1 extreme value modelling in climate science: why do it and how it can fail! aims: what the heck do...
Post on 01-Jan-2016
215 Views
Preview:
TRANSCRIPT
1
Extreme Value Modelling in Climate Science:Why do it and how it can fail!
Aims:• What the heck do we mean by “extreme”? • Summary of statistical methods used in climate science Statistics for modelling the process rather than for just making indices• Some examples of extreme value modelling:
• Problem 1: Properties of drought indices• Problem 2: Trends in extreme gridded temperatures• Problem 2: Trends in largest annual skew tides;
Professor David B. StephensonU. of Exeter
NCAR summer colloquium, 8 June 2011© 2011 D.B.Stephenson@exeter.ac.uk
2
Some wet and windy extremes
Extra-tropical cyclone
Hurricane
Polar low
Extra-tropical cyclone
Convective severe storm
3
Some dry and hot extremesDrought
Wild fireDust storm
Dust storm
4
All are complex multivariate spatio-temporal events!So to massively simplify, it is helpful to focus in on the time evolution of single variable related to the event e.g. wind speeds of major extratropical cyclones passing by London, losses to an insurers, etc.
MARKED POINT PROCESS: random times, random marks
5
What do we mean by “extreme”?Large meteorological values
Maximum value (i.e. a local extremum) Exceedance above a high threshold Record breaker (time-varying threshold
equal to max of previously observed values)
Rare event in the tail of distribution (e.g. less than 1 in 100 years – p=0.01)
Large losses (severe or high-impact)(e.g. $200 billion if hurricane hits Miami) hazard, vulnerability, and exposure
Gare Montparnasse, 22 Oct 1895
Stephenson, D.B. (2008): Chapter 1: Definition, diagnosis, and origin of extreme weather and climate events, In Climate Extremes and Society , R. Murnane and H. Diaz (Eds), Cambridge University Press, pp 348 pp.
),()),(( txetxhVRisk
NOTE! Extremeness is not a binary property of an event but an ordering of a process
6
IPCC 2001 definitionsSimple extremes:
“individual local weather variables exceeding critical levels on a continuous scale”
Complex extremes:“severe weather associated with particular climatic phenomena, often requiringa critical combination of variables”
Extreme weather event:“An extreme weather event is an event that is rare within its statistical referencedistribution at a particular place. Definitions of "rare" vary, but an extremeweather event would normally be as rare or rarer than the 10th or 90th percentile.”
Extreme climate event:“an average of a number of weather events over a certain period of time which is itself extreme (e.g.rainfall over a season)”
X~N(0,1) Y~N(0.5,1.5)
px=rank(x)/(n+1)
7
How might extreme events change?
Changes in location, scale, and shape all lead to big changes in the tail of thedistribution.
Some physical argumentsexist for changes in locationand scale.
E.g. multiplicative changein precipitation due to increased humidity(change in scale)
Scale change impacts high quantiles!Example: Normal variable1% increase in standard deviation s shifts the 10-year return value (x0.9) by 1.28s and the 200-year return value (x0.995) by 2.58s.
8
How can we relate the tails …
9
to the bulk of the distribution?
PDF = Probability Density Function Or … Probable Dinosaur Function??
Change in scale Change in shape
10
Quantile attributionDescribe the changes in quantiles in terms of changes in the location, the scale, and the shape of the parent distribution:
0.5 0.5( )
shape changes
IQRX X X X
IQR
The quantile shift is the sum of:• a location effect (shift in median)• a scale effect (change in IQR)• a shape effect
Ferro, C.A.T., D.B. Stephenson, and A. Hannachi, 2005: Simple non-parametric techniques for exploring changing probability distributions of weather, J. Climate, 18, 4344 4354.
Beniston, M. and Stephenson, D.B. (2004): Extreme climatic events and their evolution under changing climatic conditions, Global and Planetary Change, 44, pp 1-9
11
Example: Regional Model Simulations of daily Tmax
Changes in location, scale and shape all important
T90ΔT90 (2071-2100 minus 1971-2000)
ΔT90-Δm
ΔT90-Δm-(T90-m) Δs/s
12
Statistical methods used in climate science Extreme indices – sample statistics Basic extreme value modelling
GEV modelling of block maxima GPD modelling of excesses above high
threshold Point process model of exceedances
More complex EVT models Inclusion of explanatory factors
(e.g. trend, ENSO, etc.) Spatial pooling Max stable processes Bayesian hierarchical models + many more
Other stochastic process models
13
Extreme indices are useful and easy but … They don’t always measure extreme
values in the tail of the distribution! They often confound changes in rate
and magnitude They strongly depend on threshold and
so make model comparison difficult They say nothing about extreme
behaviour for rarer extreme events at higher thresholds
They generally don’t involve probability so fail to quantify uncertainty (no inferential model)
More informative approach: model the extremal process using statisticalmodels whose parameters are then sufficient to provide complete summaries of all other possible statistics (and can simulate!)
