1 Photo: F. Zwiers (Strait of Juan de Fuca)

Eventa(ribu-on:theemergingscienceofa(ribu-ngcausestoextremeevents

Francis Zwiers PCIC, University of Victoria 13IMSC, 8 June 2016

2

Introduc-on•  Enormous interest in event attribution

–  Event and media driven –  Questions are mostly retrospective

•  Requires “rapid response” science –  Places high demands on process understanding, data,

models, and statistical methods –  Recently assessed by US National Academies of

Science •  Critical aspect of the the WCRP Grand Challenge

3 Photo: F. Zwiers (Jordan River, gathering storm)

Eventa(ribu-on

4

Eventa(ribu-on•  The public asks: Did human influence on the

climate system … –  Cause the event?

•  Most studies ask: Did it … –  Affect its odds? –  Alter its magnitude?

•  Some think we should reframe the question … –  Rather than “Did human influence …” (which requires

comparison with a counterfactual world) –  Ask “How much (eg, of a given storm’s precipitation) is

due to the attributed warming (eg, in the storm’s moisture source area)” (after Trenberth et al, 2015)

5

Moststudies

•  Compare factual and “counterfactual” climates –  Counterfactual à the world that might have been if we

had not emitted the ~600GtC that have been emitted since preindustrial

•  These studies almost always –  Define a class of events rather than a single event –  Use a probabilistic approach

•  Shepherd (2016) defines this as “risk based” –  Contrasts it with a “storyline” based approach –  i.e., analysis of the specific event that occurred

6

“Framing”eventa(ribu-onstudies

•  Event type –  Class vs individual

•  Analysis approach –  “risk based” or “storyline”

•  Event definition –  What spatial scale, duration, etc

•  Which risk-based question –  Did climate change alter the odds, or the magnitude?

•  What factors should be taken into account –  “Conditioning” –  e.g., coincident SST anomaly pattern, circulation, etc

The NAS Report (2016) struggled with these distinctions

7

Riskbasedques-ons

•  Did human influence alter its likelihood

•  Did human influence alter its magnitude

! ! !, !"#$%&' !!!"!!!(!|!,¬!"#$%&')!

!"#$ ! !"#$%&' !!!"!!!"#$(!|¬!"#$%&')!!"#$ ! !"#$%&', !!" !!!"!!!"#$(!|¬!"#$%&', !!")!

! ! !, !"#$%&', !!" !!!"!!!(!|!,¬!"#$%&', !!")!

8

China’sSummerof2013

Photo: F. Zwiers (Lijiang – Black Dragon Pool)

9

How rare was JJA of 2013?

-1

-0.5

0

0.5

1

1.5

1955

19

58

1961

19

64

1967

19

70

1973

19

76

1979

19

82

1985

19

88

1991

19

94

1997

20

00

2003

20

06

2009

20

12

1.1°C ≈ 3.5 SD above the 1955-1984 mean

1.1°C

Ano

mal

y re

lativ

e to

195

5-19

84

°C Sun et al, Nature Climate Change, 2014

•  Estimated event frequency •  once in 270-years in control simulations •  once in 29-years in “reconstructed” observations •  once in 4.3 years relative to the climate of 2013

•  Fraction of Attributable Risk in 2013: (p1 – p0)/p1≈ 0.984 •  Prob of “sufficient causation”: PS=1-((1-p1)/(1-p0)) ≈ 0.23

10

Calgaryflood,2013

Looking towards downtown Calgary from Riverfront Avenue (June 21, 2013), courtesy Ryan L.C. Quan

11

Calgaryfloods(Teufeletal,submi(ed)

46

Figure 13. Return times of (a) average May-June evapotranspiration over the northern

Great Plains, (b) maximum 1-day and (c) 3-day May-June precipitation over southern

Alberta, in present-day (red) and pre-industrial ensembles (blue). Gray horizontal

lines show (a) average evapotranspiration during the 14-21 June period, (b) average

precipitation on 20 June and (c) average precipitation during the 19-21 June period,

for the members of the CRCM5_Ref ensemble. Black dashed lines show (b) average

precipitation across the region on 20 June and (c) average precipitation during the 19-

21 June period, as estimated from CaPA.

100 101 102 1030

10

20

30

40

50

60

70

1−da

y m

axim

um (m

m)

Return period (years)

