this case.
The deviance statistic for the Madagascar precipitation observations is marginally significant
at the X% level (leftmost green bar is above the red significance line , Fig. 8b), indicating that
the inclusion of the additional model parameter α is justified. Uncertainty plots for the GPD
parameters fit to these precipitation observations are indicated in Supplementary Fig. SI-7. The
bootstrap estimates for the α parameter (Supplementary Fig. SI-8a) ranges over both negative
and positive values, indicating that it does not significantly differ from the null. None of the 35
ensemble members exceed the deviance statistic significance level. Figure 1c shows that there is
no relation between these extreme drought events and GMST. One interpretation is that this is due
to the absence of any strong effect of internal variability modes on both GMST and Madagascar
drought, in contrast with the analysis of Australian temperatures in Figure 1a. Also, Figure 8b may
suggest that the CESM Large Ensemble does not adequately capture the effect of ACC on droughts
in Madagascar.
As a reminder, the threshold parameter of the GPD, corresponding to the maximum precipitation
value considered as an extreme drought year in this attribution case, was assumed by HW2022
to vary exponentially with the GMST (T ), as u = u0 expαT u0 with u0 < 0. If droughts become
more severe with increasing GMST, we expect the threshold parameter for the distribution of the
negative precipitation to get larger (less negative) with increasing GMST, and therefore αu0 to
be negative and α to be positive. The value of αGPD calculated from observations is small and
indistinguishable from zero (Fig. 8c). The exponential fit was assumed by HW2022 in order
to ensure that the scaling factor expαT u0 applied to both u0 and σ0 is strictly positive, so the
distribution has fixed dispersion (Philip et al. 2020). A scatter plot of ERA5 precipitation against

30
 (a) 101 (b)
1.0 Empirical CDF
GPD fit bottom 20%
Lognormal fit full distribution
95th %ile
0.8 90th %ile

Deviance Statistic
0.6
CDF

0.4
100
0.2

0.0
1.5 2.0 2.5 ObsLE-5 LE-15 LE-25 LE-35
Precipitation (mm/day) Ensemble Member #
(c) (d)
MRL
1.9 0.7 Expected MRL Slope
95th %ile
2-yr Mean Daily P (mm/day)

1.8 0.6
1.7 0.5
E(P - u | P > u)

0.4
1.6
0.3
1.5
0.2
1.4 0.1
GPD = 0.028
1.3 0.0
0.00 0.25 0.50 0.75 1.5 2.0 2.5
4-yr Running GMST anomaly ( C) u [Precipitation Threshold]
Fig. 8. Madagascar Drought: (a) Evaluating GPD and log-normal fits to rainfall data. An empirical CDF
of rainfall is shown (red line, with 90th and 95th percentile confidence intervals shown by red shadings), based
on 24-month running means of daily precipitation over the Madagascar area defined in HW2022 using ERA5
precipitation. A GPD fit to the bottom 20th percentile precipitation (black) and a log-normal fit to the full
distribution (blue). (b) The deviance statistic between a GPD with and without a GMST component in the CESM
Large Ensemble (blue) and observations (green). (c) Scatter plot of the 2-year running mean of ERA5 daily
precipitation over the Madagascar area defined in HW2022 against the 4-year running mean of annual GMST
for the lowest 20 percent precipitation data. The dashed blue line shows the exponential GPD log-likelihood fit
for the threshold parameter for the observations without resampling (eqn. 7) following HW2022. (d) MRL plot
(section 2e) for different precipitation thresholds. The blue shading defines a region of confidence calculated
as ± one standard deviation. The red line denotes the threshold used in the GPD fit of HW2022 (i.e., the 20th
percentile). The orange line represents the full linear equation from Equation 9 (σ uξ1 − ξ).

31
GMST anomalies looks notably linear (see the dashed blue line in Fig. 8c), consistent with the
Taylor expansion of the exponential function,

u = u0 expαT u0 ≈ u0 1 αT u0 = u0 αT .