See: Katz, R.W. (2010) “Statistics of Extremes in Climate Change”, Climatic Change, 100, 71-76
14
Furthermore … indices are not METRICS!One should avoid the word “metric” unless the statistic has distance properties! Index, sample/descriptive statistic, or measure is a more sensible name!
Oxford English Dictionary: Metric - A binary function of a topological space which gives, for any two points of the space, a value equal to the distance between them, or a value treated as analogous to distance for analysis.
Properties of a metric:d(x, y) ≥ 0 d(x, y) = 0 if and only if x = yd(x, y) = d(y, x) d(x, z) ≤ d(x, y) + d(y, z)
15
Universal Poisson process for extremes
1/
2 1
For a large number of
independent and identically
distributed values and a
sufficiently high threshold :
Pr
( ) 1
n Λ
n
z
N ~ Poisson(Λ)
Λ e(N n)
n!
zΛ t t
N=number of pointswith Z>z
t=t1 t=t2
Miraculous limit theorem for tails of i.i.d. variables!
16
Probability models for maxima and excesses
1/
2 1
1/
lim ,
Pr ( ) 1
Pr max( ) Pr ( ) 0
Pr | ( ) / ( )
1 ( )
n Λ
n z
Λ e z(N n) Λ t t
n!
Z z N z e
Z z Z u z u
z uu
Generalized Extreme Value (GEV) distribution
Gener
alized Pareto Distribution (GPD)
Note: extremal properties are characterised by only three parameters (for ANY underlying distribution!)
17
Why use these probability models? Model parameters are sufficient for providing a
complete threshold-independent description of extremal properties. All other statistics of the extremal process are a function of these three parameters.
The models provide a rigorous probability framework for making inference about extremal behaviour. Their mathematically justifiable parametric form allows more precise inference about tail properties.
Model can be used to smoothly interpolate between empirical quantiles/probabilities. Such interpolation has made efficient use of all the large values;
Model can be used to extrapolate out carefully to rarer less frequently (or never!) observed events AND provide intervals for such predictions!
18
Problem 1: Do 2 drought series have similar extremal properties?
Observed indexn=90
Reconstructed indexn=5000
Data example kindlyprovided by Eleanor Burke,Met Office
19
Do 2 drought series have similar extremal properties?
Observed indexn=90
+
Reconstructed indexn=5000
Data example kindlyprovided by Eleanor Burke,Met Office
20
Return level plots
Outlier in the extended data set? Slight kink at 2.5 in d1
),)1/(1(
periodsreturn empirical versusquantiles Empirical
][1
iyni
21
Quantile-Quantile plot
Empirical distributions similar except for the big outlier in d1
22
Modelling the excesses using GPD
i
/11
/1
Zof ceIndependen .2
,lim
support Asymptotic 1.
1
)()(
~1~1
)(
)(~
~11|Pr)(
un
uZE
uzzf
u
uzuZzZzF
:sAssumption
23
Use of mean excess to find a suitable threshold u
GPD implies linear behaviour in mean excess for u from about 0 to 1 Try fits with u=0.5 as threshold
Observedd0
Simulatedd1
1
u)X|u-E(X
u
24
Nested model approach
data extended
data obs
1
0
~~~
~~
~1~1
)(
10
10
0
0
0
/11
i
i
i
a
X
X
X
H
H
uzzf
modelContrast
model Null
25
Is there a statistically significant difference at 5% level?
• Difference in deviance 1050.4-1053.6+2*2=0.8 p=0.67• Parameter estimates 0.234/0.321=0.729 p=0.23
0.299/0.313=0.956 p=0.17
Maximum Likelihood EstimatesModel No. of
paramsAkaike Inf. Criterion
Null 2 1050.4 0.631
(0.023)
--- -0.107
(0.024)
---
Contrast 4 1053.6 0.629
(0.023)
0.234
(0.321)
-0.105
(0.024)
-0.299
(0.313)
Contrast with outlier
4 1085.2 0.599
(0.021)
0.264
(0.321)
-0.042
(0.019)
-0.361
(0.313)
pnts ~/ˆ
pnts ~/ˆ
2242 ~ DD
0~
1~ 0 1
No significant difference between the exceedances at 5% level
4.6107.0/631.05.0/~-ulimit upper Predicted 00
26
Model checking: do the quantiles match?
