Climate change, hydrodynamical models & extreme sea levels Adam Butler Janet Heffernan Jonathan Tawn Lancaster University Department of Mathematics & Statistics
The problem
Introduction Understanding the impacts of climate change upon extreme sea levels. l Understanding spatial variation in impacts. Use statistical ideas of spatial statistics and extreme value theory (Smith, 2002). l Attempt to build physically realistic models. l Applications: flood defence, offshore engineering, insurance…
Hydrodynamical models
POL models < 35km NEAC grid < 12km NISE grid V V
Observed climate inputs Run using observational climate data. Model run for period Run on NICE and NEAC grids. l Reasonable fit to observational data (Flather, 1987). Test for evidence for a linear temporal trend in extreme values.
Climate sensitivity Generate 30-year long sequences of model output under two hypothetical climate scenarios: 1“Current” CO 2 levels 2Double “current” CO 2 levels Sequences are stationary. Climate inputs generated using the ECHAM-4 climate model. We will construct parametric models. Interest is in comparing the parameter estimates obtained under the two scenarios.
Univariate extremes
The GEV distribution Blockwise maxima converge to a GEV (Generalized Extreme Value) distribution : is the shape parameter. l Conditions for convergence include: independence or weak dependence stationarity (Leadbetter, 1987).
Modelling extremes General ideas l Ignore distribution of original data. l Can model maxima using GEV distribution. l Alternative approaches to extremes exist: e.g. threshold methods (Coles, 2001). Application to the POL data l Model the annual maxima at each site. l Assume independence between sites.
Previous findings l Surge residuals l Changes (cm) in 50 year surge levels for the NISE grid. l Estimates exhibit spatial variability.
Spatial extremes
Multivariate extremes Componentwise maxima l Multivariate Extreme Value Distribution l Nonparametric or parametric estimation ? Parametric approaches Marginal and dependence characteristics. The Multivariate Logistic Distribution l Alternative parametric models (Tawn, ?) l Physically motivated subsets
Multivariate logistic distribution
Spatial extremes Assume smooth spatial variation in GEV parameters. This implies spatial coherence. Assume that observations at neighbouring sites are spatially dependent. Use a multivariate approach to extremes, with one dimension for each site. Benefits Improved efficiency in parameter estimation. l Interpretable estimates of spatial structure. l Allows extrapolation to ungauged sites. Enables regional-level estimates to be derived.
Current work
Marginal or joint estimation? Bivariate case, GEV margins, logistic dependence. l Three possible methods for estimation: joint likelihood (“joint”), product of marginal likelihoods (“marginal”) robust version of “marginal” approach Marginal approach results in reduced efficiency if there is dependence. How large is this effect ? Simulation study. l See: Shi, Smith & Coles (1992), Barao & Tawn (1999).
Conclusions Further work l “Bivariate efficiency” study l Comparison of approaches to spatial extremes l POL data: extremal trends l POL data: climate sensitivity l Scottish rainfall data…? Statistical significance l Application of modern extreme value methods in an applied context l High dimensionality
Questions ?