SR537 JOINSEA Probablility Waves Water HRWallingford

The Joint Probabilityof
Wavesand WaterLevels:
JOIN.SEA
A rigorousbut practicalnew
approach
ReportSR 537
November1998
(Re-issued
with minoramendments
May2000)
LANCASTE
UNIVERSIT r)
TheJoint Probabilityof Wavesand Water
Levels:JOIN-SEA
A rigorous but practicalnew approach
ReportSR 537
November1998
(Re-issuedwith minor amendmentsMay2000)
lnnwttingtora
AddA and RegisreredOm€ HR Wrlingrnrd Lrd. tlo*hery Park, Watlingtod, OXON OX10 EBA
Tel. +rt4 (lt) 1491835301 Fr* +44 ({l) l49l 83!231
F€grsr.r.d h Enghnd 6. 2s320€0,HRw.llifrglord i3 ! wholy dn.d .ubltdt.ry ot HRw.tlnqr6d c6up Lrd.

I***,"*.0 sR537 rW051X)
Contract
This reportdescribes
work fundedby the Ministryof Agriculture,Fisheriesald
FoodunderCommissionFD0202(predictingwavesat or nearthecoastline).
Publicationimpliesno endorsementby theMinistryof Agriculture,Fisheriesand
Foodof thereport'sconclusions
or recommendations.
The HR WallingfordJobNumberwasCCS14W.Thework at HR Wallingford

wascarriedout by Dr PeterHawkes,Mr BenGouldbyandMr Andy Yarde. The
work at Lancaster
Universitywascaniedout by Prof Jonathan
Tawnand
MissPaolaBonot. Mr MichaelOwenwaschairmanof the advisorygroupfor the
project.TheProjectManagerwasDr Hawkes.TheMinistry'sNominatedProject
OfficerwasoriginallyMr A C Polson,andthenMr J R Goudie.HR'sNominated
ProjectOfficerwasDr S W Huntlnston.
Prepared
by ....,P Hq,tr.keE
r'f*...1.,
(name)
?q*l tl
(Title)
Approvedby
(name)
?rq*or
(Title)
D","19.-IlU.Z.aes..
O Ministryof Agriculture,FisheriesandFood,Copyright2000
Er**,"*.0 tll
nf"-
WallirEford IV
Summary
The Joint Probability of Waves and Water l,evels: JOIN-SEA
A rigorous but practical new approach
Report SR 537
November 1998
(Re-issuedwith minor amendmentsMay 2000)
A common application ofjoint probability in coastalengineenngis to the

simultaneousoccurrenceofa large wave height and a high still water level, and its
consequenteffect on seadefences. If two or more variablesare either completely
independentor completely dependent,then the joint probability calculationsare
relatively easy:however,this is rarely a good approximationin practice. In the
past,the fitting and extrapolationofthe dependencefunction betweenlarge wave
heightsand high water levels has involved complicatedand./orsubjective
approachesunsuitablefor use by non-specialists.
We presenta new approachto the joint probability of large wave heightsand high
waterlevelswhich:
- removesmost of the subjectivity from the presentmethods;
- allows the distribution of wave period to be included in calculations,

by treating wave steepnessas a third variablepartially dependentupon
wave height;
- allows long-term simulation of the combinedwave height, wave

period and water level variables,in tum allowing more accurate
calculationof the effects on seadefences.
This researchwas funded by the Ministry of Agriculture, Fisheriesand Food, and

was undefiakenjointly by HR Wallingford and LancasterUniversity. The first
part of the report, written by HR Wallingford, summarisesthe new developments
and tleir rangeof applicability, and teststhem againsta number of field data sets.
The secondand longer part, written by LancasterUniversity, gives full technical
details ofthe new methodsand a more rigorous comparisonof altemative
approachesbasedon syntheticseastatedata. A shortercompanionreport
(HR, 2000) is intendedfor day-to-dayapplication as a user manualfor the
associatedJOIN-SEA joint probability software. A paper on the developmentand
testing of the new methodswas presentedat the MAFF Conferenceof River and
CoastalEngineers(Owen. 1997,1.
Availabilitvof Drosramsand reDorts

In principle, the methods,reportsand programsare freely available,sincethey
were developedentirely with MAFF funding. However, adoptionof the methods
and programsrequiressomefamiliarity with numerical modelling andjoint
probability methods,and there are costs in terms of training, follow-up advice,
software library licencesand copying of documentation(and of coursethe time
serieswave and water level data requiredas input). MAFF funded an initial
!"**n,.rro"
Summarycontinued
distributionof theprogramsto industryspecialists in the form of a
lor beta-testing,
briefingworkshopin February2000.at whichthe issuesrelatingto dissemination
wereclarified.
computerprograms
For furtherinformationon this reportandthe associated
pleasecontactDr PeterHawkesof the CoastalDepartment at HR Wallingford.
!Hn*" tngrora V1
Contents
Title page I
Contract iii
Summary
Content:t vii
Part I HR Wallingford Report
1. I n t r o d u c t i o .n. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . I. . . .
1.1 Background......... . . . . . . . . . . . . .I. . . .
1.2 Definitionsusedinjoint probabilityanalysis............... .................2
1.3 Dependence . . . . . . . . . . . . . . . . . . .3. . . . . .
1.4 Review of altemativemethods........ .................5
1.5 O u t l i n eo f t h en e w m e t h o d. . . . . . . . . . . . . . . . ...............1
I .6 R e s e a r cohu t p u t s . . . . . . . . . . , . . . . . ...............................7
2. T h en e wa p p r o a c h . . . . . . . . . . . . . ................8
2.1 I n p u td a t ar e q u i r e m e n.t.s. . . . . . . . . . . . . . . . ..................8
2.2 Fitting of statisticaldistributionsto the marginal variables...........9
2.3 Statisticalmodelsfor dependence .................. l0
2.4 L o n g - t e r ms i m u l a t i o n ..................... . . . . . . . . . . . . l.0. . .
2.5 Analysis ofjoint exceedanceextremesand structuralresponse
functions . . . . . . . . . . . . . . . . . . . . .l.l. . . . . .
2.6 S o f t w a r ei m p l e m e n t a t i o n . . . . . . . . . . . . . . . . . . . . . . . . . . . . .I .I. .
3. Applications...................... ..............l2
3.1 l3
Exampleapplicationsto syntheticdata........................................
3.2 Casestudiesusingfield data.... 16
.......................
3.3 Example applicationusing both inshoreand offshorewave data20
3.4 Issuesraiseddurins analvsisof otherdatasets...........
.................24
4. Discussion ...........26
4.1 Pastpracticefor useof resultsin design-..................................... 26
4.2 A d v a n t a g eosf t h en e w a p p r o a c h . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 1
4.? New possibilitiesbasedon long-termsimulation........................28
4.4 Recommended optionsforjoint probabilityanalysis..................29
4.5 Performance of the varioustechniques........................................31
4.6 Availability of methodsand stafftime involved......................... 33
4.'7 F u t u r ed e v e l o p m e n t s . . . . . . . . . . . . . . . . . . . . . . . . ...............34
5. References........................,. ............35
Tahles
Table I Direct hindcastingald extrapolationof overtoppingrate,
a p p l i e dt o d a t as e tS i m l . . . . . . . . . . . . . . . . . .................1. .5. . . .
Table 2 Direct hindcastingand extrapolationof overtoppingrate,
appliedtodatasetSimz................. .......................15
appliedto datasetSim3 ...........
16
Table 4 Direct hindcastingand extrapolationof oveltopping rate,
a p p l i e dt o d a t as e tS i m 4. . . . . . . . . . . . . . . . . .................1. .6. . . .
!"**,"o.* vl1 sR 53?04r'05/00

Contentscontinued
a p p l i e dt o d a t as e tS i m 5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1. .6. . .
Table 6 Comparisonbetweenovertoppingrates(mr/s/m)
for differentanalysismethodsat Christchurch.................................. 18
Table 7 Comparisonbetweenovertopping rates(mr/s/m)
s e t h o d sa t D o v e r . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .l .8. . . . . . . . . . .
f o r d i f f e r e nat n a l y s im
Table 8 Comparisonbetweenovertopping rates(mr/s/m)
for differentanalysismethodsat Dowsing...................---.--.-......-...... 19
Table 9 Comparison between overtopping rates (m'/s/m)
for differentanalysismethodsat Cardiff (0-l90"It[).......................... 19
Table l0 Comparisonbetweenovenoppingrates(m'/s/m)
for differentanalysismethodsat Cardiff (190-360"N)...................... 19
Table I 1 Comparison between overtopping rates (m'/s/m)
for differentanalysismethodsat North Wales................................... 20
Table l2 100yearovertoppingrates1ml/s/m)for Somerset
( n o l i m i t o n w a v es t e e p n e s s ) . . . ....-.........................22
Table 13 100 year armour size (m) for Somerset
( n o l i m i t o n w a v es t e € p n e s s ) . . . . . . . . . . . . ......................22
Table 14 Overtoppingrates(m'/s/m) for Somerset
(minimumwave steepness of 0.020) .....................23
Figures
Figure t Jointexceedance andjoint structuralprobabilities..............................
3
Figure 2 Probability density contoursshowing negativedependence
betweensurgeandwave height(off Hythe) ........................................4
Figure 3 Probability densitycontoursshowing positive dependence
betweensurgeandwaveheight(Christchurch 5
Bay)............................
Figure 4 Joint exceedancecontoursofH, (m) and water level (mOD)
for a sitewith very low dependence.. 13
....................
Figure 5 Joint exceedancecontoursofH, (m) and water level (mCD)
for a sitewith high dependence 14
............................
Appendices
Appendix I Comparisonsbetweenjoint exceedanceextremesfor different analysis
methods
Part II Lancaster University Report
!***"u,nuo* vlll
Part I HR Wallingford Report
lHnwat'nsfo.a
1. INTRODUCTION
1.1 Background
Meaning of ioint probabilitv
Joint probability typically refers to two or more partially relatedenvironmentalvariablesoccurring
simultaneouslyto producea responseof interest. Examplesare:
- Iargewave heightsand high water levels;
- large river flows and high sealevels;
- Iarge surgesand high astronomicaltidal levels.
Importanceof ioint nrobabilitv

Damageto seadefencesis often associatedwith times when large wave heightsand high water levels
occur simultaneously,even if neither one is exceptionalin its own right: henceit is necessaryto estimate
theirjoint probability. Thereare a rangeof techniquesavailable:from an intelligentchoiceofa singlewater
level to use with establishedwave conditions,to a rigorousjoint probability assessmentusing long time
seriesdata;from a generaloffshorestudywhich might be valid over a wide area,to a site-specificprediction
of overtoppingvolumes. MAFF has been funding researchon this topic for severalyears (HR, 1994;Coles
and Tawn, 1994;POL, 1997) and is increasinglyexpectingthe resultsto be applied in schemeassessment.
Predictionof ioint Drobabilitv

Comparedwith other types of oceanographicvariable,there is a large volume of high quality tide data
aroundthe UK. This, togetherwith the fact that spatialvariations in tide are smooth in some areassuch as
the eastcoastof Britain, meansthat extremewater levels can be predictedquite reliably for much of the
UK. Conversely,the lack of long-term wave measurementsin many cases,and the variability of wave
conditions in coastalwaters,meanthat prediction of exffeme wave conditions is more difficult and
uncertain. Meanwhile, estimatesof the level of serviceof UK coastaldefencescan be nearly as sensitive
to uncenaintiesabout the dependencebetweenlarge waves and high water levels as they are to
uncertaintiesaboutextremewave heights. HR (l994) includesan examplecalculationfor a location with a
small positive dependencebetweenwaves and water levels, and deep water at the structuretoe. The
assumptionof completedependencebetweenwave heightsand water levels could lead to a seawall crest
level up to two metrestoo high. Conversely,the assumptionof independencecould lead to a crest level
half a metre or so too low. It thereforeseemsreasonableto put about the sameamount of effort into
assessingthis dependenceas into assessingwave conditions.
The Start-upWorkshoo
In August 1994,MAFF hosteda Workshop on joint probability (MAFF, 1995) at which MessrsHawkes,
Owen and Tawn presentedtheir own preferredapproaches.As a result of the Workshop, MAFF agreedto
fund a two-yearjoint researchproject at HR Wallingford and LancasterUniversity. The intention was to
combine the best ofthe existing methods,and to develop,test and eventuallydisseminatea rigorous and
practical approachto the joint probability of large wave heights and high water levels.
The prpsell]L_Brqieal
MAFF provided funding for the developmentand testing of new joint probability methodsfrom April 1995
to March 1997,although in practicedevelopmentcontinueduntil September1998. Most of the
developmentand validation of the methodsand computerprogramswas carried out at Lancaster
University, but the processcontinuedwhen programcode was transferredto HR Wallingford in
April 1996. In principle, the computer programsand reportsare freely available,sincethey were
developedentirely with MAFF funding. However, in practicethere are additional costsassociatedwith
training, follow-up advice, NAG licencesand copying of documentation.
t"* Wallingford
1.2 Definitionsused in joint probabilityanalysis
Marginal probability and return period
Marginal probability refers to the distribution of a single variable, for example wave height or water level.
Return period refers to the averageperiod of time betweenoccurrencesof a panicular high value of that
variable. So, for examplethe | 00 year retum period still water level is the level equalledor exceeded
once, on average,in eachperiod of 100 years.
Record interval and event duration

Usually,in the contextof seadefences,only conditionsat high water are of interest,and typically peaksurges
and wave conditionspersistfor lessthan half a day. Therefore,eachhigh water (706 per year) can
convenientlybe taken as an independent'record', which is assumedto persistover the durationof high
water. Therefore,1 year and | 00 year retum period 'events', for example,haveprobabilitiesof occurence of
l/706 and 1/(100x706)and are assumedto persistover the durationof high water.
Joint structuralprobability and return pqtied

Joint probability refers to the chanceof two or more partially relatedvariables(x, y) occurring
simultaneously. The joint structural probability refers to the occurrenceof a particular response(such as
overtopping)which in turn dependson the joint occurrenceof those variables. The blue and red curves in
Figure I illustratethe typical shapeof contoursof equal response,with the shadedareasabove the curves
indicating the outcomeswhich give occurrenceof the two responses.Different types of responsemay
occupy different parts of the wave and water level distribution. In this example,the equal overtopping
curve lies towards the bottom right ofthe diagram where the water level is higher, whilst the equal force
curve lies towardsthe top left where the wave height is higher. If the responsevalue(s)is/are chosento
correspondto structuralfailure, then the probability (or risk) of failure is obtainedby summing the
probabilitiesof all the outcomesofthe responsewhich give structuralfailure, i.e. summing the
probabilitiesof (x = x6, y = y6) for all (x6,ys) in the appropriateshadedregion. Therefore,in determining
probabilitiesof structuralfailure in this way it is necessaryto estimatethe joint probability density
(x = xj, y = y0) for extremevaluesof (x0, y0) for the two variables. This may be either in the form of an
extrapolatedprobability distribution or a long-term simulation.
Joint exceedanceorobability and retum period

Joint exceedanceprobability combinationsof wave heightsand water levels with a given chanceof
occurence are defined in terms of seaconditions in which a given wave height is exceededat the same
time as a given water level being exceeded(x > xs andy > y0). The black curve in Figure 1 illustratesa
contour of equaljoint exceedanceprobability for wave heightsand water levels, with the brown points
indicating particular exampleswhich might be testedin design. The greenand yellow areasillustrate
rangesof wave height and water level with the given joint exceedanceprobability. Theseareas,and the
probability they represent,provide an approximationto the red and blue failure regions shown in Figure I
and the probabilitiesthey represent.
The discreoancybetweenjoint exceedance andjoint stxctural probabilities

In current practice, a number of combinations of waves and water levels are denved with a given joint
exceedance returnpenod (theseare represeniedby the brown dots in Figure l). Only one ofthese will be a
worst casein termsof structuralresponse,and it may not be the sarneonefor eachresponse(these'worst
cases' are representedby ringed brown dots in Figure 1). The probability of occurrenceof the structural
responsefunction (eg overtopping or force) calculated from the worst casecombination of wave height and
waterlevel will be higherthan thejoint exceedance probability. In other words,joint exceedance returnperiod
seaconditionswill tend to under-predictresponses if the responses afe assumedto havethe sameretum
period. This is becausethe same strucfural responsefunction value might be obtained by other seaconditions
in which only one or other of wave height and water level takes a very high value. This is illustrated in
Figure I by the difference between the green and red areasand between the blue and yellow areas. h cunent
practice, a small margin of safety is addedto the joint exceedanceprobability predictions to try to offset this
discrepancywith the retum periodof the response.The new method,conversely,works in termsof thejoint
south-€ast at Dover). In this situationtheremaybe a very low or evennegativedependence betweenhigh
surgesandlargewaveheights. However,this dependence (whichmaynot be particularlyhigh to beginwith)
is maskedin considerationof waveheightsandtota, waterlevels,because the astronomicaltidal componentis
unrelatedto the weatherconditions.This leadsto thebulk of waveandwaterlevel databeingvery modesdy
dependent(andperhapssuggestsa futureapproachin whichthedependence betweenwavesandsrurges is
assessed andextrapolated,beforeaddingin deterministictides). However,the mostextremeconditions
(usuallyincludinga high surgecomponent)will tendto be moredependent, particularlywherethe surgeto
tide ratio is larger.
Scatterdiagramsto illustratethe degre€of dependence

An easyway of demonstrating the apFoximatedegreeof dependence betweentwo variablesis by meansof a
scatterplot. Many pairsof valuesof VariablesA andB areshownaspointson a diagramwhich has
VariableA on its x-axisandVariableB on its y-axis. Well scatteredpointsareindicativeof low dependence.
Pointslying approximatelyon a line with a positiveslopeareindicativeof a strongpositivedependence.
Pointslying approximatelyon a line with a negativeslopeareindicativeof a strongnegativedependence.In
thecaseof wavesandwaterlevels,anydependence is usuallymodestandtypically showsitself asa slight
upward(or occasionallydownward)trendin the scatter.To illustratetheideaandto emphasise the
dependence, Figures2 and3 arescatterdiagramsof waveheightagainstsarge. Figure2 containsdatafor a
sitedemonstrating negativedependence betweenwaveheightandsurge,whilst Figure3 showsdatafor a site
with positivedependence.Differenttypesof dependence anddegreesof correlationarealsoillustratedin
Section3.3.1of the LancasterUniversitypartof this report.
SURGERESIDUAL LEVEL (M)

-1,25 -1.00 -0.?5 -0.50 -0.25 0.00 0.25 0.50 0.?5 1.00
-1.00 -0.?s -0,50 -0,25 0.00 0.25 0.50 0.75 1,00 1.25
NO- IN
(METRES) ROl4
4.00- 4.50 0 0
3,50- 4.00 0 3
3.00- 3.50 0 lz
2,50- :.00 0 11
2.00- 2.50 0 248
1.50- 2.00 0 588
1.00- 1.50 0 1 79 5
0.50- 1.00 I 4071
0,00- 0-50 ) 5 1 00
No DATA/CALMS O 1 O 1212 211 11 I0 0 C o n t o u or sf

n o .o f e v e n t s
NO. rN coLUMN 2 16 l0 s09 5254 5320 581 724 22 2
1
10
100
1 00 0
RECORDSANALYSED FROM DAY 1 OF MONTH 2 OF L9'17 PO DAY 31 OF MONTH 3 0F 1990

IIUMBER OF RECORDS ANALYSED 72449
Figure 2 Probability density contours shorying negative dependence between surge and wave
height (off Hythe)
!"^*",'*"* sR 537o,t/ottx)
SURGERESIDUAI LE"VEL {M)
-1-00 -0.?0 -0.40 -0.10 0.20 0.50 0.80 1,10
-0.?0 -0,40 -0.10 0.20 0.50 0-80 L-10 1-40

HS cALltSl NO. rN
(METRES ) NO DAT ROW
5 . 5 0- 7 . 0 0 o2
5.00- 6.50 02
s . 5 0- 6 . 0 0 02
5 . 0 0- 5 . 5 0 0'7
4.50- 5,00 026
4 . 0 0- { . 5 0 0 6'7
3 . 5 0 - 4 , 00 19 0 113
3.00- 3.50 30 0 161
2 . 5 0- 3 . 0 0 33 0 405
2-OO- 2 -50 27 0 628
1 . 5 0 - 2 . 00 17 0 833 C o n t o u rosf
1 .0 0 - 1 , 5 0 13 0 1504
0 . 5 0- 1 . 0 0 15 no.of events
0.00- 0.50 0 4258 1
NO DATA/CAI,MS 10
100
NO. IN COIJUMN 10 1850 7223 1669 1000
RECORDS ANALYSED FROM DAY 1 OF MONTH 1 OF 1978 To DAY 31 OF MONflj 3 0F 1990

NUT.{BEROF RECORDS ANAI,YSED 11790
Figure 3 Probability density contours showing positive dependencebetweensurge and wave

height (Christchurch Bay)
1.4 Reviewof alternativemethods

Scooeof the review
Thereareseveraljointprobabilityanalysismethodsavailable,dependingupontime,budgetandanount of
field dataavailable,andupontheendpurposeof the calculations.Evenwhenit is not consideredexplicitly,
andthe structuralresponseis primarily dependent uponthe incidentwaveconditions,a representativehigh
waterlevel will oftenbe requiredfor calculations.In principle,anynumberof partially dependentvariables
canbe analysed,althoughin practiceboth the calculationsandthe resultsbecometoo unwieldy whenmore
thanthreeareused. Thereareseveralmethodsof presentingthe results,dependingupon the endpurposeand
the enduser. Most of the assessment methodsarecompatiblewith mostof the presentationmethods.A
moredetailedreview andtestingof the earliermethodsis given in HR (1994).
Independentanddependentcases
The dependentcaseis trivial, assumingthat the statisticsof wavesaloneandof waterlevelsaloneareknown,
sinceone only hasto combinewavesandwaterlevelseachwith the samematginalprobabilitiesasthe
response.The independentcaseis easywhenworkingwithjoint exceedance probability(which providesan
approximationto the responsewith the sameprobability)asit is just the productof the two marginal
probabilities.As the independentanddependent casesaresimpleto calculate,it maybe worth deriving them
earlyin anyproject. Treatingtheseasthe 'mostoptimistic' and 'most pessimistic'scenarios,respectively,
mayhelp in judging the valueof anymoredetailedjoint probabilityanalysis.
Inhritivejoint probabilit]'assessment
The simplestmethodfor assessment of dependence is an intuitive onebasedon generalexperienceandthe
shapeandsizeof the seaareaaroundthe predictionpoint. The assessment may concludethat thereis a
modestdependence, andfor example,that high wavesandhigh waterlevelswith a 100yearjoint exceedance
retum periodare 10 to 100timesmorelikely to occurtogetherthanthe assumptionof indeoendence would
suggest.It may concludethat thereis a strongdependence, andfor exarnple,that high wavesandhigh watEr
levelswith a 100yearjoint exceedance returnperiodareonly 1Oto 100timeslesslikely to occurthanthe
assumptionof dependence would suggest.(Theprobabilityratio betweendreindependentanddependent
g* Wallingford
casesis 706 x 100.) This is the basisof ajoint probability methoddescnbed in the BeachnuTtxagenwnt
manual (C}F.IA, 1996, Section3.5.3).
Full analvsisof dependenceandjoint orobabilities

A betterapproach,if constraintson time, budgetand data permit, is to examineseveralyearsof simultaneous
wave and water level datain order to assessdependenceand to derivejoint probability extremes. This allows
the analysisto be performedin a similar objectivescientificmannerto that which would be appliedto the
separatewave and water level predictions. The dependenceis determined by analysis of pairs of values of
wave height and water level at eachhigh water over a period of severalyears. Before JOIN-SEA became
available,extremeswere determinedby extrapolatingwave heightsfor successivelyrarer water levels,and
water levelsfor successivelyrarer wave heights. By combiningtheseextrapolations,contourswere
consfixctedjoining combinationsof wave height and water level with equaljoint exceedanceretum periods.
The normal way of incorporatingwave direction was to undertakea seriesof 'conditional' analyses,one for
eachdirection sectorof interest. The normal way of incorporatingwave period was to assumethat it is
completelydependentupon wave height,and could be determinedfrom an assumedwave steepness.
Extrapolation of ioint orobabilitvdensitv

Joint probability densitycan be visualisedas a three-dimensionalhistogran; in which the vertical axis shows
the likelihood of occurrenceof combinationsof the variablesshownon the two horizontalaxes. If a bi-
variateprobability distributionmodel can be fitted to the observedjoint density,then the fitted distribution
can be extrapolate.dto extremes. In the past, the main difficulty with this approach was in representationof
the dependencebetween the variables in the fitted distribution (although as stated earlier, the independent and
dependentcaseswould be trivial). An altemativeway of calculatingextremedensities,which hasbeenused
in the past,involvesnumericaldifferentiationbetweenpreviouslycalculatedcontoursofjoint exceedance
retum period (seeprevious paragraph). Extrapolation of probability density is perhapsthe most flexible
methodfor presentationof extremespredictions,and permitsa direct estimationof the probability of a given
responseor structurevariablesuchasfailure due to high overtopping. This conceptis a fundamentalaspect
of the JOIN-SEA method.
Direct hindcastingof designvanables

An altemativeapproachcan be usedwherea very specificlocal result is required,for example,run-up or
overtoppingof a particularshoft stretchof seadefence. Site-specificshallow water wave datadirectly at the
point of interest,togetherwith sequentialwater level data,could be usedin a site-specificrun-upor
overtoppingformula to hindcastlong-termrun-up or overtoppingconditions. The resultingsingle variable
datacould tlen be extrapolateddirectly to extremes,avoidingthe complicationsof a truejoint probability
approachin which wavesand water levels are kept separatethroughout. This singlevariableapproachhas
the advantageofusing all the availableinformationon individual wave conditions,for exampleincluding the
wave period, which is importantfor run-up and overtopping. However, it relies on the extrapolatedvariable
being of the sameform as within the body of the distribution. It would not be applicable,for example,where
wave overtopping is expected to give way to weir overtopping or to structural failure at higher levels. The
advantagesof this approachare availablewithin the JOIN-SEA analysismethod,with the addedadvantage
that it can be appliedto a long-termsimulationof wave and water level data.
Offshoreand inshoreextremes
Joint probability extremescan be calculatedand presentedeither offshoreor inshore. Offshoreresultsare
more generallyapplicableover a largerarea,but may needto be transformedinshorebeforefurther use.
Inshore results are more site-specific, and can take better account of any hydraulic interactions, but may be
applicableonly in one smalI area. Wave predictionsandjoint probabilitiesare often calculatedas a function
ofdirection. This is important,sincegeneralexposureto waves,the relationshipbetweenhigh wavesand
high water levels,and transformationinshore,may all be dependentupon stormdirection.
Depth-limitedconditions
In coastalengineeringapplications,wave heightsmay be depthJimited. For example,theremay be only a
few metres of water at the toe of a structure even in extreme conditions. so that the full force of the waves
!***,"oo.o
may not impact directly upon the structure. In extremecasesofthese circumstances,only the highestwater
levels need to be considered. A possible general approachfor depth limitation is to determine the extreme
water levels, and then to check the probability that the depthJimited wave height could occur at the same
time, However,it is worth noting in this situationthat swell waves,which do not usually havethe largest
wave heights,may be a worsecasefor structuraldesignthan depth limited wind waves.
Presentationof results
Joint exceedanceresultsare normally calculatedand presentedas 'contours' of combinationsof wave heights
and water levelswith givenjoint exceedanceretum periods,from which a 'worst case' can be determinedas
and when required for any particular situation. The sameinformation might be shown in the form of a table
of pairsof valuesof wave height and water level (drawn from the 'contours'). This approachwould allow
'worst
the case' to be identified in terms of run-up (or other designvariables)at the site of interest,but
usuallyrequiresthat severalpotential 'worst cases'be tested. Altematively, resultscan be presentedas
extrapolatedprobability densities,in which 'contours' indicatethe likelihood of occurrenceof particular
combinationsof wave heightsand water levels.This approachis useful for risk analysis,wherethe total risk
is found by integratingprobabilitiesover a wide rangeof wave conditionsand water levels in which 'failure'
rnay occur. Either results format can be converted to other relevant variables, for example run-up,
overtoppingor force on a coastalstructure.Pleasenote that althoughthe joint densityapproachcan provide a
direct estimation of the retum period of the response,tbejoint exceedanceexfemes contours can provide
only an approximationto this probability.
1.5 Outlineof the new method

Main elements
The new methodconsistsofthe five main elementslistedbelow and describedin more detailin
Sectrons2.1-2.5.
1 . Preparationof input data"consistingof many independentrecordsofwave height,wave period and water

level.
Fitting of statisticaldistributionsseparatelyto the wave heights,the water levelsandthe wave steepnesses,
3 . Fitting the dependencebetween wave heights and water levels, and between wave heights and
steepnesses.
4 . Simulation of a large sample of wave height, wave period and water level data, using the fitted
distributions.
5. Extremes analysis of a range of responsevariables basedon the simulated data.
Main develonments
The main developmentsof the new method and its apptications are in elements 2 to 5. It is more objective,
pafiicularly in its modelling of dependence,than most existing approachesto joint probability assessment,and
allows wave period (or wave steepness) to be includedas a variable. With carefuluseits long-tem simulation
approachcan deliver more accuratepredictions ofjoint extreme values and structural responsefunction values
(eg overtopping,run-up,force) than methodsalreadyin commonuse. It will also assistin a gradualmove
towardsrisk-baseddesignof seadefences.
1.6 Researchoutputs
Products
Therearethreemaintypesof outputfrom this researchproject,namelyrcports,computerprogramsand
workshops,plusof courseimprovedjoint probabilityanalysisin consultancy
studies.
!***,,'**o sR 5-170.1/0tr)l)
JOIN-SEA software
The software consistsof FORTRAN computer programs, incorporating numerical statistical subroutines
(in NAG). The programs are research-basedand have not been testedto normal commercial standards,but are
available,initially on a trial basis,to appropriateusers.
JOIN-SEA reports
The first part of this document gives a brief description of the new approachand its application in coastal
engineering. The accompanying staiistical methodology and further casestudies are given in the technical
reportwhich forms the secondpart of this document.A companionuser manual(HR, 2000)describinginputs,
outputs and options available to the user is intended for day-to-day use with the new programs.
Workshoosand dissemination
A specialistworkshopwas held at HR Wallingford on 3 December1998,at which this reportand the
associateduser manual were releasedin draft form. The workshop was attendedby those directly involved in
the researchand by a number of invited coastalengineersfamiliar with joint probability applications. It was
hoped that the workshop discussion would guide future dissemination, usageand continued development of
joint probability methods,and ultimately MAFF policy in regard to their use. A subsequentbriefing workshop
was held at HR Wallingford on 4 February 2000. This was attendedby about a dozen industry specialists
interestedin receiving copies of the programs and training in their use. It is hoped that the subsequentbeta-
testingprogamme will yield usefulfeedbackon continueddevelopmentand usageofthe methods.
2. THENEWAPPROACH
2.1 Input data requirements
Prepara[qro:f]hetinputdata
Thefirst of thefive stagesinvolvespreparation of theinputdata.Generation andpre-processing of theinpur
waveandwaterlevel datais probablythe mosttime-consuming part of thejoint probabilityanalysis
procedure.Eachinput recordconsistsof a waveheight,a waveperiodanda waterlevel (or altemativelya
surge)preferably usingidenticalmeasurement or predictionlocationsfor bothwavesandwaterlevels.The
datacancomefrom measurements or hindcasts,but for eachrecordthe valuesshouldrepresentconditionsat a
particularpoint andtime, A convenientway of satisfyingtherequirementfor the recordsto be both
temporallyindependent, andrelevant,is to useonly thoserecordsrepresenting conditionsat the peakof each
tidalcycle(ie onerecordevery12or 13hours).At leastthreeyearsof data"notnecessarily continuous but
representativeof thetype of seastatesof interest,areneededtojustify theeffort involvedin applyingthe new
approach.
Stagesin datapre-pr@qa$sirg
Typicallythefollowingpre-processing
of the datawouldbe carriedout:
- combining of the separatewave ard waier level data into a single sequentialdata file;
- exfacdon and retention of only one record, closest to high water, per tidal cyclei
- optionally,division of the datainto two or threeseparatepopulations(datasets)corresponding

(for example)to differentwave directionsectomor seasons,or to wind-seaand swell;
- removal of obviouslyfaultydata,and
(optionally)recordswith zerowaveheightandrecordswith very low

wavesteepness (althoughbothof theseareautomatically
trappedby the
programssincetheycauseproblemsin the lateranalysisandlong-term
simulationof wavesteepness);
!nnw"rrrgrora sR 53?o4nvll)
- running of the preliminary diagnosticprogram TESTDATAto ^ssessthe type of dependence
and whether independenceor dependencewould be adequateassumptions.
The TESTDATA diagnostic proeram

It is assumedthat userswill maketheir own arrangements for any pre-processing considerednecessary:in
terms of the computer programs required, only the TESTDATA progran forms part of the JOIN-SEA package.
The purposeof this progmm,which can be run immediatelybeforethejoint probabilityanalysis,is to provide
guidance on the degreeof dependencewithin the data, and whether it changeswith exceedancethreshold. For
example, it may indicate that complete dependenceor complete independencewould be a reasonable
assumption,or that dependenceneedsto be modelled carefully. The test statistic, T (z), is derived from the
conditional probability of the secondvariable exceedingthe marginal T year level for that variable given that
the first variable exceedsits own marginal T year retum level. Here z is related to T by z = -lilog(l - 1/706T)
and a rangeof T valuesareconsidered.It is definedin Section3.3.2of the LancasierUniversity part of this
report,with examplesand adviceon interpretationofthe resultsbeing given in Section7.1.2. The rate of
changeof T (z) with log (z) shouldbe (l-p) / (1+p), wherep is the correlationcoefficientfor that threshold. A
constantrate of changeis indicativeof a constantlevel of dependence, with a constantT (z) valueof zero
indicatrng complete dependenceand a gradient of one indicating independence. A varying rate of change
of T (z) with log (z) would indicatethat dependence vaneswith threshold.
2.2 Fittingof statisticaldistributionsto the marginalvariables

Statisticalmodelsfor the marginal variables
The secondstageinvolves the fitting of statisticalmodelsto wave heights,water levels and wave
steepnesses.GeneralisedParetoDistributions are fitted to the top few percentof the primary marginal
variables,ie wave heights and water levels; and the distribution of wave steepnessis modelledthrough a
combination of its empirical distribution and a regressionrelationship with high wave heights, For eachof
the three marginal distributions,a threshold,specifiedin terms of the proportion of data lessthan, is used
for fitting. Below the threshold,the distribution is representedempirically, and abovethe thresholdby the
fitted distribution. Details are given in Chapters3 and 4 of the LancasterUniversity part of this report and
a brief summaryis given below.
The GeneralisedPareto Distribution (GPD)