No! The null model underestimates the empirical quantiles
1)Pr(~
)Pr())(1(
1
)Pr()|Pr(
1
)Pr(
1
uZTuz
uZzF
uZuZzZ
zZT
T
value Return
period Return
27
Model checking: are estimates stable?
No! Constant up to u=1.7 but then trends for larger values?!
28
Model checking: uniform in time?
Uniform distribution in time and exponential between events
29
Problem 2: Extremes in surface temperatureCoelho, C.A.S., Ferro, C.A.T., Stephenson, D.B. and Steinskog, D.J. (2008): Methods for exploring spatial and temporal variability of extreme events in climate data, Journal of Climate, 21, pp 2072-2092
Observed surface temperatures 1870-2005 Monthly mean gridded surface temperature (HadCRUT2v)
5 degree resolution Summer months only: June July August Grid points with >50% missing values and SH are omitted.
-150 -100 -50 0 50 100 150
02
04
06
08
0
Maximum temperature
0 5 10 15 20 25 30 35 40Celsius
Maximum monthly temperatures
30year
Tem
pera
ture
(Cel
sius)
2001 2002 2003 2004 2005 2006
-50
510
1520 a)
Long term trend in mean
75th quantile (uy,m = 16.2ºC)
2003 exceedance
Excess (Ty,m – uy,m)
Non-stationarity due to seasonality and long term trends
Example: Grid point in Central Europe (12.5ºE, 47.5ºN)
31
GPD scale and shape estimates1
0 1 0
Pr( | ) 1
log
z uZ z Z u
x
Shape parameter is mainly negative suggesting finite upper temperature.
Spatial pooling has been used to get more reliable less noisey shape estimates
Scale parameter is large over high-latitude land areas AND shows some dependence on x=ENSO.
0e
1e
32
How significant is ENSO on extremes?
Null hypothesis of no effect can only be rejected with confidence over tropical Pacific and Northern Continents
33
Use of covariates in models
“with four parameters I can fit an elephant and with five I can make him wiggle his trunk.” - John von Neumann
34
Model can be used to estimate return periods
-150 -100 -50 0 50 100 150
020
4060
80
a) August 2003: Excesses above 75% threshold
0 1 2 3 4Celsius
-150 -100 -50 0 50 100 1500
2040
6080
b) August 2003: Return period
1 5 10 50 150 500years
return period of 133 years for August 2003 event over Europe
Excess for August 2003
Return period for the excessfor August 2003
35
Spatial poolingPool over local grid points but allow for spatial variation by including local spatial covariates to reduce bias (bias-variance tradeoff).
For each grid point, estimate 5 GPD parameters by maximising the following likelihood over the 8 neighbouring grid points:
0,
,,0,,
,,
1
11
,
,
)()(log
),;(
jijjii
jjjyjiiii
xjijijjii
jjiijjii
ij
jjiiij
yyxx
yfL
No spatial pooling: 2 parameters from n data valuesLocal pooling: 5 parameters from 9n data values
Coelho et al., 2008:Methods for Exploring Spatial and TemporalVariability of Extreme Events in Climate Data, J. Climate
36
-150 -100 -50 0 50 100 150
02
04
06
08
0
b) Chi bar (75th quantile) Central Europe
-0.4 -0.1 0.1 0.4 0.7 1
Teleconnections of extremesBivariate measure of extremal dependency:
2 log Pr( )1
log Pr(( ) & ( ))
Coles et al., Extremes, (1999)
Y u
X u Y u
association with extremes in subtropical Atlantic
37
Problem 3: Is there a time trend in extreme skew tides?10 largest skew tides for each of n=149 years
Is there a time trend in the extremes?