IRPIRaPIRbPIRc

100 101 102 1030

20

40

60

80

100

120

3−da

y m

axim

um (m

m)

Return period (years)

IRPIRaPIRbPIRc

(a)

(b) (c) Distribution of annual May-June maximum 1-day southern-Alberta precipitation in CRCM5 under factual and counter-factual conditions (conditional on prevailing global pattern of SST anomalies)

Frequency doubles (~25-yr à ~12 yr)

Magnitude increases ~10%

Southern Alberta MJ max 1-day precip

FAR=PN≈0.5 PS≈0.04

12 Photo: F. Zwiers (Marsh Wren)

Someunresolvedissues

13

Someunresolvedissues•  Event characterization

–  Class vs individual, risk-based vs storyline –  Individual is not completely synonymous with storyline –  Data assimilation approach of Hannart et al (2016)

•  Event definition •  Dependence on models •  Counterfactual state specification uncertainty

when conditional approach is used •  Selection bias

–  Need objective event selection criteria •  Communications

–  At each stage of the media and disaster response/recovery cycle

14

Retrospec-vevsprospec-ve?•  Most studies are prompted by specific events •  For the risk-based approach, we could study

pre-defined events

!" #$% &'()*+ #! %,-./.01.-2 3!-0/.45 !" 6710!4 3.!8.3% 6!496%4#17#.!4 .: 70!-# +;< => .;%; #!?713: #$% /!? %43 !" 6!2@7170/%'AB)*:; *.#6$%// %# 7/; CDEE<7> 0F 7/:! :$!? $!? .46/-:.!4 !"7%1!:!/ %G%6#: .2@1!H%: #$% 751%%2%4# !" :.2-/7#%3 :@7#.7/ @7##%14:!" :-1"76% 6/.27#% ?.#$ $.:#!1.67/ !0:%1H7#.!4:;

+;+ I%5.!4:

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

A-1 67/6-/7#.!4 !" 1%5.!47/ 7H%175%: !" &'()*+ 743!0:%1H%3 37#7 6!4:.:#: !" #$% "!//!?.45 :#%@:N

D; A-1 +D 1%5.!4: 71% 3%T4%3 -:.45 7: 07:% #$% )I^ _!<! 51.3;&%46%> ?% T1:# .4#%1@!/7#% #$% 2!3%/ 37#7 "1!2 #$% 6!71:% &'(9)*+ 51.3 #! #$.: T4% 51.3; L!1 #$.: @-1@!:%> ?% -#./.O% 7 #$.13 !13%1@!/P4!2.7/ .4#%1@!/7#.!4 :6$%2% .4 /7#.#-3% 743 /!45.#-3% C%;5;Y$.//.@: 743 M7P/!1 DEEZF; B.H%4 #$% :2!!#$4%:: !" #$% &'()*+T%/3:> #$.: .4#%1@!/7#.!4 @1!6%3-1% @1!3-6%: !4/P 2.4.27/ 2!3.T967#.!4 !" #$% !1.5.47/ T%/3:> 0!#$ "!1 #%2@%17#-1% 743 @1%6.@.#7#.!4C4!# :$!?4F; A4/P /743 @!.4#: 71% -:%3 .4 #$% .4#%1@!/7#.!4 743> 7:

7 1%:-/#> 4!# 7// !" #$% !1.5.47/ _!<! )I^ 51.3 @!.4#: .46/-3% .49#%1@!/7#%3 &'()*+ H7/-%: C%;5; ?$%4 7 )I^ @!.4# .: :-11!-43%30P !6%74 &'()*+ @!.4#:F; M$%:% 51.3 @!.4#: 71% 4!# .46/-3%3 .4#$% 1%5.!47/ 7H%175%:;

+; '"#%1 #$% &'()*+ 37#7 71% .4#%1@!/7#%3 !4#! #$% )I^51.3> 1%5.!47/ 7H%175%: 71% @1!3-6%3 -:.45 #$% .4#%1@!/7#%3 _!<!