The small value of α found for observations indicates that the increase in GMST has had little
effect on the statistics of two-year-average precipitation in Madagascar. Whereas this is also the
conclusion reached by HW2022, we found that the data can be fitted with a log-normal distribution
rather than GPD and without GMST dependence (Fig. 8a). Further, the deviance statistic only
marginally justifies adding GMST dependence for observations and does not justify adding such
dependence for model output (Fig. 8b). The fact that only 14 data points are involved in the
fit seems much too small to draw any meaningful conclusions on an extreme value distribution.
Supplementary Fig. SI-9a indicates that even if using the equivalent of 35 times as much data, the
uncertainty range for α includes both positive and negative values. One could, therefore, conclude
that there is no clear relationship between GMST and low rainfall in this region.

4. Conclusions

We were inspired in this work by a series of attribution studies by the “World Weather Attribution
Project” (WWA) and evaluated the part of their methodology that involves a fit of extreme value
distribution functions to observations. The full protocol used by this effort (Philip et al. 2020)
involves additional tools and model analyses, including selecting and defining extreme events,
analyzing trends in observational data, using multi-model ensembles to simulate event scenarios
with and without human-induced climate change, performing quality checks on model accuracy, and
synthesizing results to determine climate change’s influence on the event’s intensity and likelihood.
A central WWA method nevertheless involves fitting an extreme value distribution function to an
observed record, with the distribution parameters depending on global mean surface temperature
(GMST), whose variation is taken to represent ACC. For example, the location parameter of an
extreme value distribution (GEV) is sometimes made a linear function of GMST, T , as µ = µ0 α×T .
The GMST dependence included in the function is then used to calculate the change in return time
of extreme events. This method was applied to many extreme events such as heatwaves, cold
spells, floods, droughts, and more (e.g., van Oldenborgh et al. 2016; van der Wiel et al. 2017; van

32
Oldenborgh et al. 2017; Ciavarella et al. 2021; van Oldenborgh et al. 2021; Harrington et al. 2022).
We examined the robustness of the results of three example WWA studies, demonstrated when the
results may be unreliable due to the effects of internal variability, and suggested several additional
statistical tests, reviewed below, that may increase confidence in the results of such an analysis.
We began by examining the GMST dependence of extreme heatwave events in Siberia (following
CC2021) and Australia (following OK2021), as well as drought in Madagascar (following OK2021),
using data from a CESM preindustrial simulation. We found a significant GMST dependence in the
first two cases, indicating that such dependence can result from internal variability rather than only
represent the effects of ACC. The GMST dependence, in these cases, is a result of regional internal
variability such as ENSO or other slowly varying modes, which affects the extreme events being
examined, and also weakly affects GMST. This points that the value of α calculated in the more
realistic case where both effects are present includes the effects of internal variability in addition
to those of ACC. The value of α in the Australian case is especially large, and could naively be
interpreted as indicating that every degree increase in GMST is associated with a 5-degree increase
in Australian heatwaves. Heuristically, the large values of α can be thought of as a result of internal
variability (possibly ENSO) affecting GMST by 0.2 degrees for every 1-degree effect it has on
Australian heatwaves. As noted earlier, this issue could be mitigated in part by including indices
for modes of internal variability as covariates in the model fitting.
We tested if an extreme value distribution is indeed needed by plotting the empirical cumulative
distribution function (CDF) with bootstrapping error bars (Figs. 2a, 6a, 8a). We found that
fits using a standard distribution (normal or log-normal) fall within the error bars of the empirical
distribution, indicating that the observations do not empirically allow for distinguishing between the
characteristics of extreme value and normal distributions. The differing tails of these distributions,
however, lead to different extreme event return times. Similarly, the standard distributions seem to
fit the data well without a GMST dependence, indicating that such dependence may not be justified
by the data. We further tested if the addition of a GMST dependence of the distribution parameters
is justified statistically using the log-likelihood ratio test, also known as the deviance statistics.
When adding a parameter — in this case α — one expects the fit to improve, but the question is
whether the improvement is beyond what is expected simply due to the addition of one more model
parameter, so that one is convinced that the additional parameter represents a genuine effect in the