The GPD is describedin Section 3.1 of the LancasterUniversity part of this report. Although in practice the
GPD is fitted only above a chosenthreshold, a feature of the distribution is that (if it is a good fit to the data) it
is invariant to the threshold. Another feature of the distribution is that, unlike the Weibull distribution for
example,it doesnot necessarilyincreaseroughly in proportionto the log ofthe retum period. Insteadit can
alsoeitherincreasemore rapidly in the tail or it can level off towardsan absolutemaximum value. Numerical
testscarriedout using severalof the projectdatasetssuggestthat extremespredictionsare not sensitiveto the
thresholdchosen,andthat 0.95 is usuallya reasonablevalueto use,so that the GPD is fitted to the top 5% of
the data.
Importanceof wave period

In some situationswave period can be as important as wave height in calculatingeffects at the coastsuch
as overtoppingor armour damage,especiallywhen wave heightsare depth-limited at a seawall. Although
good information on extremewater levels and extremewave heights will usually be obtarnedor derived
during a coastalmodelling study, the marginal distribution of wave period is rarely consideredbeyond the
estimationof a representativewave steepness.However, in the JOIN-SEA and structurevariable methods
the distribution of wave periods is built in tlrough respectivelya model for the distribution of wave
steepnessand empirically.
Incorporation of impleyeclplqeligliellfunarginal extremes

Joint probability analysisis basedon simultaneousinformation on the variablesof interest. It is quite
likely that there will be additional non-simultaneousdata on at least one ofthe variables,with which to
!***u.oo.o sR53?&/05/00
refine the extremespredictionsfor that one variable. For example,there may be 20 years of water level
data but only l0 yearsof wave data; accurateextremevaluesfor one or other variable may have already
beenestablished(eg from spatialanalysisof water levels (POL, 1997)); or there may be anecdotal
evidenceof severeseaconditions outsidethe period ofthe measurements.Good joint probability analysis
usesthis additional information. This might involve modification of the parametersof the fitted
distribution(s)or scaling of the predictedextremesto achievebetter agleementwith the refined marginal
predictions. The presentmethod incorporatesany refinementsby scaling during the long-term simulation
of data,thus permanentlybuilding this information into the synthesisedseastatedata to be usedin
subsequentstructuralanalysis.
2.3 Statisticalmodelsfor dependence

The modelling procedure
The third stageinvolves conversionto Normal scales,and fitting of a dependencefunction to the bulk of
the wave height and water level data. Simple diagnostictestshave beendevelopedto assesswhetherfull
dependenceor independencemodelsare adequateapproximations. For situationswhen these
simplifications cannotbe made,two altemativepartial dependencestatisticalmodelshave been developed
to representthe dependencebetweenwave heightsand water levels. Theseconsistof a single Bi-Variate
Normal (BVN) Distribution and a mixture of two BVN'.s. Thesemodels were chosen,sincethe
dependenceand extremescharacteristicsof the BVN are well understood,and togetherthesetwo models
are consideredto provide a sufficiently flexible family of models
Selectionof statisticalmodel
The choice betweenone and two BVNt is determinedby the relative goodnessoffit to the data, which can
be assessedwith referenceto the varying degreeof correlation,expressedas a function of exceedance
level- The single BVN may be adequatefor a location at which all the wave conditions belong to a single
population,althoughfrequently the correlationwill increaserapidly towards the tail and the mixture model
will be needed. However, where tle wave conditions belong to more than one population, for example
wind-seaand swell, the mixture model is always likely to be neededto capturethe different dependences
in the two populations.
User input
In this and in the previous stagesthe user retainssome control over the process,primarily by having the
choice of selecteddependencemodel, and secondarilyby being able to selectboth the thresholdsabove
which the fitting will be applied,and the staning valuesfor optimisation of the fits: this is assistedby
referenceto diagnosticsto assessthe fits.
2.4 Long-termsimulation
The simulation procedure
The fourth stageinvolves simulation of a large sample of synthetic records of I{", T* and water level, basedon
the fitted distributions, and with the same statistical characieristics as the input data. At this stage it is
possibleto re-scalethe marginal extremedistributionsto more establishedvalues,if they are known. For
example,concurrentwave and water level data may be availablefor a period often years,and by
extrapolationof this data the marginal distributionsof the two variablesare calculatedwithin the joint
probability software. However, if either or both of the data setsextendsbeyond the concurrentperiod of
ten years,it is preferableto usethe additional information containedwithin the extra data. Thereforethe
marginaldistribution can be derived from the longer data set and incorporatedinto the data simulation
processthrough the use of the re-scalingoption. This is achievedby altering individual affected wave
heightsand water levels from one scaleto the other betweensimulation from the fitted distnbutions and
writing permanentlyto an output file. Thus, basedon all availabledata,thousandsof years worth of sea
conditions can be simulatedwith fitted distributions,extremesand dependencesfor wave height, water
level and wave period. This provides greaterflexibility in the subsequentanalysisofthe seastatedata.
!"**rroo,.o l0 sR5tt 04/r)5/{11)

The simulation product
Output from this stagedetailsthe number of eventsper year, the number of recordsin the file, as well as
the individual seastaterecordsand the marginal extremes. Checkscan be made on the return period
marginal conditions that are within the time spanof the input data. For example,if the input data consisted
of a time spanof five years,the I year marginal extremeshould be approximatelyequal to the fifth highest
record in the input file.
2.5 Analysisof joint exceedanceextremesand structuralresponsefunctions

Analysis of the simulated data
The fifth stageinvolves analysis of the large simulated sample of data to produce extreme values for use rn
designand assessment of seadefences.Thesecan take the form of extremewave heights(and associated
periods), extreme water levels, or extreme combinations of the two. In addition, any structural response
function(eg overtopping,run-up,force) which can be definedin termsofconstants(eg wall slope,toe depth,
crest elevation etc) and variables H,, T. and water level, can be synthesiseddirectly for every record in the
simulated data sample. Direct analysis of the distribution and extremesof the structural responsevariable is
then relativelyeasy.
'count-back'
The extremesanalysis method
Rather than fitting probability distributions to the synthesiseddata, extreme values are estimatedfrom the
appropriateempincal exceedanceprobability in the synthesiseddata. This is achieved by 'counting back'
through the highest values within the simulated data, a method based on the literal definition of retum period
and bestdescribedby meansof an example. If, for a 2000 year simulation,the 100year retum period value
(ie the level equalledor exceededon averageonceevery hundredyears)is required,then this is given by
'counting
back' to the 2000/100= 20th highestvalue, which is then assigneda returnperiod of 100years.
Joint exceedanceextremesare detemined in a similar way, for exampleby 'countingback' throughthe
highestwave heights,but only consideringthoserecordsabovea certainthresholdof water level. Thrs
intuitive approachis reliablefor retum periodsup to aboutone quarterof the simulationlength. Henceit is
necessaryto synthesiseat leastfour timesas much dataas the highestretum period of interest,and in practice
ten times would be a more rypical ratio.
Direct analysisof structuralresponsevariables

At present,as well as marginal andjoint exceedanceextremes,four simplified structuralresponsevariables
are included in the computerprogramsfor demonstrationpurposes,for which more details of the formuiae
usedand the input variablesare given in the user manual (HR, 2000):
- overtoppingrate on a smooth slope;

- run-up on a smooth slope;
- fbrce on a vertical wall;
- armour size for a seawall.
Analvsis of more comBlgl_ylli4blgs

More complex variables, for example the simultaneousoccurrenceofa high force and a high overtopping ratre,
which would havebeendiffrcult to assesspreviously,can now be routinely studied. Sensitivityand altemative
designscan also be assessedrelatively easily, by making appropriate changesto the structural response
tunction(s) analysedwithout the need for addrtronalstatistical analysis of dependenceor marginal variables or
a new long-termsimulation.
2,6 Softwareimplementation
Introduction
This section briefly describesthe software packageand the functions of the five main programs for the joint
probability analysis: BVN, MIX, SIMBVN, SIMMIX and ANAZYSIS. For a more detailed description and
instructionson the use of the programs,a companionmanual(HR, 2000) is available. In normal usein
!"**r,*'.0 11 sR 53? 04/{t5ll)t)

consultancy studies,the user would need to write some additional programsto carry out the necessary
pre-processingof the input data, and to compuie actual site-specific structural responsevariable(s),
Fittine of statisticalmodels(BVN andMIX

Two programs are usedto fit statistical models to samplesof data, one basedon a single Bi-Variate Normal
distribution (BVA) and one basedon a mixture of two Bi-Variate Normal's (MIX). Both programs also fit the
marginal distributions of wave height, of water level, and of wave steepness,each above a threshold selected
by the user. Input datato the two programs,in fie dntafile, consistsof a largenumberof recordsof wave
height, water level and wave period. Output from BVN consistsof three files'. rhovalues, diagrnstic ancl
transfer. Rhovalues anddiagnostic provide information on the fitting of the distributions and the conelation at
different probability thresholds,whilst rrander is usedto input this information into the subseqaentSIMBVN
program. The next step is to run either MIX or SIMBVN, dependenton the information on correlation detailed
in the diagnostic and rhovalues frles. Essentially M/X performs the sametask as BVN but will provide a better
fit to data where there is a large variation in correlation with probability threshold. Output from MIX consists
of a diagnostic {ie and a transfer file that perform the samefunction as the output from ByN.
Simulation of long data samoles(SIMBVN and S/MMID

Two programs, SIMBVN and SIMMIX, areusedto synthesisemuch longer samplesof data basedon the two
aliemative fitted distributions. S/MBVN is run subsequentto ByN, andSIMMIX is run subsequentto M1X.
Although the two programsperform the samefunction, the inptt transfer fies (output from ByN or M/X)
differ. An option of re-scalingthe marginalextremes,if more establishedvaluesareknown, is availablein
both programs. The output files (r imdata) from SIMBVN andSIMMIX arc identical in form and consist of a
largenumberof individual wave height,water level and wave periodrecords. 'Importancesampling'can be
applied at this stage,such that only the higher records are retained, so as to reduce the eventual size of the
simdata fiIe.
Extremesfrom the simulateddata (ANALISIS)

The last program, ANALYSrc,is usedto derive rnarginalextremes,joint exceedanceextremesand extreme
valuesof demonstrationstructurevariable(s)from the simd.atafrle. Thresholdwave heights are specified
for which extremewater levels are derived, and vice versa,thusjoint exceedancecombinationsare
obtained, For the structurevariablesit is necessaryto specify parameterssuch as toe depth, wall slope and
crest level for individual structures.
User inpul
The ptocedureis largely automated,but requiresquile substantialcomputertime, memory and storage
capacity. Although there is a large number of questionsput to the user during operationof the programs,
the majority can be answeredby acceptingthe default values. The user should be preparedto check for
mixed populationsor faulty recordsin the original data,and to re-run parts of the procedureif indicatedto
do so by the diagnosticinformation. It is sometimesnecessaryto take a view on the importanceofone,
two or three outliers,that is recordsof severeseastateswith apparentlyhigh dependencebetweenwave
height and water level in an otherwiseuncorrelateddata set. As in any joint probability analysismethod,
the eventualresultscan be quite sensitiveto the importancegiven to theserecords,and it may be worth
checkingthe validity of the individual recordsconcemed. More specific advice is given in the JOIN-SEA
user manual (HR, 2000).
3. APPLICATIONS
The test data sets
During validationand comparisonof resultswith altemativejoint probability methodsseveraltestdatasets
were assembled. Some were basedon mea-suredand hindcast wave and water level data - these had the
advantagesof not coming from pre-determinedstatistical distributions and of being representativeof data sets
to be usedin consultancystudies;somewere synthesisedusing specifiedprobabilitydistributionsloosely
basedon measureddata - these had the advantaeethat the 'true' extremesand ioint orobabilities were known.
t"- Wallingford t2
3.1 Exampleapplications to synthetic data
The simulateddatasets
Five setsof waveand water level datawith known statisticaldistributions,but broadlyrepresentativeof
actualconditionsat different points aroundthe coastof EnglandandWales,were synthesised.Details of
the derivationof thesedatasetsaregiven in Chapter5 of the LancasterUniversity part of this report.
Comparisons madebv Lancaster University

Combinationsof waveheightsand waterlevels with givenjoint exceedance retum priods werecalculated
using HR Wallingford's existingJOINPROBsoftware(HR, 1994)andalso usingthe new methods. The
resultswerecomparedwith the 'true'joint exceedance extremesfor returnperiodsof 10 years(equalto the
lengthof the original dataset)and 1000years. Two of the comparisonsareshownin Figures4 and5
(which aretypical of the wider rangeof comparisonsgiven in Chapters8 and 9 of the LancasterUniversity
part of this report). In thesedia$ams thejoint exceedanceextremespredictionshavebeenadjustedto give
marginalextremesin agreementwith targetvalues,so asto highlight the modelling andextrapolationof
'true') values' The
dependence.In both figures,the 'simulation Model' lines representthe target(or
'EstimatedModel' lines showresultsfrom the new method,with marginalextremespredictionsscaledto
give exact agreementwith targetvalues. The 'HR Estimates'lines showresultsfrom HR Wallingford's
JOINPROBmethod,herewith marginalextremespredictionsscaledto agreewith earter slightly different
targetmarginalextremes.As expected,both methodsperfr:rmwell for a joint returnperiod equalto the
length of the dataset,but the new methodgivesmorefobustpredictionsat muchhigher returnperiods.
(o
- EstimatedModel
--.- Simulation
Model
--- HR Estimates
f<o
C{
0123456
WL
Figure 4 Joint exceeilancecontours of H, (m) and water level (mOD) for a site with very low
dependence
t"* \4hllingford t3 sR53? 0:V05Xr

N
tr)
.!4 0
=+;+J.
r- - - StiCIeiBglt%#l
t-tH Es males
4 5 6 7 .8 9
WL
Figure 5 Joint exceedancecontours ofH" (m) and water level (mCD) for a site with high
dependence
Someconclusionsfrom the LancasterUniversity comparisons

Overall,the comparisonspresentedin the LancasterUniversity part of this report suggestthat the new
approach,working directly in tems ofthe structuralresponsefunction, is moreconsistentandreliable than
altemativemethods. The importanceof waveperiod is shownin severalways:the useof waveperiods
whicharetoo low (perhapsdueto usinga constantwavesteepness whichis too high) do€snot affectthe
joint exceedance extremesdirectly but it doesleadto an underpredictionof overtoppingrate;the useof a
variablewave steepness, ratherthana constantwave steepness, tendsto increaseovertoppingpredictions.
The return periodsof structuralresponsevariablescalculated(ashasusuallybeendonein the past)from
joint exceedance extremestendto be abouthalf the sizeofthe retumperiodsofthejoint extremes
themselves.The new long-termsimulationapproachallows a direct and accurateassessment of the 'risk'
with a given return period,without resortto the fairly arbitrary small marginsof conservatismcurrently
usedin consultancy studies.
Additionalcomoarisons usinqdirecthindcasting
Onefurther analysismethodwas appliedto the simulateddatasetsaftercompletionof the Lancaster
University (LU) part of the report,to illustrate an analysismethodusedsatisfactorilyby HR Wallingford
severaltimes in the past. The remainderof this sectioneffectively forms a post-scriptto LU's Chapter9.
It involves direct hindcastingand exhapolationof overtopping,aiplied in the way that hasbeenusedin
previousHR consultancy studies.This is comparable with LU's 'SVM appliedto Q.', referringto the
StructureVariable Method appliedto overtoppingrate,which suggestedthat the rnethodwasflawed due to
its greatsensitivity to threshold,but that the SVM appliedto log (eJ wasreasonable.
["**,roo* ).4 sR 53?04/05/(x)

Extremesanalysisusing direct hindcastins
For each of the simulateddata sets(Sim1-Sim5) in hrm, eachrecord of wave height, water level and wave
steepnesswas converted to an equivalent rate of overtopping using the same formulae and parameter
values as describedin LU's Chapter 9 (althoughobviously the vast majority of the valueswere zero). For
eachdata set an empirical distribution of overtoppingrate (Q") was assembledfor extrapolationto extreme
valuesusing HR's standardmethods. This involved fitting a threeparameterWeibull distribution (see
LU's Section 4.1.2), not to individual overtoppingrecordsbut to equally spacedthresholdsof overtopping.
In other words, tle information usedin fitting the Weibull probability model consistedof, for example, the
percentages of datalessthan Qc (mlVm) = 0.005,0.010,0.015...,or0.010,0.015... Severaldifferent
thresholdswere testedfor eachdata set, and the predicted l, 10 and 100 year overtopping rateswere
comparedwith the 'true' values(not availablefor the 1 year retum period) from the simulation model (see
LU's Chapter9).
Resultsand conclusionsfrom direct hindcasting

The extreme overtopping rate predictionsgiven in the Tables 1-5 below do not show much sensitivity to
thresholdand are quite close to target values. To put theseresultsinto the contextof the comparisonsmade
in Chapter 9 of the Lancaster University part of the repolt, take averagepredictions for each column in Tables
1-5 and convert to equivalent 'true' retum periods. The accuracy of the direct hindcasting approachis
comparable with that of the JOINPROB analysis: a little better than JOINPROB at the 10 year retum period
(within the data) but less reliable at the 100 year retum period. This suggeststhat HR's current
implementation of the structure variable method is sound in this application, although some care is required in
preparation of the data and checking of the results, particularly for data setscontaining just one or two high
overtopping results. Given the simplicity of the approach,and the fact that it is not subject to the uncertainties
associatedwith the fitting and exnapolation of dependence,the method could be retained for use in prediction
of specific responsevariables where severalyears of input data are available.
Table I Direct hindcasting and extrapolation of overtopping rate, applied to data set Siml
Threshold(ie the PredictedextremeovenoDpingrateO" (mYs/m)

proportion of data I yearreturn 10yearretum 100yearretum
ignoredin fitting the
distribution)
0.00000 0.00064 0.0059 0.034
0.99800 0.00063 0.0065 0.021
0.99860 0.00082 0.0065 0.030
0.99914 0.00103 0.0063 0.031
'Tme'
values Not available 0.0075 0.030
Table 2 Direct hindcastingand extrapolationof overtoppingrate, appliedto data set Sim2
Threshold(ie the Predicted extreme overtopping rate Q" (m'is/m)

proponion of data '100
1 year rehrm 10 yearretum year retum
ignored in fitting the
distribution)
0.00000 0.56 l.z-1
0.97240 0.61 1.12 t.67
0.98800 0.61 t.12 r.63
0.99390 0.61 1 . 1I 1.63
0.99643 0.62 1 . 1I 1.57
0.99814 0.67 1.08 1.45
0.99829 0.71 1.07 1.38
'True' values Not available l.l) t.87
t"* Wallingford 15 sR53?lM^)tu)

Threshold(ie the Predictedextremeovertopoinqrate O. (mls/m)

proportion of data 1 yearretum 10 yearretum 100yearretum
distribution)
0.00000 0.62 r.77 4.O
0.97910 0.64 1.75 3.5
0.99130 0.63 r;77 -r.6
0.99510 0.62 1.78 3.9
o.99670 0.63 r.79 4.4
o.99't57 0.63 l;79 4.5
o.99829 0.54 1.81 3.8
'True' values Not available r.4l z.o
Threshold(ie the Predicted extreme oveftoppins rate O" (m'/Vm)

proportion of data 1 yearretum 10yearretum 100yearre m
ignoredin fitting the
distribution)
0.00000 J_O 9.2 18.8
0.97900 4.1 /.o tt.2
0.99110 4.1 7.5 11.0
0.99370 4.r 1.5 10.8
'J /1
0.99600 4.2 10.5
0.99730 4.3 7.3 10.1
0.99800 4.) 7.3 9.8
'Tme'
values Not available tl.7 27.2
Table 5 Direct hindcastingand extrapolationof overtoppingrate, appliedto data set Sims
Threshold(ie the rate O. ( m'/sim)

PredictedextremeovertoDDinq
proportionof data I yearretum 10 yearretum 100yearretum
distribution)
0.00000 2.5 6.7 1.4.1
0.98120 3.0 5.7 8.5
0.99130 3.0 5.7 8.5
0.99390 3.0 J./ 8.4
0.99-560 3.0 5.7 8.4
0.99714 3.0 5.6 8.2
0.99'711 3.0 5.6 8.2
'Tme'
values Not available 8.3 19.5
3.2 Casestudies using field data

The field data sets
The compzrisons between methods describedin the LancasterUniversity part of this report are basedon five
simulated data sets,loosely basedon five setsof field data from different parts of England and Wales, namely
Christchurch,Dover, Dowsing, Cardiff and North Wales. This sectionof the reportdescribesequivalent
compaLrativeanalysesbetween JOINPROB and JOIN-SEA using the original long time seriesfield data sets
underlying the synthetic data sets. (The word 'field' is perhapsa little misleading since in each casethe wave
!Hnw"rr*sro.a l6 SR53?l).Vl'sX)
data came from a numerical hindcasting model, but the sequentialwind and water level data are from actual
measurements.) Details of the data setsare given in Table 5.1 of the Lancaster University part of the report.
Two data popUlationsfor Cardiff

The 2S-yearCardiff data set containedtwo distinct populationsof wave conditions: swell waves with an
offshore directlon between 190 and 360"N and locally generated waves with an offshote direction between
0 and 190"N. The two populationswere separatedat the data preparationstageand were treatedas two
separate28-yeardata setsduring the subsequentanalysiseven though eachcontainedonly about half as
many recordsas the original combined set. The direction-dependentmarginal andjoint extremeswere
then defined such that there will be one extremeoccurrence,on average,in eachdirection sectorin each
return period.
The structuralresponsefunction
For the purposesof this study the samestructuralresponsefunction was used as in the Lancaster
University comparisons, namely sea wall overtopping. Idealised seawalls were devised and the
overtopping dischargewas calculatedusing the formulae of HR (1980) at eachof the sites. Further details
of the formulae and the seawall parameterscan be found in Table 9.1 of the LancasterUniversity part of
this report. However, it is not so much the valuesof overtoppingcalculatedthat are of interest,more the
comparisonbetweenthe different methods. Tables 6-11 give overloppingdischargeratesusin-gthe
different analysismethodsdescribedbelow. All dischargesin the tablesbelow are given in m'7s/m,but
pleasenote that the idealisedseawalls usedin the testsdo not correspondto the actual seadefencesat any
of the five sites.
JOI{PROB analvsis
The JOINPROB combinationscolumn in Tables 6-l l refers to the existing HR analysismethod whereby
different combinationsof wave height and water level, with the samejoint exceedanceretum period, are
extractedfrom a j oint exceedanceprobability contour plot. Thesecombinationsare then applied to the
structuralresponsefunction and the 'wofst case' is found. The wave steepnessis calculatedfor the top few
percentof wave conditions and is assumedto have a constantvalue, which is then applied to the extreme
conditions to calculatethe wave oeriod.
JOIN-SEA analysis
The new joint probability methodswere applied in threedifferent ways in order to test different aspectsof
the procedure. Two involved the joint exceedancecombinationsapproachfor comparisonwith eadier
methodsand one involved direct simulation of the struchrralresponsefunction. The mean wave period
(T) for the JOIN-SEA analysiswas calculatedin three ways:
I Assuming the samesteepness(HR steepness)as usedfor the JOINPROB calculationsin assigning

a wave period to eachwave height and water level combination.
Usingthe averagesteepnessnotedby theJOIN-SEAmodelfor thetop 5Voof waveheights(not

necessarily
the sameasthe HR steepness,whichwasdeterminedmanually)in assigningwave
periods.
duringthe simulationofthe long time seriesdataset,as

Allowing a naturalvariationin steepness
in the normaloperationof JOIN-SEA.
Empirical extremes
Column 6 in eachof Tables 6-11 gives e.,ttremevaluesderived directly from the original data. Each
record, in terms of significant wave height, meanwave period and water level, was convertedto an
equivalentrate of overtopping. The I year return period extremevalue was determinedfor eachlocation
using the 'count-back' method describedin Section2.5 (so, for example,in a 28 year data set, the required
value would be the twenty-eighthlargestovertoppingvalue). This is perhapsthe most accurateexlreme
E***u"*.0
value estimatein eachtable. The 10 year, and in somecasesthe 20 year retum period values,were also
estirated from the data, but theseare subjectto greatuncertainty.
ComparisonbetweenJOINPROB and JOIN-SEA joint exceedanceextremes

Method 1 allows a direct comparisonbetweenthe traditional seastatecombinations(JOINPROB) method
and the new JOIN-SEA method,applied so as to obtain designseastatesdefined by specific wave heights
and water levels. (NB: the calculationof overtoppingusing the HR steepnessin the JOIN-SEA model is
for comparisonpurposesonly and is not a routine practice). If both methodsworked perfectly then the
resultsin Columns 2 and 3 ofthe tablesshould be identical. Discrepanciesare due to differencesin
marginal extremespredictionsas well as to the representationof dependence.
Comparisonbetweenjoint exceedanceextremesand simulation aBple@hg!

JOIN-SEA analysismetlods, numbered2 and 3 above,give a direct comparisonbetweenthe traditional
method of applying design seastatecombinations,and the new method of structuralresponsesimulation.
Differences here highlight the discrepancybetwe€nestimatingextremevaluesof structuralresponsefrom
joint exceedanceprobabilities (without applying the usual small margin of conservatism)and from joint
structuralnrobabilities.
Table 6 Comparison betweenovertopping rates (m3/s/m)for different analysismethods at

Christchurch
Returnperiod JOINPROB JOIN-SEA JOIN-SEA JOIN-SEA Empirical

(yearsl joint Jolnt joint structural (directly from
exceedance exceedance exceedance response original data)
(HR steepness (HR steepness (model simulation
of 0.06) of 0.06) steepness
of 0.053)
1 't.45
l.l3 t.23 1.81 1.68
10 2.29 LOI 3.07 3.55 2.75
20 2.63 3.40 3.88 4.20 Not available
100 Not available 4.85 5.49
200 5.31 5.75
Table 7 Comparison betweenovertopping rates (m3/s/m)for different analysismethods at Dover

(years) joinr joint joint structural (direct\ from
of 0.06) of 0.06) srcepness
of 0.045)
1 0.17 0.20 0.35 0.53 0.46
l0 0.51 0.48 0.75 0.99 1.1+0.4
20 0.67 0.56 0.86 1.14 Not available
200 1.86 1.96
E"* Wallingford l8 sR5J70.V05/{x)

Table 8 Comparison betweenovertopping rates (m3/Vm)for different analysis methodsat
Dowsing
Retum period JOINPROB JOIN-SEA JOIN-SEA JOIN-SEA Empirical

(years) Jornt joint Jornt structural (directly from
of 0.053)
1 | .14 0.99 1.19 2.20 2.6
l0 3.40 3.35 3.83 5.76 7.3+2.2
20 4.85 4.26 4.84 7.19 Not available
200 9.84 12.5
Table 9 Comparisonbetwe€novertoppingrates(m3/s/m)for different analysismetlods at

Cardiff (0-190"N)
Return period JOINPROB JOIN-SEA JOIN.SEA JOIN-SEA Empirical

(years) joint joint joint structuml (directlyfrom
exceedance exceedance exceedance response originaldata)
of 0.05) of 0.05) sreepness
of 0.06)
0.000003 0.000006 0.000003 0.00005 0.000015
t0 0.001I 0.00033 0.00019 0.0013 0.0013
20 0.0067 0.00075 0.00045 0.0037 0.002iil.0005
100 Not available 0.0026 0.011 Not available
200 0.0083 0.017
Table 10 Comparisonbetweenovertoppingrates(m3/s/m)for different analysismethodsat

Cardiff (190-3601V)

(yearsJ joint joint joint structural (directlyfrom
of 0.039)
0.0001 0.00025 0.00000008 0.000008 0.00014
l0 0.016 0.0081 0.00006 0.00089 o.oo?7
20 0.049 0.016 0.00018 0.0023 0.0056+0.0012
100 Not available 0.0032 0.024 Not available
200 o.oo77 0.089
!"**u,n*'. 19
Table 11 Comparison betweenovertopping rates (m'/Vm) for different analysismethods at North
Wales

(yearsl joint joint joint structural (directlyfrom
of 0.047)
I 0.26 0.38 o.57 1.04 0.82
t0 1.26 1.43 1.86 2.59 l.55
20 2.04 1.89 2.44 Not available
200 4.56 5.28
Discussionof the direct comparisonbetweenJOIMROB and JOIN-SEA

The comparisonin columns 2 and 3 of Tables 6-l l, betweenJOINPROB (HR steepness)and JOIN-SEA
(HR steepness)using thejoint exceedancemethod of determiningthe worst caseovertoppingevent and
assumingthe samevalue of wave st€epness,showsgood agreementat all sites. This suggeststhat the
method of marginal extrapolationat varying thresholds,for both variables,to constructcontoursof equal
joint exceedanceretum period, gives similar resultsto those derived from the extrapolatedjoint probability
density obtainedfrom JOIN-SEA. This is bome out by plots in Appendix I showing contoursof equal
joint exceedanceretum period (inespectiveof wave steepness)for eachof the test sites,derived by the
JOINPROB and JOIN-SEA methodsrespectively.
Discussionof the sensitivitv to assumedfixed wave steepness

Comparisonofthe two JOIN-SEA combinationscolumns gives insight into the potential errors o{
manually calculating the mean steepnessfor the top 57oof wave conditions,as opposedto the more
rigorous calculation of steepnesscarried out routinely in the model. The resultsshow the difference in
dischargeratesto be slightly higher (no more than a factor of two, in terms of return period) using the
model steepness,for all sites,with the exceptionof Cardiff. The differenceof the results at Cardiff is
probably explained by the separationof wave conditionsby direction as opposedto steepness.Some
longer wave period (lower steepness)wave conditions were presentin the 0-190"N (higher steepness
locally generatedwave conditions) data, and vice versa.
Discussionof differencesbetweenthejoint exceedanceextremesandjoint nrobabilit], densitv approaches

The simulation of the structuralresponsefunction gives ovedopping dischargesthat are a factor of 2-3
times greater,in terms ofthe retum period, than thosecalculatedusing the combinationsof marginal
values(JOIN-SEA (model steepness)).This is as previously expectedand is explainedby the fact that the
joint exceedancemethod doesnot accountfor the entire range of combinationsthat can contribute to the
strxctural responsevalue for a given retum period.
Comoarisonwith empirical extremes

With the exception of the very low overtoppingrates(lessthan 0.002m'/Vm) predictedfor some retum
periods for the Cardiff data set,the JOIN-SEA resultslisted in Columns 4 and 5 ofthe tables agreewell
with the empirical values. The discrepancyin the low overtoppingratesfor Cardiff is thought to be due to
the difficulty in representingthe distribution of wave period correctly, as can be seenin the spreadof
results acrosseachof the rows in Tables 9 and 10.
3.3 Exampleapplicationusing both inshoreand otfshorewavedata

Reasonsfor choosinsthe Somersetsiae
In addition to comparing results from different analysis methods,it is interesting to compare results from
i) offshore and nearshore(ie after wave ffansformation modelling) input wave data, and ii) depth-limited and
!"**",r"o.0 20
deep water conditions at the structure. To illustrate the differences an example study was carried out for a
hypotheticalsite on the Somersetcoast. This locationwas chosenfor a numberof reasons:it has a high tidal
range; it is exposed to a mixhrre of locally generatedwaves and waves arriving from the Atlantic; wave
refraction effects cause a significant increasein T* as the waves prcpagatefrom offshore to nearshore;there is
increasing dependencebetween wave heights and water levels as wave height increases;and 28 years of
simultaneousmeasuredwater level data and hindcast wave data were available.
The input wave and water level data

Input water level data were derived from houdy measurementsat Avonmouth over a period of 28 years,
transfbrmed for use at the study site. Input wave data wa-sderived from a site-specific wave hindcasting model
driven by hourly wind mea-surements at Cardiff Airport over the sameperiod of 28 years. After
pre-processingof the data as describedin Section 2. I , about 20,000 records of I{", T," and high water level
remained as input to the joint probability analysis. Two altemative t)?es of wave data were used:
- offshore wave data representativeof conditions severalkilometres away from the coast in
about 20m water denth:
- nearshorewave data (after transformationusing a wave refraction model) just outsidethe

breaker zoneat hieh water level.
The analysis methodg

The mixhrre of locally generatedand swell waves at the site suggestedthat the mixture of BVN's would be a
more appropriate probability model than the single BVN. Therefore the mixture model was run to produce
bothjoint exceedance(JE) probabilityseastatesand directpredictionsof the structurevariable(SV). The
BVN model was also mn for comparison purposes,but was used only for direct predictions of the structure
variable. HR's existingJOINPROB methodwasusedto producejoint exceedance probabilityseastates.To
summarise,four altemative j oint probability analyseswere used:
- JOINPROB_JE analysis(HR, 1994) as used in consultancystudiesin recentyears - this

producesmultiple combinationsof wave condition (with a fixed wave steepness)and
water level, with a given joint exceedanceprobability, which can then be convertedto
structuralresponsevariable predictionsi
- MD(_JE analysisusing two fitted BVN'S - i) producing multiple combinationsof wave

condition (again with a fixed wave steepnessfor comparisonwith JOINPROB_JE) and
water level, with a givenjoint exceedanceprobability, which can then be convertedto
structuralresponsevariable predictions;
- MIX-SV analysisusing two fitted BVN's - ii) working directly in terms of the structural
responsevariable, and now using a variable wave steepnessderived from the orjginal data;
- BVN_SV analysisusing a single BVN - working directly in terms ofthe structural

responsevariable,and using a variablewave steepnessderived from the onginal data.
The coastal structures

Two hypothetical structureswere tested:
a plain seawall with a crest level 2,8m aboveMIIWS and deepwater at the toe of the
wall:
a plain seawall with a crestlevel 2.8m aboveMHWS and a toe elevation2.7m below
MHWS.
!"**n,"ro.o 21 SR53?0,U05/00
The structural responsevariables
Two structural responsevariables were tested:
- overtopping rate using the SWALLOW equationsin HR (1980) for a wall slope of l:4;
- armour size using equations(5.44) to (5.46) in CIRIA/CUR (1991) for a wall slope of l:2.
Note on wave steeDness

For illustrative purposes,no upper and lower cut-offs were applied to wave steepnessduring the initial
long-term simulation used to derive the results shown in Tables 12 and 13 for a rehrm period of 100 yean. For
the offshorc wave data, this made little difference. However, for the nearshorewave data, which contained a
wider range of lower wave steepnesses,this had the effect of producing a small number of quite high waves
with much longer mean periods than in the original sample. For this example application, the longer period
waves demonstratethe importance of wave penod and the sensitivity of overtopping rate to uncertainties
therein. Il a subsequentsimulation,a lower limit of 0.020was imposedon wave ste€pness, as would be done
in a consultancy study. The new extreme oveftopping predictions, for a wider range of retum periods, are
given in Table 14,
The effect of correlation coefficient

For illustrativepurposes,two additionalseriesof BVN_SV resultsare shownin Table 14. Whilst the 'Best'
sub-column contains overtopping rate predictions for the best estimateof correlation direcdy from the data, the
'P = 'P
0' and = I' sub-columnsshow equivalentresultsassumingindependence andcompletedependence,
respectively.
Table 12 100 year overtopping rates (m3/s/m) for Somerset (no limit on wave steepness)
Wave data Seawall JOINPROB-JE MD(-JE MIX SV BVN SV

