1

STATISTICS OF EXTREMES IN CLIMATOLOGY AND HYDROLOGY PART II: RECONCILING THEORY WITH OBSERVATIONS

Rick Katz Institute for Study of Society and Environment National Center for Atmospheric Research Boulder, CO USA

email: rwk@ucar.edu Home page: www.isse.ucar.edu/staff/katz/ Lecture: /staff/katz/docs/pdf/sc2epfl.pdf

Quote

Climate change undermines a basic assumption that historically has facilitated management of water supplies, demands, and risks.

Stationarity is dead: Whither water management Milley et al. (Science, 2008)

Outline

(1) Modern Perspective (2) Interpretation of Tail Behavior of Climate/Hydrologic Extremes (3) Unified Approach (Extremes/Non-Extremes) (4) Complex Extreme Climate/Hydrologic Events (5) Risk Communication under Climate Change

(1) Modern Perspective

Traditional Engineering Design -- No longer only need for extreme value theory Complex Extreme Events -- e. g., heat waves Unified Approaches -- Need to model both extreme and non-extremes values (e. g., generation of climate change scenarios)

Non-Stationarity -- Specter of climate change -- Lack of use of extreme value theory Trend detection (least squares or nonparametric) -- Risk communication Meaning of return level under non-stationarity?

(2) Interpretation of Tail Behavior of Climate/Hydrologic Extremes

Ultimate Extreme Value Theory -- GEV distribution as limiting distribution of maxima X1, X2, . . ., XT with common distribution function F MT = Max{ X1, X2, . . ., XT} Penultimate Extreme Value Theory -- Suppose F in domain of attraction of Gumbel type (i. e., = 0) -- Still preferable in nearly all cases to use GEV as approximate distribution for maxima (i. e., act as if 0)

-- Expression for shape parameter T Hazard rate (or failure rate): H(x) = F'(x) / [1 F(x)] Then one choice of shape parameter is: T = (1/H)' (x) |x=u(T) where the characteristic largest value u(T) = F 1(1 1/T) Here T 0 as block size T

Example: Exponential Distribution -- Exact exponential upper tail (unit scale parameter) 1 F(x) = exp( x ), x > 0 -- Shape parameter for penultimate approximation is Hazard rate: Shape parameter: H(x) = 1 , x > 0, T = 0, T = 1, 2, . . .

So no benefit to penultimate approximation

Example: Normal Distribution -- Fisher & Tippett (1928) proposed Weibull type of GEV as penultimate approximation F Normal distribution (with zero mean & unit variance) Hazard rate: H(x) x (for large x) u(T) (2 log T )1/2 (for large T)

Characteristic largest value:

Penultimate approximation is Weibull type with T 1 / (2 log T ) For instance: 30 0.15, 365 0.085

10

Example: Stretched Exponential Distribution -- Traditional form of Weibull distribution (Bounded below) Note: Weibull extremal type is reflected version 1 F(x) = exp( x c ), x > 0, c > 0 where c is shape parameter (unit scale parameter) -- Shape parameter for penultimate approximation is: T (1 c) / (c log T ) (i) Superexponential (c > 1) (ii) Subexponential (c < 1) T 0 as T T 0 as T

11

Apparent Upper Bound -- Wind speed Weibull type of GEV advocated as penultimate approximation (in engineering applications) -- Maximum possible hurricane intensity Estimate with possible trend due to global warming -- Thermostat hypothesis Hypothesized upper bound on sea surface temperature in tropical oceans (Implications for impact of global warming on coral reefs)

12

GPD Fit
estimated upper bound threshold 31 sst 28 0 29 30

1000

2000

3000 Time

4000

5000

6000

13

Apparent Heavy Tail -- Precipitation (i) Penultimate approximation Frchet type of GEV can be obtained with F stretched exponential distribution (Shape parameter c < 1) (ii) Physical argument Wilson & Toumi (2005) gave heuristic argument for universal shape parameter of c = 2/3 for stretched exponential distribution for extreme high precipitation