33
data. We found that in the Australia and Madagascar cases, there is no justification for the addition
of α.
The Madagascar drought attribution (HW2022) was based on only fourteen of the lowest two-
year-averaged precipitation rates, hardly sufficient to construct a meaningful empirical CDF or
to constrain the four parameters of the GPD distribution used for that purpose. In addition,
computing two-year averages of precipitation and applying an extreme value distribution may
lead to inconsistencies a priori given that averages converge to a Gaussian distribution (under the
Central Limit Theorem). The GPD fit to the record, in this case, is meant for high values above
a threshold, and the expected mean needs to vary linearly with the threshold chosen. We follow
Coles et al. (2001) in suggesting that a mean residual life (MRL) plot can generally be used to
test the self-consistency of using a GPD. In this case, it was not clear that these drought events
are well represented by a GPD, but with the aforementioned overarching concern regarding the
number of data points. To deal with the loss of data associated with this approach and difficulties in
choice of an optimal precipitation threshold, Naveau et al. (2016) suggest instead using a statistical
model fit to the full distribution of precipitation data, which has the added benefit of following the
rules of extreme value theory for both high and low precipitation events. While our results—that
attribution of two-year-averaged precipitation is not possible—agree with those of HW2022, we
emphasized our different perspective: it is unlikely that attribution would be possible based on such
a small number of data points, as well as the other difficulties pointed out above. In contrast, other
approaches evaluating changes in the seasonal cycle of precipitation have led to the attribution of
shifts in seasonal rainfall patterns in Southern Madagascar to climate change (Rigden et al. 2024).
An important issue raised by this work is the discrepancy between the modeled and observed
findings, as seen, for example, in the deviance statistic results for the observations versus CESM
Large Ensemble (Figs. 2b, 6b, 8b). This may be due to a particular sequence of internal variability
in the observations, measurement error in the observational record, or issues with the simulation
of extreme weather in the CESM Large Ensemble. It would be helpful to test if these discrepancies
also occur with other models. The use of multiple models by the WWA (Philip et al. (2020),
OK2021), may help mitigate against this kind of problem.
Overall, our results suggest uncertainty in the interpretation of the dependence of the parameters
of extreme event distributions on the global mean surface temperature as reflecting ACC. The

34
proposed and demonstrated deviance statistic, MRL plot, and a careful analysis of uncertainty
ranges, especially for α, may help avoid misguided confidence in attribution results. As noted
in the introduction, WWA does not exclusively rely upon the empirical methodology evaluated
here in order to reach conclusions, and also evaluates other historical data, model simulations,
and the dynamical context. Philip et al. (2020) also highlight some caveats associated with their
study, including the necessity of testing for other probability laws (e.g., Gaussian, Gamma, etc.)
and covariates (ENSO, PDO, etc.), among other issues. Nevertheless, the observations-based
component of the WWA methodology examined here still figures prominently in their analyses,
and we have highlighted several caveats with its interpretation, foremost that internal variability can
give rise to the appearance of strong covariance between GMST and local extremes. It is important
to study when the ACC signal emerges distinctly relative to internal variability on a case-by-case
basis. The ACC signal is known to have emerged in most regions of the world for temperature
extremes, while this may not be the case for variables like precipitation that have much higher
variability. Time of emergence is a factor that should always be considered when discussing the
results of an attribution study. Our results do not provide evidence against a relationship between
GMST and local extremes but, rather, that internal variability needs to be taken into account when
fitting a distribution function to the record. As currently implemented, the WWA approach can lead
to first-order biases in some cases. It, therefore, seems that it may be best to use a multi-method
approach to ensure a robust attribution approach, including the storyline approach that can add
insight about the change to extreme events (e.g., Shepherd et al. 2018; Leach et al. 2024; Robin
and Ribes 2020b; Sippel et al. 2024).
We suggest two key takeaways. First, controlling for internal variability, possibly by making the
distribution parameters depend on additional covariates such as ENSO, appears important in order
to accurately estimate the mean ACC trend. This improvement upon the current method should be
further developed. Inclusion of additional covariates is vital as a way to avoid biases in the estimate
of ACC effects on extreme event statistics. Second, empirical attribution techniques should be
examined by applying the approach to simulations with and without changes to anthropogenic
emissions. Application to such simulations would make it much easier to identify the likely
influence of other factors on the local response in the absence of ACC, and so to ensure that all
relevant modes are included in the statistical model, assuming the models are right in this respect.