Offshore Deep at toe 1.10 0.72 0.85 0.69
(steepnessof
0.050for JE) Depthlimited 0.46 0.30 0.42 0.3r
at toe
Nearshorc Deep at toe 1.34 1.08 9.0 10.7
(steepnessof
0.029for JE) Depthlimited 1.10 0.83 8.9
at toe
Table 13 100year armour size(m) for Somerset(no limit on wavesteepness)
Wave data Seawall JOINPROBJE MIX JE MD( SV BVN-SV

Offshore Deep at toe t.27 1.27 1.n 1.31
(steepnessof
0.050for JE) Depth limited 0.82 0.82 0.83 0.82
at toe
Nearshore Deep at toe r.45 1.45 1.43 1.31

(steepnessof
0.029for JE) Depth limited 0.90 0.90 0.96 0.94
at toe
E"o*"u'*"'.0 22 sR5t7 04 rtrx)

Table 14 Overtopping rates (m'/Vm) for Somerset(minimum wave steepnessof 0.020)
Wave data Sea wall MIX_JE MIX SV B\,'}I-SV

Returnperiodsare1, 10,20, 100and200years
Best estimate of p P=0 Best P=1

Offshore (steepness Deepat toe.-\- 0.03 0.10 0.06 0.09 0.41
of0.050 for JE)
0.23 0.37 o.2r 0 . 3 1 0.99
0.34 0.50 o.z5 0.38 1 . 2 1
0.72 0.85 o.47 0.69 1 . 8 1
0.95 1.04 0.58 0.85 2.(n
Depth limited at toe 0.03 0.05 0.03 0.05 0.27
0.15 0.19 0.l0 0 . 1 6 0.57
0.21 0.25 0.13 o.2l 0-6s
0.30 o.42 o.22 0 . 3 1 0.87
0.41 0.48 o.27 0.37 0.98
Nearshore Deep at toe 0.04 0.15 0 . 1 1 0 . 1 4 0.68
(steepnessof0.029
for JE) 0.35 0.73 0.54 0.67 2.M
o.5l 1.01 o.77 0.90 2.45
1O8 1.72 1.44 1 . 6 1 3.50
t.25 2.25 r.)cr t ; 7 I 4.31
Depth limited at toc 0.04 0.11 0.07 0.09 0.67
0.31 0.53 0.34 0.48 1.96
0.41 0.73 o.46 0.65 z.z2
0.83 | .zo 0.76 0.99 2.78
0.91 1.78 0.84 2.94
Discussionof the Somersetcaseshrdv

All of the methods appearedto work satisfactorily, but there are some interesting differences. Roughly the
samearmour siz€ (which is only a little dependentupon wave period and not directly dependenton water
level) was calculated whatever method was used. However, overtopping, which is dependenton H,, T. and
water level, was quite dependenton which method and which sourceof wave data was used. As with the
synthetic data analysedin the Lancaster University part of this report, slightly higher values were derived
directly from the structure variable than from the joint exceedanceextremes. As expected,higher overtopping
was predicted usin gthe nearshore wave data,becauseof the increasein T- in the approachesto the coast, and
slightly lower armour sizes and overtopping rates were calculated in the casesof waves being depth-limited at
the wall, compared to those with unlimited depth. The existing methodsgave slightly higher overtopping rates
than the new method, where both worked with ioint exceedanceextremesand an assumedconstant wave
steeDness.
E* Whllingford 5R5-r7ov0t/(Xr
Lower llmit on wave steeDness
The importance of wave period was dramatically demonstrated,in the form of much increasedovertopping
rates for the struchlre variable calculations using nearshoredata where wave steepnesswas allowed to vary
over a wide range about its mean value, basedon the variability seenin the original data. To some extent this
is a genuine effect, reflecting the existenceof swell waves at the site, but it is over-emphasisedin the results
seenin Table 12. However, in practice this would be corrected by setting a lower limit on wave steepnessin
the synthesiseddata as seenin the much improved results in Table 14. The variability of the results illustrates
the care that should be taken in selecting input data and methods of analysis and interpretation.
Sensitivitv to correlation coefficient

The three altemative BVN_SV results shown in the three sub-columnson the right hand side of Table 14
illustratethe sensitivityto conelationcoefficient. The lowestpredictionsare for'p = 0', correspondingto
independencebetween waves and water levels, and the highest are for 'p = 1', corresponding to complete
dependence,with the best estimatesbeing only a little higher tian for independence. The independent
predictions fof retum periods of l0 and 200 yearscorrespondroughly to best estimate predictions for 5 and
50 year retum periods, respectively. A factor of four difference in retum period at this site correspondsto a
difference of about a quarter of a metre in still water level, and perhapsa half a metre difference in seadefence
level (allowing for the larger waves which could reach the wall at a higher water level. Conversely, the
dependentpredictions for retum periods of 1 and 20 years correspondroughly to best estirate predictions for
10-50 and 800-3000 year retum periods, respectively. This degreeof overprediction for the design retum
period at this site would conespond to a difference of about two thirds of a metre of water level and perhapsa
metre or so in seadefence level.
3,4 lssues raisedduring analysisof other data sets

The new software has alreadybeen applied on a number of consultancystudies. This sectiongives an
insight into some of the experiencegainedin applying the new methods.
Tees Bay
The overtopping hazard was assessedat three different locationsaroundthe TeesEstuary. Wave
conditions were calculatedat the three locationsusins numerical models. Waler level data from North
Shields was convefted to the sites of interestusing faitors basedon differencesin the tidal range. The
converslonofthe water level data to the TeesEstuaryraisedseveralimportant issuesregardingwind and
wave set-up.
Wind set-upis an increasein water level that occursover a large areadue to the direct effect of wind stress
on the water surface. Wave set-upis a much more localisedeffect that causesincreasedwater levels in the
surf zone as waves break. It was thought that wind set-upwould havebeen presentin the measuredwater
level data at North Shields,as it is a widespreadphenomenon. However, the extent to which it should be
included in the converted water level at the TeesEstuary was unclear. Wave set-upwas thought to be
excludedfrom the measurements,sincetide gaugesare generallyplacedwith a view to excluding wave
set-up.
As a result of this assessmenttidal flow modelling was carried out to assessthe extent of wind set-upat
Tees Bay in relation to North Shields. An empirical formula was derived that allowed the water level data
at Tees Bay to include the effects of wind set-up;additionally an allowancefor wave set-upwas made. As
it transpired,the inclusion of theseprocesseshad little effect on the joint probability extremesas the wind
was always from the west or south-west(ie off the land) at the time of the maximum measuredwater
levels. However, there was a small positive corelation betweenlarge wave heightsand overall water
levels in the modified data,which was usedin the subsequentanalysis.
Lyme Bay
JOIN-SEA was also used at a location in Lyme Bay where flood risk was to be assesse.d
over an area
comprising of severaldifferent seadefencestructures. The JOIN-SEA method of exffapolatingthe joint
!"**r,r"",0 u sR 5t7 0t/05^xl

probability density and simulating a very large data set was beneficialhere in treating the different risk
elementsin a consistentway. As in the Tees Bay study, considerableeffort was concentratedon
accuratelyassessingthe water levels at the site of interest(POL, 1997,which provides extremewater level
estimatesfor the whole of the UK was not availableat the time of the study).
The nearestmeasuredwater levels to Lyme Bay (with sufficient length of data) are at Devonport, which is
a considerabledistanceaway. The tidal range changessignificantly along this stretchof coastline,causing
uncertainty in the interpolationof water level data to Lyme Bay. To overcomethis problem, extremesat
Devonport and Weymouth (to the eastof Lyme Bay) were derived, and also variationsin the spring range
considered,before calculating factors to be appliedto the measuredDevonport data. The converteddata
was then compafed to severalmonths of measuredwater level data at the site of interest,which showed
good agreement. However, this study highlighted the problemsthat can arisein 'moving' measure.dwaler
levels along a stretchof coastlinewhere the tidal range varies significantly, and reinforced the preference
for having measureddata directly at the site of interest.
An interestingpoint to make in passingis that it is not strictly necessaryto 'move' the time seriesdata,
either on waves or water levels, from the measurementor prediction point to the point of application. It
would be possibleto determinethe dependencebetweenlarge wavesand high water levels using data for a
nearby location and then, as in the intuitive approach(seeSection 1.4) to apply it at the new point of
interest. However, it would still be necessaryto determinethe marginal extremesat the new point.
Korea
A consultancyproject off the west coastof Korea provided a good oppomrnity to use the new software in
environmentalconditions that were considerablydifferent to UK conditions. The wave and water level
data were split into two seasons,summerand winter, ald consideredas two separatecases,since there is a
distinct difference in climate betweenthe two. The highestwater levels were expectedto occur in summer
and the highest wave conditions were expectedto occur in winter. The analysisfor eachof the two
different data setsrevealeda negativedependencebetweenwave heightsand water levels. This was
thought to be feasible,if unexpected. However, as a conservativemeasurefor the design calculationsthe
relationshipbetweenthe two variableswas taken to be independent.
This study highlighted the requirementfor a relatively long simultaneousdata set. As the original three or
four years of data was split into two populations,the lenglh ofthe time serieswas effectively halved, thus
creating a relatively short period of data on which to basethe fitted distributions. There was thereforea
Sreaterthan normal degreeof uncertaintyin the calculateddesignconditions,prompting the inclusion of a
higher than averagemargin of safety allowancein thejoint probability results.
Truro
Until now the only primary variablesconsideredhavebeen wave height and water level. However, there is
no reasonwhy JOIN-SEA cannotbe usedfor different variables. With MAFF funding, HR Wallingford is
currently involved in researchinto the prediction of extremewater levels in estuaries.In parts of many UK
estuariesthe combined effects of wave conditions,tidal surgesand river dischargesneed to be considered.
As part of the researchJOIN-SEA has beenused to assessthe dependencebetweenriver flows and wave
conditions, and river flows and tidal surges,as well as wave conditionsand tidal surges.
Knowing the dependencebetweeneachpair of variablesenablesa three dimensionalsurfaceof

combinationsof wave height, tidal surgeand river discharge,with a specifiedjoint exceedancerefirm
period, to be plotted. Tests are currently under way to assessthe applicability of such a method to
determineextremeestuarinewater levels using three-variablejoint exceedanceextremes.
Another method being considered,which is perhapsmore in keepingwith the JOIN-SEA approach,is to

determinethe correlation betweeneachpair of variablesand to use the resultsin a long-term simulation
using the Tri-Variate Normal distribution. However, whichever approachis used,conversionof original
!*o*"r,*"'.0 25
data on flow rate or wind speedto an equivalentincreasein water level remainsas a separateproblem
probably best addressedusing hydraulic modelling.
Humber Estuarv
As part of a major strategystudy, HR Wallingford undertookwave hindcastingat sixteenpoints within the
Humber Estuary (basedon | 4 years of winds measuredat Spurn Point) and a j oint probability analysis
basedon water levels measuredwithin the Humber. A number of interestingpoints are noted.
Although there are many tide gaugeswithin the Humber, HR's client neededto undertakea spatialjoint
probability analysis(of tides and surges)similar to that describedin POL (1997) in order to achievethe
desiredresolution in the water level data.
Although river flow and wind set-upwill affect the water levels within the Humber, they were not
explicitly accountedfor in thejoint probability analysis,since they would alreadybe presentin the
measurementsand in the spatialanalysisbasedon those measurements.
The great majority of the recordsshowedno dependencebetweenwave heightsand water levels, but for
most of the prediction points a small handful of recordsshowedlarge wavesoccurring simultaneouslywith
high water levels. This ratheruncertainindication of a significant correlationin the high tail of the
distribution meantthat some manual smoothingof the derived conelation coefficients was necessaryto
maintain spatial consistencybetweenprediction points.
Anzrlysisof extreme water levels basedon over forty years of water level data in the Humber gave
sufficiently different extremevaluesthat refined valueswere determinedby HR's client for use in
re-scalingduring the simulation (otherwisebasedon only fourleen yearsof data).
4, DISCUSSION
4.1 Past practicefor use of resultsin design
Joint exceedancecombinations
In the past,joint probability analysishas usually resultedin a rangeof combinationsof wave heights and
water levels, eachwith the samejoint exceedancereturn period (and these 'design seastates'are a valuable
output from tle new method). Each combination is expectedto be equalledor exceededonce, on average,
in each retum period. In designing or assessinga structure,one would needto ensurethat it could
withstand every 'design seastate' for the return period being used. In other words, for eachstructural
responsevariable of interest,eachcombinationof extremewater level and wave condition should be
tested,to determinethe worst casefor that resDonsevadable. The retum Deriodof the value thus derived
for the responsevariable(eg overtoppingtwiligenerallybe lessthanth" r.tu- periodof thejoint
exceedanceextremes,as illustrated in Figure I and discussedin the LancasterUniversity part of this
report. This discrepamcybetweenthe two retum periods is usually addressedby inclusion of a small
margin of conservatismsomewhereelse in the assessment.A convenientmetiod sometimesusedin
previous studiesis to use marginal extremewave conditions with the standard3-hour event duration, as
opposedto the 12-hourduration which is more strictly correct for extremesconditional upon being at high
tide. The resulting wave conditions are then appropriatelyconservative.
Application of resultsat different locations

Ifthe extremesare derived for a location other than the point at which they are to be applied, some
adjustmentof values may be necessary.The most obvious caseis the needto modify wave conditions
calculatedoffshore to allow for shallow water transformationsorior to their arrival at coastaldefences.
However, where the extremeshave been calculatedfor a large irea, such as ChristchurchBay, it will also
be necessaryto modify the water levels from one end of the bay to another. The wave transforrnationmay
require a site-specificnumerical model, but the water levels can usually be adjustedwith referenceto
published valuesof tidal ramgesat different locations. This remainsa considerationin the new method.
!unw"rr,n6o.a 26 sR 53t 04/05/00

Allowance for trends and/or unrepresentativedata
If the wave conditions or water levels are known to be subjectto any long-0ermvariations or if the period
of measurementsis known to be unrepresentative,then some allowanceshould be made for this. Short-
term or long-term variations could be filtered out from the data before analysisbegins. Altematively,
derived extremescould be adjustedto offset known unrepresentativeness. The most obvious example of
long-term changeis the expectedrise in mean sealevel due to global warming. The rate of rise is expected
to be about 5mm per year for the foreseeablefuture, which shouldbe incorporatedinto design and
assessmentof coastaldefences. Other trendsinclude increasingwave heightsin the Atlantic, the tendency
for the highest tides to occur at certain times of day and certain times of year, and the 18.6year cycle for
predictedtides. Again this remainsa considerationin the new method,but one which is handled more
neatly than in past approaches.
Analvsis of wave period

In the past,except where direct hindcastingof the responsevariablehas been possible,wave period has not
normally been consideredas a separatepartially independentvariable. Any treatmentof wave period has
usually been basedon the assumptionthat wave height and wave period are strongly dependent,perhaps
being related by some standardwave steepness.(2EH,igT.1 is calculatedfor approximatelythe top one
per cent of the wave data and the samesteepnessis assumedto be valid for the extremewave conditions.
Analvsis of wave direction

Any treatment of wave direction is usually basedon a 'conditional analysis', the condition being that the wave
records analysed have directions within a particular angular sector. In other words, wave data within different
direction sectorsare considered as being membersof different populations (perhapswith different wave
steepnesses)which can be analysedseparately. Ideally, the directional sectorsshould be chosento correspond
with the different populations of waves expectedto occur at the inshore site of inierest, taking accounl of
aspectssuch as bathymetry, fetch lengths, headlands,dependenceupon water depth etc. Each direction sector
then provides separateand effectively independentseaconditions to be consideredin any subsequent
assessment.This approachto the use of wave directionis retainedin the new method.
4.2 Advantagesof the new approach

The main advantages
The new joint probability approachis potentially more objective, flexible and accuratethan a.ltemative
methods currently used in consultancy studies,although preparation of the input data and assessmentof the
diagnostics for the statistical analysis requires experience. The new approachcan provide all of the outputs
available from alternative methods,eg marginal distributions and extremes,joint exceedanceextremes,joint
probability density, direct calculation of the distribution and extremesof structural responsevariables, and
parametersof all fitted distributions. It can thus provide input to both 'design seastate' and 'risk-based'
analysesof seadefences. Generally it performs at least as well as currently used methods,and in some tests
gave significantly better estimatesthan existing methods,particularly at very high retum periods. Specific
advantages, relevantin the majority of applications,are:
1. Incorporation of the variability of wave period (or wave steepness)both in the analysis and in the
simulationof long samplesof data.
2. Assimilation into the analysis of improved marginal predictions basedon additional (non-simultaneous)
data on any of the vanables.
3. More direct analysis of struchrral responsefunction(s) such as run-up or overtopping, thereby addressing
the risk of failure or damagemore directly. Oncethe long-termsimulationhasbeencarriedout, several
!"**n,"u*o 27
different structural responsefunctions and/or different designscan be testedconsis0entlyand relatively
easily.
More accuratesensitivity testing

Sensitivity to uncertainty in the environmentalvariablescan be testedin a similar way. For example,
assumethat all wave heightscould be under-predictedby 15Vaand all wave periodsby 107o:sensitivity
could be calculatedeasily and directly by repeatingthe structuralresponsevariable(s)calculationswith
suitablemultipliers applied to all wave heightsand periods. Uncertaintyin the dependencelevel could be
testedin a similar way, Although not so easy,the presentapproachwould permit direct assessmentof the
sensitivity to a distribution of uncertaintyin one or more of the input variablesand/or the combined effect
of uncertaintiesin more than one partially dependentvariable.
Users and usageof the new anproach

The methodsare intendedfor use both by practisingengineersand by researchers,although training,
software and input data will be required. The amount of effoft, experienceand input data required is
comparablewith mnning a faidy sophisticatedwave transformationmodel. The potential benefit of using
the new joint probability approach(or somethingsimilar), as opposedto an intuitive estimate,is also
comparablewith the benefit of using a sophisticatedwave model, as opposedto a simpler method.
The requirementfor spesialistdata and orogtams

One disadvantage,common to all analytical approachesto joint probability, concernsthe need for reliable
simultaneousdata on waves and water levels, the cost of which may not bejustifiable in some applications.
Another factor that rnay limit widespreadadoptionofthe new methodsis their relianceon specialist
computer programsand training.
4,3 New possibilitiesbasedon long-termsimulation

'design
The present seastate'approach
A small number of design seastatesfor a particular location are often used in design and assessmentof sea
defences. These seastatesmight correspondto wave conditions with specified retum periods or to
combinationsof wave conditionsand waterlevelswith givenjoint exceedance rehrmperiods. Thesedesign
sea statesrnay be applied to the calculation of different variables (eg overtopping, armour size etc) and to
different types of seadefence. However, the same(or even higher) values of the struch.ral responsevariables
might occur, for example, during swell wave conditions with a lower wave height but with a longer wave
penod. The return periods of the calculated structural responsevariables will therefore not necessarilybe the
sameas thoseof the conespondingdesignseastates.
An alternativeapJ@4qLLaseddirectlv on the structuralresponsevariable(s)

A potentially better approachis possiblebasedon long-term simulation of a wide range of seastates,from
which the distribution and extremesof the structuralresponsevariable(s)can be determineddirectly. Once
the long-term seastatedata has been simulated,severalstructuralresponsevariablesand/or defencescould
be assessedquite quickly, without needingto know the correspondingseastates:the probability of
simultaneousfailure mechanismsoccurring could also be evaluated.
Al altemative risk-base.dapproach
Design and assessmentcould thus be done directly in terms of any structural responsevariable(s) of interest.
Using this approach,it would not be necessaryto detemine a designseastate,but insteadthe designwould be
basedmore directly on the overall risk of structural damageand/or unacceptableovertopping etc. A further
point is that if a number of altemative designsand a number of different failure modes are to be analysed,then
the risk-based approach may be quickest. This would imply a major change in design practice, but it is
consistentwith a gradualmove towardsmore risk-baseddesign.
t* VVallinglord
28
4.4 Recommendedoptionsfor joint probabilityanalysis
There are a range of techniques available: from an intelligent choice of a single water level to use with
establishedwave conditionsto a rigorousjoint probability assessment using long time seriesdata; from a
general offshore study which might be valid over a wide area to a site-specific prediction of overtopping
volumes.
Oualitv of the input data

As well as being consistentwith the overall projectvalue,the effort put into any joint probability assessment
is dependenton the quantity and quality of relevantfield data. For example,thereis little point in carefully
'actual'
assessingthe dependencebetweenhigh wavesand high water levels,if the input datais of fairly low
quality, perhapsbeing drawn from intermittentanecdotalevidence.Around most of the UK coast,thereis
sufficient good quality wind/wave/tide data for a proper joint probability assessmentto be carried out. To
justify a rigorous approach(as opposedto an intuitive assessment), involving analysisof long time series
data, about three or more years of good quality wave and tide data should be available. This amount of
recorded wave data is very rarely available, but all around the UK there are several years of good quality
sequentialwind data,which can be usedin wave hindcasting.However,somepartsof the UK coastare so
far from the nearestlong-term tide recorder that it may be impractical to derive reliable local long-term water
level data. Similarly, there are some parts of the UK where wave hindcasting may be rather uncertain,
perhapsdue to the mix of wind-sea and swell, or due to the complicated coasdine. In these circumstancesthe
effort involve.din a full joint probability study may not be justified.
Simplerioint probabilitvmethods
In mostcoastalengineeringprojectsit is worth giving somethoughtto joint probability,evenif it doesnot
justify a fbrmal study. At the simplestlevel, an engineermight decideto test the designwave conditionsat
Mean High Water Springs (if waves and water levels are almost uncorrelated) or at Highest Astronomical
Tide (ifthey are well correlated).Altematively (or additionally)the independentand dependentcasesfor
waveszmdwater levels might be consideredto determinethe maximum rangeof uncertaintyin water level
for use with any given extreme wave condition. A better, but still empincal, approach would be to determine
the degreeof dependencebetweenwavesand water levelsand then to apply the intuitive methoddescribedin
the third paragraphof Section 1.4. The necessary'correlationfactor' can be estimatedfrom general
experienceor with referenceto earlier studiesin the samearea. (In view of the uncertainty,a conservative
value should be taken,as the methoditself is not inherentlyconservative).Armed with this 'factor', extrcme
combinationsof wavesand water levels (expressedin termsof their marginalrehrm periods)can easily be
determined. These approachescarry a wide margin of uncertainty in terms of joint return period, but they are
betterthan not taking any accountofjoint probability andcheaperthan a more rigorousstudy.
Riqorousjoint probabilitv methods

The main distinguishing feature of a 'rigorous' joint probability analysis is that the dependencefunction, and
its variability with threshold,is determinedanalyticallyfrom simultaneousdataon the variablesconceme.d.
For wavesand water levels,aboutthre€yearsof simultaneousdataare necessaryto justify the additional
effort of a rigorous analysis. (Regrettably,the cost of purchasingthis data,evenwhereit alreadyexists in
accessibleformat, can be prohibitive,so this shouldbe checkedat an early stage.) The presentJOIN-SEA
approach described in Chapter 2 analysesthe marginal distributions of the separatevariables, and the
dependencefunction(s)linking them. It then simulatesa much largerquantity of data with the estimated
marginal and joint distributions, effectively exkapolating the joint probabllity density. From this data,
extremejoint exce€danceprobability combinations,and extremevaluesof structuralresponsevariables,can
be determined, The format also providesappropriateinput to risk analysismethods.Extensionof the
JOIN-SEA approach to three variables has been demonstratedduring ongoing researchon exffeme water
levels in estuaries.HR's earlierJOIMROB analysismethodwas lessflexible and a little lessobjectivein
derivingjoint exceedanceextremes,and is no longer in frequenluse.
!"* *"u"oo.o 29 sR 53?r)4Jt)5n0

Better knowledge of marginal extremes
The three basic inputs to joint probability analysis are the distribution of the fimt variable, the distribution of
the secondvariable,and knowledgeof the dependencebetweenthe two. There may be additional
information on one or more ofthese threeaspects,beyondthat which is useddirectly in thejarnl probability
analysis. The most common exampleis a betterknowledgeof extremewater levels,basedon previous
analysisof a longer period of data,than is availablefrom analysisof the period for which thereis both wave
and water level data. The most convenientway of incorporatingthis additionalinformation is by re-scaling
of the results. For example if the shorter period of data is thought to under-predict extremesby 10cm, then
all extremevaluesare simply increasedby lftm. JOIN-SEA allows this to be done smoothly,interpolating
between different amounts of adjustment specified by the user, during simulation of the very large sample of
sea statedata.
Offshore and inshore joint probabilities

When specifyingajoint probability study,the coastalengineershould,of course,haveregardto the intended
rangeof use of the results. If all of the existingcoastaldefenceswithin a large bay are to be assessed, then a
general offshore wave and water level study is appropriate, resulting in combinations of water levels and
offshore wave conditions with given renrm periods. The offshore conditions will then be transformed to the
variousinshorelocationsas and when needed.However,if only one inshorelocation is of interest,it is
probably better to transform the time series wave data to the inshore location and to carry out the joint
probability study directly at the inshore position. This has the advantagesof assessingthe dependence
between high waves and high water levels directly where they are needed,and of providing wave condition
results which are directly useful without further work. In some situations it may also be necessaryto make
local allowance for the effects of wind set-up, wave set-up, river flow and seiching, where this is not already
representedin the water level data used for the analysis. It may also be necessaryto consider the dependence
of wave transforrnation on water level. A common example of this is where waves are strongly depth-limited
at a seadef-enceand where more extreme water levels will pemit higher waves to reach the structure.
Inclusion of wave period and direction

For design seastates,expresse.d in termsofthejoint exceedanceprobability of wavesand water levels,wave
conditions are usually representedsolely by wave height, with wave period and dircction being treated as
secondary variables. Wave periods can be assignedbasedon a standardwave steepness,although it may be
worth checkingthe sensitiviryof the end calculationsto wave period. The new JOIN-SEA approachanalyses
the distribution of wave steepnessin the original data and allows wave period to vary as a function of wave
height. This variation is then automaticallyincludedin JOIN-SEA's long-termsimulationof seaconditions.
If relevant, wave direction can be incorporated by meansof separateanalysesfor each wave direction sector
of interest.
Direct calculationof structuralresponsevariable

If only one particular structural responseis of interest, for example overtopping or force on a wall at one
peuticular location, then that variable can be hindcast and extrapolated directly. This has the advantagesof
including wave period and direction as separateindependentvariables, and ofproviding results in
'single variable'
format exactly wherethey are needed.Similarly, the long-termsimulationpart of
JOIN-SEA will incorporatethe variability of wave period,which can then be usedin the calculationof
structuralresponsessuch as ovenopping. However,the engineershouldbe awarethat it is not possibleto
infer the more generalresultsfrom the more specificresults.
Note on the use of multi-valuedresults

It should be emphasisedagainthat wherejoint exceedanceanalysisyields multiple combinationsof wave
conditionsand water levels,eachwith the sameretum period,then the worst of thesepotential 'design sea
states'should be adoptedfor designpurposes.The particularcombinationmay vary from one loeationto
anotherand from one responseparameterto another.
E*^ Walling{ord 30 sR5l7 lt4/[v(x]

Note on climate chanse
The most convenient and most common way to make allowance for expecte.dfuture climate changesis to
perform thejoint probability calculationsat today's valuesald then to modify the resultsas and when
requiredfor designcalculations. For example,to allow for 0.25m of firture sealevel rise, simply add 0.25m
to all water level resultsderivedfrom thejoint probability analysis. Trendsare harderto identify and model.
A possibility is to eliminatetrend by convertingall recordsto today's valuesbeforej oint probability analysis,
and then to incorporate trend by modifications to the results as suggestedfor future climate change.
Summary
To summarise, consideration of the dependencebetween high waves and high water levels should never be
neglected, where both variables are important. However, the scope of eachjoint probability assessment
shouldbe decidedon its own merits,in termsof input dataavailable,the intendedend use,and the potential
benefitsto be derived. HR's presentpracticein consultancyshrdiesinvolvingjoint probability analysis
follows the recommendations above. Just over half of the studies are carried using the one of the methods
outlined in the paragraphheaded'Simplerjoint probability methods',with most of the remainderbeing
undertaken using JOIN-SEA. Onshore or offshore wave data, wave period ald/or direction, better
knowledge of marginal extremes,joint exceedanceextremes and/or struchrral responsevariables, and
allowance for clinate change, are included as and when appropriate.
4,5 Perlormanceof the varioustechniques

Overall conclusions
Several example calculations and casestudies are presentedin both parts ofthis repot, demonstrating the
relative merits of the different approachesto joint probability assessment.All of the methods seemedto work
satisfactorily, subject to their own limitations, mainly relating to the manner in which dependenceand wave
period were assessed (if at all). Overall, JOIN-SEA performedbest,perhapsbecauseof its rigorousanalysis
of dependenceand its incorporation of variability in wave period, but still some care is required in operation
and interpretation of results. Conclusions from individual comparative tests are given in Chapter 3 of this
part of the report (field data) and in Chapter 10 of the Lancaster University part of this report (simulated
data). Some of the main points are repeatedhere.
Relative importance of different uncertainties

General experience suggeststhat uncertainties in wave conditions, uncertainties in water levels and
uncettainties in dependenceare of roughly equal importance in determining the overall uncefiainty in sea
defence design. However, recent advancesin extreme water level prediction, together with the large amount
ofhigh quality data available around the UK, mean that uncertainties in water levels are probably lower than
in wave conditionsor dependence.A conclusionin the LancasterUniversity part of this report,basedon
analysis of simulated data sets,is that refinement of extreme wave conditions is more important than
refinementof dependencemodelling. However,this is partly due to the direct use of offshorewave
conditions in structural responsevariable analysis, when in reality shallow waler effects would usually have
moderated them before arrival at seadefences. Where wave heights are depth-limited by breaking at sea
defences,refinementof extremewater levels is probablythe most importantofthe threeaspectsofjoint
probability analysis,followed by dependenceassessment.Where shallow water effectsarc importantbut
waves are not depth-limited by breaking, refinement of extreme waves and dependenceare probably of about
equal importance.
Dependencemodelling
JOIN-SEA's statisticalmodelshave sufficientflexibility and robusmessto capturethe featuresofthejoint
distribution of seaconditions producing extreme sea statesfor a range of observational and simulated data
sets. As with all methods,experienceis requiredin handlingdata setswhich do not conform to expectations:
examples are an unusually wide distribution of wave steepnessor more commonly a very small number of
severeseastatesexhibiting high dependencewithin al otherwiseuncorrelateddata set. With care,it is
possibleto treat suchdata setslesssubjectivelyusing JOIN-SEA than with other methodstestedduring this
proJect.
!"**"u.oo.o 3l
Wave period modelling
The importance of wave period was demonstratedin both parts of this report. Where a standard wave
steepnesswas applied to obtain the wave period to use with extreme combinations of wave height and water
level, significantly different overtopping rates were predicted for a site-specific value of wave steepness
compared to a 'global' value. The use of a global value tended to under-predict the oveftopping rate, due to
its sensitivityto wave period. The facility to allow a distributionof wave steepness within JOIN-SEA
worked well, but somecare is neededto preventover-predictionof low steepness wavesand consequent
over-predictionof overtoppingrate.
The discrepancv between the iqillgxgeedalee aruLk)int density approaches

In the joint distribution of wave height and water level there is a discrepancy between the probabiliry
associatedwith thejoint exceedanceextremesexpressedin termsof combinationsof wave height and water
level, and the failure region associatedwith struchrralresponsevariables.The probability of the
combinations (the green zmdyellow squareareasin Figure 1) is smaller than the probabilities of the
correspondingresponses(the red and blue roundedareasin Figure l). This meansthat a worst case
combinationwith a givenjoint exceedanceretum period will under-predictthe responsewith the sameretum
period. This effect is well known and, without being quantified,is usually assumedto be the equivalentof a
factor of 2-3 in retum period and to be offset by conservativeassumptionselsewherein the analysis
procedure. Tests with the simulated data in the Lancaster University part of this repoft suggesta factor of
lVz-2 in rchsm peiod, whereas tests with field data in this part of the report suggesta higher vahte of 2-4 .
Direct use of structuralresponsefunctions,whereknown, avoidsthis difficulty but introducesothers.
Direct comparisonbetweenJOINPROB and JOIN-SEA

The comparisonof JOINPROB with JOIN-SEA in Section 3.2, using the combinationsmethod of
determining the worst caseovertoppingevent and assumingthe samevalue of wave steepness,shows good
agfeementat all sites. This suggeststhat the method of marginal extrapolationat varying thresholds,for
both variables,to constructcontoursof equaljoint exceedanceretum period, gives similar resultsto the
extrapolationof the joint density, as usedin JOIN-SEA.
Comparison between.ioint probabilitv and structure variable methods

In general, the newly developedjoint probability methods are better than structure variable methods. This
was not necessarilythe casefor existingjoint probability methodsas,unlike structurevariablemethods,they
ignored variation in wave period. Analysis presentedin Chapter 9 of the Lancaster University part of this
report shows that it is possible to produce very erroneousextreme values by direct extrapolation of the
structuralresponsevariable. However,similar analysisin Section3.1 ofthis part ofthe report suggeststhat
HR's existing approachto direct hindcastingand extrapolationof the structuralresponsevariableis
reasonablyrobustand reliable in this example,althoughnot as good as rhe new JOIN-SEA method.
The imponanceof site-specificnearshoredata

The Somersetcasestudy in Section3.3 illustratedthe importanceoflocal variationsin wave conditionsand
dependencewhich might not be representedin ajoint probability analysisusing offshoreconditions.
Different overtopping rate predictions (in this case higher, due to the increasedwave period) were obtained
using nearshore wave data. Reduced overtopping predictions were obtained when waves were assumedto be
depthJimited by brezking before arrival at the seawall. A similar pattem was seenin the armour size
calculations, but differences between nearshoreand offshore wave data were less, due to its lower sensitivity
to wave period,
A practicalillustration of the conclusions

Some of the conclusionscan be summarisedin more practicallerms in the form of the following example.
For locationswith a small positivecorrelationbetweenwavesand water levels,the assumptionof complete
dependencebetweenwave heightsand water levels could lead to seadefencecrestlevelsbeing designedup
to abouttwo metrestoo high. Conversely,the assumptionof independencecould lead to seadefencecrest
levels being set too low. More realistically a poor but deliberately conseruativeestimate of dependence
t"- \,/allingford sR53? 04/05/{X'

might resultin a seawall beingdesigne.d
abouthalf a metrehigherthanif a morereliableanalysisofjoint
probabilityhadbeenundertaken.
4.6 Availabilityof methodsand staff time involved