14

-- Simulation experiment Generate observations with stretched exponential distribution (with shape parameter c = 2/3) Use block size of T = 100 to simulate maxima M100 (Corresponds to daily precipitation occurrence rate about 27%, ignoring variation in number of wet days) Penultimate approximation: Should produce GEV shape parameter of 100 0.11 Fitted GEV distribution (40,000 replications): Obtained estimate of 100 0.12

15

16

-- Aggregation Issue Apparent decrease in shape parameter of GEV or GP distribution Stretched exponential should be capable of resolving (at least qualitatively) Simulation experiment: Sum of two independent stretched exponentials (each with c = 2/3) Use block size of T = 100 to simulate maxima M100 Fitted GEV dist. (40,000 replications): Estimate 100 0.065 But note that precipitation really a random sum

17

(3) Unified Approach (Extremes/Non-Extremes) Approaches to Modeling Entire Distribution -- Example of precipitation intensity (gamma distribution popular) (i) Replace gamma with more flexible distribution -- e. g., stretched exponential (ii) Mixture of distributions -- e. g., mixture of exponentials (iii) Hybrid approach -- e. g., gamma for low to moderate values GP for high values

18

(i) Replace Gamma with Stretched Exponential -- Threshold Selection for Stretched Exponential Extremely difficult: Should not be too low Should not be too high (unlike case of GP distribution) -- Try hourly (instead of daily) precipitation intensity?

19

(ii) Mixture of Distributions -- Mixture of two stretched exponentials Fit to all precipitation intensity, not just high values -- Example Mixture of exponential (i. e., c = 1) & stretched exponential (c = 2/3) c.d.f. F (x; w, 1 , 2 , c) = 1 (1 w) exp[ (x/1)] w exp[ (x/2)c ], x, 1 , 2 , c > 0, 0 < w < 1

20

21

(iii) Hybrid Approach -- Given gamma distribution (fit to all data) -- Replace with GP distribution above high threshold -- How to tie together two pdfs at threshold u? -- Adjust scale parameter of GP distribution = 1 / H(u) where H is hazard rate for gamma distribution -- Fort Collins July precipitation intensity, u = 0.5 in

22

23

Unified Treatment of Covariates -- Alternatives to focusing solely on extremes Trends: Estimate on basis of all data (extreme & non-extreme)? Seasonality: Remove seasonal cycle from all data before modeling extremes? Geophysical covariates: Alternative modeling procedure such as quantile regression?

24

(4) Complex Extreme Climate/Hydrologic Events

Climate/Hydrologic Extremes -- Many events have complex structure (e. g., spells) Definition of Heat wave / Hot spell -- Recall runs de-clustering algorithm -- More complex definition (e. g., multiple thresholds)? Lack of Use of Extremal Models -- e. g., paper by Meehl & Tebaldi (Science, 2004) Example (European heat wave, 2003)

25

26

27

Model clusters (instead of declustering) -- Rate of occurrence of clusters Modeled as Poisson process (rate parameter ) -- Distribution of cluster length? Geometric distribution for cluster length with mean 1/ -- Intensity of cluster Cluster maxima modeled as generalized Pareto (GP) distribution with shape parameter and scale parameter -- Model of dependence among excesses within cluster? Conditional GP distribution for temporal dependence of excesses within cluster

28

29

30

-- Let Y1, Y2, . . ., Yk denote excesses over threshold within given cluster / spell -- Model conditional distribution of Y2 given Y1 as GP distribution with scale parameter depending on Y1: e. g., (y) = 0 + 1 y, given Y1 = y > 0 Hold shape parameter constant (y) = y Similar model for conditional distribution of Y3 given Y2 (etc.) -- Drawbacks Need to identify link function (y) Unconditional distribution of Y2 no longer exactly GP

31

32

Conditional distribution of Y2 given Y1 = y -- Conditional mean [increases with (y)] E(Y2 Y1 = y) = (y) / (1 ), < 1 -- Conditional variance (increases with mean) Var(Y2 Y1 = y) = [E(Y2 Y1 = y)]2 / (1 2 ), < 1/2 -- Conditional quantile function F 1[p; , (y)] = [(y) / ] [(1 p) 1], 0 < p < 1 Increases more rapidly with (y) for higher p