35
Acknowledgments. PS is supported by the Harvard Global Institute. ET is supported by DOE
grant DE-SC0023134, and thanks the Weizmann Institute for its hospitality during parts of this
work. PH is supported by NSF Award 2123295. The CESM project is supported primarily by
the National Science Foundation (NSF). We acknowledge the CESM Large Ensemble Community
Project and supercomputing resources provided by NSF/CISL/Yellowstone. We also acknowledge
instructive conversations with Angela Rigden.

Data availability statement. All data used in this paper are available in public repositories. The
climatological output from the CESM Large Ensemble and CESM preindustrial runs used in this
paper are publicly available using the Intake-ESM Python package. All codes and data used
here are available under the public open science foundation depository https://osf.io/fycu2/, DOI
10.17605/OSF.IO/FYCU2.

36
References

Allen, M., 2003: Liability for climate change. Nature, 421 (6926), 891–892.

Christidis, N., P. A. Stott, and F. W. Zwiers, 2015: Fast-track attribution assessments based on
pre-computed estimates of changes in the odds of warm extremes. Climate Dynamics, 45 (5),
1547–1564.

Ciavarella, A., and Coauthors, 2021: Prolonged Siberian heat of 2020 almost impossible without
human influence. Climatic Change, 166 (1), 1–18.

Clarke, B., F. Otto, R. Stuart-Smith, and L. Harrington, 2022: Extreme weather impacts of climate
change: an attribution perspective. Environmental Research: Climate, 1 (1), 012 001.

Coles, S., J. Bawa, L. Trenner, and P. Dorazio, 2001: An introduction to statistical modeling of
extreme values, Vol. 208. Springer.

De Paola, F., M. Giugni, F. Pugliese, A. Annis, and F. Nardi, 2018: Gev parameter estimation and
stationary vs. non-stationary analysis of extreme rainfall in african test cities. Hydrology, 5 (2),
https://doi.org/10.3390/hydrology5020028, URL https://www.mdpi.com/2306-5338/5/2/28.

Eyring, V., and Coauthors, 2021: Human influence on the climate system. In Climate Change 2021:
The Physical Science Basis. Contribution of Working Group I to the Sixth Assessment Report of
the Intergovernmental Panel on Climate Change, Cambridge University Press, https://doi.org/
10.1017/9781009157896.005.

Faranda, D., S. Bourdin, M. Ginesta, M. Krouma, R. Noyelle, F. Pons, P. Yiou, and G. Messori,
2022: A climate-change attribution retrospective of some impactful weather extremes of 2021.
Weather and Climate Dynamics, 3 (4), 1311–1340, https://doi.org/10.5194/wcd-3-1311-2022,
URL https://wcd.copernicus.org/articles/3/1311/2022/.

Gettelman, A., and Coauthors, 2019: High climate sensitivity in the community earth system
model version 2 (cesm2). Geophysical Research Letters, 46 (14), 8329–8337.

Gillett, N. P., and Coauthors, 2016: The detection and attribution model intercomparison project
(damip v1. 0) contribution to cmip6. Geoscientific Model Development, 9 (10), 3685–3697.

37
Harrington, L. J., and Coauthors, 2022: Limited role of climate change in extreme low rain-
fall associated with southern madagascar food insecurity, 2019–21. Environmental Research:
Climate, 1 (2), 021 003, https://doi.org/10.1088/2752-5295/aca695, URL https://dx.doi.org/10.
1088/2752-5295/aca695.

Hegerl, G., and F. Zwiers, 2011: Use of models in detection and attribution of climate change.
Wiley interdisciplinary reviews: climate change, 2 (4), 570–591.

Hendon, H. H., D. W. Thompson, and M. C. Wheeler, 2007: Australian rainfall and surface
temperature variations associated with the Southern Hemisphere annular mode. Journal of
Climate, 20 (11), 2452–2467.

Hersbach, H., and Coauthors, 2020: The era5 global reanalysis. Quarterly Journal of the Royal
Meteorological Society, 146, 1999 – 2049.