37#7 ?%.5$#%3 0P #$% 6!:.4% !" #$% /7#.#-3% !" #$% 51.3 @!.4# /!679

!"#$ % I%5.!4: -:%3 .4 #$% 747/P:.: @1%:%4#%3 .4 #$.: ?!1Q C:%% M70/% + 743 #%8#F

&'()* + `.:# !" 1%5.!4: -:%3 .4 #$.: :#-3P; A4/P /743 51.3 @!.4#: 71%-:%3 .4 #$% 747/P:.:

[72% '61!4P2 `7#.#-3% C!F `!45.#-3% C!F

'-:#17/.7 '^] V<]aDD] DD_\aD<<\'27O!4 J7:.4 '*b +_]aD+[ W+KaSVK]!-#$%14 ]!-#$ '2%1.67 ]]' <Z]a+_] cZKaV_K)%4#17/ '2%1.67 )'* D_[aS_[ DDZKaWSKK%:#%14 [!1#$ '2%1.67 K[' S_[aZ_[ DS_KaD_SK)%4#17/ [!1#$ '2%1.67 )[' S_[a<_[ D_SKaW<K\7:#%14 [!1#$ '2%1.67 \[' +<[a<_[ W<KaZ_K'/7:Q7 '`' Z_[ac+[ Dc_KaD_SKB1%%4/743 BI` <_[aW<[ D_SKaD_K*%3.#%1174%74 J7:.4 *\( S_[aVW[ D_KaV_\[!1#$%14 \-1!@% [\^ VW[ac<[ D_KaV_\K%:#%14 '"1.67 K'L D+]aDW[ +_Ka++\\7:#%14 '"1.67 \'L D+]aDW[ ++\a<+\]!-#$%14 '"1.67 ]'L S<]aD+] D_Ka<+\]7$717 ]'& DW[aS_[ +_KaZ<\]!-#$%7:# ':.7 ]\' DD]a+_[ E<\aD<<\\7:# ':.7 \'] +_[a<_[ D__\aDV<\]!-#$ ':.7 ]'] <[aS_[ Z<\aD__\)%4#17/ ':.7 )'] S_[a<_[ V_\ac<\M.0%# MXJ S_[a<_[ c<\aD__\[!1#$ ':.7 ['] <_[ac_[ V_\aDW_\

Dc+ B.!15. 743 L1746.:6!N ^46%1#7.4#.%: .4 1%5.!47/ 6/.27#% 6$745% @1%3.6#.!4

Fast-track attribution assessments

1 3

estimates of the control temperature anomalies discussed in the previous paragraph and compute the FAR. Repeat-ing this for all the pairs, we end up with a sample of FAR estimates from which the best estimate (50th percentile) and the uncertainty range (5th and 95th percentiles) are calcu-lated. As a simpler alternative to the FAR, we also discuss our results in terms of changes in the probability of events.

4 Results

We first carry out an optimal fingerprinting analysis for each model individually by regressing simulated decadal temperature anomalies against the observed anomalies, as described in Sect. 3. Optimal fingerprinting, as applied here, decomposes the observations between the forced ANTHRO and NAT responses and internal variability. Figure 6 illustrates the ANTHRO and NAT scaling factors from analyses of changes in the annual mean, JJA and DJF temperature. The anthropogenic fingerprint is detected in all cases (scaling factors do not include zero). In most cases

the models under- or over-estimates the ANTHRO response (as implied by scaling factors greater or less than unity) and the fingerprint needs to be scaled up or down to best match the observations. This scaling introduces observational con-straints into our analysis expected to provide a more realis-tic representation of the climate response than the one from unscaled model patterns. The weaker NAT fingerprint is not detected in the observations with the exception of seasonal temperature analyses with the GISS-E2-R model, though the signal detectability in this case is sensitive to the exact EOF truncation employed in the analysis. NAT scaling fac-tors have greater uncertainties as the signal is weaker and hence more obscured by the effect of internal variability. The scaling factors from DJF analyses have moderately larger uncertainties compared to the ones for JJA and the annual mean, because of the greater land-area extent in the Northern Hemisphere, characterised by greater winter variability.

Jones et al. (2013) used several CMIP5 models (includ-ing six of the models used here) to further partition the anthropogenic response between the greenhouse gas

Fig. 8 As in Fig. 7, but for the JJA temperature. The record temperature anomaly in ANT (4.83 K) lies outside the plotted range

JJA mean temperature

°C

N. Christidis et al.

1 3

agreement, except for colder regions where the signal to noise ratio is lower.