Usageofjoint probabilitv analvsiswithin coastalengineeringstudies
Joint probability has change.dover the last ten years from being a major element of a coastal desk study to
being a fairly routine tool for use in most coastal studies. A coastal engineer who is capable of deriving
extreme wave conditions and extreme water levels should also be capableof making an intelligent estimate
as to how likely they are to occur together, once the principles have been understood. A more detailed
assessmentof dependenceandjoint probability requires more experience, specialist knowledge, and good
data on waves and water levels. However, coastal engineerswho are familiar with wave predictions and
numericaltechniquesshouldbe capableof carrying out theif own assessments. A rigorousjoint probability
analysis using long time series data requires a large volume of data, experience, specialist software, and
expertisein interpretationof the results.
Relativecostsof different analysismethods

The cost of ajoint probability assessment
obviously variesdependingupon how much input data has to be
purchasedand upon the particular circumstancesof the study. However, to give an idea of the effort
involved, the following numbersof days of experiencedstaff time are typical, including time for assimilation
of data,checkingof marginalextremes,and for interpretationand reporting:
. Intelligent estimate of dependence,leading to sensible water levels at which to test derived extreme
wave conditions:% day
. Intuitive assessment ofjoint probability extremes,leadingto a rangeof combinationsof wavesand

water levels with given joint exceedanceretum periods: 2 days
. Rigorous assessment ofjoint probability extremesbasedon long time seriesdata,leadingto a range

of combinations of waves and water levels, from different direction sectors,with given joint
excee.dancerehrm periods:5 days
' Rigorous site-specific derivation and extrapolation of force or overtopping, based on long time senes
data on wavesand water levels: 5 days
The times given above refer to analysis at the first point within a study area. Each subsequentanalysis point
within the same areaand samerepoft would add only about one third to the staff time for the first point.
Acquisitian of data, and wave generation and transfornwtion modelling are rct included in the times given.
ApprOqche$laru used at HR Wallinsford

Joint probability analysis is a fairly standardrequirement in UK coastal defence studieswhere overtopping
may be an issue. It is a less common requirement in overseasstudies. Around two thirds of HR's joint
probability studies are undertaken using desk study methods (even in caseswhere wave or tidal flow models
may be neededelsewhere in the project). For larger studies and for final design, a rigorous analysis is
desirable, but the data purchasecosts will often be prohibitive. A full analysis will not consistendy either
increaseor decreasethe best estimate of the required height for a seawall, but it should reduce the uncertainty
by 10-20cm. If the potential saving in construction cost is high comparedto the cost of buying or generating
the reliable time series data neededfor a firll analysis,then this would seemto be a cost-effective option.
Since about 1998,HR Wallingford has stoppedusing the earlier JOIMROB programs, and has switched
entirely to the new JOIN-SEA methods. Resultscontinue to be usually quoted in terms of joint excervdance
extremes,but direct simulationof responsessuchas overtoppingis sometimesused.
4t-- \ hllingford 33 sR 53? {}4J0yd)

4.7 Futuredevelopments
Ongoing research
The methods are continuing to be developedand applied within other MAFF-funded researchat
HR Wallingford. For example, the same statistical methods have been adoptedfor use within a project on
extreme water levels, flows and waves in eshraries. As part of another ongoing project, the programs have
beenreleasedfor specialistbeta-testingwithin the industry,which hopefu\ will leadto new developments
amdapplications.
Application to othervadables
Most of this report deals with joint probability as it affects large waves and high water levels. However, it is
worth mentioning that, in principle, the methods have wider application, for example to waves and winds,
swell waves rmd stom waves, waves and currents, and water levels and river flows. However, it remains
necessaryto have good data on the marginal extremes,and sufficient data to estimatethe dependencebetween
the two variables. Also, the new progams are specificto wave heightsand waterlevels,with wave periodsas
a related third variable, so some re-programrning would be neededbefore use with other variables.
Move from 'designseastaie' to risk-baseddesign

In the longer-term, the methods offer the chanceto make a significant change to the way in which seadefences
are designedand assessed.An accuratejointdistributionof wave height,wave period and water level can now
be constructed, applied and analysedwith relative ease. It would be feasible to work directly in terms of the
retum period for different structural responsefunctions (eg overtopping or force) on a range of seadefence
options,without the needfor the lessdirect 'designseastale' approachusedat present.The risk-based
approachmay even be quicker to apply than the more traditional methods where a number of altemative
designsand failure modesareto be considered.
!"**"u,*",0 34
5. REFERENCES
CIRIA (1996). Beachmanagementmanual. CIRIA Report 153.
CIRWCUR (199I). Manual on the use of rock in coastal and shoreline engineering. CIRIA Special
Publication83.
S G Coles and J A Tawn (.1994).Statistical methodsfor multirariate erdremes:an application to

structural design. Applied Statistics.
HR Waflingford (1980). Designof seawallsallowingfor waveovertopping.HRReportEX 924.
HR Wallingford (1994). Valifution of joint probability methodsfor large wavesand.high water levels.
HR ReportSR 347.
HR Wallingfbrd (2OOO).IOIN-SEA user manual. HR Report TR 71.
Ministry of Agriculture, Fisheries and Food (1995). Joint probabilities - a workshop to further research.
MAFF Flood and CoastalDefenceNewsletteqJuly 1995.
MWOwen,PJHawkes,JATawnandPBortot(1997). Thejoint probabilityof wavesarulwaterlevels:A

rigorous but practical new approach. MAFF Conferenceof River ard CoastalEngineers, Keele.
Proudman OceanographicLaboratory (1997). Estiftntes of extremesea conditions: Spatial analysesfor the

UK coast. POL Intemal DocumentNo 112.
E** W dlingford 35 SR 53? l"l/05/d)

Appendices
lHa w"ttinsrora
AppendixI
betweenjoint exceedance
Comparisons eKremesfor differentanalysismethods
FigureAl: Ckistchurch
FigureA2: Dover
FigureA3: Dowsing
FigureA4: Cardiff (0-190"N)
Figure,{5: Cardiff ( 190-360'N)
FigureA6: North Wales
Notes:l) Dotsshoworiginaldata
2) Predictionpointsshowjoint exceedance with given
extremes
retum oeriods
!*^
Christchurch
a.1
v)
E
tn4
0)
E
C)
6
F
0
0.0 0.5 1.0 1.5 2.0
'lVater
Level (m ODN)
FigureA1 Christchurch
l!"***noo-
Dover
tr
->4
a
E
bo
'6
o
>2
=
/?
'lVater
Level (m oDN)
Figure A2 Dover
lHnw"nngora
Dowsing
rl
E
m^
a4
uo
0.)
o
(!
=
4 o
Water Level (m oDN)
Figure A3 Dowsing
!***","rr*
Cardiff (0-tOO aeg)
(4
a'1
bD 1.v
C)
T
0)
d
=
0.5
0.0
4
Water Levei
FigureA4 Cardiff (G.190'N)
[*^*n,*o,"o
Cardiff (t0o-zzo aeg)
a'i
a
E
a'i
u0
a)
:T:
G)
>
(n n
' ' '6
=
4 6
'lVater
Level ( m o D N )
Figure .4.5 Cardiff (190-3611'N)
E*
North Wales
.1
a':
un=
q)
E
o
6
=
tz
Water Level (m cD)
FigureA6 North Wales
tno*,'*o*
Part II LancasterUniversity Repoft
!"^*r,"0"'.0
JOINT PROBABILITY METHODS FOR
EXTREME STILL WATER LEVELS AND WAVES
Paola Bortot and Jonathan A. Tawn

D epartmentof Mathemat'icsand,Stat'ist'ics,
Lancaster Un'iaersr,ty,
LancasterLA1 LYF.
In collaborationwith
HR Wallingford.
1997.
September
CONTRACT
This report describeswork funded by the Ministry of Agriculture, Fisheriesand Food

(Flood and CoastalDefenceDivision) under CommissionFD0202with HR Wallingford.
HR's Nominated OIficel was Dr S. W. Huntington. The Ministry's Project Oficer was
Mr A C Polson.Publicationdoesnot imply endorsementby the Ministry of the report's
conclusionsor recommendations.
@Copyright Ministry of Agriculture, Fisheriesand Food 1997.

Contents
Introduction I
1.1 Frameworkof the Study 10
L2 CurrentImplementation 13
1.3 Commenton the Current Implementation t4
1.4 Objectives t5
1.5 Outlineof the Report 16
I.6 GoodStatisticalPractice 17
I Theoretical Basis of the Joint Probabilitv Method 19
Outliue Statistical Methods 2L

2.1 StructureVariableMethodISVM) 22
2.2 Joint ProbabilitiesMethod (JPM) oa
2.3 Comparison
of Methods 24
Extreme Value Methods 27

3.1 UnivariateMethods 27
3.1.1 Annualmaximum. 27
3.7.2 Thresholdmethods 29
3.1.3 Propertiesof the GPD . . . 30
3.1.4 Thresholdselectionfor GPD 31
3.1.6 Temporal dependenceand non-+tationarity . 32

3.2 Illustrative DependenceExamples ea
3.2.1 The structurefunctionof the minimum

3.2.2 The structurefunctionof the maximum 34
3.3 MultivariateMethods JO
3.3.1 Models. 35
3.3.2 Diagnostics 40
COI\iTEIVTS
3.3.3 Illustrativeexamples 4l
3.4 TechnicalCalculations 42
3.4.1 Independence 43
3.4.2 Bivariate extreme value dependence 43
Statistical Models used for JPM 45

4.7 Marginal Statistical Models 45
4.1.1 Still Water Level 46
4.LZ SignificantWaveHeight 46
4.1.3 Steepness DI
4.1.4 WaveDirection 52
4.1.5 WavePeriod 52
4.2 Dependence Models. 56
4.2.1 Modelsfor denendencebetweenSl4/L and Hg 59
4.2-2 Modelsfor dependence betweenI1e and ,9 . . 62
4.3 Evaluationof the Probability of Failureusing the JPM 63
II Examples of Data Analysis using the Joint Probability

Method 67
Study Data 69
5.1 Observational data 69
5.2 Simulated data 70
5.3 Principal differencesbetweenttre simutateaO*u ,.*
5.4 Comparisonof observedand simulateddata 76
Marginal Estimation 89
6.1 Still Water Level 89
6.1.1 ObservationalData 89
6.1.2 SimulatedData o<
6.2 Surge. 96
6.2.2 SimulatedData 101
6 . 3 SignificantWaveHeight 101
6.3.2 SimulatedData 120
6.4 Steepness 120
6.5 Wave Direction 124
CONTEAI"S 5
Dependence Estimation 727

7.I (SWL, Hs) Dependence . r27
7.1-l Observationaldata . t27
7.1-2 Simulateddata . 136
7.2 (^9,fls ) Dependence
for SimulatedData . . 148
Simultaneous Extremes of (SW L, Hs) 161

8.1 Marginal distributionsknown . 163
8.1.1 Comparison with currentimplementation . 170
8.2 Marginal distributionsunknown
8.2.1 Comparisonwith current implementation . 182
III Applications to the Estimation of the Probability of Fail-

ure for an Existing Design 185
Analysis of Overtopping Discharge Rates 187

9.1 BackgroundMethods 1BB
9.2 Applicationto Siml . 190
9.3 Applicationto Sim2. 193
9.5 Application to Sim4 . 100

9.7 Commenton CurrentImolementationResults 207
9.8 Other Figures 207
10 Comparisons and Conclusions 217
L1 References ?7s
A Background information 221.
A.1 Joint Distributions 22L
223
A.2.1 UnivariateNormalDistribution 223
A.2.2 MultivariateNormalDistribution
A.3 The Probability Integral tansform 226
A.4 Likelihoodsfor Dependence Models 226
A.4.1 BivariateNormalModel 227
A.4.2 BivariateNormalThresholdModel 227
COIf?EAITS
A.4.3 Mixture of Bivariate Normals .....229

CONTENTS
Notation
Structure parameters
A: the structurefunction,whichlinks the seaconditionvariablesto the variableof interest

(the structurevariable).
Y2:the structure variableat time f.
Qg: lhe overtoppingdischargerate (an examplestructure variable).
X1: the (multivariate)seaconditions,at time f, for generalitytaken to be d-dimensional,
but usuallyconsistingof (SWL, Hs,T2, d, S) Iistedbelow.
A,: failureregionfor X, corresponding to the set {x:A(x) > u}.
Input variables
.9W-L:still water level.

,.9zrge:non-tidal componentof still water level.
f15: significantwaveheight.
T2: (zerc meancrossing)waveperiod.
d: (predominant)wavedirection.
,5: wave steepness.
Distributious, densities and probabilities
ft(c): the (marginal)densityfunction of X evaluatedat z.

Fx(r): the distributionfunctionof X evaluatedat z, i.e. Pr{X < r}.
O(z): the distribution function of a standardNormal (Gaussian)random variable.
ft: the joint densityfunctionof X.
fx: the estimatedjoint densityfunction of X.
Fx(x): the joint distributionfunction function of X : (Xr, . . . , Xa), i.e. Pr{X1 <
ry,...,Xa1ral.
Fx(x): the joint survivorfunctionof X, i.e. Pr{X1 > 11,.. . , X1> r1}.
Pr(,4): the probability of event,4 occurring.
Pr(A): the estimatedprobability of event,4 occurring.
Statistical Models and parameters
GPD(o, {): the generalizedParetodistribution with scaleparametero and shapeparam-

eter {, usedas a model for excesses over a threshold.
GEV(p, a {): the generalizedextremevaluedistribution with locationparameterp, scale
parametero and shapeparameter{, usedas a model for maxima.
gp: the return level with return period of 1/p.
COfVTEf,/TS
1lo:the estimated return level.

BVN: a bivariate normal distribution which has correlation parameter p.
z: a threshold used in both marginal and dependencemodelling.
Threshold BVN model: a model for the joint extremes which is of bivariate norma.l form
with correlation, p,, which depends on the threshold used to define a joint extreme.
Mixture of BV N model: a model for the dependencestructure which is a mixture of two
separate BVI{ models with potentially different correlations. The probabilities ofthe two
BV/y' models in the mixture arc pM and,| - pM.
TYansformed variables
S*: wavesteepness variabletransformedto follow a standardnormal distribution.

I{: significantwaveheightvariabletransformedto follow a standardnormal distribution.
Diagnostics
T(z): a samplebaseddiagnosticstatistic, which is used to assesswhen a conservative

dependence estimate)of completedependencebetweenthe variables,can be usedwithout
substantialover-estimation.
Chapter t
Introduction
Since 1991, MAFF have funded a study of joint probability methods at HR Walling-
ford. HR have developedpractical methodsof estimatingthe degreeof correlationbe-
tween wavesand water levelsfor practising engineersto apply to the designof coastal
flood defences.Under a separateMAFF-funded project, the Proudman Oceanographic
Laboratory and LancasterUniversity have developedstatistically rigorousmethods for
estimating the joint probability of tides and surges,and more generally,techniquesfor
handling joint probabilitiesof extremesof any environmentalvariables. However,these
rigorousmethodscan only be appliedby specialiststatisticians.
In previous work by Michael Owen, whilst at HR, joint probability methods were
identified as being important in estimatingthe probability that coastalflooding occurs,
asthis is determinedby overtoppingdischargerates (givenby sea-levels and waves)rather
than by sea-levelsalone.
In August 1994 MAFF hosteda Workshopon joint probability chaired by Michael
Owen. As a result of the WorkshopHR and LancasterUniversity submitted a joint
researchproposalto MAFF entitled The use of joi,nt probabili,tgdata on wauesand,water
leuelsi,n coastaleng,ineering This proposeda 2 year study starting April
appli,cati,ons.
1995.
o During the first yearLancasterUniversitywouldconcentrateon developinggenerally

applicablestatistical techniquesand software,suitable for eventual use by non-
specialistsfor extrapolationin joint probability problems. Thesemethodswere to
be tested on synthesiseddata and the resultscomparedwith those obtained using
HR's existingmethods.
r The secondyear would see further refinementof these techniquesand extensive

applicationin practical studiesat HR.
10 CHAPTER 1. INTRODUCTION
MAFF funded this project with the aim of bringing together the best aspectsof ex-
isting approaches,combiningstatistical rigour with engineeringpracticality, to provide
more powerful tools for routine use by practisingengineersand consultants.This report
describesthe work in the first year of the study focusingon the progressat Lancaster
University into producingrigorousyet applicablemethods.
1.1 FYamework of the Study

When considering the design of a new sea-wall,or other sea defence,or when assessingthe
safety offered by an existing design, a key step is the estimation of the probability that
the design fails to protect against extreme sea conditions. This probability assessmentis
a fbur stage process:
1. the selection of a design (either the existing design or one basedon some preliminary
analysis),
2. the identification of all possible types of failure for that design - the modes offailure,
3. for each mode of failure, the identification of the combinations of sea condition
variables which cause failure,
4. the estimation of the probability ofthese combinations which give failure.
There are a number of modes of failure. such as
o overflow when the water level exceedsthe level of the crest of the defence,
r overtopping - when the combined effect of waves and water levels results in waves
running up and breaking over the defence,
r structural failure such as severedamage to rock armour revetment or erosion of an

embankment leading to the formation of a breach
For any particular mode of failure we term the associated variable of interest the
structure variable. Generallv the structure variable is related to the sea conditions
o tide,
. Sulge,
significant wave height,
wave period (heretaken to be zero mean crossingperiod),

FRAMEWORK OF THE STUDY l1
. wavedirection (heretaken to be predominantwavedirection),
via the function A, which we term the structure function. Thus
structure variable : A( tide, surge,significantwaveheight,

waveperiod.wavedirection)
A more useful notation is that, at time t, we }ave seaconditions,X;, and the structure
variableX, with
Yr = A(&) for all t.
HereX1 is multivariate(5-dimensional) , (Xr,r,... , X5,1),with X1,1denotingthe iide, X2,1
denotingthe surge,X3,sdenotingsignificantwaveheight,Xa,1denotingwaveperiod, and
X5,1denotiugwavedirection.
Throughout the study the methods that are developedapply to any mode of fail-
ure/structure function. However,to illustrate the methods in application we consider
only overtoppingas the modeof failure and take the structurevariableto be the overtop-
ping dischargerate. For this examplethe structure function is
A(X) : o19X3Xaexp{- a2(u- X2 - X)llxa(sh)1/211, (1.1.1)
for a sea-wallof height a and where g is the accelerationdue to gravity. This form
was suggestedby HydraulicsResearchStation (1980)basedon wave tank studiesfor a
simple type of sea-wall.Here the sea-wallis a slopingdesignand the constantsa1 and a2
in equation(1.1.1)dependon the characteristics of the sea-walldesign.From equation
(1.1.1)it is clearthat overtoppingwill occuronly if at leastoneof the variables
r still water level,Xr + &,
. significantwaveheight, or
r waveperiod
is sufficientlylarge.
Aside: note that Jor thi,ssturlg the structure function ignoresthe directi,onalityof
the uaues but waued,irectionts retai,nedin the sea conrl,iti,on uectorX2 to ai,d,the latter
stati,sticalmodellingof thejoi,nt distri,butionof the otherseacorul,itr,on
uari,ables.Srmi,larly,
the decompositi,on of the sti,ll uater leuel into tide and,surgecoTnponents i,snot essential
other than to ar,dthe statr.sticalmod,elli,ng.
Now, the probability of somecritical dischargerate, u say,beingexceededat time f is
given bv
Pr{Y1> u}.
12 CHAPTER 1. INTRODUCTION
In terms of the seaconditionsX; we can re-write this probability as

P.{4>u}: P r { A ( X 1 )> z }
: Pr{Xr € -4"},
wherethe set A, correspondsto all combinationsof the X1 variableswhich lead to values
of the structurevariablewhichexceedz. i.e.
A": {x: A(x) > u}.

We term the set,4., the failure region. Note that everycombinationin the failure region
is not equally likely, for examplemoderatelylarge still water levelsand wavesmay give
the samedischargerate as an extremestill water level with small waves,but one of these
combinationswill be more likely than anotherdependingon the joint distribution of X.
Motivated by the two forms for expressingthe probability of the structure variable
exceedinga critical level,i.e. a designfailure occuring,two statisticalmethodshavebeen
proposedfor estimating this probability, or the designparameters,in practice. These
methodsare:
1 . Structure Variable Method
The sea condition data are used to construct a seriesof structure variable data.
Thesestructure variabledata are extrapolatedto more extremelevels,via a statis-
tical model for the distribution of the structure variable.
2 . Joint Probabilities Method

The seacondition data are themselvesextrapolatedto more extremecombinations.
This is achievedvia a statisticalmodelfor the joint distribution of the seaconditions.
The probabilitiesofinterestfor the structurevariablearethen inferredby integrating
the joint distribution over the failure region.
There are three key distinctions between the methods:
1. The variable extrapolated: the structure variable in the structure va,riablemethod,

and the sea conditions for the joint probabilities method.
2. The use of the structure function beyond the data: this knowledge is ignored by
the structure variable method but, if known, can be fully exploited by the joint
probabilities method.
3. The generality of the analysis: the statistical extrapolations for the structure vari-
able are specific to the site, design, and mode of failure under consideration, whereas
fbr the joint probabilities method the design and site-specific features are excluded
from the extrapolation as they are features of the structure function. Thus the
statistical analysis for the joint probabilities method applies to any design analysis.
1.2. CURRENT IMPLEMENTA'IION 1t
A consequence of thesefeaturesis that the joint probabilitiesmethod is generallythe

preferredapproachby both statisticiansand designengineersbut has the drawbackof
additional complexity,specificallythe dimensionalityof the extrapolationof the seacon-
dition variables,which requiresboth
e extrapolationof the separateseaconditionvariables
r extrapolationof the dependence

betweenthe seacondition variables.
In this study we will developboth thesemethodsfor problemsinvolvingthe estimation

of the probability that overtoppingdischargeratesexceedcritical levels.
L.2 Current Implementation

HR and consultantcoastalengineerscurrently implementversionsof both the structure
variable and joint probability methodsfor clients and for internal resea.rch
in a routine
fashion(Hawkesand Hague,1994).
Structure Variable Method
Applicationofthis methodinvolvesthe useofthe structurefunction,e.g. equation(1.1.1),
which is viewedto be the most appropriatefor the designunder consideration.Data on
the structure variable a,reused to extrapolateto more extremeeventsusing statistical
methodssimila,rto thoseusedin this report.
Joint Probabilities Method
There are a number of levelsof sophisticationat which this method is currently applied.
Herewe review only the two most rigorous.
The approachcurrently usedinvolvesa numberof stages.
1. The joint distribution of still water level,SW L, and significantwaveheight, IIs, is

evaluatedby estimationof the probabilities
Pr{SWL>r,Hs>A}
for all r and g in which either r or A ate la.rgelevelsof the respectivevariables.
2. The wavesteepness(a function of waveperiod and significantwaveheight) is taken

to be a constantvalue. This constantis estimatedas the averagewave steepness
valueof the largest1% of significantwaveheights(or somestandardvalueis taken).
3. The structurevariablevalueassociatedwith each(r, g) pair aboveis now evaluated.

This evaluationrequiresthe useof a waveperiod which can be inferredfrom y and
the estimatedconstantwavesteepness value.
14 CHAPTER]. IN"RODUCTIOAI
4. The probability of the structure variableexceedingthe level a is taken to be esti-

mated by
maxPr{SWtr>r,Hs>U}
where the maximum is taken over all combinationsof (2, g) suchthat the resulting
structurevariablevalueis z.
A more refined versionof this method follows the format abovewith certain steDs
modified as below:
1/. The probabilityF(z,g) -PI{SWL) n,Hs > g} is estimatedas aboveand from this
the probabilityof (SWL,11s)fallingin a rectangularcell [16,r1j x lAo,yt]is obtainedby
P r { 2 6< S W L < r t , y o < H s < A t } : F ( " o , a o ) + F ( q , y ) -F(ro,Ar) -F(rt,yo)
2'. Sometimesvariation of the steepness of the wavesis accountedfor. Usually by trying

differentvaluesof S.
3'. Step3 is unchanged.
4r. The probability of the structure variableexceedingthe level u is given by the sum of
the cell probabilitiesover thosecellsfor which the corresponding(SW L, Hs) pair, when
combinedwith the constantwavesteepness value,give a structurevariableexceedingz.
1.3 Comment on the Current Implementation

There are three featureswhich may potentially limit the accuracyof results obtained
using the joint probabilitiesmethod as currently implemented.Thesefeaturesare not a
restriction for the structure variablemethod.
1 . The estimationof the joint probability Pn{SWL } r, Hs > gr} is essentiallybased

on statisticalestimatesfor the separatevariablescombinedby an empiricalestimate
of the dependencebetweenthe variables.When both n and g are large it is well-
known in statistical theory that suchempiricalestimatescan be poor.
2 . Wave period is taken to be given exactly by the significantwave height and the
estimated constantwave steepnessvalue. Waveswith longer periods for a given
wave height can occur and thesewill typically producelarger valuesof the struc-
ture variable. Sometimesin applicationsof the refined method this restriction is
overcomeby incorporationof variationsin S.
a A feature of the more basic current approachis that the failure region used to
estimate the probability of failure, Pr{X e ,4,}, is a subset of the true failure
region ,4,. Thus, even if the joint distribution of the seacondition variableswas
1.4. OBJECTIVES li)
accurate,the probability of failure would be under-estimatedby current methods.

To illustrate this feature,ignoringwaveperiod, the current methodtakesthe failure
regionto be approximated \, {SWL} z,Hs > y} for somesuitablychosen(c,g).
All the elementsof this set givea structurevariablevaluewhich exceedssomelevelz
say. However,this set doesnot contain aIl (SW L,11s) valueswhich give a structure
variablegreaterthan z. For someSW L > r with 115( g, and for someSW L < x
with Ils > g, the structurevariablewill exceedti. Thesepotential failure valuesare
excludedfrom the failure regionleadingto a biasedestimate.
I'rom work prior to the commencement of this project it was unclearhow important any
ofthese deficiencies/approximationsin the existingmethodswere. They may canceleach
other out, or they may causeadditive errors. However,evenif the errorsare additive, the
approximationmay be suflicientlygood to obtain the optimal design.
The only evidenceprior to the project comesfrom inconsistentoutputs from current
implementationsof the structure variablemethod and the joint probability method. Dif-
ferent approacheshavebeenusedfrom time to time, dependenton budget,availability of
input data and intendedend useof the results. However,when differenceshavearisenit
has not beenclearwhetherthe disprecancylies in the estimatesprovidedby the structure
variable method, due to its poor extrapolationfeatures,or is due to the weaknesses in
the implementationof the joint probability method. This feature has resultedin some
doubts about the eiven estimates.
L.4 Objectives
This study proposesto examineeachof the two statistical methods as aids to design
assessment. The methodswill be developedgenerallyand applied,as an exampleapplica-
tion, to the estimationof the probability of overtoppingdischargeratesexceedingcritical
Ievels.In so doing, the study objectivesare:
r To developmore fundamentallyrigorousmethodsfor analysingthe dependence be-

tween still water level and wave extremesand improve the handling of the wave
steepnesscomponentof the statistical model. Thesemodificationsshould remove
the statistical inadequaciesin the current 'practical' approachand also lend them-
selvesto applicationby practitioners;
r To removeany inconsistencies betweenthe approaches in terms ofthe derivedprob-

ability of failure, i.e. through using the samestructure function/failure region for
eachmethod;
16 CHAPTER 1, INTRODUCTION
To identify and commenton situationswherethe two most often usedcurrent ver-

sionsof the joint probability method work well and thosewherethey are poor;
To incorporate the most up-to-date information on the separatedistributions of

extremestill water level and significantwaveheight;
To providestatisticaloutput directly usableby coastalengineersto assess

probabil-
ity of failure for the mode of failure under consideration;
To validate the methodson simulatedand observational

/hindcast data sets.
The way that we shall approacha simple solution to the problem is to adopt the
rigorousframeworkof Colesand Tawn (1994)but insteadofusing their complexstatistical
modelsfor dependence in the extremejoint tail of the distribution we shall explorea much
simpler but more flexible approachbasedon useof the multivariate normal distribution
for the seacondition variablesafter suitabletransformation.This differs from the Coles
and Tawn approachwheremultivariateextremevaluedistributionalmodelsare usedafter
a different marsinal transform.
1.5 Outline of the Report

The report comprisesthree distinct parts. In Part I the theoretical basis for the joint
probability method is given. In Part II the estimationof the joint distribution of the sea
condition variablesis consideredfor observationaland simulateddata. Finally in Part III
we illustrate the use of the joint probability method to give the distribution of extreme
valuesofthe structurevariableby applicationto the problemof estimatingthe probability
of the overtoppingdischargerate exceedinga critical level.
More explicitly, Part I containsthe following: Chapter 2 gives a generalreview of
the two statistical methods(structurevariableand joint probability) and comparestheir
properties.In Chapter3 backgroundinformationis giveninto the statisticalmethodsused
for extreme value modelling in univariate and multivariate problems. For multivariate
problemswe examinethe importanceof the degreeof dependenceand the form of the
failure region in determiningthe probability of failure. Chapter 4 containsthe details
of the marginal and dependencestatistical model componentsof the joint probability
method. Also in Chapter 4 the evaluationof the probability of interest from the joint
probability method is discussed.
Part II starts by introducing the study data sets in Chapter 5. Observationaldata
from 6 sitesand 5 data setssimulatedfrom knownstatisticalmodelsare described.These
data are studied throughout Chapters6-8, where the methods developedin Chapters
GOOD STATIS'IICAL PRACTICE 17
2-4 are applied to thesedata. Specifically,for both types of data set, in Chapter 6 each
of the separateseacondition variablesis analysedwith differentstatistical modelsbeing
considered;in Chapter 7 the dependence betweenextremesof the seaconditionvariables
is analysed;and in Chapter 8 the joint extremesof still water level and significantwaves
are estimated.
Part III comprisesChapter 9 where the distribution of the overtoppingdischarge
variable is derived from both joint probability and structure variable methods. A key
featurein this context is the assessment of the impact of extra knowledgeon the separate
sea condition variableson the estimation of the distribution of the structure variable.
This chapter comparesthe structure variable and joint probability methods relative to
eachother and against the methods usedby coastalengineers.For the simulateddata
the joint distribution of the seacondition variablesis known (givenin Chapter 5) so the
true distribution of the structure variable (e.g. the overtoppingdischargerate) can be
inferred. Thus the true performanceof the methodscan be assessed for the simulated
data. In Chapter 10 someconclusionsare presented.Throughout,all technicalstatistical
details have been kept to a minimum, with details when necessarygiven in a technica,l
aooendix.
1.6 Good Statistical Practice

There is a recognisedstandard frameworkfor good statistical practice when modelling
data using statistical techniques.This frameworkis
o Useof valid data
r Use of all availableinformation
Incorporatlonof scientificknowledge
Rationally chosenstatistical models
Efficient method of inference
r Assessmentof sensitivity
o Quantificationof uncertainty
e Variety of waysof communicatingresults
In many applicationsthe quality of the data, or the purposeof the investigation,

sometimeslead to deviationsfrom this framework.In this studv when we havedeviated
18 CHAPTER1, INTRODUCTION
from the frameworkwe have tried to follow the rationa.lebehind the framework. The
particular area where this study is weak, is on the quantificationof uncertainty. There
are three sourcesof uncertainty:
1. the data (as wavesare typically hindcast)
2. the choiceof statistical model
3. the estimationof the parametersin the statisticalmodel.
The latter is given by standard errors of parameterestimates. The secondis partially

assessedby comparisonof different models, but we make no attempt to addressthe
former. Sincethe study was aimed at assessing whether there were grosserrors in the
existing methods,rather than whetherthey are closeto optimal in terms of meansquare
error criteria, we havenot given standarderrorsfor most parameterestimates.
Part I
Theoretical Basis of the Joint

Probabilitv Method
Chapter 2
Outline Statistical Methods
When assessingthe probability of failure for a singlesea condition variable, the events
which give failuresare easilycharacterisedas failuresof the structure causedby extreme
valuesofthe variable,i.e. failuresoccurwheneverthe variableexceedssomelevel. When
the sea condition variable is inherently multivariate evaluation/estimationof the prob-
ability of failure is more problematicas there is no generalordering which determines
which are extremevalues.However,giventhe context ofthe problem,a natural approach
is to basethe orderingon the associatedstructure variable,sincea failure is deemedto
have occurred,in a multivariate problem, if the structure variable,A(X), exceedssome
critical level.
An added complicationto the problem is that a designmay have to satisfy several
di{ferentdesigncriteria, i.e. withstand the extremesof severalstructure variables. For
simplicity we will focus on there being a singledesigncriterion. The work in the second
year of the project will extendthis frameworkto more generalcases.
Thereforefor a structurefunction A, wherefailuresoccurwhenA(X) ) z, the extreme
valuesof X are the set ,4r, given by
Au : {x: A(x) > u}.
In this project A will be takento be a generalstructurefunctionuntil Part III of the

reportwhereit is takento be givenby equation(1.1.1)for overtoppingdischarge rates.
Now, in general,the probability of failure at sometime f say,is given by
P r { X 1e A " ) : P I { A ( X 1 )> z }
: pr{y, > z},
where Y1- A(Xr) is the structure variable. The left and right hand sidesof the above
equation provide alternative formulationsfor the probability of failure. Starting from
eachof theseexpressionsin turn providesthe basisfor the derivationof the two statistical
27
CHAPTER2, OUTLINE STATISTICALMETHODS
approachesto the estimationof the probability of failure in multivariate problemswhere

a designexists or has been specified.Working from the right hand side determinesthe
structure variable method; whereasworking from the left hand side determinesthe
joint probabilities method.
In describingthesemethodsbelowwe will supposethat
r the observationsof each of the processesunder study are identically distributed
through time, that is the distribution of observationsremainsthe samewhatever
the time of the year.
o completeobservationsare availableon the vector of seaconditionsat eachtime.