IPCC, 2021: Summary for policymakers. Climate Change 2021: The Physical Science Basis.
Contribution of Working Group I to the Sixth Assessment Report of the Intergovernmental Panel
on Climate Change, Cambridge University Press, Cambridge, United Kingdom and New York,
NY, USA, 3–32, https://doi.org/10.1017/9781009157896.001.

IPCC, 2023: Section 2: Current status and trends. Climate Change 2023: Synthesis Re-
port. Contribution of Working Groups I, II and III to the Sixth Assessment Report of the
Intergovernmental Panel on Climate Change [Core Writing Team, H. Lee and J. Romero
(eds.)]. IPCC, Geneva, Switzerland, pp. 35-115, Cambridge University Press, https://doi.org/
10.59327/IPCC/AR6-9789291691647.

Kay, J. E., and Coauthors, 2015: The Community Earth System Model (CESM) large ensemble
project: A community resource for studying climate change in the presence of internal climate
variability. Bulletin of the American Meteorological Society, 96 (8), 1333–1349.

King, A. D., A. J. Pitman, B. J. Henley, A. M. Ukkola, and J. R. Brown, 2020: The role of climate
variability in Australian drought. Nature Climate Change, 10 (3), 177–179.

Knutson, T., J. Kossin, C. Mears, J. Perlwitz, and M. Wehner, 2017: Detection and attribution of
climate change. Climate Science Special Report: Fourth National Climate Assessment, Volume
I, U.S. Global Change Research Program, Washington, DC, USA.

38
Leach, N. J., and Coauthors, 2024: Heatwave attribution based on reliable operational weather
forecasts. Nature Communications, 15 (1), 4530.

Lenssen, N., G. Schmidt, J. Hansen, M. Menne, A. Persin, R. Ruedy, and D. Zyss, 2019: Im-
provements in the gistemp uncertainty model. J. Geophys. Res. Atmos., 124 (12), 6307–6326,
https://doi.org/10.1029/2018JD029522.

Lobeto, H., M. Menendez, and I. J. Losada, 2021: Future behavior of wind wave extremes due to
climate change. Scientific reports, 11 (1), 7869.

Lorenz, E. N., 1969: Atmospheric predictability as revealed by naturally occurring analogues.

Journal of Atmospheric Sciences, 26 (4), 636–646.

Lott, F. C., N. Christidis, and P. A. Stott, 2013: Can the 2011 east african drought be at-
tributed to human-induced climate change? Geophysical Research Letters, 40 (6), 1177–1181,
https://doi.org/https://doi.org/10.1002/grl.50235, URL https://agupubs.onlinelibrary.wiley.com/
doi/abs/10.1002/grl.50235.

Min, S.-K., W. Cai, and P. Whetton, 2013: Influence of climate variability on seasonal extremes
over australia. Journal of Geophysical Research: Atmospheres, 118 (2), 643–654.

Miralles, O., and A. C. Davison, 2023: Timing and spatial selection bias in rapid ex-
treme event attribution. Weather and Climate Extremes, 41, 100 584, https://doi.org/https:
//doi.org/10.1016/j.wace.2023.100584, URL https://www.sciencedirect.com/science/article/pii/
S2212094723000373.

National Academies of Sciences, 2016: Attribution of extreme weather events in the context
of climate change. National Academies Press, Division on Earth and Life Studies and Board
on Atmospheric Sciences and Committee on Extreme Weather Events and Climate Change
Attribution.

National Center for Atmospheric Research, 2024: Known issues - Large Ensemble Com-
munity Project (LENS). Community Earth System Model, URL https://www.cesm.ucar.edu/
community-projects/lens/known-issues, accessed: August 29, 2024.

Naveau, P., R. Huser, P. Ribereau, and A. Hannart, 2016: Modeling jointly low, moderate, and
heavy rainfall intensities without a threshold selection. Water Resources Research, 52 (4),

39
 2753–2769, https://doi.org/https://doi.org/10.1002/2015WR018552, URL https://agupubs.
onlinelibrary.wiley.com/doi/abs/10.1002/2015WR018552, https://agupubs.onlinelibrary.wiley.
com/doi/pdf/10.1002/2015WR018552.