There are a few interesting points to make on the results shown in Figs. 10, 11 and 12. Firstly, it is evident that model uncertainty is more prominent in the FAR estimates for DJF, as there are more cold regions in this season. (We refer to “cold regions” in the paper, meaning either polar regions, or mid/high latitude regions during wintertime). The DJF estimates are also generally smaller, because the temperature distributions for the actual and natural climate are less separated (Fig. 9), which means there is a smaller discrepancy in the likelihood of exceeding thresholds with and without the effect of human influence. The smaller sep-aration of the distributions results from the greater variabil-ity in colder regions and the consequently smaller signal to noise ratio. In some cases it is not possible to estimate the FAR for DJF at high thresholds, as these events are so rare that extreme statistics can no longer provide reliable probability estimates. Figure 11 shows that in most regions the GISS-E2-R model yields notably lower estimates of the FAR for JJA. While all models produce a weak natural warming over the globe of 0.05–0.15 K during 2003–2012

(possibly due to the recovery from preceding volcanic activity), it is only the NAT response of GISS-E2-R that is scaled up with positive scaling factors (Fig. 6b), which results in a warmer natural world and hence smaller FAR estimates. Finally, it is interesting to note that in the AMZ region, the FAR does not saturate to unity at high thresh-olds, but may even decrease. While in other regions the PDFs with and without the anthropogenic effect have a similar shape, the natural world PDF in AMZ is broader. The FAR is therefore determined not only by a shift of the distribution to warmer temperatures, but also by the change in the PDF shape, which leads to a decrease at higher thresholds.

We have so far only looked at best estimates of the FAR for each GCM corresponding to the 50th percentile of the FAR distribution as discussed in Sect. 3. We will now consider uncertainties in the FAR represented by the 5th–95th percentile range. We examine changes in the odds of a climatological 1-in-10 year event, which would provide a useful attribution assessment, for example, for the pur-poses of adaptation planning. We derive temperature anom-aly thresholds for the 1-in-10 year event in each region by

Fig. 11 As in Fig. 10, but for the JJA temperature

Fraction of Attributable Risk

N. Christidis et al.

1 3

agreement, except for colder regions where the signal to noise ratio is lower.

There are a few interesting points to make on the results shown in Figs. 10, 11 and 12. Firstly, it is evident that model uncertainty is more prominent in the FAR estimates for DJF, as there are more cold regions in this season. (We refer to “cold regions” in the paper, meaning either polar regions, or mid/high latitude regions during wintertime). The DJF estimates are also generally smaller, because the temperature distributions for the actual and natural climate are less separated (Fig. 9), which means there is a smaller discrepancy in the likelihood of exceeding thresholds with and without the effect of human influence. The smaller sep-aration of the distributions results from the greater variabil-ity in colder regions and the consequently smaller signal to noise ratio. In some cases it is not possible to estimate the FAR for DJF at high thresholds, as these events are so rare that extreme statistics can no longer provide reliable probability estimates. Figure 11 shows that in most regions the GISS-E2-R model yields notably lower estimates of the FAR for JJA. While all models produce a weak natural warming over the globe of 0.05–0.15 K during 2003–2012

(possibly due to the recovery from preceding volcanic activity), it is only the NAT response of GISS-E2-R that is scaled up with positive scaling factors (Fig. 6b), which results in a warmer natural world and hence smaller FAR estimates. Finally, it is interesting to note that in the AMZ region, the FAR does not saturate to unity at high thresh-olds, but may even decrease. While in other regions the PDFs with and without the anthropogenic effect have a similar shape, the natural world PDF in AMZ is broader. The FAR is therefore determined not only by a shift of the distribution to warmer temperatures, but also by the change in the PDF shape, which leads to a decrease at higher thresholds.

We have so far only looked at best estimates of the FAR for each GCM corresponding to the 50th percentile of the FAR distribution as discussed in Sect. 3. We will now consider uncertainties in the FAR represented by the 5th–95th percentile range. We examine changes in the odds of a climatological 1-in-10 year event, which would provide a useful attribution assessment, for example, for the pur-poses of adaptation planning. We derive temperature anom-aly thresholds for the 1-in-10 year event in each region by

Fig. 11 As in Fig. 10, but for the JJA temperature

N. Christidis et al.

1 3

agreement, except for colder regions where the signal to noise ratio is lower.