Dots show largest values
Line is linear fit to the meanof the 10 values
Data example kindlyprovided by Tom Howard,Met Office
38
r Largest Order model
0
0
10
1
0
)(
/1
12
!1)(Pr)max(Pr
1)( Pr
,lim
X
es
rzNzZ
zttΛ
n!
eΛn)(N
zn
r
s
sr
Λn
model Trend
ondistributiOrder Largest r
39
• Estimates for null model are similar for r=1,5,10• Estimates get more precise for larger r• Null model has slightly better AIC than trend model• Trend model has trouble estimating trend parameter
Could either constrain shape=0 and/or pool over more data
Maximum Likelihood EstimatesModel No. of
paramsAIC
Null
r=10
3 -6882.1 0.661
(0.0096)
--- 0.147
(0.0065)
0.033
(0.025)
Null
r=5
3 -2469.2 0.659
(0.0099)
--- 0.146
(0.0068)
0.050
(0.034)
Null
r=1
3 -91.5 0.658
(0.014)
--- 0.146
(0.0099)
0.031
(0.059)
Trend
r=10
4 -6880.4 0.754
(0.0095)
-4.57E-5
(???)
0.147
(0.0042)
0.032
(0.0025)
01 0 0
40
Model checking: null model r=10
Model slightly underestimates largest r=1 and r=2 quantiles
41
Model checking: trend model r=10
Including a time trend does not improve the r=1 and 2 fits
42
Summary Sufficiently large values of an independent
identically distributed variable can be described by a 3-parameter non-homogenous Poisson process;
This leads to simple parametric forms for the distribution of maxima and r-largest values (GEV) and exceedances above a high threshold (GPD);
MLE can be used to estimate the parameters (but estimates are often sensitive to individual values);
Non-stationarity can be accounted forby making model parameters systematic functions of covariates;
Spatial pooling can be used to obtain more precise estimates but covariates have to be included to avoid bias
43
Some outstanding questions … 1. What do extreme indices really tell us about extremes?
2. How best to develop well-specified extreme value models that account for non-stationarity (non-identical distributions) caused by natural and climate change processes?
3. How to deal with large sampling uncertainty due to the rarity of events and shortness of available observational records? Robust estimation in the presence of outlier events?
4. What can imperfect climate models tell us about real world extremes? How to bias correct model errors in extremes?
5. How to develop and test well-specified inferential frameworks for prediction and attribution of real world extremes from multi-model ensembles?
44
ReferencesStephenson, D.B. (2008): Chapter 1: Definition, diagnosis, and origin of extreme weather and climate events, In Climate Extremes and Society , Cambridge University Press, pp 348 pp.Definitions of what we mean by extreme, rare, severe and high-impact events
Ferro, C.A.T., D.B. Stephenson, and A. Hannachi, 2005: Simple non-parametric techniques for exploring changing probability distributions of weather, J. Climate, 18, 4344 4354. Attribution of changes in extremes to changes in bulk distribution
Beniston, M. and Stephenson, D.B. (2004): Extreme climatic events and their evolution under changing climatic conditions, Global and Planetary Change, 44, pp 1-9 Time-varying attribution of changes in heat wave extremes to changes in bulk distribution
Coelho, C.A.S., Ferro, C.A.T., Stephenson, D.B. and Steinskog, D.J. (2008): Methods for exploring spatial and temporal variability of extreme events in climate data, Journal of Climate, 21, pp 2072-2092GPD fits to gridded data including covariates. Spatial pooling and teleconnection methods.
Antoniadou, A., Besse, P., Fougeres, A.-L., Le Gall, C. and Stephenson, D.B. (2001): L Oscillation Atlantique Nord NAO: et son influence sur le climat europeen, Revue de Statistique Applique , XLIX (3), pp 39-60 One of the earliest papers to use climate covariates in EVT fits – NAO effect on CET extremes
Stuart Coles, An Introduction to Statistical Modeling of Extreme Values, Springer. Excellent overview of extreme value theory.
45
There are worse things than extreme climate …
e.g. extreme ironing!
Thanks for your attentiond.b.stephenson@exeter.ac.uk
46
Tubing Boulder Creek on Sunday?
Sunday noonwhitewatertubing.com
See me today if you are
interested.
47
Proposed taxonomy of atmospheric extremes
Rareweather/climate
events
Rare and Severe events
Rare and Non-Severe Events
Rare, Severe, Acute events
e.g. hurricane in New England
Rare, SevereChronic events
e.g. European blocking
Rare, Non-severe, Acute events
e.g. hurricane over the South Atlantic ocean
Rare, Non-severe, Chronic events
e.g. Atlantic blocking
Rarity
Severity
Rapidity
Acute: Having a rapid onset and following a short but severe course.Chronic: Lasting for a long period of time or marked by frequent recurrenceStephenson, D.B. (2008): Chapter 1: Definition, diagnosis, and origin of extreme weather and climate events,
In Climate Extremes and Society , R. Murnane and H. Diaz (Eds), Cambridge University Press, pp 348 pp.
top related