Otto, F. E., S. Philip, S. Kew, S. Li, A. King, and H. Cullen, 2018: Attributing high-impact extreme
events across timescales—a case study of four different types of events. Climatic change, 149,
399–412.

Otto, F. E. L., and Coauthors, 2020: Challenges to understanding extreme weather changes in lower
income countries. Bulletin of the American Meteorological Society, 101 (10), E1851 – E1860,
https://doi.org/10.1175/BAMS-D-19-0317.1, URL https://journals.ametsoc.org/view/journals/
bams/101/10/bamsD190317.xml.

Pascale, S., S. B. Kapnick, T. L. Delworth, and W. F. Cooke, 2020: Increasing risk of another cape
town ”day zero” drought in the 21st century. Proceedings of the National Academy of Sciences,
117 (47), 29 495–29 503.

Pendergrass, A. G., 2018: What precipitation is extreme? Science, 360 (6393), 1072–
1073, https://doi.org/10.1126/science.aat1871, URL https://www.science.org/doi/abs/10.1126/
science.aat1871, https://www.science.org/doi/pdf/10.1126/science.aat1871.

Perkins, S., A. Pitman, N. J. Holbrook, and J. McAneney, 2007: Evaluation of the ar4 climate
models’ simulated daily maximum temperature, minimum temperature, and precipitation over
australia using probability density functions. Journal of climate, 20 (17), 4356–4376.

Philip, S., and Coauthors, 2020: A protocol for probabilistic extreme event attribution analyses.
Advances in Statistical Climatology, Meteorology and Oceanography, 6 (2), 177–203.

Pinto, I., R. Barimalala, S. Philip, M. Zachariah, S. Sengupta, and M. Vahlberg, 2023: Extreme
poverty rendering madagascar highly vulnerable to underreported extreme heat that would not
have occurred without human-induced climate change. Tech. rep., Grantham Institute-Climate
Change and the Environment.

Proistosescu, C., and P. J. Huybers, 2017: Slow climate mode reconciles historical and model-based
estimates of climate sensitivity. Science Advances, 3 (7), e1602 821.

40
Rigden, A., C. Golden, D. Chan, and P. Huybers, 2024: Climate change linked to drought in
southern madagascar. npj Climate and Atmospheric Science, 7 (1), 1–9.

Robin, Y., and A. Ribes, 2020a: Nonstationary extreme value analysis for event attribution
combining climate models and observations. Advances in Statistical Climatology, Meteo-
rology and Oceanography, 6 (2), 205–221, https://doi.org/10.5194/ascmo-6-205-2020, URL
https://ascmo.copernicus.org/articles/6/205/2020/.

Robin, Y., and A. Ribes, 2020b: Nonstationary extreme value analysis for event attribution com-
bining climate models and observations. Advances in Statistical Climatology, Meteorology and
Oceanography, 6 (2), 205–221.

Seneviratne, S., and Coauthors, 2021: Weather and climate extreme events in a changing climate.
Climate Change 2021: The Physical Science Basis. Contribution of Working Group I to the Sixth
Assessment Report of the Intergovernmental Panel on Climate Change, Cambridge University
Press, Cambridge, United Kingdom and New York, NY, USA.

Shepherd, T. G., and Coauthors, 2018: Storylines: an alternative approach to representing uncer-
tainty in physical aspects of climate change. Climatic change, 151, 555–571.

Sippel, S., and Coauthors, 2024: Could an extremely cold central european winter such as 1963
happen again despite climate change? Weather and Climate Dynamics, 5 (3), 943–957.

Smith, C. A., and P. D. Sardeshmukh, 2000: The effect of enso on the intraseasonal variance of
surface temperatures in winter. International Journal of Climatology: A Journal of the Royal
Meteorological Society, 20 (13), 1543–1557.

Stott, P. A., N. P. Gillett, G. C. Hegerl, D. J. Karoly, D. A. Stone, X. Zhang, and F. Zwiers, 2010:
Detection and attribution of climate change: a regional perspective. Wiley interdisciplinary
reviews: climate change, 1 (2), 192–211.

Stott, P. A., D. Stone, and M. R. Allen, 2004: Human contribution to the european heatwave of
2003. Nature, 432 (7017), 610–614.