There are a few interesting points to make on the results shown in Figs. 10, 11 and 12. Firstly, it is evident that model uncertainty is more prominent in the FAR estimates for DJF, as there are more cold regions in this season. (We refer to “cold regions” in the paper, meaning either polar regions, or mid/high latitude regions during wintertime). The DJF estimates are also generally smaller, because the temperature distributions for the actual and natural climate are less separated (Fig. 9), which means there is a smaller discrepancy in the likelihood of exceeding thresholds with and without the effect of human influence. The smaller sep-aration of the distributions results from the greater variabil-ity in colder regions and the consequently smaller signal to noise ratio. In some cases it is not possible to estimate the FAR for DJF at high thresholds, as these events are so rare that extreme statistics can no longer provide reliable probability estimates. Figure 11 shows that in most regions the GISS-E2-R model yields notably lower estimates of the FAR for JJA. While all models produce a weak natural warming over the globe of 0.05–0.15 K during 2003–2012

(possibly due to the recovery from preceding volcanic activity), it is only the NAT response of GISS-E2-R that is scaled up with positive scaling factors (Fig. 6b), which results in a warmer natural world and hence smaller FAR estimates. Finally, it is interesting to note that in the AMZ region, the FAR does not saturate to unity at high thresh-olds, but may even decrease. While in other regions the PDFs with and without the anthropogenic effect have a similar shape, the natural world PDF in AMZ is broader. The FAR is therefore determined not only by a shift of the distribution to warmer temperatures, but also by the change in the PDF shape, which leads to a decrease at higher thresholds.

We have so far only looked at best estimates of the FAR for each GCM corresponding to the 50th percentile of the FAR distribution as discussed in Sect. 3. We will now consider uncertainties in the FAR represented by the 5th–95th percentile range. We examine changes in the odds of a climatological 1-in-10 year event, which would provide a useful attribution assessment, for example, for the pur-poses of adaptation planning. We derive temperature anom-aly thresholds for the 1-in-10 year event in each region by

Fig. 11 As in Fig. 10, but for the JJA temperature

N. Christidis et al.

1 3

agreement, except for colder regions where the signal to noise ratio is lower.

There are a few interesting points to make on the results shown in Figs. 10, 11 and 12. Firstly, it is evident that model uncertainty is more prominent in the FAR estimates for DJF, as there are more cold regions in this season. (We refer to “cold regions” in the paper, meaning either polar regions, or mid/high latitude regions during wintertime). The DJF estimates are also generally smaller, because the temperature distributions for the actual and natural climate are less separated (Fig. 9), which means there is a smaller discrepancy in the likelihood of exceeding thresholds with and without the effect of human influence. The smaller sep-aration of the distributions results from the greater variabil-ity in colder regions and the consequently smaller signal to noise ratio. In some cases it is not possible to estimate the FAR for DJF at high thresholds, as these events are so rare that extreme statistics can no longer provide reliable probability estimates. Figure 11 shows that in most regions the GISS-E2-R model yields notably lower estimates of the FAR for JJA. While all models produce a weak natural warming over the globe of 0.05–0.15 K during 2003–2012

(possibly due to the recovery from preceding volcanic activity), it is only the NAT response of GISS-E2-R that is scaled up with positive scaling factors (Fig. 6b), which results in a warmer natural world and hence smaller FAR estimates. Finally, it is interesting to note that in the AMZ region, the FAR does not saturate to unity at high thresh-olds, but may even decrease. While in other regions the PDFs with and without the anthropogenic effect have a similar shape, the natural world PDF in AMZ is broader. The FAR is therefore determined not only by a shift of the distribution to warmer temperatures, but also by the change in the PDF shape, which leads to a decrease at higher thresholds.

We have so far only looked at best estimates of the FAR for each GCM corresponding to the 50th percentile of the FAR distribution as discussed in Sect. 3. We will now consider uncertainties in the FAR represented by the 5th–95th percentile range. We examine changes in the odds of a climatological 1-in-10 year event, which would provide a useful attribution assessment, for example, for the pur-poses of adaptation planning. We derive temperature anom-aly thresholds for the 1-in-10 year event in each region by

Fig. 11 As in Fig. 10, but for the JJA temperature

N. Christidis et al.

1 3

agreement, except for colder regions where the signal to noise ratio is lower.