Theseare reasonableassumptions,as
1. Observationsare approximatelyindependentas we focuson only high water levels,
so observationsare separatedbv 12 hours;
2. Observationsare approximatelyidentically distributed, at least over the winter

storm season. This is a better assumptionfor the surge variable than the still
water level as the long term non-stationarityinduced by the astronomicaltide is
still present(e.g.the nodalcycleof 18.61years);
3. Completeobservationsare availableas the wave data are hindcast over the same
time periodsas the tide and surgedata are available.
2.I Structure Variable Method (SVM)

The structure variablemethod is an approachto the estimationof the probability that
a given design fails. It is basedon the statistical analysisof derived observationson
the structure variableof interestfor the given design,i.e. supposethere is a variableof
interest Y which is a function A, the structure function, of the vector of sea condition
variablesX,
Y,: A(Xr) (2.1.1)
for t:1,...,n., wheren is the numberof observations.In practicewe haveobservations
of X; but not Y1.Given A, the approachinvolvestwo steps:
1. Createthe structurevariable,Y, of interestusingequation(2.1.1)for the entire
time seriesof seacondition observations.
2. For the time series{Yr, t : 1,...,n} ol the structurevariable,use a suitable

statistical model to extrapolatethe seriesto the appropriatereturn level. This can
involve either of two oossibilities:
2.2. JOINT PROBABILITIESMETHOD(JPM) .re
. Fitting a probability distribution to the entire sampleofthe structurevariable.

o Fitting a statistical tail model to the extremevaluesof the structure variable,
using techniquesof the form outlined in Chapter 3.
The choiceof which of these approachesto use dependson how simple the true
distribution of the structurevariableis. Clearly what is neededin practiceis a good
fitting statistical model which can be accuratelyestimatedand extrapolateswell.
The method to chooseis the approachwhich most adequatelymeetsthesecriteria
in practice.
Of the two approachesdiscussedabove,generallythe secondis preferable,as the former

is most influencedby observationsin the bulk of the distribution which haveno influence
on the form of the tail of the distribution, and so is liable to give poor fits to the tail.
An alternativeviewpoint is that fitting a statisticalmodel to the extremesof the sample
only is wastefulof data. This is certainly true if a good statistical model for the whole
distribution can be found.
For the examplestructurevariablewe considerin this study, we usethe tail modelling
approachfbr the structure variableof overtoppingdischargerate. A key reasonfor this
choiceis that for most, we haveY, :0, i.e. no discharge, so it is only valuableto model
the statistical distribution of dischargeconditionalon there being a discharge.
Once a distributional model has been fitted to the structure variabledata then ex-
trapolationscan be obtainedfrom the fitted statisticalmodel in the standardmethod by
solvingthe equation
P r { } ' > A p }: 1 - P
for the return leuelye, given a suitably chosenexceedance
probability p.
2.2 Joint Probabilities Method (JPM)

The joint probabilitiesmethodis an approachfor estimatingthe probability of a structure
variableexceedinga critical level,basedon the joint analysisofthe seaconditionvariables,
X, from which the distribution of the structure variablecan be inferred. The estimated
joint distribution of the sea condition variables,X, is of much wider use at the design
stageof a sea-wall/structureas this can be usedto aid the selectionof a provisionalform
for the designwhich onceselectedis subjectedto a more detailed probability of failure
analysisusingthe full JPM. Herethe focusis on the useof the estimatedjoint distribution
of seaconditionsfor evaluatingthe probability of failure for a selecteddesign.
If ft is the joint densityofthe seaconditionsand ,4,,the failure regionofX (e.g. when
the structure variable is overtoppingdischargerates this correspondsto sea condition
24 CHAPTER 2. OUTLINE STATISTICALMETHODS
combinationswhich give overtoppingdischargerates exceeding u), then the approach

involvesthree steps:
1. Estimation of the joint density ft of the sea condition variables. This requiresa
selectionof a statistical model for
a the distributionof eachof the d (d:5) separatevariables,X;, e =

a the dependence
betweenthe componentsof X.
We denotethis estimateof ft by f11.
2. Evaluationof the estimatedprobabilityof failure,Pr{X1 e 4,,}, from the estimated

joint density,fi6, using the relationship
er1x,e A,j: [ /r;(x)dx,

J 4,,
i.e. integrationof the estimatedjoint density over the set 4,.
3. Conversionof this estimateof the probability of failure from an observationalscale

to the annualscale.Sincethere are nu, (no,: 705) independentobservationsof X
over a year, we have that
Pr{ no structure variable exceedsz in a year} P r { ) t 1I A u f o rt : 1 , . . . , n r , }

t. / \ r \?,rr
- "
[r Jr, /x(x)ox/
2.3 Comparison of Methods

The structure va,riable(SVM) and joint probability methods(JPM) outlined abovecan
be used to addressexactly the samedesignquestions,howeverthey are basedon quite
differentassumptions.The validity of theseassumptionscomparedwith the simplicity of
usedeterminewhich of theseapproaches shouldbe usedin practice. Specifically,features
of the methodswhich influenceselectionof the anoroachare:
o Simolicitv.
Extrapolation properties,
Flexibility to coverdifferentdesigns,
Requirementsand exploitationof data,
Usefulinput and output information.

2.3. COMPARISOI{OF METHODS t<
In more detail these features are:
Simplicity
The SVM is much the simplerto useas the statisticalcomponentof the methodis routine
relative to the JPM. For the SVM the analysisand extrapolationis for one variableonly,
the structure variable,whereasfor the JPM analysisextrapolationis requiredfor eachof
the separateseaconditionvariablesand for the dependencebetweenthe separatevariables.
Extrapolation properties
When extrapolating the distribution of the structure variable the SVM makesno as-
sumption about the form of the structure function beyondthe observeddata. Therefore
information about more extremestructure variablevaluescan only be inferredfrom the
observeddistribution of the largeststructure variablevalues,i.e. a purely statistical ex-
trapolation. By comparisonthe JPM extrapolatesthe seacondition variablesand then
builds in knowledgeof the stmcture function beyondthe data when integratingthe joint
density. This omissionof the knowledgeof the structure function from the extrapolation
of the SVM can influenceto the extrapolationas shownby the followingexample.
If A1 and A2 are two structurefunctions,which a.resuchthat A1(x1) : A2(x1) for all
t : I,... , n then the associated
observationson the two structurevariablesare identical.
Applying the SVM to the two data setsgivesidenticalextrapolationseventhough ,A1and
A2 maj be quite different outsidethe observedrangeof the data. Thesedifferencesin
A1 and A2, if known, are incorporatedinto the JPM through the use of different failure
regionsover which the commonjoint densityof seaconditionsis integrated.
Flexibility to cover different designs

If a number of structure functions are to be considered covering a range of designs for
one site or neighbouring sites, or different modes of failure for a given design, then the
whole statistical analysis (i.e. including the extrapolation of the structure variable) has
to be repeated for each structure function for the SVM, whereas a single statistical anal-
ysis is required by the JPM with different integrations of the fitted joint distribution
(/r. j1(x)ax) as the failure region,4,, will be different for each desiga (but ft is indepen-
dent of the design). As the JPM separatesstatistical extrapolation from the specification
of the structure function/failure region it is much more economical on effort when many
designs are considered. Additionally, at the stage of the selection of a provisional form of
26 CHAPTER 2, OUTLINE STATISTICAL METHODS
a designa detailedstatistical analysisis too computationallyintensivefor eachpossible

design.Typically, infeasibledesignsare eliminatedfrom considerationby applyingmeth-
ods such as those describedin Section1.2. Here the estimatedjoint distribution of the
seacondition variablesis required,hencethe JPM estimateof this joint distribution can
be used,so aspectsof this method can be usedin designselection.
Requirements and exploitation of data

To be able to createdata on the structurel'ariable,s.imultaneous
observations
are required
on eachof the seacondition variables.If, for any time period, we haveobservationson a
subsetof the seacondition variablesthen thesecannotbe utilised by the SVM, however
theseincompletedata may providevaluableinformationabout the separateseacondition
variableswhich the JPM is able to exploit. An extremeexampleis when there are no
simultaneousobservationsof the sea condition variables,so no structure variable data,
yet the distribution of separateseaconditionvariablescan be estimated,and the depen-
denceinferred from other neighbouringsites, so the JPM can still be usedto assessthe
probabilitv of failure.
Useful input and output information

For both wavesand still water levelsmuch independentresearchhas provided accurate
informationon the distribution of theseseparateseaconditionvariables.This information
cannot be exploited by the SVM but is a valuableinput to the JPM. A by-product of
the JPM is that considerableadditional information about the processes under study is
obtainedin the form of return levelsfor eachof the separateseaconditionvariables,and
knowledgeof the dependenceEtructurebetweenthe seacondition variables.
In conclusion,the SVM is much the simpler of the two methodsto apply but has
drawbacksconcerningthe requirementto have simultaneousdata on the sea condition
variables,and the necessityto assumethat the form of the structure function doesnot
changebeyond the data. The JPM is dificult to implement which has often led to
independenceor completedependencebeing assumed.Howeverif it is properly applied
the JPM is preferablesinceit separatesthe statisticalanalysisfrom the oceanographic and
engineeringcomponentof the assessment of the probability that the structure variableis
extremeand hencehas wider applicability.
Chapter 3
Extreme Value Methods
In this chapterwe briefly overviewthe extremevalue methodsand ideasthat provide a

basisfor the joint probability and structure variablemethodsof Chapter 2. Specifically:
uniwariate methods for analvsisof
. structure variabledata in the SVM,

r the separatemarginalvariablesin the JPM;
multivariate methods for the dependence

analysisin the JPM.
There is much literature on the subject but the most suitable reviewsare Davison
and Smith (1990),Srnith (19B9)and Thwn (1992)on univariateextremes,and Colesand
Tawn (1994)and Ledford and Tawn (1996)for multivariate extremes.
3.1 Univariate Methods

Consider a sequenceof random variables which have the same distribution at each time
point, i.e. they are identicallv distributed. Here we will consider the distribution of
the annual maximum of these variables and the distribution of exceedancesof a high
threshold.
3.1.1 Annual maximum

The distribution of the maximum, after normalisation,of such a sequenceof size n, as
n -+ oo, is the generalisedextremevalue distribution, GEV. This characterisationof the
limit distribution is subject to the minimal requirementthat observationsof the process
which arewell separatedin time areapproximatelyindependent(seeLeadbetter,Lindgren
and Rootzen,1983). Theseconditionswill typically be satisfiedby all the seacondition
27
28 CHAP'|ER3. EXTHEMEUALUEMETHODS
variables(andhencealsothe structurevariable).If Y fotlowsthe GEV(p,a, {) distribution

it has distribution function
+ e@- Alrl;'/e),
Pr{Y < y} : exp{-11
wherethe notations.. denotesma"x(s,0).The threeparameters

of this distributionare:
r ;r a location parameter,
r o a scaleparameter(" > 0),
r { a shapeparameter.
The level, 9p, exceededwith probability p, i.e. which satisfiesPr{Y > yo}
bv
'u.p
= tt+ o{[- log(t - d]-t - I]l€,
so yo is the return level for return period l/p time units.

The key property of the GtrV for applicationsis that it is the asymptoticdistribution
of the maximum of a samplewhateverthe distribution of the original observations.Thus
if interest is only in the extremevaluesthen this whole family can be fitted instead of
trying many candidatedistributions to fit the original observations. Furthermorethe
theoretica.ljustification for the GEV providesa basisfor extrapolationbeyondthe data
to long return period events
The application of the GEV distribution typically involvestaking the annual maxi-
mum valuesin a sequenceof relevantvariablesas following a GEV distribution. Using
annualmaximumdata, the threeparametersofthe distribution are fitted usingmaximum
likelihood,or someequivalentstatisticalmethodof fit, to giveestimates,u,d and {. The
justification for using annual maximum data is that there are a large number of obser-
vations within a year which justifiesthe useof the limiting distribution, the distribution
has been found to work well in practiceprincipally due to its flexibility, and becauseit
removesthe potential for seasonalbias. The estimatedreturn level, with return period
l/p years,is then
1 t , : P + o [ [ - I o g ( ]- p ) l - ( - 1 ] / € .
The biggestdrawbackin using the GEV applied to annual maximum data is that it
is wastefulof data: the numberof data points to basethe estimationon correspondsto
the numberof yearsof data (oftenlessthan 10). An alternativeapproachis to useall the
largevaluesin the sequence,not just the annualmaximumobservations.In the following
sectionwe review suchapproaches.
3.1. UNIVAHIATEMETHODS
3.1.2 Threshold methods

In this section a brief descriptionof one of the most commonlyused methods for the
statistical analysisof extremevaluesis given. This method is referredto in the literature
as the ThresholdMethod (Davison and Smith, 1990,for example)and is found to be
both flexible and widely applicable.In essence, the ThresholdMethod consistsof flxing
a threshold,u say,and fitting a suitabledistribution to the valueswhich exceedz, while
ignoring observationsbelowz. The main componentsof this procedureare the threshold
u and the distribution of the exceedances of u.
Let {X;,i : 1, ..,"} denotethe sequence which are assumedto be
of observations,
independentand identicallydistributed,@denotethe vectorof unknownparameters,and
g4 denotethe densityfunction ofthe distribution G4 adoptedto describethe exceedances
over u. Under theseassumptions,the marginal distribution of X (a typical observation
from the sequence)is
P(x <t):G6@)P(x > u)+{r-P(x > z)}, f o rr > u ' .
Letting ) : P(X > u), it followsthat the likelihoodfunction associatedwith the Thresh-
old Method is given by
,(I, +) : (1- ,i;"-r"^r"

Jl o4(&), (3.1.1)
Ltl
where1: {1 < i 1n X; > rr}, N is the numberof elementsof 1, i.e. the number
of exceedances of the thresholdu. The parametersS and ) are estimatedby maximum
Iikelihood,whichinvolvesmaximizingthe likelihood,L(^, d, with respectto the unknown
parameters.Sometimes,,\ is unconstrainedand treated as a parameterof the statistical
model (seebelow); in this case,the maximum likelihood estimateof ,l is -nf/n,,i.e. the
proportion of exceedances of the threshold. In other situations,) is explicitly linked to
the parametersof the exceedance distribution(cf. Section4.1.2).
So far no mention has been made of how to chooseG
6, and, in fact, no unique
specificationfor G4 can be given,unlessthe true distribution is known.
However,basedon essentiallythe sameasymptoticargumentsthat justified the use
of the GEV for the annualmaximum above,it can be arguedthat the natural family of
distributions to use in order to describethe excesslevelsof observationsabove a high
thresholdis the generalisedParetodistribution (Pickands,1975).That is, if the threshold
z is sufEcientlyhigh, the conditionaldistribution of a randomvariableX given X > z,
F(z) - F(z)
P(X <rlX >u) : T)M, (3.1.2)
I-F(u)'
30 CHAPTER3. EXTREMEUALUEMETHODS
with F being the true distribution of X, can be well approximatedby the generalised
Paretodistribution,GPD(o,{), with distribution function
1 - {t + {(r - u)loll'/4, forz ) (3.1.3)
wheres-; : max(s,0). The two parameters

of this distribution
. o (o > 0) is a scaleparameter
r { is a shapeparameter.
A specialcaseof this distributionis the GPD(a,0), i.e. when{ : 0: this is the exponential
distribution with distribution function
1- exp{-(r - forr> u,.

")1"},
To fit this statisticalmodelwe let Gq@) be the distributionfunction givenby equation
(3.1.3)with 6: @,0 and we leave) unconstrained. Maximumlikelihoodestimatesof
o and { are obtainedby maximisingequation(3.1.1).
Despitethe GPD havinga theoreticaljustification asa distribution for the exceedances
of a high threshold,this doesnot excludethe possibility of employinga different para-
metricstatisticalmodel,aswill be donein Section4.1.2,wherethe thresholdapproachis
appliedto waveheights,with G6 givenby the TruncatedWei,bullDistri.buti,on.
3.1.3 Properties of the GPD

A key property of the GPD is that it is invariant to the thresholdlevel.
Let X denotethe valueswhich exceedz, and supposethat X - u followsa GPD(o, {).
Then,forz>0.
+r- u)lollltq
Pr{X-u<tlX >t}: t- 11+{(d
[1+
-l
lr + fulo.l+I/E (3.1.4)
whichcorresponds to a GPD(a-,{), whereo* : o + t(u - u). Thusexcesses of the higher

threshold.{i. are alsoGPD.
There is a direct relationshipbetweenthe GPD and the GEV distributions. Davison
and Smith (1990)showthat if there aren observationsin a year,eachwith the probability
) of exceedingthe thresholdu, and where excessvaluesof the thresholdare GPD(o, {)
then
Pr{ annualmaximum I r} : exp{-)rz[l + {(z - for r > tr.

") l"]*t/t}
3.1. UNIUAHIATEMETHODS e1
This is a GEV distributionwith location,scaleand shapeparameters
i, + o{(n.\)€- 1}/{, o(n.\){ and {
respectively.Comparingparameters,we seethat the GEV and GPD shapeparametersare

identical.This propertyhelpsto explainthe potentialbenefitsof the ThresholdMethod
as the parametersof the GPD give the GEV parametersyet can be estimatedfrom a1l
large valuesrather than just annualmaximumvalues. Consequentlythe GPD approach
shouldprovideimprovedprecisionof estimatesoverthe GEV as all relevantdata are used
in the estimation.
3.1.4 Threshold selection for GPD

The drawbackwith the thresholdapproachis that the parameterestimatesare dependent
on the subjectiveselectionof a threshold. The choiceof the thresholdhas to be made
with considerablecare,as Davisonand Smith (t990) show:too high a thresholdand there
are insufficientexceedances to estimatethe GPD parameterswith requiredaccuracy,too
low a thresholdand the asymptoticjustification for the GPD will no longerhold, so it is
likely the GPD will not providea good statisticalmodel for the thresholdexcesses.
For the GPD, a tool which helps in the selectionof a suitable threshold z is the
so called mean resid,uallife plot (Davisonand Smith, 1990). Followingthe property of
thresholdinvariancein equation(3.1.a),i.e. if X is suchthat X - ulX > u is GPD(o,{),
then for all thresholdstt, > u, X - tlx > u is GPD(o + t!t'- z),{), it followsthat
provided { < 1, for ii suchthat o + {(u, - u,) > 0,
o+(iL-u)t
E(X-ulx>u): (J.r.J.l
1-f
Equation (3.1.5) showsthat the mean excessof X over d is a linear function of z. This
result suggeststhe construction of a graph in which the empirical mean excessesof are
plotted against fr, with a taken to be the smallest value of d over which the mean residual
lile plot exhibits approximately a linear behaviour.
3.1.5 Goodness-of-fit for GPD

Once a thresholdz has beenselectedand a suitableparametricdistribution fitted to the
exceedances of z via maximum likelihood,the goodnessof-lit of the statistical model to
the observeddata needsto be verified. Someeasychecksmay be carried out graphically
usingprobability plots (P-P plots) and quantileplots (Q Q plots). Both are basedon the
samelogic,that is, to comparethe fitted and the empiricaldistributionsofthe exceedances
32 CHAPTER3. EXTREME VALUE METHODS
either on a probability or a quantilescale,respectively.Explicitfu for the fitted statistical

modelwith distributionfunctionG(z), andorderedsamplevaluesrG) < re\ I .. . 4 n61:
in the P P olot we olot
G ( r 1 ; ;a
) g a i n st t/ ( n + t ) f o r e = 1 , . . . , n ;
and in the Q-Q plot we plot
r1;yagainstG-r(il(n + 1)) for i : l, . ..,n.
For eachplot departuresfrom the s : 9 line suggestsmodel inadequacy.For assessing

extremevalue modelsthe Q Q plot is more informativesinceit highlights discrepancies
in the uppertail.
3.1-.6 Temporal dependence and non-stationarity

In the previous sectionsit was assumedthat the observedtime seriesare independent
and identically distributed. However,seacondition data depart from theseassumptions,
demonstratins:
a short-range dependence,leading to clusters of extreme values;
non-stationarity, generally due to seasonality,trends and tides.
As far as temporal dependenceis concerned,a declusteringprocedureis requiredto

producesequences of independentobservations.In multiva,riatestudies,the identification
of a suitable declusteringprocedureis generallya difficult step (Colesand Tawn, 1994).
However,for sea condition data a natural way of declusteringis to use only concurrent
meas[rementsat times of high water. At U.K. sitesvariation in the sea-levelprocessis
dominatedby semi-diurnaltidal variation. Consequently, eventswhich threaten to cause
coastalflooding occur only at times of high tide, approximatelyevery 12 hours a,nd26
minutes. This form of declustering
r is consistentwith current practicein oceanographicanalyses(Hawkesand Hague,

1994),
o doesnot differ substantiallyfrom studieson still waterlevel (Dixon and Tawn, 1994,
iee5).
There are additional benefits of this declustering scheme as it also largely removes the
still water level (or surge) variable non-stationarity due to tide, and it enables data sites
which only possesshigh water level data to be analysed.
ILLUSTHATIVE DEPENDEATCEEXAMPLES
3.2 Illustrative Dependence Examples

In this sectionwe motivate the needfor dependence modelsfor extremeevents. In par-
ticular we showthe impact of the degreeof dependence and the form of the failure region
on the probability of a structurefailing.
For a bi yariateproblemwith two identicallydistributedvariables(Xr, &) we consider
two structure variables
Y: min(Xr,Xr),
Y - ma-r(Xr,X:),
these are idealised casesto illustrate extreme design conditions. They are given to aid the
illustration of the methods in a problem where we can explicitly evaluate the probability
of failure, and do not necessarilyrepresent realistic structure va,riables.For example, the
secondstructure function correspondsto failure from two unrelated failure types, in which
an extreme of one variable and/or the other leads to failure irrespective of the associated
value of the other variable.
We also consider three forms of dependencebetween (Xt, &)
o complete positive dependence,
o independence,
r complete negativedependence.
Again these are idealisedand are presentedwith illustration, rather than practice, in
mind.
Throughout we take the structure to fail when Y > u, for large z, and to simplify
notationlet Pr{X1 > z} - Pr{& > u} : p.
3.2.1 The structure function of the minimum

Here we will evaluate the level of protection of the design for the structure variable,
min(X1, X2), for the three different dependencestructures. Writing the probability of
failure in terms of an event for (X1, Xr) we have
Pr{Y > u} : Pr{min(X',&) > u}

: Pr{Xr } u.,X2 } u}
: Pr{Xr > z} Pr{X2 > ulXl > u} (3.2.1)
,,1
CHAPTER3. EXTREMEVALUEMETHODS
wherethe secondterm in equation(3.2.1)is the conditionalprobabilityof X2 > z given

X1 ) u, seeAppendix A.1. The dependence between(Xr, Xz) influencesthe value of the
conditionalprobability with
1 completepositivedependence
under
Pr{& > ulXl > u} - Pr{X, > u} independence
under
0 under
completenegativedependence.
(3.2.2)
Combiningthe expressions
in equations(3.2.1)and (3.2.2)givesthe probabilityof desrgn
failure to be
tp under completepositivedependence
Pr{)' > u} : \ p ' under independence (3.2.3)
Io under completenegativedependence.
To illustrate what this meansconsiderthe situation when z is taken to be the 10 year

return level for eachmarginal variable,and there are 705 independenteventsper year,
the return period (in years)of designfailure is
10 under complete positive dependence

70500 under independence
oo under complete negative dependence,
seeSection3.4 for working. Similarly,for the designto havea 100year return period, u
must be chosento give marginal return periods (in years)for X1 and X2 of

0.38 under independence
Aside: note that Jor complete negati,uedependenceang leuel is adequatefor the marginal
return leuel as d,esignfai,Iure i,s impossi,ble.
3.2.2 The structure function of the maximum

Now considerthe max(X1,X2) structurevariablefor the three differentdependence struc-
tures. Writing the probability of failure in terms of an eveutfor (X1,X2) we have
Pr{Y > u} : Pr{max(X', &) > u}

: Pr{Xr > u and/or Xz > u}
: Pr{Xr > u} +Pr{X, > u} - Pr{XL > u,X2> u}. (3.2.4)
MULTIVARIATE METHODS 35
The joint probabilityin the final term in equation(3.2.\ wascalculatedin Section3.2.1.

It fbllowsthat the probability of designfailure is
under completepositivedependence
Pr{Y > z} :
{;r - p) under
under
independence
completenegativedependence.
(3.2.5)
To illustrate this, considera, to be taken as the 10 year return level for eachmarginal
variablethe return period (in years)of designfailure is
10 under completepositivedependence
5 under completenegativedependence,
seeSection3.4 for working. Similarly,for the designto havea 100year return period z
must be chosento give marginal return periods(in years)for X1 and X2 of

200 under complete negative dependence.
3.3 MultivariateMethods
3.3.1 Models
This sectiongivesa simpleintroductionto dependence modellingin multiva.riateextremes.
We focuson bivariateextremaldependence only and restrict attentionto Pr{X2 > ulXr >
u) where (Xr, Xr) are identicallydistributed randomvariables.In Section3.2 we consid-
eredthree specialcasesof dependence betweentwo variables(Xr, &), and found that for
la.rgez
( t under completepositivedependence
P r { X 2 > z l X r> u } = j p r t x r > u } u n d e r independence
[ 0 under completenegativedependence.
Clearly many diflerent statistical models for intermediate degreesof dependencebetween
(Xr, &) exist, however for joint probability methods it is the dependencebetween the
extreme values of X1 and X2 that a,reof interest, and the form of this dependenceis well
identified by considering how Pr{X2 > ulxl > z} behavesfor large z, i.e. how does the
knowledge that X1 is big influence the probability that X2 is big?
Here we will give this conditional probability for two statistical dependencemodels
that are often used in ioint nrobabilitv studies.
36 CHAPTER3. EXTREMEUALUEMETHODS
Bivariate Extreme Value Distribution

Colesand Tawn (1994)proposethe useof this statistical model for dependencebetween
extremevalues. After transformationof eachmarginal variableto a unit Fr6chetdistri-
bution (i.e. with distributionfunctionPr{X1 < r} : Pr{& < r) : exp(-l/r) for
r > 0) the form of the dependencebetweenthe two va,riables(X1, X2) is taken to foilow
a bivariate extremevalue distribution (with logistic dependence
structure) given by
P r { X ; < z 1 , X 21 r z } : exp{-(:1""+*;'t"r"t, (3.3.1)
where 0 ( a ( 1 determinesthe degreeof dependence.For general (identically dis-

tributed) marginal variableswith this form of dependence
structure
Pr{X1 < u, X2 I u} : (Pr{Xr < ,})'" ,
where 0 < a ( 1 determinesthe degreeof dependence

with
o a: 1 corresponding
to independence,
. o -+ 0 correspondingto completedependence,
o decreasinga leadingto increaseddependence,
o negativedependence
beingimpossible.
For this joint distribution it followsthat
Pr{x1
>u'x2>u} <u'x2<u}
: l;"ii; ;1, i;.4:= l;""'t"'
- (2 - 2") pr{x1 > u} + z'-1(2" - t) pr{X1 > u}2, (J.3.2)
for large z since:
Pr{xr < u}2' - (t - P.{x, > u})"" * r -2"pt{X1 > u} +2"-1(2" - 1)Pr{X1 > z}2.
Hencethe conditionalprobability is
P.{& > zlXl > z} x (2-2") +2a-1(2a- t)Pr{& > z} for Iargez.
This statistical model givesa very specificclassof dependencemodelsbetweenthe vari-

ableswhich leadsto dependencebetweenthe extremevalues. To clarify this, note that
given the variableX1 is extreme(i.e u -+ oo so that P.{X, > z} -+ 0) there is a reason-
able probability, (2 - 2"), that X2 will be equally extreme.The degreeof dependencein
the extremevaluesis determinedby the value of a, so that when o : 1 the first term
3.3. MULTIUAHIATEMETHODS a1
of this conditional probability is zero and the secondterm reducesto Pr{X, > u}. For
completedependence(i.e. when o -+ 0) the conditionalprobability is 1.
FinaJIy,in Figure 3.1 we give the joint density contoursfor the bivariate logistic de-
pendencestructure whenthe marginaldistributionsfollow standardnormal distributrons.
The figure showshow dependence, particularly in the joint tail region,growsa"str is de-
creased.When a : 1 the variablesare independentand herethe contoursare circles. For
a < 1 the contoursare elliptical in form but with a distinct pointednesson the diagonal
in the joint upper tail.
Bivariate Normal Distribution

The joint distribution and dependence structure for the bivariate normal distribution
are given in detail in Appendix A.2.2. For general (identically distributed) marginal
variables with this form of dependencestructure, Ledford and Tawn (1996) show that to
a reasonableapproximation
Pr{X1 > u.,X2 > u) *t C, (Pr{Xy > u})2/(1+e)
for large z, where C, is a constant and -1 < p < I determines the degreeof dependence
with
o 0 ( p< 1 correspondingto positive dependence,
. p :0 correspondingto independence,
o -1 ( p < 0 correspondingto negativedependence.
. It follows that
P.{& > zlXl > z} = Co(Pr{X1 > u})G-e)/o+e\for largeu. (3.3.3)

From equation (3.3.3) the conditional probability decreasesto 0 as u ircreases (whatever
p), with the conditional probability increasing as p increases(for fixed u). This distribu-
tion has dependencein the extremes, but unlike the bivariate extreme value distribution,
the probability of (X1, Xr) being simultaneously extreme tends to zero as more extreme
values are considered.
Finally, in Figure 3.2 we give the joint density contours for the biva,riate normal de-
pendencestructure when the marginal distributions follow standard normal distributions.
The figure shows that the contours are exact ellipses whatever the correlation coefrcient,
p. When p is negative the contours are identical to contours for -p after reflection a"round
the g: z line. When p:0 the variablesare independentand the contourscircular, as
for a: 1 for the bivariate logistic dependencestructure above.
38 CHAP'IER3. EXTREMEUALUEMETHODS
Figure 3.1: Joint density contoursfor the bivariate logistic dependencestructure when
the marginaldistributionsare standardnormal,i.e. Z;: O-1(exp(-l/Xa)) for X; a unit
Frdchetvariableand i : 1, 2. The dependence parametera : 1,0.9,0.8,0.7,0.5,0.3.
alpha=1
-3 -2
alpha=0.8 alpha=0,7
-3 .2
alpha=0.5 alpha=0.3
-3 -2 -1
3.3. MUI.JTIUARIATE
METHODS 39
Figure 3.2: Joint density contoursfor the bivariate normal dependencestructure when
the marginaldistributionsare standardnormal,i.e. Z - O-1(exp(-l/Xr)) for X; a unit
Fr6chetvariableand i : 1,2. The dependence parameterp: -0.5, -0.2, 0.0,0.2,0.5,0.8.
rho=-o.5 rho=-o-2
rho=o.0 rho=0.2
rho=o.5
40 CHAPTER3. EXTREME VALUEMETHODS
3.3.2 Diagnostics
In applicationsit turns out to be important to distinguishbetweenthree forms of depen-
dencestructure illustrated by
1. the bivariate extremevaluedistribution
2. the bivariate normal distribution
3. independence.
The underlyingreasonis that thesedependence modelshavedifferentbehaviourofPr{& >

ulXl > ttl; for large u.
If, as z -+ co, Pr{X2 > olXr > tl} -+ c > 0, where c is a constant,as is the casefor
the bivariate extremevalue distribution, then precisedependencemode)lingis often not
important, as subsequentinferenceis relativelyinsensitiveto the valueof the constantc.
This suggestsadopting a conservativeapproachof taking c: 1, i.e. taking the variables
(Xr, Xz) to be completelydependent.To clarify this, for the logisticmodelin the bivariate
e x t r e m ve a l u ef a m i l yc, : 2 - 2 o ,s o c : 1 w h e n a : 0 .
In contrast,for other dependence structures,suchasthe bivariatenormal distribution,
Pt{Xr > zlXl > a,} -+ 0 as u -+ oo, with the rate of convergence to zero determinedby
the degreeof dependence.For this classof dependence structuresit is important to model
the dependence structure. However,a specialcaseofthe bivariatenormal distribution, for
which no explicit dependence modellingis required,is whenthe variablesare independent
( i . e .p : 6 1
Distinguishingbetweentheseclassesof dependencestructure is important as it en-
ablesus to identify when we can approximatethe dependence by completedependence,
independence,or need to carefully model dependence.The following diagnostic,7(z),
enablesus to identify from the data to hand which approachto take.
Specificallyconsideridentically distributed random variablesX1 and X2 with each
followinga unit Fr6chetdistribution,i.e. Pr{X1 > u} - Prt& > u} x u-\ for largetr,.
Definethe diasnosticstatisric
Pr{X1 > uz, X2 } uzlz

rtrl:-r"s{ (3.3.4)
Pr{X1 > 'u,X2> uj
for large z and all z > 1. The diagnostictest is to plot ?(z), estimatedby replacing
the probabilitiesby the associatedempirical proportions,againstlogz for a range of .a.
If Pr{X2 > ulXr > u} -+ c > 0 then 7(z) is approximatelyzero for aIl z, whereas
for Pr{X2 > ulXr > z} -r 0, we have fhat T (z) is linearly increasingwith log z, with
the gradient related to the rate at which P.{& > zlXl > u} -+ 0. Specifically,the
3.3- MULTIVARIATE METHODS 4I
faster the convergence of the conditional probability to zero the steeperthe gradient,
with independencecorrespondingto a gradientof one.
To illustrate this theoreticallyfirst considerthe bivariate extremevalue distribution
with logistic dependence structure. For largez this distribution gives
Pr{X1 > u, X2 > u} ='

;t"
Then for large u
r(z)x -:'s{@{#+} : -.,1'y: o

Similarly, for the bivariate normal distribution, with Fr6chetma,rginaldistributions,
Pr{X1 > u,X2> u} = C ou-2/(r+o)
Then for large u
?(z)r-rog : -rog1,to-'111t+o)}
-fi^"", (3.3.5)
{ ,,*e*+#}
so is linearlyincreasing in logz, with gradient(t - p)lQ + p) for all p < 1. When p - 0,
i.e. independence, then expression (3.3.5)givesthe gradientto be one.
In Section7.1.2we usethis form of diagnosticstatistic plot prior to dependence anal-
ysis to identify the line of attack that can be taken.
3.3.3 Illustrative examples

We now return to the illustrative examples of Section 3.2 a"nd examine the impact of
the dependence models discussed in Section 3.3.1 which provide intermediate degrees
of dependence to the extreme cases (complete positive dependence, independence and
complete negative dependence)consideredthere.
Structure function of the minimum

For the structure function Y : min(Xr, &) the probability Pr{Y > z} is givenby equa-
tion (3.2.1).Usingequation(3.3.2),for a bivariateextremevaluedependence structure,
with logistic form,
Pr{Y > u) : (z - 2")p+ 2'-1(2" - l)p2,
wherePr{X1 > u} : Pr{& > uJ : p. Similarly,usingequation(3.3.3),we have,for the

bivariatenormaldependence
structure,lhat
Pr{Y > u}:Cp'p?/(t+o)

42 CHAPTER3. EXTREME VALUEMETHODS
To illustrate theseresultswe focuson the bivariateextremevaluelogisticmodel casewith

a : 0.8 (the value used for one of the casestudy data sets in Chapter 5). When u is
taken to be the 10 year return level for eachmarginalvariable,then the return period of
the designis 38.6years.Similarly,for the designto havea 100year return period z must
be chosento give marginal return periodsof 26 years(working is given in Section3.4).
Structure function of the maximum

For the structure function Y : max(Xr, X2) the probability Pr{Y ) z} is givenby equa-
tion (3.2.4).Usingequation(3.3.2),for a bivariateextremevaluedependence structure,
with logistic form,
Pr{Y > u} : 2'p-2"-1(2" -t)p2

x 2"p fot Iargeu.
Similarly,usingequation(3.3.3),for the bivariatenormaldependence

structure
Pr{Y> u} = 2p-Cnp2l(t+o)
x 2p for large u.
Interestingly,for this structurefunction, whateverthe degreeof correlationfor the bivari-

ate normal dependence structure,the probability of failure iends to zero (with decreasing
p) in the sameway asifthere wereno dependence betweenthe variables(i.e. as if p: a)
To illustrate theseresultswe focuson the bivariateextremevaluelogistic model case
with cr - 0.8. When a is taken to be the 10 year return level for eachmarginal va,riable,
then the return period of the designis 5.7 years. Similarly,for the designto have a 100
year return peiiod, u must be chosento givemarginalreturn periodsof 174years(working
is givenin Section3.4).
3.4 Technical Calculations

In this section we derive the mathematical equations through which the probabilities
of joint events in Sections 3.2 and 3.3.3 are converted into annual return periods. We
consider the casesof independent and bivariate extreme value (logistic dependencemodel)
separately. Throughout we assumethere are nq :705 independent events per year, and
z is such that
Pr{ annual maximum Xr < u} : Pr{Xr I u}nt' - 1 - Oo,

3.4. TECHNICALCALCULATIONS
i.e. a is the p;l year return levelfor variableX1 . As Xr and X2 areidenticallydistributed,

u is the return level for X2 as well. It followsthat
Pr{X1 < z} : Pr{X, S : (l - pA)'/"u'.