Stott, P. A., and Coauthors, 2016: Attribution of extreme weather and climate-related events. WIREs
Climate Change, 7 (1), 23–41, https://doi.org/https://doi.org/10.1002/wcc.380, URL https://

41
 wires.onlinelibrary.wiley.com/doi/abs/10.1002/wcc.380, https://wires.onlinelibrary.wiley.com/
doi/pdf/10.1002/wcc.380.

The World Weather Attribution project, 2024: https://www.worldweatherattribution.org/.

Trevino, A. M., K. A. McKinnon, and P. J. Huybers, 2020: Extremely wrong: When limited data
leads to biases in general extreme values distribution estimates. AGU Fall Meeting Abstracts,
Vol. 2020, A147–0010.

Tziperman, E., 2022: Global Warming Science: A Quantitative Introduction to Climate Change
and Its Consequences. Princeton University Press, 336 pp.

Ummenhofer, C. C., M. H. England, P. C. McIntosh, G. A. Meyers, M. J. Pook, J. S. Risbey,

A. S. Gupta, and A. S. Taschetto, 2009: What causes southeast Australia’s worst droughts?
Geophysical Research Letters, 36 (4).

van der Wiel, K., and Coauthors, 2017: Rapid attribution of the august 2016 flood-inducing extreme
precipitation in south louisiana to climate change. Hydrology and Earth System Sciences, 21 (2),
897–921, https://doi.org/10.5194/hess-21-897-2017, URL https://hess.copernicus.org/articles/
21/897/2017/.

van Oldenborgh, G. J., F. E. L. Otto, K. Haustein, and K. AchutaRao, 2016: The heavy precipitation
event of december 2015 in chennai, india. Bulletin of the American Meteorological Society,
97 (12), S87 – S91, https://doi.org/10.1175/BAMS-D-16-0129.1, URL https://journals.ametsoc.
org/view/journals/bams/97/12/bams-d-16-0129.1.xml.

Van Oldenborgh, G. J., A. Van Urk, and M. Allen, 2012: The absence of a role of climate change
in the 2011 thailand floods. Bull. Amer. Meteor. Soc, 93, 1047–1049.

van Oldenborgh, G. J., and Coauthors, 2017: Attribution of extreme rainfall from hurricane
harvey, august 2017. Environmental Research Letters, 12 (12), 124 009, https://doi.org/10.1088/
1748-9326/aa9ef2, URL https://dx.doi.org/10.1088/1748-9326/aa9ef2.

van Oldenborgh, G. J., and Coauthors, 2021: Attribution of the Australian bushfire risk to anthro-
pogenic climate change. Natural Hazards and Earth System Sciences, 21 (3), 941–960.

42
Vautard, R., and P. Yiou, 2009: Control of recent european surface climate change by atmospheric
flow. Geophysical Research Letters, 36 (22).

Yamaguchi, M., J. C. Chan, I.-J. Moon, K. Yoshida, and R. Mizuta, 2020: Global warming changes
tropical cyclone translation speed. Nature Communications, 11 (47).

Yiou, P., R. Vautard, P. Naveau, and C. Cassou, 2007: Inconsistency between atmospheric dynamics
and temperatures during the exceptional 2006/2007 fall/winter and recent warming in europe.
Geophysical Research Letters, 34 (21).

Yoshida, K., M. Sugi, R. Mizuta, H. Murakami, and M. Ishii, 2017: Future changes in tropi-
cal cyclone activity in high-resolution large-ensemble simulations. Geophysical Research Let-
ters, 44 (19), 9910–9917, https://doi.org/https://doi.org/10.1002/2017GL075058, URL https:
//agupubs.onlinelibrary.wiley.com/doi/abs/10.1002/2017GL075058.

Zwiers, F. W., X. Zhang, and Y. Feng, 2011: Anthropogenic influence on long return pe-
riod daily temperature extremes at regional scales. Journal of Climate, 24 (3), 881 – 892,
https://doi.org/10.1175/2010JCLI3908.1, URL https://journals.ametsoc.org/view/journals/clim/
24/3/2010jcli3908.1.xml.

43