There are a few interesting points to make on the results shown in Figs. 10, 11 and 12. Firstly, it is evident that model uncertainty is more prominent in the FAR estimates for DJF, as there are more cold regions in this season. (We refer to “cold regions” in the paper, meaning either polar regions, or mid/high latitude regions during wintertime). The DJF estimates are also generally smaller, because the temperature distributions for the actual and natural climate are less separated (Fig. 9), which means there is a smaller discrepancy in the likelihood of exceeding thresholds with and without the effect of human influence. The smaller sep-aration of the distributions results from the greater variabil-ity in colder regions and the consequently smaller signal to noise ratio. In some cases it is not possible to estimate the FAR for DJF at high thresholds, as these events are so rare that extreme statistics can no longer provide reliable probability estimates. Figure 11 shows that in most regions the GISS-E2-R model yields notably lower estimates of the FAR for JJA. While all models produce a weak natural warming over the globe of 0.05–0.15 K during 2003–2012

(possibly due to the recovery from preceding volcanic activity), it is only the NAT response of GISS-E2-R that is scaled up with positive scaling factors (Fig. 6b), which results in a warmer natural world and hence smaller FAR estimates. Finally, it is interesting to note that in the AMZ region, the FAR does not saturate to unity at high thresh-olds, but may even decrease. While in other regions the PDFs with and without the anthropogenic effect have a similar shape, the natural world PDF in AMZ is broader. The FAR is therefore determined not only by a shift of the distribution to warmer temperatures, but also by the change in the PDF shape, which leads to a decrease at higher thresholds.

We have so far only looked at best estimates of the FAR for each GCM corresponding to the 50th percentile of the FAR distribution as discussed in Sect. 3. We will now consider uncertainties in the FAR represented by the 5th–95th percentile range. We examine changes in the odds of a climatological 1-in-10 year event, which would provide a useful attribution assessment, for example, for the pur-poses of adaptation planning. We derive temperature anom-aly thresholds for the 1-in-10 year event in each region by

Fig. 11 As in Fig. 10, but for the JJA temperature

N. Christidis et al.

1 3

agreement, except for colder regions where the signal to noise ratio is lower.

There are a few interesting points to make on the results shown in Figs. 10, 11 and 12. Firstly, it is evident that model uncertainty is more prominent in the FAR estimates for DJF, as there are more cold regions in this season. (We refer to “cold regions” in the paper, meaning either polar regions, or mid/high latitude regions during wintertime). The DJF estimates are also generally smaller, because the temperature distributions for the actual and natural climate are less separated (Fig. 9), which means there is a smaller discrepancy in the likelihood of exceeding thresholds with and without the effect of human influence. The smaller sep-aration of the distributions results from the greater variabil-ity in colder regions and the consequently smaller signal to noise ratio. In some cases it is not possible to estimate the FAR for DJF at high thresholds, as these events are so rare that extreme statistics can no longer provide reliable probability estimates. Figure 11 shows that in most regions the GISS-E2-R model yields notably lower estimates of the FAR for JJA. While all models produce a weak natural warming over the globe of 0.05–0.15 K during 2003–2012

(possibly due to the recovery from preceding volcanic activity), it is only the NAT response of GISS-E2-R that is scaled up with positive scaling factors (Fig. 6b), which results in a warmer natural world and hence smaller FAR estimates. Finally, it is interesting to note that in the AMZ region, the FAR does not saturate to unity at high thresh-olds, but may even decrease. While in other regions the PDFs with and without the anthropogenic effect have a similar shape, the natural world PDF in AMZ is broader. The FAR is therefore determined not only by a shift of the distribution to warmer temperatures, but also by the change in the PDF shape, which leads to a decrease at higher thresholds.

We have so far only looked at best estimates of the FAR for each GCM corresponding to the 50th percentile of the FAR distribution as discussed in Sect. 3. We will now consider uncertainties in the FAR represented by the 5th–95th percentile range. We examine changes in the odds of a climatological 1-in-10 year event, which would provide a useful attribution assessment, for example, for the pur-poses of adaptation planning. We derive temperature anom-aly thresholds for the 1-in-10 year event in each region by

Fig. 11 As in Fig. 10, but for the JJA temperature

N. Christidis et al.

1 3

agreement, except for colder regions where the signal to noise ratio is lower.