33

Incorporation of non-stationarity / climate change -- Cluster rate Trend in mean of Poisson rate parameter (t), year t -- Cluster length Trend in mean of geometric distribution 1/(t), year t -- Cluster maxima (or first excess) Trend in scale parameter of GP distribution (t), year t

34

Procedure for fitting trends -- Cluster rate and cluster maxima (or first excess) Obtain from fitting trend in point process intensity (indirectly if use GEV parameterization) -- Cluster length Use generalized linear model (glm) (negative binomial family)

35

Hot spells to heat waves -- Merge clusters -- Minimum cluster length -- Higher threshold -- Cluster functionals Measures of heat wave intensity (e. g., mean or total excess) -- Trends in hot spell characteristics Convert to corresponding trends in heat wave characteristics

36

(5) Risk Communication Under Climate Change

Return Period / Return Level -- e. g, so-called 100-year flood Return level has 1% chance of being exceeded in given year Return period (or average recurrence interval) corresponding to this probability of 0.01 is 1/0.01 or 100 yr -- Intended to more effectively convey risk of rare event (Instead of simply stating probability) -- Common to misinterpret (Under unchanging climate / stationarity)

37

38

Example (Mercer Creek, WA) -- Period of rapid urbanization (starting about 1970) Effects on runoff within watershed of land-use changes Model for Non-Stationarity in Annual Peak Flow -- GEV distribution with time varying parameters Piecewise linear trends in & ln(): (i) Constant 1956 1970 (ii) Linear 1971- 1985 (iii) Constant 1986 2006

39

40

41

Interpretation of return level x(p) with return period T (under stationarity) Pr{X > x(p)} = p, where p = 1/T (i) Length of time T for which expected number of events = 1 1 = Expected no. events = T p, so T = 1/p (ii) Expected waiting time (assume temporal independence) Waiting time W has geometric distribution: Pr{W = k } = (1 p)k1p, k = 1, 2, . . ., E(W) = 1/p = T

42

One Approach (under non-stationarity) -- Effective return level (i. e., conditional quantiles) Permit return level to vary from one time period to next (Hold probability of occurrence constant) -- Like moving flood plain from one year to next Impractical for many long-term planning purposes

43

Alternative Approach (under non-stationarity) -- Retain one of two interpretations under stationarity Notation: Let pt (u) = Pr{Xt > u}, year t = 1, 2, . . . (i) Expected number of events [e. g., M. Nogaj, Ph. D. thesis] -- Let NT (u) denote number of events during t = 1, 2, . . ., T E[NT (u)] = p1 (u) + p2 (u) + + pT (u) Given specified return period T: Set E[NT (u)] = 1 & solve for return level u Then u satisfies p1 (u) + p2 (u) + + pT (u) = 1 In other words, average of T probabilities is 1/T

44

Frequency-Based Return Level (u 871 cfs for T = 50 yr)

45

(ii) Expected waiting time [e. g., Olsen et al., Risk Analysis, 1998] Assume temporal independence (not just stationarity) -- Let W(u) denote first time t that Xt > u Pr{W(u) = k } = { t=1,k1 [1 pt (u)] } pk (u), k = 1, 2, . . . Given specified return period T: Set E[W(u)] = T Solve for return level u Not possible to obtain analytical expression for E[W(u)] (or for threshold u )

46

Waiting Time-Based Return Level (u 871 cfs for T = 50 yr)

47

Return Levels for Mercer Creek Example -- Stationarity assumption Fit single GEV distribution: u 1036 cfs for 50-yr return level -- Frequency-based definition u 871 cfs for 50-yr return level -- Waiting time-based definition u 871 cfs for 50-yr return level Note: Two definitions would generally produce different return levels

48

Abandon Concept? -- Provide information (e. g., in form of table) on probability of one or more events (for given threshold) as function of length of time Decision-theoretic justification -- Examples on internet (under stationarity) Educate consumers about need to purchase flood insurance Tables of such probabilities (e. g., probability of 100-year flood within next 30 years)