"}
3.4.L Independence
First consider]' : min(Xr, &). Then
Pr{Y < z} : Pr{Xr < u} +Pr{X, < z} -Pr{X1 < u,X2 < u},
and from this
Pr{ annualmaximumY < ul : -'y,oyt"'" - (7 - poltt"u'1""'.

fzg
Now considerthe Binomial approximation
( 1- p a ) " x ' l - a p + p 2 a ( a - 1 ) 1 2 f o r s m a l l p .
Applying this approximationto the above
Pr{ annualmaximumY < u} r - r'ol x I - p2af

ny,.,
lt "'0,f""
which implies that the return period (in years)of the structure variable exceedingz is
nn,fpza.For the exampleof pa : 6.1 the return period is 705 x 102years.
Now considerthe structurevariableY : max(Xr,&). Then
Pr{Y < z} : Pr{Xr 1 u, X2 I u'
and
P r { a n n u am
l a x i m u mY < u ) : l l - p o 1 ' t " " " 1 " ": ( t - p o ) " x L - 2 p n ,
which implies that the return period (in years)of the structure variable exceedingz is
(2pn)-t. For the exampleof po :0.1 the return periodis 5 years.
3.4.2 Bivariate extreme value dependence

First considerY : min(Xr, Xz). Theo
Pr{Y < ?r}: Pr{Xr < z} +Pr{X2 < u} -Pr{X1 < u,X2 < u},
and
Pr{ annualmaximum Y < u} : - 'po1tt""'- (l - ,o12"/",'1"-
lZlt
44 CHAPTER3. EXTHEMEUALUEMETHODS
Applying the Binomial approximation
Pr{ annualmaximumY < ul x [1 - (2 - 2*)ptlnn,]""'ru 1- (2 - 2")po,
whichimpliesthat the returnperiodofthe structurevariableexceeding z is l(2-2")lrA]-1.

For the example of pa :0.1, with a : 0.8,the return periodis 38.6years.
Now considerthe structurevariableY : max(Xr,X2). Then
Pr{Y < u} : Pr{Xr < u, X2 ! u},
and
Pr{ annualmaximumY <u):l$-ne1t"/"',1"" = (1 - pA)2"Nl-2"po,
which implies that the return period (in years)of the structure variableexceedingz is
(z"pA)-t. The exampleof pA : 0.1 givesa return periodof 5.7yearsfor the design.
Chapter 4
Statistical Models used for JPM
The three componentsof the JPM are
r marginal statistical models
. statistical modelsfor dependence
o evaluationof the probability of failure.
In Sections4.1 and 4.2 we developstatistical modelsfor both marginal and dependence

featureswhilst in Section 4.3 we examine methods for evaluating the integral of the
estimatedjoint densityoverthe failure regionrequiredfor the calculationof the estimated
probability of failure.
4.r Marginal Statistical Models

In this sectionwe developstatisticalmodelsfor eachofthe separateseaconditionvariables
listed below:
. Still \,VaterLevel (tide and surge)
r SignificantWaveHeight
o Steepness
r Wave Direction
. WavePeriod (via statisticalmodelsfor signilicantwaveheight and steepness).
The suitability of thesestatistica.lmodelsis assessed

in Chapter6 wherethey are applied
to all the observational and simulated data sets.
46 CHAPTER 4, STATISTICALMODELS USEDFOR JPM
4.L,1 Still Water Level

The still water level (SWL) is the sea-levelafter waveshavebeenaveragedout. It is the
compositionof the meansea-level,the astronomicaltidal level (tide) and the meteorolog-
ical surgelevel (Surge). The meansea-levelcomponentis treated as a fixed constantin
this work; therefore,variationsin still water levelsare considereda consequence of the
tide and surgecomponents.
There are two approachesto the analysisof extremestill water levels: direct and
indirect (Tawn, 1992; Dixon and Tawn, 1994). The direct methods analyseexplicitly
the observedextremesof the SWL process,whereasthe indirect methodsdecomposethe
SWL into the two constituent components,analyseeachof these separately,and then
re-combineresults to producethe distribution of extremestill water levels. There are
reasonswhy the indirect methodsshouldbe preferred.For example,they keepseparate
the deterministictidal component,for which no extrapolation is required as it can be
predicted exactly,from the stochasticsurgecomponent,which needsto be statistically
modelled. However,indirect methodsrequirethe dependence betweentide and surgeto
be statisticallymodelled.
As we are primarily interestedin the relationshipbetweenextremewater levelsand
wavesin this study we essentiallyadopt a direct approachbut consideroneexamplewhere
a simpleversionofthe indirect style of approachis adoptedfor comparisonpurposes.The
statistical model we use for the still water level (in the direct approach)and the surge
(in the indirect approach)is basedon the Threshold Method, as describedin Section
3.1.2,with the distributionof exceedances modelledparametricallythrough the GPD.
Additionally, in the indirect approachwe take
the tide to be given by the empiricaldistribution of high tide values,
the high tidesand surges(at high tides)to be independent.
4.1.2 Significant Wave Height

The significantwave height (/Js) is definedas the meanheight of the highest 1/3 of the
wavesin a period of a given duration (generally20 minutes). The standard distribution
adopted for statistically modelling significantwave height, Ils, is the three parameter
Weibull distribution, which has distribution function
W(r): Pr{IIs 1 r} : r - exp{-[(o - for x ] a. (4.1.1)

")lb]]
Here the parametersa,bar:dcare respectivelylocation,scaleand shapeparameters.Both
b and c are positive parametersr and since significant wave height is a positive random
4.1. MARGINAL STATISTICAL MODELS 47
variable,we shouldexpecta to be positivealso. This distribution is widely usedfor wave

data and has generallybeenfound to fit reasonablywell throughout the entire rangeof
observeddata (Carterand Challenor,1981).
A procedureoften adoptedby HR for fitting this statistical model is the Threshold
Method discussed in Section3.1.2. After fixing a suitablethresholdu, they fit to the
exceedances a TYuncatedWeibull distribution, definedbv
P(Hs < rlI{s > u) :

w(r) -w(u)
1 - W(u)
r-"*n (T)"], rorx > u, (4.r2)
[- (?)".
wherethe parameters(4, b,c) satisfythe conditionsrequiredfor the Weibull distribution,
i.e. equation(4.1.1),togetherwith the additionalconstraintthat z > a. This statistical
modelis fitted hereusingthe thresholdlikelihood(3.1.1)to the observationsof I15 which
exceedz. Here the densityof the valueswhich exceedthe thresholdis
( c /:,"-a\"-t
'"P1-[/ r - a \ ' /z*0\'l ^
rorr>u
a /*[ a i i;1, a /
and.\, the probabilitythat the thresholdis exceeded,
is linkedto the parametricdistri-
bution for exceedancesby
^ = r - w(u):."o
' t {- (' I ")'} .
\ b t)
Three motivationsfor this approach/fittingmethod are:
e If extremewavesare of primary interestthen fitting the wavedistribution with most

attention on the quality of the fit paid to the largestobservationsis natural. The
choiceof the Weibull distribution is more questionable,but is an obviousstarting
point when moving awayfrom the standardapproach.
. Supposewavescomefrom a numberof differentpopulations,e.g. differentclassesof

wavesdeterminedby their generation which is apparentthrough different direc-
tions or steepnesscharacteristicsof the observedwaves. Then if the wave heights
from eachpopulation follow a standardWeibull form, the mixing of the different
populationsproducesa complexdistribution for all wave heights. However,if one
population produceswaveheightswhich dominatethe extremesin the sample,then
the tuncated Weibull distribution will fit well evenwhen the Weibull distribution
doesnot.
r Hindcast significantwave height data are sometimesgeneratedonly for extreme

storms over the period of interest,for examplefor all stormswhich havesignificant
48 CHAPTER 4, STATISTICAL MODELS USED FOR JPM
wave heights over some threshold level z. Consequently,if all significant wave
heightsfollow a Weibull distribution, then the distribution of theseobservedvalues
is a truncated Weibull distribution.
Relativeto the standa,rdWeibull distribution the TruncatedWeibull is more flexible ow-

ing to the introduction of the threshold, u, which is essentiallya fourth parameterof
the distribution. However,there is no theoreticalargumentto suggestthat the Trun-
catedWeibull shouldbe preferredto other distributions(Carter and Challenor,1981).
By contrast, in Chapter 3 statistical modelsfor thresholdexceedances which have some
theoreticaljustification were discussed.For this reason,we also considerthe Threshold
Method with the GPD usedas the distribution for exceedances for Ifs.
To compare the Weibull and GPD threshold approachesmore generallywe assess
them on the basisof their stability with respectto differentthresholds.As remarkedin
Section3.1.4, a property requiredfrom tail modelsis the non-sensitivity of the results
to threshold selection. To assessthe impact of making this choicewhen in fact the
Weibulldistributionis the true distributionof the population,10000samplesof size6000
(a typical samplesizeof the observationaldata) weresimulatedfrom a three-pa.rameter
Weibull distribution with parametersa, 6 and c. A rangeof valuesof c wereselectedto
representtypical estimatesderivedfrom applicationsto data. Without lossof generality,
we set a:0 and b: 1, sincethe corresponding behaviourfor all otherlocationand scale
parameterscan be derivedfrom this case.For eachsample,the GPD and the tuncated
Weibull distributions were fitted to exceedances of the 95% empirical quantile. Bias
and mean squarederror of the estimatesof variousextremequantilesfor the two fitted
distributions are givenin Tables4.1-4.4. As expected,the bias of the estimatesprovided
by the Weibull distribution is practicallv zero. For the fitted GPD the bias is low, but
the mean squarederror, especiallyfor the most extremequantiles,is consistentlybigger
than for the TruncatedWeibull. This impliesthat evenif the underlyingtrue distribution
werethe Weibull, the GPD shouldprovideacceptableresultsin terms of bias, but at the
cost of greatersamplingvariability.
Our proposedapproachin practical studiesis to fit both the Tluncated Weibull and
the GPD via the ThresholdMethod and to assess the two methodsbasedon their ability
to provide good estimatesof return levelsin the tail of the observedsample,the realism
of the extrapolationsto long return periods,and the quality of fit of the models. For this
project, sincethe GPD tail model is robust to the underlyingdistribution and unbiased
when the true population is Weibull, we adopt it in preferenceto the Tiuncated Weibull
distribution. Henceonly the GPD will be consideredfor the simulateddata, eventhough
both are assessed for the observationaldata.
4.1. MARGINAL STATISTICAL MODELS
Quantiles r-Ll:.0? 1-11103 r-Llro4 1-11105

Bias -0.0005 -0.0286 -0.098 -0.1989
GPD
Mean SquaredError 0.0053 0.0395 0.2549 0.9491
Bias -0.0045 -0.0085 -0.0177 -0.0144
Weibull
Table4.1: Resultsfrom simulating10000samplesof size6000from a Weibull distribution

with knownparameters a:0, b: 1 and c: 1.2.
Quantiles 1-11102 r-1/103 1-11104 1-rlr0s

Bias 0.0004 -0.0108 -0.0595 -0.1453
rI DT'\

Bias -0.0026 ,0.0042 -0.0054 -0.0062
Weibull
Table4.2: Resultsfrom simulating10000samplesof size6000from a Weibull distribution

w i t h k n o w np a r a m e t e ar s: 0 , b = 1 a n d c : 1 . 5 .
Quantiles r - r l : . 0 2 1 - 1 1 1 0 3r - L 1 r c 4 1 - r l t 0 5
Bias 0.001 -0.0107 -0.06 -0.1451
GPD
Bias -0.0023 -0.0033 -0.0041 -0.0045
Weibull
Table 4.3: Resultsfrom simulating10000samplesof size6000from a Weibull distribution

with knownparameters a:0, b = 1 and c: 1.7.
50 CHAPTER 4, STATISTICALMODELS USEDFOR JPM
Quantiles t-rlr02 1-1.1fi3 r-11704 r-11105

Bias -0.0064 -0.0269 0.1476 0.7735
GPD
Bias 0.0079 0.0331 0.081 0.153
Weibull
Table 4.4: Resultsfrom simulating10000samplesof size6000from a Weibull distribution

with knownparametersa: 0, b: 1 and c : 0.7.
4.1. MARGINAL STATISTICAL MODELS 51
4.1.3 Steepness
We model the joint distribution of (1{s,S) insteadof (Hs,Ts) for the wavecharacteristics.
Wavesteepness, S, and Tz arc rclatedby the followingequation
q- 2nHg
--_-=- (4.1.3)
st2
whereg denotesthe gravitationalconstant,so a statisticalmodel for (Ils, S) determines
alsothe distributionof Q.
There are a number of reasonswhich support an analysisbasedon the variable S
insteadof ?2.
o It is commonpracticein oceanographic studiesto estimateextremesof waveperiod
by fixing a typical value of steepnessand employingthe distribution of ffs and
equation(4.1.3)to derive?2 (Alcock,1984). The average,3", of the observed
steepness of the 1% largestwaveheightsis calculatedand the valueof Q associated
to an estimatedextremelevel of I{* is obtainedfrom
rr1_ - (a't 4\
r For most of the sites analysedin this work the relationshipbetweenHs and Tz
tends to assumea quadraticform, whereasthe relationshipbetween1{s and ,5 is
more linear, and hencestatistically modelledmore easily.
e Steepnessis a variableof interest in its own right, rega,rdless

of its relationshipto
11
r There are theoreticalrestrictionsto the maximum valuethat S can assume:theory

predicts that randomwaveswith S biggerthan approximately1:10must break.
No parametricstatisticalmodel is proposedfor the marginaldistribution of steepness.

The statistical model we haveadoptedis basedsimply on the observeddata; that is, the
empirical distribution. A restriction in the use of empirical modelsis the inability to
extrapolatebeyondthe observedrangeof data, but this is not an important issuewhen
modelling S, as we are mainly interestedin wavescharacterisedby high valuesin both
Hs and,T2, which producemid-range/typicalvaluesof steepness.Onesituation in which
extrapolationwith respectto S might be relevantis in the presenceof swellwaves,which
haverelatively low significantwaveheightsand long periods.Historically,however,swell
waveshavebeenomitted in studiesof probability of extremewavesbecause:
r observationson swellwavesare of limited extent, as valuesfor this variableare not
producedby standardhindcaststudiesusing the local wind climate;
I swell waves are only important at some exposed sites;
r swell waves alone are rarely believed to pose a significant threat and, furthermore,
they are deemed to be independent of the other sea condition variables.
4.L.4 Wave Direction

As with steepness, there is no requirementto extrapolatethe wavedirection variable,so
againwe estimatethis distribution empirically.The empiricalestimateof the densityfor
wavedirection is given by the proportion of the observed/hindcastwave directionsthat
fall in eachdirection sector.
In practice,if wave direction is not explicitly in the structure function for the design
under consideration,we shouldtake the direction sectorsas large as possible,subject to
simple statistical modellingof the joint distribution of the other seacondition variables
over this sector. This is the casefor overtoppingdischargerates,seeequation(1.1.1).
Experienceat HR suggeststhat sectorsof widths lessthan 30" are impractical as there
are insufficientdata on the other variableswithin eachsectorfor an adequatestatistical
analysis.
4.1.5 Wave Period

Wave period is a key variablein the seacondition vector. Its behaviourin this study is
describedby the joint behaviourol (fu, S). So the discussionof the choiceof statistical
model here should be treated as expositoryrather than as a basisfor subsequentfitting.
Two possibleapproachesfor developinga statistical model for the marginal distribu-
tion of waveperiod are:
. the useof distributions basedon theoreticalwaveprocesses;
. empirical fbrms basedon observeddata.
The reasonthat neither approachis adoptedin practiceis that the theoreticalmodelsdo

not capture important featuresof observationaldata and that the empirical modelsdo
not provide extrapolationsto waveperiodsbeyondthoseobservedin the data.
Instead,current practiceis to developstatisticalmodelsfor waveperiod throughstatis-
tical modelsfor the conditionaldistribution of waveperiod giventhe associatedsignificant
waveheight measurement.The current approachHR recommendis to take waveperiod,
ft, as being completelydeterminedby the significantwaveheight, .FIs,i.e. the value of
Tz car' be determinedgiventhe associatedIls. The statisticalmodel recommendedhere
is to link thesevariablesvia wavesteepness,S. Taking a typical valueof steepness,

3€,as
the averagesteepness of the observed1% la,rgestsigniflcantwaveheights,then
, - 11.
l 2 t rH s \ ' ' ' (4.1.5)
'
\gs")
Consequently, the implied marginal distribution of wave period is
'Pr{Ts < :
r} Pr{(2rnslks.))1/2 5 x}
: Pr{tls < r2s3"l(2r)}.
Thus the distribution ol Ts is determinedby the distribution of I15. To illustrate this, if

1Jsfollowsa Weibulldistribution,seeequation(4.1.1),then
pr{72<r}:r- e ' r { - ( r 2 g s ' / ( ? ? T } - ' ) " } * , r > ( 2 n asf B , ) L / z . (4.1.6)

t \ b ))
Similarly, if IIs is GPD(a,1) abovethe thresholdu, then
/ x 2 o s . / Z n -) u \ - r l (
Pr{Ts< o}: r - ^ (t **t"t#)', rorr> (Zrufs3")t/2. (4.r.7)
Thesederivationsassumethat,9 is constant.In practiceS is variableand is dependent

on 115. If a statisticalmodel for the joint distributionof (fls,S) were availablethen
the distributionof T7 could be obtainedfrom relationship(a.1.5). To show this let
Fxy(r,A): Pr{X 1 r,Y 1g} denotethe joint distributionfuncticnof generalvariables
X and Y. It followsfrom joint distribution resultsthat
PlTz<d : 1*}Fr":fu's)o,
Jo ds
: /- dFs",s bsr,zlQlT), s) , (4.1.8)

Jn ar ""'
Dependence aspectsof the joint distributionof (Ils, S) are considered in Section4.2.2.
Clearlythe impact of variation in 5, and the dependence of S on fls, are differentfactors.
Here we illustrate the impact of the former by producing a statistical model for the
distribution of wave period using representation(4.1.8) under the false assumptionof
independence of I1s and S. We estimatethe marginaldistribution of 5 usingthe empirical
distribution for ,9. For the marginaldistribution of Ils we usethe ThresholdMethod with
the GPD for valuesgreaterthan the threshold(taken as the 95%empiricalquantile) and
the empirical distribution function for valuesbelowthis threshold. By comparisonwith
equation(4.1.7),where,9 is takento be constant,we have
P{rz
s"}::8","(#) (4.1.e)
54 CHAPTER4, STATISTICALMODELS USEDFOR JPM
wheren is the numberof observations, si, i :1,. . . , n denotethe observedvaluesof S,

and Fs, is the distribution function of I!s, which is given by the GPD for valuesgreater
than the thresholdand by the empiricaldistribution function for valuesbelow z. This is
quite different from equation(4.1.7),whereS was taken to be constant.
To illustrate thesedistributionalmodelsFigure 4.1 showsestimatesof the distribution
function of Ts for observationaldata for Christchurch(describedin Chapter 5). The
estimatesshownare basedon differentapproaches.Theseare:
r the assumptionof independence and S, i.e. equation(4.1.9);

of .F15
o the empirical distribution function of 72;
r the distributionfunctionbasedon the constant,9assumption,

i.e. equation(4.1.7).
The distribution of ft obtainedby fixing S to be S", showsmore substantialdepartures

from the empiricaldistribution for typical valuesof waveperiod,but hasgoodperformance
in the tail, reflectingthe presenceof dependencebetweenhigh valuesof I15 and S. As
this featureis not presentin the statisticalmodelbasedon independence of S and 1J5,the
tail is quite different in that case. However,the introduction of the variation of ,9 gives
a better fit in the bulk of the distribution. This suggeststhat once we have developed
a suitable statistical model for the dependencebetween(Ils,5) we should have a good
descriptionof the whole distribution of ?s.
Figure 4.1: Estimation ofthe distribution function of Tz at Christchurch:estimatesfrom

the empirical distribution of 72, constant,9 and independence of (f/5,,S).
Tz
Christchurch
oo CHAPTER 4. STATISTICAL MODELS USED FOR JPM
4.2 Dependence Models

In this section we describe statistical models for the inter-relationships between the sep-
arate sea condition variables. Typically the analysis of the joint distribution of the sea
condition variables will be for an offshore location (a 20m depth is the usual depth limit
to which waves can be hindcast without the use of a site-specific wave transformation
model). At such a location the sea condition variables generally exhibit, or are taken to
have, the following relationships:
r Surges and Waves ale dependent.
e Significant wave height, wave period and direction are dependent.
o Tides and Surges a,redependent processes.
r Tides and Waves are independent.
There are essentially two physical mechanismswhich generate the sea condition variables
1. astronomical
2. meteorological.
Variables generated by the different physical mechanisms are independent unless they
interact with each other. Offshore, the water depth is sufficient for the waves to be
effectively independent of tides (not exactly independent becauseof the effect of currents),
but the surges typically depend on the tide. In shallower water, waves become dependent
on the tide, as the water depth influences the speed of the waves and through depth-
dependent frictional effects, causesthem to break.
That the tides and surgesare dependent processesin shallow water areasis well known.
Prandle and Wolf (1978) discuss the nature of observed interaction, which subsequent
numerical model results show to be remarkably well captured by the known dynamical
processesof depth basedfriction effects (Flather, 1987). However, if only high water levels
are considered then, to a reasonableapproximation, the associatedtide and surge levels
are independent.
Since surge and waves (significant wave height, wave period and wave direction) are
both influenced by the wind climate, it should be expected that there is a relationship
between these variables. The form of dependenceat a site dependson the coastal location
of the site, its associated fetch length and direction. For some sites positive dependence
between surge and significant wave height is expected. At other sites negative dependence
(or independence) between these variables is expected.
DEPENDENCEMODELS
Often the dependencebetweenstill water level and wave characteristicsis required

slnce the still water level is not decomoosedinto its constituent comoonents of tide and
surge. Dependenceis anticipated betweenwavesand still water levels, but is almost
certain to be closer to independencethan the dependencebetweensurgesand waves,
sincethe addition of the 'independent'tidal level masksthe relationship.
We needto describethe full joint distribution of the seaconditionvariables,X. Each
separatevariablehas beenwidely studied,seeSection4.1, and, as discussedabove,phys-
ical understandingexistsfor thc presenceof dependencebetweenSWLf Surge and IIs.
However,little appearsto be known about the pairwiserelationshipsbetween
o IIs and S,
o SWLf Surge and S,
o Surge and wavedirection(9),
or of the higher dimensionalrelationships.Through somelimited studieswe haveidenti

fied:
t . strongerdependence
betweenthe pairs (SWL,Hs) and (ffs,,9) than betweenthe
pair (^9I7.L,
S);
2. strongerdependence
betweenthe pairs (SWL.,Hs) and (IIs, d) than betweenthe
pair (SWL,0).
Thesefeaturessubstantiatethe argumentthat surgescan only be linked to wave steep-

nessand directionthrough the occurrenceof waves,i.e. both pairs (Surge,S) and
(Surge,0)arerelatedthroughI15. This corresponds between
to conditionalindependence
SWL/Surge and,9 given Ifs, and betweenSWLISurge and d given I15. A graphical
representationof the claimeddependencestructure is as follows:
tides surges - wave height - wave steepness
Here linesjoining the variablesshowa direct relationshipbetweenthat pair of variables;

and a variableIinkedto a secondonly througha third variableis conditionallyindependent
of the secondvariablegiventhe third variable.
Wave direction is omitted from this graphicalrepresentation.For purposesof sta-
tistical modellingwe examinethe abovestructure for separatewavedirection sectors,so
separateanalysesare performedfor eachdirectionsectorwhich appearsto havea different
impact on the distribution of the other variables.
A key adrantageof decomposingthe joint distribution of X into the conditionalin-
dependencestructure shownaboveis that to examinethe ioint distribution we onlv need
5B CHAPTER 4. STATISTICALMODELS USEDFOR JPM
to examinethe joint distribution of linked pairs, i.e. conditiona,llyon the wa,vedirection

we estimatethe joint distributions of
o (SWL,Hs)
o (fIs, S).
In this casewe may write the joint densityof (S\UL, Hs, S) as
z)
f swt,ur(n,a)f ur,s(Y,
fo"(a)
wherethe subscriptsindicate the marginaland joint distribution variables.Equivalently,
the joint density can be rewritten as
f swt',a (t, Y)f Ea (zlY),

" "
where /s1s" is the conditionaldensityof Sl//s. In Sections4.2.1and 4.2.2we exam-
ine statisticalmodelsfor estimatingthe joint distributionof (SWL, Hs) and (Hs, S)
respectively.More specificallyin the latter casewe describethe conditionaldistribution
,Slfls. Together,thesedistributionsprovidethe full joint distribution of the seacondition
variables.
In the followingsectionsit is easiestto describethe dependence
betweenthe variables
ifthe marginalvariablestake a standardform, suchas a standardnormal distribution (see
Appendix A.4.1). It is possibleto transformany continuousrandomvariableto any other
by repeateduse of the probability integral transform (seeAppendix A.3). Specifically,
the random variableX with distribution function F71can be transformedto a standard
normal random variableX* bv
x-: o-1(rx(x)),
whereO-r is the inverseofthe distribution functionofa standardnormal randomva.riable.

Thus we havethat
SWL- a-l(FswL6wL)), (4.2.1)

!u
s* o 1(Frrs(r1s)), (4.2.2)
r) o-'(F"(s)), (4.2.3)
are standard normal versions of the random variables SW L, Hs and S respectively.

4.2. DEPENDENCEMODELS
4.2.L Models for dependencebetween SW L and f/5

In this section three statistical models for the dependencestructure of the variables
(SWL,Hs) are proposed.Thesedependence modelsare for the marginal variablesaf-
ter transformationto standardnormalform (SWL.,H!):
the bivariate normal model;
the bivariate normal threshold model;
the mixture of bivariate normals model.
The second and third models are both extensions of the bivariate normal model.
Bivariate Normal Model

Herewe assume(SWtr-,14) - BVN(O,D) (for notationseeAppendixA.4.1)where
'- t pl
"-\o/ ti
This statisticalmodelis fitted usingthe likelihoodgivenin AppendixA.4.1. The maxi-
mum likelihood estimatorof p, the singleparameterin this statisticaldependence
model,
is approximatelythe samplecorrelationbetween(SI4zI-, Hi) pairs.
Bivariate Normal Threshold Model

The bivariate normal model aboveassumesthat there is the samedegreeof dependence
betweenSW Lr and 1{i whatevervaluesthesevariablestake. However,the dependence
betweenstill water levelsand wavesis expectedto changeas more extreme eventsare
observedsincethe larger the surgeeventthe lessthe maskingeffectof the tide and hence
for large still water levelsthe true relationshipbetweenthe surgeand wavesis identified.
This is becausetides a,reindependentof the other variablesand so weakensany form of
dependencebetweenthe surgeand waves.
Thus, to capturethis potentialfor changingcorrelation,we adopt a stat.isticallymodel
consistingof a bivariate normal dependence structure in a regionwhereboth SWL. and,
IJi are sufficientlylarge. Explicitly, for SWL. ) u and IIi > u for somethresholdlevel
u, we take the joint density for (SW L.,11.i) to be identical in form to that of bivariate
normalrandomvariables,i.e. BVN(0,D) where
/"\
,- I r P" I
.--\r,
r )
60 CHAPTER4. STATISTICALMODELS USEDFOR JPM
This model is only completelyexplicit in this joint extremeregion. The known marginal
distributions and information derivedfrom this joint extremeregionimposestructure on
the positionsof observationsin regionswhereone or other variableis large,and also the
region whereneither is large
For a given threshold this statistical model is fitted using the threshotd likelihood
given in Appendix A.4.2. The maximumlikelihoodestimatorof p,,, the singleparameter
in this statisticaldependence model,is essentially the correlationbetween(SWL-,Htr)
pairs which are simultaneouslyabove the threshold z in each variable. However,the
estimation of this parametertakes into account the number of observationsin each of
the four regionsdeterminedby the thresholds,and for the regions(,9W.L.> u, Hfi > u),
(SW L- > u, HI < u,) and (SI4z.L.< u, HI > u,) the sizesof the thresholdexceedance
valuesare accountedfor.
We can treat p,, as a function of the common marginal threshold, a. For u less
than the minimum data value this estimateis approximatelythe estimateof p for the
bivariate normal model, i.e. the correlationcoefficient,whereasfor large z this measures
the dependencebetweenthe extreme pairs. If in fact the data were from a bivariate
normal distribution then p,, would be independentof a.
A thresholdhas to be chosenfor this statisticalmodel. What is requiredis that u is
high enoughso that the degreeof correlationdoesnot changeabovethat threshold. A
suitablethresholdcan be identifiedfrom the estimatedp.,function, by taking the smallest
z abovewhich the function is constant.
Mixture of Bivariate Normals Model

A disadvantageof the bivariate normal thresholdmodel is that the dependenceis taken
to be constantabovea thresholdlevel. If the correlationfunction p,, showsno evidence
of constancythen fixing dependence to be constantabovea thresholdcan lead to bias in
extrapolationsof the dependence structure. Thus there is a needto model the changein
the correlationfunction with threshold,to increaseconfidencein extrapolations.
One idealisedexamplein which the correlationfunction r.arieswith thresholdlevel is
as follows. Suppose
r there are two distinct 'meteorologicalmechanisms'which producestill water level
and waye eyents.
each different mechanismproduces different Ievelsof dependencebetween still water

level and waves.
from the (SW L, H s) data only, we cannot identify which of the different mechanisms
produced a particular data pair.
4.2, DEPENDENCEMODELS
A crude example of this is given in situations where
1. regionally generated waves and still water levels are dependent, but
2. locally generated waves and still water levels are independent.
A consequenceof such a structure is that the dependencebetween still water level and
significant wave height is a mixture of two different dependenceforms. If one component
of the mixture describesthe dependencebetween low values and the other between high
values then the correlation function pu will change with z to give the correlation of the
second form for the largest events.
The statistical model here is more complex than for the above cases,being given by
Appendix A.4.3. For each form of dependencea bivariate normal distribution is used to
describe the variations. The overall statistical dependencemodel has seven parameters,
which can be fitted by maximum likelihood:
r one parameter pM: determining the proportion ofthe data of each dependencetype
(pu : 0 or pt,r : 1 suggests that there is only one type of dependence and a
bivariate normal model would fit well);
. two parameters p1 and p2: which measure the correlation associatedwith each type
of dependenceform; and
o four parameters p, = (gzt, Fzz),ozt, o22: which explain the differencesin mean level
and variation between events generatedby the two mechanismsafter transformation
to the standard normal marqinal scale.
This statistical model has two maior benefits over the bivariate normal threshold model:
1. the model is fitted with all the data having equal weight contrasting with the thresh-
old model where the most extreme values are given most weight;
2. the model fits and extrapolates caseswhere the dependencecontinues to change

throughout the data and beyond, so the arbitrary specification of a threshold is
avoided.
On the negative side the main disadvantageis that the statistical dependencemodel
is more complexto fit and parameterestimatesmore dificult to interpret. Critical to
its application is testing whether two, or one, bivariate normal dependenceforms exist.
Unlessthere is significantevidencefor two forms, only one (i.e. the standard bivariate
normalmodel)shouldbe used.
62 CHAPTER 4. STATISTICAL MODELS USED FOR JPM
4.2.2 Models for dependence between Hs and. S

In this sectionwe developstatistical modelsfor the distribution of S conditionalon the
variable115.The current approachusedby practisingengineersis to take this distribution
as degenerate,i.e. the steepnessis given a fixed estimate,3., irrespectiveof the value of
11s. The estimate6" is given by the averagesteepness of the 1% largestsignificantwave
height values.
We proposea model for the distribution of ,9 given lls which
. dependson the value of I1s,
r has variation in S valuesfor a givenI15 value,
r containsthe existingmodel as a specialcase.
In the transformedstandardnormal (Gaussian)randomvariablespace(I13,,9-),seeequa-

tions (4.2.2)and (4.2.3),we statisticallymodelthe conditionaldistributionof 5-1119sub-
ject to Ifi being sufficientlylarge. Our reasonfor focusingon large fli, say f1j > z, is
that we are only interestedin 5 conditionalon IIs being large. Specifically,we take
s - l ( H ; - h ) - l r ( o + b ( h - u ) , o z ) ,f o rh > u (l) L\
i.e. a linear statistical regressionmodel for ,9- with explanatoryvariable h - u, having

intercept,gradientand varianceparameters(a,b,o2) respectively,and a Normal distribu-
tion for the error variable.
This method of describingthe dependenceis strongly related to the assumptionof
a bivariate normal dependencemodel (seeAppendix A.4.1) between,9- and I1j. For
example,if (S-,14) follow a bivariatenormaldistributionwith correlationp, then the
parametersof equation(4.2.4)would take the form
a:0
b-- p
tt2 : I-p2.
Thus the statistical model we propose is slightly more general than a bivariate normal
dependencestructure as our model does not require a zero intercept to the linear model
or a link between the residual variation and the gradient of the linear model.
Finally we have the problem of threshold selection again. The choice of threshold is
less important than for other aspects of the modelling of the joint distribution as results
are reasonably insensitive to threshold choice here. To aid threshold choice we produce
plots of
E(S-lIlj > z) and Var(S-lI1} > z)
EVALUATION OF THE PROBABILITY OF FAILURE USING?HE JPM 63
using equivalentsamplebasedmoments,for a range of potential thresholdvalues. We

look for approximatelinearity in the expectationand constancyin the varianceas indica-
tors that a sufficientlyhigh level has beenused.Theseplots are best usedfor identifying
whether the dependenceis positive,negative,or independence.Positive/negativedepen-
denceis indicated by rising/falling valuesfor E(.9-ll{i ) u), whereasfor independence
both the aboveplots showconstancy.
Strictly theseare not the most usefulplots for thresholdselection,sincefor the linear
regression modelto be appropriate, we needE(S-1113 : u) to be linearand Var(S-lfl! :
u) to be constant. Plots basedon thesefeaturesrequirecommonI{ valuesto be pooled,
so herewe havethe problemofthe choiceofthe degreeof pooling. In Chapter7 we adopt
the simpler, though lessinformativeapproachwhereconditioningis basedon 11$> z.
4.3 Evaluation of the Probability of Failure using the

JPM
As describedin Section2.2 the final step in the implementationof the JPM is the evalu-
ation of the integralexpression
Pr{xe A,l: [ ft(x)dx,