There are a few interesting points to make on the results shown in Figs. 10, 11 and 12. Firstly, it is evident that model uncertainty is more prominent in the FAR estimates for DJF, as there are more cold regions in this season. (We refer to “cold regions” in the paper, meaning either polar regions, or mid/high latitude regions during wintertime). The DJF estimates are also generally smaller, because the temperature distributions for the actual and natural climate are less separated (Fig. 9), which means there is a smaller discrepancy in the likelihood of exceeding thresholds with and without the effect of human influence. The smaller sep-aration of the distributions results from the greater variabil-ity in colder regions and the consequently smaller signal to noise ratio. In some cases it is not possible to estimate the FAR for DJF at high thresholds, as these events are so rare that extreme statistics can no longer provide reliable probability estimates. Figure 11 shows that in most regions the GISS-E2-R model yields notably lower estimates of the FAR for JJA. While all models produce a weak natural warming over the globe of 0.05–0.15 K during 2003–2012

(possibly due to the recovery from preceding volcanic activity), it is only the NAT response of GISS-E2-R that is scaled up with positive scaling factors (Fig. 6b), which results in a warmer natural world and hence smaller FAR estimates. Finally, it is interesting to note that in the AMZ region, the FAR does not saturate to unity at high thresh-olds, but may even decrease. While in other regions the PDFs with and without the anthropogenic effect have a similar shape, the natural world PDF in AMZ is broader. The FAR is therefore determined not only by a shift of the distribution to warmer temperatures, but also by the change in the PDF shape, which leads to a decrease at higher thresholds.

We have so far only looked at best estimates of the FAR for each GCM corresponding to the 50th percentile of the FAR distribution as discussed in Sect. 3. We will now consider uncertainties in the FAR represented by the 5th–95th percentile range. We examine changes in the odds of a climatological 1-in-10 year event, which would provide a useful attribution assessment, for example, for the pur-poses of adaptation planning. We derive temperature anom-aly thresholds for the 1-in-10 year event in each region by

Fig. 11 As in Fig. 10, but for the JJA temperature

N. Christidis et al.

1 3

agreement, except for colder regions where the signal to noise ratio is lower.

There are a few interesting points to make on the results shown in Figs. 10, 11 and 12. Firstly, it is evident that model uncertainty is more prominent in the FAR estimates for DJF, as there are more cold regions in this season. (We refer to “cold regions” in the paper, meaning either polar regions, or mid/high latitude regions during wintertime). The DJF estimates are also generally smaller, because the temperature distributions for the actual and natural climate are less separated (Fig. 9), which means there is a smaller discrepancy in the likelihood of exceeding thresholds with and without the effect of human influence. The smaller sep-aration of the distributions results from the greater variabil-ity in colder regions and the consequently smaller signal to noise ratio. In some cases it is not possible to estimate the FAR for DJF at high thresholds, as these events are so rare that extreme statistics can no longer provide reliable probability estimates. Figure 11 shows that in most regions the GISS-E2-R model yields notably lower estimates of the FAR for JJA. While all models produce a weak natural warming over the globe of 0.05–0.15 K during 2003–2012

(possibly due to the recovery from preceding volcanic activity), it is only the NAT response of GISS-E2-R that is scaled up with positive scaling factors (Fig. 6b), which results in a warmer natural world and hence smaller FAR estimates. Finally, it is interesting to note that in the AMZ region, the FAR does not saturate to unity at high thresh-olds, but may even decrease. While in other regions the PDFs with and without the anthropogenic effect have a similar shape, the natural world PDF in AMZ is broader. The FAR is therefore determined not only by a shift of the distribution to warmer temperatures, but also by the change in the PDF shape, which leads to a decrease at higher thresholds.

We have so far only looked at best estimates of the FAR for each GCM corresponding to the 50th percentile of the FAR distribution as discussed in Sect. 3. We will now consider uncertainties in the FAR represented by the 5th–95th percentile range. We examine changes in the odds of a climatological 1-in-10 year event, which would provide a useful attribution assessment, for example, for the pur-poses of adaptation planning. We derive temperature anom-aly thresholds for the 1-in-10 year event in each region by

Fig. 11 As in Fig. 10, but for the JJA temperature

Distribution of annualJJA temperature in the 2000’s relative to 1961-90 in East Asia with and without ANT forcing

Chr

istid

is e

t al,

2014

°C

15 Photo: F. Zwiers

Questions?