JA"
where / is the estimated joint density of X and A., is the failure region. This may seem
simple, but the joint density of the sea condition variables is (4/5)-dimensional and the
set over which the integration is required (the failure region) is a complex set, so this
integration requires some care.
No analytical expressionis obtainable in these problems so numerical methods of some
form need to be used. There are two approaches:
e numerical integration;
o simulation.
There are many numerical methods for integration, but in practice these are not ideal
when the dimension of the integral is relatively high and the form of the failure region is
non-linear. Therefbre the simulation approach is proposed, which involves a three stage
procedure:
1. simulation of a large number, n, say, of pseudo observations from the fitted joint
distribution of X;
2. evaluation of the structure variable, A(X), for each pseudo observationl

3. evaluationof the integral asthe proportionof the pseudoobservationson the struc-

ture variablethat exceedthe critical level u. Letting X, denotethe sth simulation
from the joint density f, then the Monte Carlo estimateof the integral is
numberof the points X,, s : 1, . . . , rz, in the set ,4.,

'vs
Step 1 aboveis generaland independentof the form of the structure function. Thus if
a wide rangeof designsare to be consideredthis step only needsto be undertakenonce.
The output in the form of a seriesof eventsso is of interestfor a rangeof designstudies.
By itself this is a valuableoutput from the analysisfor coastalengineers.
Increasingz, improvesthe precisionof the evaluationof the estimatedprobability of
failure. In practicen, shouldcorrespond to 10-100times the length of the return pe-
riod of interest,so the largestsimulationsamplesizesare requiredfor the most extreme
extrapolations. This can be computationallyintensivealthough straight forward to im-
plement. More generally,methodsexist for improvingthe precisionwithout increasingn".
Thesemethodsuse importancesamplingin the simulationstage,which involvessimulat-
ing a dis-proportionatenumber of extremeseacondition observationsat the simulation
stageand appropriatelydown-weightingthe proportionof failuresin the estimationstage.
More specifically,pseudoobservationsare generatedfrom a suitably chosenjoint distri-
bution with density,9(x), giving a disproportionatenumber of extremeX values. The
importancesamplingfactor for a simulatedvalueX is
/x(x)
e(x)
i.e. the fitted joint density relative to the joint density used for the simulation of the
pseudoobservations.So Step 1 involvesgeneratingn, valuesX;, from g, with importance
samplingfactorsft(X6)/9(X;) for i: I,. .., n,. The estimatedprobabilityof failure,i.e.
Pr{X e ,4"}, is then givenby
1i i1x,e A,)ft(x;)/e(x,).
where
if X,€A,
IIY.cA.'_I'
.,.-,-,,,,_
I 0 if X,#,4,.
This method is less computationallyintensivethan the direct simulation method but
requiressomeexperiencein the selectionof an appropriatejoint densityfunction g.
In this report we haveusedthe importancesamplingtechniquesin Chapters7 and 8,
but for applicationsof the JPM to the estimationof the probability of failure in Chapter
EVALUA'TION OF THE PROBABILITY OF FAILURE USING THE JPM
I we found the direct approachwas suitable. In other applications,where the structure

function is too computationallyintensiveto evaluatemanythousandsof times,importance
samplingmay be necessary.
CHAPTER4. STATISTICALMODELS USEDFOR JPA/I
Part II
Examples of Data Analysis using the

Joint Probabilitv Method
67
Chapter 5
Study Data
Within the project two formats of data wereconsidered:
1. Observationa.l
data consistingof measureddata and hindcastdata;
2. Simulateddata - consistingof syntheticallygenerateddata from a knownstatistical

model.
In this chapterwe describethe two differenttypes of data setsusedin the study, and for
the simulateddata give the statisticalmodel usedto generatethese.
5.1 Observational data

A numberofsites wereidentifiedashavinginterestingcharacteristics relevantto the study.
Primary interestwas in obtainingsiteswhich had a wide rangeof dependence structures,
so sites with negativedependence,independenceand positive dependencebetweenstill
water levelsand waveswereselected.Siteswith the still water level separatedinto tides
and surges,and where the directionality of the wave processesis influential to the joint
distribution ofthe other seaconditionvariables,werealsoselectedto enablethesefeatures
to be examined.An additionalcriterion wasthat eachsite shouldhavea numberof years
of data available.
Six sites which possessedthe relevantpropertieswere Cardiff, Christchurch,Dover,
Dowsing,North Walesand Shoreham.Table 5.1 givesdetails of the data for thesesites.
In eachcase,the wavedata are hindcastvalues. Other siteswheremeasuredwave data
are also availablewerestudiedto a lesserextent.
69
70 CHAPTER5, STUDYDATA
Site Variables Sampling Years

Cardiff SWL, Surge,Waves High Waters 60-87
Christchurch SWL, Surge,Waves Hourly 7B-90
Dover SWL, Surge,Waves Hourly I I IY
Dowsing SWL, Surge,Waves Hourly 78-87

North Wales SWL, Surge,Waves High Waters 70-83
Shoreham SWL, Surge,Waves High Waters B 1 9 1
Table 5.1: Information about the observationaldata sites: SWL - still water level,Waves
significantwaveheight, waveperiod, wavedirection.
5.2 Simulated data

Five data setshavebeensimulatedto reproducethe main featuresof five of the six main
data sets(excludingShoreham).The true statisticalmodelfor the joint distributionof
the five simulateddata sets (Sim1- Sim5) is describedbelowfor eachdata set.
In eachcasewe havesimulated7000independentpoints, approximatelyequivalentto
10 yearsof separatehigh water levels.Marginal aspectsof the joint distribution are taken
as either
the empiricaldistribution of the variablefrom oneof the observationaldata sets,or
basedon parametricmodelsfitted to the observationaldata.

Dependence aspects are not as representative of the characteristics of the observational
data sets as the marginal features. A number of dependencemodels were selectedto have
different, but simple, dependenceforms and a range of dependencelevels.
For each simulated data set we generated complete data on the vector
r still water level (SI4/Z). (In addition for Sim3, the tide and surge constituents were
separately generated);
r significant wave height (I15);
o wave direction (d);
r wave period (ft).
In the simulations wave period was obtained by generating Ils and the waue steepness.,
,5, and using the relationship
^1z: 1znl:.,rltl'
| -----;-
gJ
| ,
\ J
5.2. SIMULATEDDATA
whereg is the acceleration

dueto gravity.
Siml data
Marginal distributions
The marginal distribution of d is taken to be the sameas the equivalentempiricaldistri-
bution of the Cardiff data. For both SW L anclHs a mixture of the empiricaldistribution
and parametricmodelsareused.For ,9172belowthe 95%empiricalquantileof the Cardiff
SW L data, the empirical distribution is used;abovethis threshold,the distribution fol-
lows a generalisedParetodistribution(seeChapter3, equation(3.1.3))with parameters
o-0.362and{:-6.264.
Both ^9and f15 havedistributionswhich dependon d:
oWhen0'<d<110":
For f15: belowthe 95%empiricalquantileofthe Cardiff ffs data from this direction
sector,the empirical distribution is usedwhereasabovethis threshold,the distri-
bution follows a generalisedPareto distribution with parameterso : 0.1958and
€ : -0.102.
For ,9 the empiricaldistribution of the Cardifl data in this sectoris used.
o When110'<d<360':
For ff5: belowthe 95%empiricalquantileof the Cardiff Ils data from this direction
sector,the empirical distribution is used,whereasabovethis threshold,the distri
bution follows a generalisedPareto distribution with parameterso : 0.1417and
( : -0.0604.
For S, the empiricaldistribution of the Cardiff data in this sectoris used.
Dependence Structure
After transformationof eachmarginalvariable(Hs, SW L, S) to a standardnormal distri
bution, the form of the dependence
betweenthe three variablesis a trivariate normal dis-
tribution with zero-meanvector (seeAppendix A.2.2). Howeverthe variance-covariance
matrix, X, dependson wavedirection,with D givenby
1 0.3 0.62
0.3 1 0.3x0.62
0.62 0.3x 0.62 1 )
whend<1l0",and
1 0.3 0.742
0.3 | 0.3x 0.742
0.7420.3x 0.742 )
72 CHAPTER5. STUDYDATA
whend > 110'.

This dependence structurecorrespondsto positivedependencebetween(Hs,SWL)
and between(I1s,S), but with (SWL,S\ only relatedvia their relationshipwith I15.
The directionality of the wavesinfluencesthe dependencestructure by increasingthe
dependencebetween.FIsand ,Swhen g > 110".
Sim2 data
The marginal distributions of d and S are taken to be the sameas the equivalentempir-
ical distributions of the Dover data. For both SWL and F1sa mixture of the empirical
distribution and the parametricmodel is used. Specifically,for SW L belowthe 95% em-
pirical quantileof the DoverSW L data,the empiricaldistribution is used,whereasabove
this threshold,the distribution followsa generalisedParetodistribution with parameters
o : 0 . 1 8 6 6 a n d€ : - 0 . 0 7 6 4 .
Significantwaveheight has a distribution which dependson the wavedirection:
For0'(d<110":
belowthe 95%empiricalquantileof the DoverIls data from this directionsector,the
empiricaldistribution is used,whereasabovethis threshold,the distribution follows
a generalisedParetodistributionwith parameters o :0.731 and { : -0.4215.
For110"<d<360':
below the 95% empiricalquantile of the Dover fls data from this direction sector,
the empirical distribution is used whereasabove this threshold, the distribution
followsa generalised Paretodistributionwith parameters o :0.46 and { - -0.112.
The direction,d, only influencesthe marginaldistributionsof the variables,but not their
dependencestructure which is assumedto be constantacrosssectom.'Ihe (Hs.,SW L, S)
variablesare taken to be dependent.After transformationof eachmarginal variable to
a standard normal distribution, the form of the dependencebetweenthe (11e,SWL, S)
variablesis a trivariate normal distribution with zero-meanvectorand variance-covariance
matrix, X, given by
1 -0.2 -0.1
F- -0.2 1 -0.2x-0.1
- 0 . 1 - 0 . 2x - 0 . 1 1
This dependencestructure correspondsto negative dependencebetween (Hs,SWL) and

5.2. SIMULATEDDATA /.)
between(I{s, S), but with (SWL, S) positivelydependent throughtheir relationshipwith

Hs.
Finally to ensurethat no artificially large valuesof T2 ate producedas a result of a
high .FIsvalueoccurringwith a small S value,\Memodify the simulateddata set by setting
Ils to be the minimum observedf{s valuewhen ,9 < 0.02.
SimS data
The marginal distributions of d, ,9 and tides are taken to be the sameas the equivalent
empirical distributions of the North Wales data. For both Su,rge arrd /1s variablesa
mixture of the empirical distribution and the parametricmodel is used. Specifically,for
botrhSurge and IIe belowthe 95%empiricalquantileof the associatedNorth Walesdata,
the empiricaldistribution is used,whereasabovethesethresholdsthe distribution follows
a generalised Paretodistribution,with parameters o :0.L446 and { : 0.0803for ,9urge
and o: 0.6119and {: -0.1583for I/s.
The wavedirectionand tide are takento be independent ofthe (Hs,Surge,S) variables,
which are taken to be dependent. After transformationof eachmarginal variable to a
standardnormal distribution, the form of the dependence betweenthe three variablesis a
trivariate normal distribution with zero-meanvector and variance-covariance matrix, D.
given by
( t 0.16 o.o49T \
x:l 0.16 1 0.16xo.oaezl.
\ 0.04970.16x 0.0497 | l
This dependencestructure correspondsto positive dependencebetween (Hs,Surge) a,nd.
very weak positive dependencebetween (fls, S), and with (Surge, S) almost independent.
Sim4 data
The marginaldistributionsof I and S are takento be the sameas the equivalentempirical
distributions of the Christchurchdata. For both SWL and./{s valuesa mixture of the
empirical distribution and parametricmodelsis used. Specifically,for both SW L and
,I/s below the g5% empiricalquantile of the associatedChristchurchdata, the empirical
distribution is used,whereasabovethesethresholdsthe distribution followsa generalised
Paretodistribution,with parameters o : 0.126and {: -0.15 for SWL and o:0.6685
and { : 0.15for fls.
The wavedirection and ,9 variablesare taken to be independentof the (Hs,SWL) vari-
ables,which are taken to be dependent.After transformationof eachmarginal va.riable
to a unit Fr6chetdistribution,(i.e. with distributionfunctionPr{X < r} : exp(-I lr)
for r > 0) the form of the dependencebetween the two variables (X, Y), is taken to be a
bivariate extremevalue distribution (with logistic dependence
structure) grvenby
Pr{X < r,Y < a} - exp{-(r-1l" +a-t1o1"}, (5.2.1)
where0<a{ldeterminesthedegreeofdependence.Wetooka:0.8.Thisdependence
structure ccrresponds to a strong form of positive dependencebetween (Hs,SWL) wifh
particularly high levels of dependencebetween the most extreme values of each variable,
seeSectiou3.3.1.
Finally, to ensure that no artificially large values of Ts ate produced as a result of a

high /1s value occurring with a small ,9 value, we modify the simulated data set by setting
.I1s to be the minimum observed I{s value when S < 0.014.
Simb data
The marginal distributions of d and ,9 are taken to be the sameas the equivalentempir-
ical distributions of the Dowsingdata. For both ,SWtr and f15 valuesa mixture of the
empiricaldistribution and parametricmodelsis used. Specifically,for both SWL and Hg
belowthe 95% empiricalquantileof the associatedDowsingdata, the empiricaldistribu-
tion is used,whereasabovethesethresholdsthe distribution followsa generalisedPareto
distribution, with parameters o:0.2091 and{: -0.0711for SWL and o:0.8091 and
{ - 0.1121for I1s.
The wavedirection and S variablesare taken to be independentof the (Hs,SWL) var\-
ables,which are taken to be dependent.After transformationof eachmarginal variable
to a standardnormal variable,the joint distribution of thesevariablesis given by by the
mixture of bivariatenormaldistributionsdescribedin AppendixA.4.3. The parameters
of this dependencemodel are
p2: ) ,2 1- - o z z : 1 , p : 0 . 9 ,
(0.3,0.3o h:0, a1dpz
5.3. PRINCIPAL DIFFEREATCESBETWEEN THE SIMULATED DATA SETS 75
5.3 Principal differences between the simulated data

sets
The key differencesbetweenthe simulateddata setsare summarisedin Table 5.2. These

differencesare largely self explanatory,with
c the SWL and 1{5 tails determinedby the valueof the shapeparameter,{, usedto
simulatevaluesfrom the GPD distribution for the tail of the variable,
r the distorted tails for Sim4 are a consequenceof the way that the tails of /{s and
Sll/.L have been deliberatelylengthenedand shortenedrespectivelyfrom values
estimatedat Christchurchto illustrate how the SVM and JPM comparewhen the
two variableshavequite diflerent tail forms. The simulateddata do not reproduce
the hindcastdata well at this site.
data key feature SW L tajl I{s tail Dependence

Siml directionality important short medium * medium
Sim2 directionality of fls medium short/medium - medium
Sim3 tide and surgedata long short * medium
Sim4 distortedtails short Iong + high
Sim5 complexdependence medium long medium/high
Table 5.2: Differencesin the statisticalmodelsfor the simulateddata setsSiml-Sim5.

5.4 Comparison of observed and simulated data

Figures5.1-5.22 showvariousaspectsofthejoint distributionofthe seaconditionvariables
for both the observationaland the simulateddata setsin the report. Specifically,joint
scatter plots of the data are shownfor
o still water level (SWL) and significantwaveheight (fI5)
. snrgeand 115
e -I15and waveperiod (77)
r SWL and 72.
In addition, for both Cardiff and Siml data sets,wavedirectionis plotted againstF/s and
Featuresof interestare:
For Cardifr and SimL LargerTs valuesoccurin Siml than the observationaldata. The
reasonfor this is that the simulationmodel for dependence betweenf/s and ,9 has
not sufficientlyrestrictedsmall ,Svaluesto occur only with small I15 values.Other
than this the simulateddata reproducesthe key characteristicsof the Cardiff data
well, possiblywith the exceptionof not giving largeenoughTz v^ltes for directions
lessthan 100".The probableexplanationof this is our useof 110"ratherthan 190'
as a basisto separatedirectiontypes whensettingup the Siml data. Later analysis
of the Cardiff observationaldata takesthe split at 190'.
For Dover and Sim2 The agreementhereis very good. Figure5.8showssomeappaxent

differencein the relationshipof (IIs,Ts) but this is primarily due to ties in the
observationaldata which appearas a singlepoint on the plot.
North Wales and Sim3 There is a good agreementwith the observationaldata. Note
the outlier in the observationalsurgedata.
Christchurch and Sim4 Here the key differenceis in the scaleof the 7z upper tail,
which is much longer (almosttwice as long) for Sim4 than the observationaldata.
This is due to the (fls,,9) valuesbeingtaken to be independent,thus allowingsmall
,9 valuesto occur with largeIls values(to producevery largeT2) in the simulation
model. The ft valuesare unrealisticallylarge for Sim4. Sim4 was the simulated
data set wherethe ,[1stail wasdeliberatelylengthened.However,this doesnot seem
to havemadea notablechangefrom the observationaldata set (as seenfrom Figure
5.15). As a result of thesefeatures,Sim4 cannotbe considered to be a realistic
representationof seaconditionsat Christchurch.
COMPARISOAI OF OBSERVED AAID SIMUIATED DATA 77
For Dowsing and Sim5 The main difference between these sets is that the simulated
data have a few much larger values of f15 (almost twice as large) than appear in
the observational data. Correspondingly, these observationSproduce a few larger
ft values for Sim5 than the Dowsine observational data.
78 CHAPTER5, STUDYDATA
Figure 5.1: Cardiff observationaldata and Siml data: fls vs SWL

Observationaldata Simulateddata
2468 234567
SWL swL
Figure 5.2: Cardiff observational data and Siml d.ata: Hs vs Tz

:
g.
t
l,
t!. i3;
o2468 2 4 6 I 10
1z
5.4. COMPARISONOF OBSERVEDANDSIMULATEDDATA 79
Figure 5.3: Cardiff observationaldata and Siml data: SWL vs ft

=J
2468 2 4 6 I 10
Figure 5.4: Cardif observational data and Siml data: Ils vs d

50 t00 150 200 50 100 150 200

Dil Dir
80 CHAPTERS, STUDYDATA
Figure 5.5: Ca,rdif observational data and Siml data: T2 vs 0

qll:---*
100 150
Figure5.6: Cardiff observational data: ffs vs Surge

Cardiff
.1.0 -0.5 0.0 0.5

Surge
5.4. COMPARISONOF OBSERVEDANDSIMULATEDDATA 81
-F'igure5.7: Dover observationaidata and Sim2 data: fls vs SWL

01234 01234
swL swL
Figure 5.8: Dover observational data and Sim2 d,ala: Hs vs Tz

Observationaldata Simulateddeta
TN
2468 2 4 6 I t0
B2 CHAPTERS. STUDYDATA
Figure 5.9: Dover observationaldata and Sim3 data: SWL vs ?z

Observational
data Simulateddata
a., ;N
2468 2 4 6 I 10
fz
Figure 5.10: Dover observational data: I1s vs Surge

Dover
-1 0t
Surge
5,4. COMPARISONOF OBSERVEDAND SIMLILATEDDATA 83
Figure 5.11: North Wales observationaldata and SimS data: IIs vs SWL
11 12 13 10 11 12
swL SWL
Figure 5.12: North Wales observationaldata and Sim3 data: Hs vs T2

tz
CHAPTERs. STUDYDATA
Figure 5.13: North Walesobservationaldata and Sim3 data: SWL vs Tz

I'igure 5.14: North Wales observationaldata and Sim3 data: F1svs Surge
0.0 0,5 1.0 1 . 5 -0.5 0.0 0.5 1.0

Surge Surge
5.4, COMPARISONOF OBSERVEDANDSIMULATEDDATA flo
Figure 5.15: Christchurchobservationaldata and Sim4 data: f1s vs SWL

0.0 0.5 1.0 r.5 0.5

SWL swL
Figure 5.16: Christchurchobservationaldata and Sim4 data: Hs vs Tg

68
Iz
86 CHAPTERS. STUDYDATA
Fiqure5.17:Christchurch
observational
data and Sim4data: SWL vs T"
=
o;
.^ =-
I 10 14
Figure 5.18: Christchurchobservationaldata: 115vs Surge

Christchurch
Surge
5.4. COMPARISONOF OBSERVEDANDSIML|LATEDDATA 87
Figure 5.19: Dowsingobservationaldata and Sim5 data: f/5 vs SWL

45670
swL SWL
Figure 5.20: Dowsingobservationaldata and Sim5 data: H5 vs 1'7

2 4 6 I 10 2 4 6 8 10 12
fz
BB CHAPTERs. STUDYDATA
Figure5.21:Dowsingobservationai
data and Sim5data: SWL vs ft
Observational
data Simulated
data
3r4 3
2 4 6 8 10 2 4 6 I 10 12
Figure 5.22: Dowsing observational data: //s vs Surge

Dowsino
-1.0 -0.5 0.0 0.5 1.0

Surge
Chapter 6
Marginal Estimation
In this chapter we apply the models proposed in Chapter 4 to each of the marginal
variables of the sea condition vector.
6.1 Still Water Level
6.I-.1 Observational Data

The thresholdmethod of Section3.1.2is now appliedto the still water leveldata from the
observationalsites (Cardiff, Christchurch,Dowsing,Shoreham,North Walesand Dover).
To eliminate time dependence,observationsat Dowsing,Christchurchand Dover were
declusteredby meansof the proceduredescribedin Sectioa3.1.6. This step was unnec-
essaryfor time seriesat Cardiff, Shorehamand North Wales,as they consistonly of high
water data.
Threshold Choice and Estimation
For thresholdselection,a mean residuallife plot was constructedfor eachof the 6 sites.
Except at Cardiff, for which no linearity could be detected,a thresholdfixed at the 95%
quantile of the empirical distribution seemeda reasonablechoicefor all sites. Figures
6.1-6.3 contain plots of maximum likelihood estimatesof the shapepa,rameter{ of the
GPD plotted againstthe thresholdu (expressed as the non-exceedance probabiliiy, f(u),
ofz) for eachsite. For all sitesthe estimateof( is a negativevalue,which indicatesshort-
tailed distributions with finite upper end point. Again, with the exceptionof Cardiff, a
stability of the estimatescan be seento hold abovethe 95% empirical quantile. Thus,
this value of z is used at all sites with the associatedmaximum likelihood estimatesof
the GPD parameters { and o givenin Table6.1.
Goodness-of-fit
To assessthe goodness-of-fitof the fitted GPD tail modelto still water level,probability
89
90 CHAP'IER 6. MARGINAL ESTIMATION
(P-P) and quantile (Q-Q) plots were constructed.Figure 6.4 containsresultsobtained

for Cardiff and North Wales.As expectedfrom the preliminaryanalyseson the threshold,
Cardiff presentsa poor fit, as indicatedin the quantile plot by the observedupper tail
being shorter than the fitted GPD model. For North Walesthe goodness-of-fitof the
GPD is satisfactory,as the probability and quantileplots exhibit no systematicdeparture
from linearity. The other sites exhibit a quality of fit which is similar to that of North
Wales. Overall)we can concludethat the GPD model givesan acceptabledescriptionof
the observedtail of SWL at most sites,though for extrapolationbeyondthe rangeof the
data, estimatesbasedon an indirect method may be preferable,as discussedin Section
4.1.1.
Site Shape Parameter{ Scale Parameter a

Cardiff -0.264(0.0124) 0.362(0.0121)
Christchurch -0.05e3(0.0474) 0.126(0.00863)
Dowsing -0.0711(0.0371) 0.20e1(0.0141)
Shoreham -0.1324(0.0486) 0.1796(0.01325)
North Wales -0.0457(0.0482) 0.2246(0.015)
Dover -0.0764(0.0537) 0.1866(0.0151)

Table 6.1: Maximum likelihood estimatesof the parameterscf the GPD fitted to SWL
data at observationsites (standarderrorsin parentheses).
6.1. STILL WATER LEVEL 91
Figure6.1: GPD shapeparameterestimates versusthe empiricaldistributionfunctioncal-

culatedin u fot SW L data at Ca,rdiffand Christchurch.Dots denotemaximumlikelihood
estimates.Lines..denotelower and upper boundsof the 95% confidenceintervals.
Cardiff
A
o
o
i: (\l
-c
0.90 o.92 0.94 0.96 0.98
F(u)
Christchurch
o
(DY
o- c;
-c
oO
0.90 o.92 0.94 0.96 0.98
F(u)
92 CHAPTER 6. MARGINAL ESTIMATION
Figure 6.2: GPD shapeparameterestimatesversusthe empirical distribution function

calculatedin z for SW L da,taat Dowsingand Shoreham.Dots denotemaximumlikelihood
estimates.Lines denotelower and upper boundsof lhe g5% confidenceintervals.
Dowsing
c.i
rrJ
bF
!!o
=F
(! rc)
o)
o
c;
0.90 0.92 0.94 0.96 0.98 1.00
F(u)
Shoreham
-i
c
g'-
<d .-i
o;
o
-c
CfJ
0.90 0.92 0.94 0.96 0.98
F(u)
STILL WATER LEVEL

calculatedin u for SWL data at North Walesand Dover. Dots denotemaximumlikelihood
estimates.Lines denotelower and upDerboundsof the g5% confidenceinterva.ls.
NorthWales
O-i
E
O.|l)
(D-i
o'i
0.90 0.92 0.94 0.96 0.98 1.00
F(u)
Dover
{
E c\l
c
o-Y
c)v
(s
0.90 0.92 0.94 0.96 0.98
F(u)
94 CHAPTER6, MARGINALESTIMATION
Figure 6.4: Probability and quantile plots for the GPD fiited to SWL data at Cardiff
and North Wales.Also shownin eachplot is the o: grhne.
Cardiff:P-PPLOT Cardiff:Q-Q PLOT
eF.
-o c
I (d
o- a^
Y-
c) oF-
q)
_c
-c?
0.0 0.4 0.8

emp.proD.
North Wales:P-PPLOT Wales:Q-QPLOT
-ci o
a
d+
Ed
o
0.0 0.4 0.8 12.813.213.6
emp.prob. emp.quanr.
STILL WATERLEVEL 95
6.1.2 Simulated Data

For the simulatedstill water leveldata the 95% empiricalquantileis usedas a threshold
to fit the GPD model to the exceedances. It followsthat for all data setsthe exceedance
probability estimateis 0.05. Table 6.2 givesestimatesof the GPD parameterestimates
for the five simulated data sets. For Sim1. Sim2. Sim4 and Sim5 the statistical model
that is fitted is of the samestructure as the model usedto simulatethe data. For Sim3
howeverthe data weregeneratedas independenttide and surge,so the still water level is
not guaranteedto fit well abovethe selectedthreshold.
For Siml and Sim2 the parameterestimatesare very closeto the true valuesusedto
simulate the data (seeChapter 5). For Sim4 and Sim5 the two parametersdiffer from
the true values,but the changein the tail form due to the two mis-estimationslargely
cancelsout as one error lengthensthe tail while the other shortensit. Little can be said
about Sim3 at this stageexceptthat the shapeparameterestimatesfor the still water
level/surgecorrespondto an upper endpoint existing/not existing respectively.This is
consistentwith the errors found by Dixon and Tawn (1994,1995)for direct analysisof
still water levels.
Data Shape Parameter{ Scale Parameter o

Siml -0.2246 (0.041,) 0.3402(0.023)
Sim2 -0.1284(0.054) 0.1843(0.014)
Sim3 -0.11 (0.045) 0.24 (0.0167)
Sim4 -0.1811(0.05) 0.1344(0.00e8)
Sim5 -0.1166(0.049) 0.2151(0.016)
Table6.2: Maximumlikelihoodestimatesof the parameters of the GPD fitted to SW L

simulatedfor the syntheticdata sets (standarderrorsin parentheses).
6.2 Surge
6.2.1 Observational Data
The ThresholdMethod wasappliedalsoto declusteredsurgedata. Figures6.5-6.7contain
sensitivityplots on the basisof whicha thresholdequalto the empiricalg5%surgequantile
was chosenfor all sites. Maximum likelihoodestimatesare shownin Table 6.3. With the
exceptionof Christchurch,estimatesof the shapeparametertend to be higher than the
va.luesobtained from still water level. This is consistentwith the Sim3 data results
above. Dixon and Tawn (1994,1995)noticed that for sites at which still water level is
dominated by the tidal variation, extremevalue modelsapplied to extremesurge data
indicate longertail distributionsthan the samemodelswhen fitted to extremestill water
levels. In contrast, they observean agreementin the resultsfor sites in which the surge
variation is dominant. Christchurchbelongsto this latter classof sites,and as seenfrom
Tables6.1 and 6.3, the shapeparametersare of the samesign in eachcase. Graphical
assessments of goodness-of-fitfor Cardiff and Christchurchare containedin Figure 6.8.
For still water level,somedepa,rturefrom linearity is observablein the Cardiff Q-Q ploi.
At the other sites,the GPD seemsto providea reasonablefit to the surgeexceedances of
u-
Site Shape Parameter{ Scale Parametero

Cardiff 0.0123(0.0293) 0.137 (0.0061)
Christchurch -0.0954(0.049) 0.142 (0.01)
Dowsing 0.0158(0.058) 0.1473(0.012)
Shoreham 0.02691(0.0468) 0.104 (0.0076)
North Wales o.o8o3(0.048) 0.1446 (0.0097)
Dover 0.2 (0.072) 0.131 (0.0122)
Table 6.3: Maximum likelihood estimatesof the parametersof the GPD fitted to Surge
level data at varioussites (standarderrorsin parentheses).
6.2. SURGE 97
Figure 6.5: GPD shapeparameterestimatesversusthe empiricaldistributionfunction

calculatedin z for Surgelevel data at Cardill and Christchurch.Dots denotemaximum
Iikelihoodestimates.Linesdenotelowerand upperboundsof the 95%confidenceintervals.
Cardiff
a,l
6 c.'l
(E
o.
x,..i
(!'
E
0.90 0.92 0.94 0.96 0.98 1.00

F(u)
Christchurch
.:
o
(I)
E
.i
^, -':
dY
o
,n
0.90 0.92 0.94 0.96 0.98
F(u)
CHAPTER6, MARGINALESTIMATION

calculatedin u for Surgelevel data at Dowsingand Shoreham.Dots denote maximum
likelihoodestimates.Linesdenotelowerand upperboundsof the 95%confidenceintervals.
Dowsing
Cr?
(trv
XF
0.90 o.92 0.94 0.96 0.98
F(u)
Shoreham
\t
o
(I) 'i
lY nr
8c;
o-
o-
0.90 0.92 0.94 0.96 0.98
F(u)
6.2. SURGE
Figure 6.7: GPD shape parameterestimatesversusthe empirical distribution function

calculatedin z for Surgelevel data at North Walesand Dover. Dots denotemaximum
likelihoodestimates.Linesdenotelowerand upperboundsof the 95%confidenceintervals.
NorthWales
(o
:Il, o
6
(Es ( \ 1
Eo
6-' c.l
6Y
0.90 0.92 0.94 0.96 0.98
F(u)
Dover
ga
Po
c
E-
q O "-l
O6
o
0.90 0.92 0.94 0.96
F(u)
100 CHAPTER6. MARGINAL ESTIMATION
Figure 6.8: Probability and quantileplots for the GPD fitted to Surgeleveldata at Cardiff
and Christchurch.Also shownin eachplot is the r: g line.
Cardiff: P-P PLOT Cardiff: Q-Q PLOT

(\
C
(g
it'
;-i
o
_c
a
0.0 0.4 0.8 0.4 0.8 1.2
emp.prob. emp.quant.
Christchurch
: P-PPLOT h ristch
urch:Q-Q PLOT
-l
95
d oql
o ol
o
_c
=
.1
"l
OJ
zv
-c
N
0.0 0.4 0.8 0.4 0.6 0.8 1.0

emp.prob. emp.quanl.
6.3. SIGNIFICANT WAVE HEIGHT 101
6.2.2 Simulated Data

The only simulated data set with surge data is Sim3. The parameter estimates obtained
by applying the threshold method with threshold corresponding to the 95% empirical
quantile, and GPD for exceedances,are given in Table 6.4. From this table we see that
the parameter estimates are close to the true values (seeChapter 5).
Data Shape Parameter{ Scale Parametero

Sim3 -0.026(0.052) 0.167 (0.012)
Table6.4: Maximum likelihoodestimatesof the parametersof the GPD fitted to simulated

Surgelevel data (standarderrorsin parentheses).
6.3 Significant Wave Height

6.3.1 Observational Data
In this section we fit three distributional models to the hindcast fls wave data from
the sites Cardiff, Christchurch,Dowsing,Shoreham,North Wales and Dover. All six
time seriesconsistof hindcastdata wherethe temporal dependence has beenremovedby
applyingthe declustering procedureof Section3.1.6.Sincefor thesesiteswaterleveldata
are also available.The three statistical modelsconsideredare:
1 . Weibull;
2. TruncatedWeibull;
J. GPD,
the latter two being fitted through the thresholdmethods,but all fitted using maximum
likelihood. The first two modelscan be comparedby examiningtheir parameterestimates,
the latter two by comparingestimatestability to thresholdlevel. In eachcaseP-P and
Q Q plots are usedto assessgoodnessof fit.
First considerthe Weibull distribution fitted to all /1s values.Table 6.5 showsmax-
imum likelihood estimatesof the Weibull distribution parameters(o, b,c) for eachsite.
The goodnessof fit at each site is most easily assessed using the Q-Q plots shown in
Figures 6.9 6.11. The model fits well for the Dover data; for Cardiff and Christchurch
the fitted model has a slightly longer upper tail than the empirical upper tail, with the
t02 CHAPTER6, MARGINALESTIMATION
reversebehaviourobservedat North Wales. However,for both Dowsingand Shoreham

the fit is exttemelypoor. In one casethe Weibull modelover-estimates the upper tail, in
the other it under-estimates the upper tail. Clearly,for thesetwo sitesthe basicWeibull
modelis inadequate.
Now considerthe TruncatedWeibull distribution. Table 6.6 containsmaximum like-
Iihoodestimitesof the parameters(a,b,c) for the samesites,obtainedby applyingthe
ThresholdMethod with the distribution of exceedances given by the Tluncated Weibull
distribution, c.f. equation(a.1.2),with the thresholdtaken at the 95%empirical/1s quan-
tile. If the Weibull distribution werethe true underlyingdistribution, then the estimates
of (a,b,c) would be similar to those{rom the previousWeibull modelfit to the whole
data set. This only seemstrue for the Dover data. The worst agreementoccurswith
the Dowsingand Shorehamdata. Thesefindingsconfirmthe obsenationsmadefrom the
Weibull model fit above. Probability plots and quantile plots for the TruncatedWeibull
a,reshownin Figures6.12-6.74.
We can concludethat the Weibull distribution,while providingan acceptablefit to the
bulk ofthe data, is a poor descriptionofthe upper tail for lls at thesesites. In somecases,
as for Dowsingand North Wales,it leadsto shortertailed distributionsthan is indicated
by the observations,whereasfor Cardilf, Christchurchand Shoreham,it overestimates the
empiricalextremequantiles.Only Doverdemonstrates a reasonable fit. As a consequence,
for most of the sitesconsidered,the Weibull distribution will produceunreliableestimates
of valuesassociatedwith long return periods.
Now considerthe GPD thresholdmodelusingthe samethresholdas for the Truncated
Weibull model. The GPD has the potential advantageoverthe TruncatedWeibull model
that it is asymptoticallyjustified. However,this doesnot gua,rantee
that it will fit better
to data in practice, so we must assesshow well it describesthe data relative to the
TruncatedWeibull mode1.
Resultsobtainedby maximumlikelihoodestimationof the shapeand scalepa.rameters
of the GPD are givenin Table6.7. The parameterestimatesshowthat with the exception
of Dowsingwe estimatean upper ondpointto the 115distributionfor eachsite. Probability
plots and quantileplots are containedin Figures6.15-6.17.The GPD model seemsto
perform well in each casethere is a strong similarity with the associatedTruncated
Weibull plots. Discriminationbetweenthe Tluncated Weibull and the GPD, basedonly
on this graphicaltool, is difficult.
A further comparisonbetweenthe two tail modelsmay be madeon the basisof their
stability with respectto differentthresholds.As remarkedin Section3.1.4,a property re-
quired from tail modelsis the non-sensitivityofthe resultsto thresholdselection.Figures
6.18-6.23showestimates,for eachof the six sites,of two fixed quantiles,one within and
SIGAryFICAAI?WAVE HEIGHT 103
one be]'ond the rangeof the data, as functionsof thresholdunder eachtail model. For
both quantilesthe GPD seemsmore stable,giving empiricalsupport for the preference
of this tail model. Furthermore,with the exceptionof Dowsing (which was the site for
which we estimatedno upper endpoint),the GPD leadsto lowerquantile estimatesthan
the TruncatedWeibull, suggestingthat the latter is unduly conservative.
Site Location Parametera ShapeParameterc ScaleParameterb

Cardiff 0.00532(0.0293) 276(0.01) 0.3069(o.oo2oe)
Christchurch 0.00827
(0.00057) 1194(0.0106) 0.972(0.0107)
Dowsing 0.0565
(0.00158) 563 (0.0149) 1.443(0-0126)
Shoreham 0.00598
(0.001) 144(0.01286) 0.e45(0.01124)
North Wales 0.00873

(0.0006) 4e3(0.0118) 1.032(0.0077)
Dover 0.0271
(0.00141) 61 (0.016) 1.231(0.0104)
Table 6.5: Maximum likelihood estimatesof the parametersof the Weibull distribution
fitted to -FIsobservationsat varioussites (standarderrorsin parentheses).
104 CHAPTER 6. MARGIIfAL ESTIMATION
Site Location Parameter a ShapeParameterc Scale Parameter b

Cardiff 0.0861(0.133) 1.1166(0.197) 0.2227(0.088)
Christchurch 0.234(0.5145) r.25r(0.212) 1.04e6(0.3667)
Dowsing 1.417(0.486) 0.658(0.14) 0.28e5(0.1e2)
Shoreham 0.0006(0.0182) 1.585(0.0744) 7.267(0.0479)
North Wales 0.0003(0.0096) 1.363(0.0527) 1.015(0.0367)
Dover 0.0235(0.0102) 1.683(0.0821) 1.312(0.048e)
Table 6.6: Maximum likelihood estimatesof the parametersof the Truncated Weibull
distribution fitted to Hs exceedances
at varioussites (standarderrorsin parentheses).
Site Shape Parameter{ Scale Parametero

Cardiff -0.0375(0.0373) 0.1793(0.0091)
Christchurch -0.0474(0.0511) 0.6685 (0.0489)
Dowsing 0.1121(0.0693) 0.80e1(0.073e)
Shoreham -0.1703(0.044) 0.5788(0.0411)
North Wales -0.1583(0.0433) 0.611e(0.0386)
Dover -0.237(0.0513) 0.56B9(0.0435)
Table 6.7: Maximum likelihood estimatesof the parametersof the GPD fitted to lls
exceedances at varioussites (standarderrorsin parentheses).
6-3. SIGNIFICANTWAVEHEIGHT 105
Figure 6.9: Probability and quantileplots for the Weibull distribution fitted to.F/s obser-
vations at Cardiff and Christchurch.Also shownin eachplot is the r : ? line.
q
-o C
fY
o
o a o
E 9!o
!d
0.0 0.4 0.8 0.0 0.5 1.0 1.5

Christchurch
: P-PPLOT :Q-Q PLOT
Christchurch
(o
-o c
(g
o +
0) d)
E c{
o
0.0 0.4 0.8 0246
Figure 6.10: Probability and quantile plots for the Weibull distribution fitted to 115
observationsat Dowsingand Shoreham.Also shownin eachplot is the r = g line.
Dowsing:P-PPLOT Dowsing: Q-Q PLOT
ô c(6
E
a- E
o o
E
0.0 0.4 0.8 02468
emp.proD. emp.quant.
Shoreham:P-P PLOT Shoreham: Q-Q PLOT
(d+
o+ Y(")
pcn
o
0.0 0.4 0.8 012345
emp.proD. emp.quani.
6.3. SIGA/IFICANTWAVEHEIGHT
Figure 6.11: Probability and quantile plots for the Weibull distribution fitted to 115
observationsat North Walesand Dover. Also shownin eachplot is the z: z line.
NorthWales:P-PPLOT NorthWales : Q-Q PLOT
-o *r
I (!
-q
q) O'!
E Q)
0.0 0.4 0.8 01234

emp.proo. emp.quant.
Dover:P-PPLOT O-OPLOT
(f)
o-
o .N
E
0.0 0.4 0.8 01234

108 CHAPTER6. MARGIT{ALESTIMATION
Figure 6.12: Probability and quantile plots for the TruncatedWeibull distribution fitted
to lls exceedancesat Cardiff and Christchurch.Also shownin eachplot is the r : grline.
_o
(t
CD
o- q(\
.i
c) Or
o
-c
0.96 0.98 1.00 0.8 1.2 1.6
Christchurch
: P-PPLOT : Q-QPLOT
Christchurch
_o c
(! rf)
J
u
o q)
-c -c
o
0.95 0.97 0.99 3456

6.3, SIGNIFICANT WAVE HEIGHT
Figure 6.13: Probability and quantileplots for the TfuncatedWeibull distribution fitted
to FIs exceedancesat Dowsingand Shoreham.Also shownin eachplot is the r: g line.
Dowsing:P-P
PLOT Dowsing
:Q-Q PLOT
o)
.o
E r...
q.@
+
cr)
0.96 0.98 1.00 4567A9

Shoreham:P-PPLOT Shoreham:Q-QPLOT
to
€o c
YOt d
+o Y.A
o o_1
-c o(r)
-c
qt
o |r)
cri
0.95 0.97 0.99 2.5 3.5 4.5
110 MARGINAL ESTIMAI:ION
Figure 6.14: Probability and quantile plots for the TruncatedWeibull distribution fitted
to Ils exceedancesat North Walesand Dover. AIso shownin eachplot ir trhgs : E line.
NorthWales: P-PPLOT NorthWales : Q-Q PLOT
. -.1' .q
o E
(s{
-l F. qu)
O
(l)
.ri
--
-c
tc,
u?
o, (\
0.95 0.97 0.99 2.5 3.5 4.5

Dover:P-PPLOT Dover:Q-Q PLOT
-o
o-
o
0.96 0.98 1.00 2.5 3-0 3.5 4.0
6.3. SIGNIFICANT WAVE HEIGHT 111
Figure 6.15: Probability and quantile plots for the GPD fitted to I{s exceedances
at
Cardiff and Christchurch.Also shownin eachplot is the r : rr line.
c
d
Er (!
oF
0)
(\l
oq
0.0 0.4 0.8 0.8 1.2 1.6
Christchurch
: P-PPLOT : Q-QPLOT
Christchurch
llt rO
f
4) +
=
(r)
0.0 0.4 0.8 J+Cb
t12 CHAPTER6, MARGINALESTIMATION
Figure 6.16: Probability and quantile plots for the GPD fitted to I{s exceedances
at
Dowsingand Shoreham.Also shownin eachplot is the z: E line.
Dowsing:P-PPLOT Dowsing
: Q-QPLOT
o
a-
o
0.0 0.4 0.8 0.0 0.4 0.8
Shoreham
: P-PPLOT Shoreham: Q-Q PLOT
(r)
_a
o
f
q) ctd
-c C)
-c
rq
(v
0.0 0.4 0.8 2.5 3.5 4.5
6-3. SIGNIFICANT WAVE HEIGHT 113
Figure 6.17: Probability and quantileplots for the GPD fitted to IIs exceedances
at North
Walesand Dover.Also shownin eachplot is the r: y line.
NorthWales:P-PPLOT NorthWales: Q-Q PLOT

ro
@
lt c
d
:l
6
u?
(v)
o o
q)
c
ro
C!
0.0 0.4 0.8 2.5 3.5 4-5
Dover:P-PPLOT Dover:Q-QPLOT
6
_o o
o
+ {a;
o o
- v.
!t)
-i
ni
0.0 0.4 0.8 2.5 3.0 3.5 4.0
Figure 6.18: Cardiff: Estimatesof two extremequantilesunder the Weibull and the GPD
modelsversusthe empiricaldistribution function calculatedin u.
Cardiff Quantile=
99.9%
s
Weib u
- - - - - - -G P D
-:
'F (O
q(f)
(\I
0.90 0.92 0.94 0.96
F(u)
Cardiffguantile=
99.999%
e.l
C\I
c.i
OO
c\\
(d
f.n -:
Er
- WeiDU
GPD
\
0.90 0.92 0.94 0.96
F(u)
SIGI /IFICAAI? WqVE HEIGHT
Figure 6.19: Christchurch:Estimatesof two extremequantilesunder the Weibull and the

GPD modelsversusthe empiricaldistribution function calculatedin z.
= 99.9o/o
Christchurchquantile
+
|r)
Weibull
ro GPD
a?
6lo
=
ddi
0.90 0.92 0.94 0.96
F(u)
quantile=
Christchurch 99.999%
EN
0.90 0.s2 0.94 0.96
F(u)
116 CHAP'IER 6. MARGINAL ESTIMATION
Figure 6.20: Dowsing:Estimatesof two extremequantilesunder the Weibull and the

GPD modelsversusthe emoiricaldistribution function calculatedin z.
DowsingQuantile=
99.9%
C!
tr
tr
ar@
:=
c
<S re.
=-
s
- Weibull
C\l GPD
0.90 0.92 0.94 0.96 0.98
F(u)
DowsingQuantile=
99.999%
Weib u
GPD
(D
:=
(Ev
fr
0.90 0.92 0.94 0.96 0.98
F(u)
SIGATIFICAAITWAVE HEIGHT '1.r7
Figure 6.21: Shoreham: Estimates of two extreme quantiles under the Weibull and the
GPD models vercus the emDirical distribution function calculated in z.
ShorehamQuantile=
99.9%
n"
- Weibull
\| GPD
l.E
gv
c
(d
+
c',1
.V
0.90 0.92 0.94 0.96 0.98
F(u)
Shorehamguantile=
99.99%
ro
.; Weib u
GPD
o(o
ou)
ri
0.90 0.92 0.94 0.96 0.98
F(u)
118 CHAPTBR 6. MARGINAL ESTIMATION
Figure 6.22: North Wales:Estimatesof two extremequantilesunder the Weibull and the
GPD modelsversusthe empiricaldistribution function calculatedin z.
NorthWales Quantile=
99.9%
v _ weibull
+ GPD
ot o?
Esf
c
a$l
0.90 0.92 0.94 0.96 0.98
F(u)
NorthWales Quantile=
99.999%
- Weibull
1() GPD
,^
0.90 0.92 0.94 0.96 0.98
F(u)
6,3. SIGNIFICANTWAVEHEIGHT
Figure 6.23: Dover: Estimatesof two extremequantilesunder the Weibull and the GPD
modelsversusthe empiricaldistribution function calculatedin z.
DoverQuantile=
99.9%
Weib u
+ GPD
(t)^,
cv
J
+
0.90 0.92 0.94 0.96 0.98
F(u)
Doverguantile=
99.99%
q - Weibull
ro GPD
.;
o*,
=
c@
g+
+
n
s
0.90 0.92 0.94 0.96 0.98
F(u)
r20 CHAPTER 6. MARGIAIAL ESTIMATION
6 . 3 . 2 Simulated Data
For the simulated/1s data we only fitted the GPD thresholdmodel. In eachcasewe took
the 95% empiricalquantile of fls as the threshold. From Section5.2 we can seethat for
Sim3-Sim5this is the correct specificationof the distributional form, whereasfor Siml
and Sim2 the marginaldistribution of Ile is more complexthan this as, in eachcase,the
distribution of I/s dependson the associatedwavedirection. Table 6.8 givesestimatesof
the fitted GPD parametersfor eachsimulateddata set. For Siml-Sim3/Sim4 Sim5 the
IIs distribution is estimatedto havea finite/infinite upper endpointrespectively,which is
consistentwith the true simulationmodels.For Sim3 and Sim4 the estimatesare closeto
the true values(seeChapter 5) whereasfor Sim5 the shapeparameterestimateis slightly
too large,correspondingto an over-estimated upper tail. Comparisonsfor Siml and Sim2
are more difficult. For Sim2.the fitted model approximatelyis an averageof the models
for the two directional IIs distribution. This is reasonable,as eachdirection sector is
approximatelyequally likely to produceextremeI1s values. For Siml the fitted model
closelyresemblesthe lls distribution for d > 110". Again this is reasonable,as most
waves,including the largestwaves,comefrom this direction sector.
Data Shape Parameter{ Scale Parameter o

Siml -0.0727(0.061) 0.1e42(0.016)
brmz -0.2712(0.063) 0.5125(0.042)
Sim3 -0.136(0.05) 0.605(0.044)
Sim4 0.132e(0.071) 0.6506(0.057)
Sim5 0.2i08(0.067) 0.802(0.068)

Table6.8: Maximum likelihoodestimatesof the parametersof the GPD fitted to simulated
Ils thresholdexceedancesfor the syntheticdata sets (standarderrorsin parentheses).
6.4 Steepness
Figure 6.24 containshistogramsof steepness
for Cardiff, Christchurch,Dowsing,Shore-
ham, North Wales and Dover. Our approachfor modelling the distribution of 5 is to
STEEPI{ESS L21
usethe empirical distribution of ,9, i.e. the observedhistogramwith no extrapolationof

,Sbeyondthe largestor smallestobservationsof S. For the simulateddata we usedthe
empiricaldistribution of ,9 from the associatedsite (cf. Section5.2).
For all sites,the most likely valuesof steepness lie between0.04 and 0.06. With the
exceptionof Cardiff, all histogramsare uni-modal,exhibiting either positive or negative
skewness.At Cardiff, the empiricaldistribution of ,5is bimodal,with peaksapproximately
correspondingto 0.01 and 0.05. The explanationof this feature is that for Cardiff the
hindcastwavemodel predictstwo types of wave:
1. locally generated(steeperwaves),and
2. externally generated(shallowerwaves)which haveonly beenable to reachCardiff

due to refraction.
As specifiedabove,all six data setsconsistof hindcastwavedata. Despitethe hind-

castingmodel complexity,there remainsthe risk that someaspectsof the actual observed
S processmay be missed.To illustrate that the distributions of measuredand hindcast
steepness are not substantiallydifferent,histogramsof measured,9 at Morecambe,Lyme
and Boygrift are shown in Figure 6.25. They seemconsistentwith the corresponding
plots of hindcast steepness,possiblyshowingmore variability and longer tails, but this
difierencewas not consideredimoortant for this studv.
r22 CHAPTER 6. MARGINAL ESTIMATION
Figure 6.24: Histogram of Sleepnessat varioussites.
Cardiff Christchurch
(\
0.0 0.02 0.04 0.06 0.08 0.0 0.02 0.04 0.06 0.08
Dowsing Shoreham
0.0 0.02
E
,l!!t
0.04 0.06 0.08
(\l
0.0 0.o2 0.04 0.06 0.08
NorthWales Dover
0,0
,xXXXXxx,,,
0.02 0.04
S
0.06 0.08
N
0.0 0.02 0.04

s
0.06 0.08
STEEPI,/ESS
Figure 6.25: Histogramof measuredsteepness

at Morecambe,Boygrift and Lyme.
Morecambe
0.0
C\I
()
0.0 0.08
724 CHAPTER6. MARGINALESTIMATION
6.5 Wave Direction

Figure 6.26showshistogramsof wavedirectionfor Cardiff, Christchurch,Dowsing,Shore-
ham, North Walesand Dover. Our approachfor statistically modellingthe distribution
of 0 is similar to that of S, i.e. through the empiricaldistribution of d. For the simulated
data we usedthe empiricaldistribution of d for the associateddata site (cf. Section5.2).
The distribution of d seemsquite differentat Dowsingcomparedwith the other sites.
At Dowsingall directionsare is approximatelyequally likely, whereasat other sites the
distributionis bi-modal,with peaksat 80'- - 100"(thesecondary peak)and 200'- -300'
(the dominant peak). A possiblereasonfor this differenceis that Dowsingis further
offshorethan the other sites. For Cardiff, the mode around 200" includesthe externally
qeneratedwaveswith lower S values.
6.5- WAVEDIRECTION 725
Figure 6.26: Histogramof WaveDirection at varioussites.
Cardiff Christchurch
i|u.,,,,,,,J
100 200 300 100 200 300
Dowsing Shoreham
100 200 300
NorthWales
100 200 300 100 200 300

CHAPTER 6. MARGINAL ESTIMATION
Chapter 7
Dependence Estimation
In this chapterwe examinethe suitability ofthe statisticaldependence modelsproposedin

Chapter 4 when appliedto observational/hindcast data and the five simulateddata sets.
Becauseof the structureof the statisticalmodelsproposedwe can separatelyexaminethe
dependencebetween(II s, SW L) and (II5, S). In Section7.1 we examineboth types of
data when applying statistical modelsfor (Hs, SW L), whereaswe only give results for
the simulateddata setswhenexamining(Hs, S) in Section7.2.
7.L (SWL,115)Dependence
7.I.L Observational data

First we estimate the correlationfunction, p,, given by the bivariate norma,lthreshold
model of Section4.2.1. Constancyof this function abovea given level suggeststhat a
bivariatenormal modelcan be usedabovethis levelto describethe dependence structure.
However,if the function is
r constautfor all u: this suggeststhe bivariate normal dependencemodel is appro-

priate for all the values;
r not constantaboveany z: then a more complexstatistica,ldependence

model,such
as the mixture of bivariatenormalsdependence
model,is required.
Figures 7.1 and 7.2 show the correlationfunction, p", plotted against the threshold
non-exceedance probabilityp : Pr{IIs < u} : Pr{SWL < z} for (Hs,SWL) and
(Hs,Surge) respectively(note on the plots p, is shownas p(p)) The main featuresof
theseplots are:
1. the weakercorrelationbetween(Hs,SWL) than between(Hs,Surge);
727
128 CHAPTER 7. DEPENDENCE ESTIMATION
2. all sitesshowa limited increasein the (IIs,,Szrge) valuesof pu as u, or equivalently

p, is increased;
J. greaterstability of p",,with respectto z, is seenfot (Hs, SW L);
4. Christchurchhas a strong correlationof 0.6 in the extremeSurge and I1s levels,

whereasfor Dover (Hs,Surge) are generallynegativelycorrelated,with extreme
levelspossiblybeing independent;
5. in all cases, as the threshold is increased, the confidence intervals (not shown on
these plots) grow so any irregular changesin the function for large z can be ignored.
For Cardiff the direction of the waves is known to be important to the form of the
joint distribution of the other sea condition variables, so in Figures 7 .3 and 7.4 the cor-
responding correlation function plots are produced conditionally on the wave direction
variable, d, being 0 < I90" or d > 190". Here we see a distinct difference in the degree of
dependencefound, with strongest dependencefor 0 > 190', and that the feature of rising
correlation at more extreme levels observedin Figures 7.1 and7.2 is retained.
These findings suggest that the mixture of bivariate normals model should be consid-
ered as a feasible candidate at all sites. Tables 7.1 ard 7.2 give the pa.rameterestimates
of this statistical model for each site.
Although the parameters of the dependencemodel are of some interest, it is really
how they combine to produce a changing degreeof dependencewhen looking further into
the joint tail of the distribution that is important. To examine this we evaluate lhe pu
function for this dependencemodel. This cannot be evaluated in closed form so we use
simulation techniques. Figures 7.5 7.10 show the resulting pu plots for the mixture of
bivariate normals model together with the threshold model estimate of p, for compa,rison.
These plots are shown for each site and for both (Hy,SWL) and (-FIs,Surge) d.ata. ln
each case the mixture model gives a very good approximation to the pu curve. This
suggeststhe use of the mixture of bivariate normals model, as this has the added benefits
of:
1 . the inference uses all data, not just the values of threshold exceedancesand down-
weighted (censored) values of other observations;
the fit does not depend on an arbitrary/subjective selection of a threshold level;
3 . the fit provides extrapolations for dependencewhich allow for an increased, or de-
creased,degreeof dependencebeyond the data.
The disadvantasesof the mixture of bivariate normals model a,re:

7.1. (SW L,I{s) DEPEATDEIVCE 129
1. that it is indicative of a dependencestructure which is more complex than the

bivariate normal, or the thresholdbivariatenormall
2. if may be an approximationto a situation in which a mixture of diferent types of

dependenceapply, but a mixture of two types of dependence structure may be too
simplistic;
3. if doesn't provide a relationshipbetweenthe un-mixeddependenceforms and the

other wavevariables(period and direction);
4. it is quite highly parametrisedwith the parameters(for the variableswith Gaussian

marginals)not easilyinterpretablefrom practicalconsiderations.
Of the data sets,the parameterestimatesfor Dowsing,Dover,and Cardiffsuggestthat the

simpler bivariate normal dependencestructure is appropriateas a dependencemodel in
thesecases.We havesubsequentlyusedthe simplermodel for dependence, but if we had
continuedwith the broaderfamily, given by the mixture model,we would haveobtained
similar results.
130 CHAPTER 7. DEPENDENCEESTIMATION
Figure 7.1: Correlationfunction, p.,, between(Hs,,SW L) versusthe threshold(in proba-

bility of non-exceedance)for the differentsites
Rho(Hs-SWL)
,------"')
0.0 0.2 0.4 0.6 0.8 | .0

p
Figure 7.2: Correlationfunction, p",,between(Hs.,Surge) versusthe threshold(in prob-

ability of non-exceedance)for the differentsites
Rho(Hs-Surge)
D.O O.2 0.4 0.6 0.8 1.0

p
7.1. (SWL, Hs) DEPENDENCE 131
Figure 7.3: Cardiff: Correlationfunction, pz, between(Hs,SWL) versusthe threshold

(in probability of non-exceedance)separatelyfor 0 < 190" and d > 190'
Cardi{f:Rho(Hs-SWL)
Figure7.4: Cardiff: Correlationfunction,pa, between(Hs,Surge) versusthe threshold

(in probability of non-exceedance)
separatelyfor I < 190" and d > 190"
Cardiff:Rho(Hs-Surge)
{
11.)
CHAPTER7. DEPENDENCEESTIMATION
Parameters Christchurch Dowsing Dover North Wales Cardiff Shoreham

PM 0.51 0.05 0.50 0.968 0.52 0.88
Pt' -0.13 0.32 0.13 0.02 -0.05 -0.048
Pz 0.52 0.03 -0.12 0.48 -0.09 0.36
P2r 0.7 0.0 1.1 1.8 0.7
Pzz 0.6 0.6 -0.4 0.3 0.1 7.2
o2r 0.9 0.3 0.5 0.8 0.4 0.6
o22 1.1 0.3 0.7 1.2 1.0
Table 7.1: Parameter estimates for the mixture of bivariate normals applied to observa-
tional (I{s, ,9tr7.L)data.
Parameters Christchurch Dowsing Dover North Wales Cardiff Shoreham

Pu 0.47 0.33 0.27 0.25 0.45 0.76
Pt - 0 . 1 7 -0.40 -0.48 -0.28 -0.42 -0.04
P2 0.56 0.28 -0.05 0.31 0.33 0.3
ltzt 0.8 0.0 0.0 0.0 0.0 0.8
IL22 0.7 1.D t.7 1.1 1.O
621 0.9 l-2 0.7 1.3 0.5 0.6

o22 1.0 0.8 0.9 t.2 1.0
Table 7.2: Parameterestimatesfor the mixture of bivariate normalsaDDliedto observa-
tional (fls, ,9zrge)data.
7.1. (SWL, HS) DEPENDENCE
Figure 7.5: Christchurch:Estimatedcorrelationfunction,p,,, between(Hs,SWL) and

(Hs, Surge) versusthe threshold(in probability of non-exceedance)
for the fitted mixture
of bivariate normalsand the thresholdbivariate normal
Christchurch:
Rho(Hs-SVVL)
Ch slchurch:
Fho(Hs-Surg€)
Figure 7.6: Dowsing: Estimated correlation function, p.,, between (Hs, SW L) and
(Hs, Sur ge) versusthe threshold(in probability of non-exceedance)
Dowsingi
Fho(Hs-SwL)
Dowsing:Rho(Hs.SulgB)
3
e"
E:
134 CHAPTER7. DEPENDENCEESTIMATION
Figure7.7: Dover:Correlationfunction,p,,,between(Hs,SWL) and (IIs,,9arge)versus

the threshold(in probability of non-exceedance)
for the fitted mixture of bivariatenormals
and the thresholdbivariatenormal
Dover:Fho(Hs-SWL)
oover:Rho(Hs-SurgB)
Figure 7.8: North Wales: Estimated correlationfunction, pn, between(Hs, SW L) and

(Hs, Surge) versusthe threshold(in probability of non-exceedance)
NodhwalesiRho(Hs-swL)
Nodhwales:Rho{Hs'suee)
f-
7.1. (SW L, H s) DEPEI{DEI\ICE
Figure 7.9: Cardiff: Correlationfunction,p?,,between(Hs,SWL) and (I/s, Szrge) versus

the threshold(in probability of non-exceedance)for the fitted mixture of bivariatenormals
and the thresholdbivariate normal
cardift Rho(Hg-SWL)
f!
Carditl:BholHs-Surge)
Figure 7.10: Shoreham:Estimatedcorrelationfunction, p,,, between(Hs, SW L) and

(Hs.,Surge) versusthe threshold(in probability of non-exceedance)
of bivariate norma.lsand the thresholdbivariate normal
Shoreham:
Bho(Hs-SWL)
Sholeham:
Rho(Hs-sues)
s
136 CHAPTER7. DEPENDENCE
ESTIMATION
7.L.2 Simulated data

In this sectionwe considerestimationof an appropriatedependence structure for the five
Sim data sets. The sectionis split into tlvo stages:diagnosticassessmentand dependence
modelling. In the diagnosticsubsectionwe try to identify for which, if any, of the data
setswe can adopt a conservativeapproachfor by taking the variables(SWL,Hs) tobe
completelydependent,without badly over-estimatingthe dependence.When complete
dependenceis a poor representationof the dependence, we are alsointerestedin testing
whether taking the variablesto be independentis an over-simplification.
Diagnostics for dependence form
We use the diagnosticstatistic,?(e), givenby equation(3.3.4),appliedto Siml-Sim5
in turn. The statistic, with associatedpointwise confidenceintervals,is shown for the
respectivedata setsin Figures7.I1-7.I5.
For Sim1, the gradientof the diagnosticplot is approximatelyj and, judging by the
confidenceintervals, the gradient of this plot is significantlygreater than zero and less
than one. This suggeststhat careful modellingof the dependenceis requiredfor these
data. The gradient is in line with the simulationmodel, which is essentiallya bivariate
normal distribution with p : 0.3, so following equation (3.3.5) should have gradient
( 1 - 0 . 3 ) / ( 1+ 0 . 3 ): 0 . 5 4 .
The diagnosticplots for Sim2 and Sim3 are similar to that for Siml. The respective
gradientsare approximately1.5 and 0.75, and are significantlygreater than zero. For
eachdata set, the confidenceinterval includesa gradientof one, so independenceis not
inconsistentwith the observeddependence. Thesefindingssuggestthat carefulmodelling
of the dependenceis requiredfor thesedata, but that the dependenceis weak,so inde-
pendencemay be an adequateapproximation.The gradientis in line with the simulation
models,which are in essence a bivariatenormaldistributionwith p: -0.2 and 0.16,so
followingequation(3.3.5)shouldhaverespective gradients(1 + 0.2)/(1- 0.2) : 1.5 uo6
( r - 0 . 1 6 ) / ( 1 +o . 1 6 :) 0 7 2
Sim4 has a notably difierent diagnosticplot than the other simulateddata setsas the
confidenceintervalfor the diagnosticstatisticccntainszerofor all logz, and the gradientof
the plot is not significantlydifferentfrom zero(this is consistentwith the bivariateextreme
lalue dependence rnodelwe haveusedto simulatethe Sim4data). For thesedata we can
take the dependence to be completedependence without beingoverlyconservative.In the
subsequentanalyseswe will try to model the dependence for thesedata and commenton
differencesbetweenestimatesbasedon thesemodelsand the completedependence model.
Independence is a very poor descriptionof the observeddependence.
The dependencemodel usedto generatethe Sim5 data was more complexthan the
others, being a mixture of independencefor low levels, and high correlation (0.8) for
7,1. (SWL, H S) DEPENDENCE ID i
high levels.The diagnosticplot identifiesthis transition in dependence with a gradientof

approximately1 up to loge : 0.5,and a gradientof approximately0.12subsequently.The
plot has a gradient significantlydifferentfrom zero and one, so the positive dependence
needsmodelling.
Dependence Models
We start by fitting the mixture of bivariatenormalsmodel to eachsimulateddata set and
examiningthe correlationfunction p,,. For selecteddependencemodelswe continueby
evaluatingthe fitted joint densityfunction, and compareit to the true joint density (used
for the simulation)via comparisonof contoursof equaljoint densityfunction.
Table 7.3 gives parameterestimatesof the mixture of bivariate normalsmodel. For
Siml Sim3 we find p,y * 1, i.e there is only one bivariate normal dependencestructure
in the data, so that the bivariatenormal model is sufficient.For Sim4,p1ais large,but p1
and p2are quite different,suggestingthat the extremeobservations havea differentdegree
of dependencefrom the bulk of the distribution. Note, p2 : 0.7 effectivelymeasuresthe
correlationin the extremeshere. Finally, for Sim5 the fitted dependence model is in the
samestatistical family as wasusedto simulatethe data, and from Section5.2 we can see
that the dependencemodel givesparameterestimatesvery closeto the valuesused for
the simulation.
To confirm thesefindingswe now estimatethe correlationfunction, p,, for eachsimu-
lated (fls, SWL) data set. This function is shownin Figure 7.16,wherewe seestability
for Siml-Sim3,but rising correlationlevelsfor the joint extremevaluesin Sim4 and Sim5.
For Sim3we additionally estimatethe correlationfunction for (Hs, Surge), shownin Fig-
ve 7.17. This hasthe property,notedearlier,of weakerdependence between(Hs, SW L)
than between(Hs, Surge). In eachcase,for Sim3,the correlationfunction appearsto be
independent of z.
To assessthe fit of the mixture of bivariate normals model for Sim4 and Sim5, we
also evaluatethe p,, function for this dependencemodel by fitting the bivariate normal
thresholdmodelto simulateddata from the fitted model.Figures7.18and 7.19showthat
the correspondingfunctions capturethe pu curvesextremelywell. Of course,in Figure
7.19this fitted modelis of the sameform as the simulationmodel.
Based on the findings above,for the remainderof the report we consideronly the
following dependence modelsestimatedfor Siml-Sim5:
Siml bivariate normal dependence

with correlationcoefficient0.24;
with correlationcoefficient-0.21;
Sim2 bivariate normal dependence
SimS for (Hs, Surge) a bivariate normal dependencewith correlationcoefficient0.14,

whereasfor (f15,SWL) a bivariate normal dependencewith correlationcoefficient
Figure7.11:Siml: Diagnostictest for the form ofthe dependence in the extremesof IIs-
SWtr. The solid line is the test statistic and the dotted line its pointwise95%confidence
interval.
Figure7.12:Sim2:Diagnostictest for the form ofthe dependence in the extremesof f15-

SW.L. The solid line is the test statistic and the dotted line its pointwise95%confidence
interval.
log z
7.1. (SWL,Hs) DEPE TDEI'ICE
Figure 7.13: Sim3: Diagnostictest for the form ofthe dependence in the extremesof Ils-
SW L. 'Ihe solid line is the test statistic and the dotted line its pointwise95% confidence
interval.
1
log z
Figure 7.14: Sim4: Diagnostictest for the form of the dependence in the extremesof 11s-
SWL. The solid line is the test statistic and the dotted line its oointwise95% confidence
interval.
0.03;
Sim4 bivariatenormal thresholdmodelwith p,, : 0.588for z equalto the 95%threshold,

and the fitted mixture of bivariatenormals (parameterestimatesin Table 7.3);
Sim5 bivariatenormal thresholdmodel with p, - 0.335for u equalto the 97.5%thresh-

old, and the fitted mixture of bivariatenormals(parameterestimatesin Table 7.3).
Figures 7.20 and 7.21 show the fitted and true/simulationjoint density contoursfor
Siml and Sim2respectively.In eachcasethe bivariatenormalmodelis of the correctform
for the dependencemodelling,so inconsistencies in the estimatesare due to parameter
estimation rather than mis-specificationof the statistical model. The agreementis very
good for Siml, and reasonablefor Sim2. In the latter the upper tail of the fls variableis
a bit short in the estimatedjoint densitymodel.
For SimS we have dependercemodelsfor borh (H5,SWL) and (I15,Szrge). Joint
densitycontoursare shownin Figures7.22 and7-23. "the agreementis goodin both cases,
with the only obviousdisagreementarisingfrom the estimatedjoint distribution having
a shorter upper tail for Surge and SWL than the simulatedvalues.
For Sim4we havetwo fitted statisticalmodelsto compare.Figures7.24and 7.25show
joint densityestimatesfor the thresholdand mixture bivariatenormal modelsrespectively.
For the former the estimation of the dependencestructure is extremely poor, with the
contoursin poor agreementover most of the distribution. By comparison,the mixture
model is in good agreement,with the exceptionof the SW L upper tail, which is under-
estimated.
Finally considerestimatesfor SimSshownin Figures7.26and7.27. Againthe thresh-
old baseddependence modelfits poorly, and the mixture modelis a markedimprovement.
However,whateverdependencemodel is fitted the joint density estimateis poor as the
marginal modelsover-estimates/under-estimate the tails of .FIsand SW L respectively.
In ChaptersB 9 we examinehow important thesediscrepancies are when considered
from the designperspective,and by taking the marginsto be both known and estimated,
we examinethe importanceof marginal and dependence modelling/estimation.
7.1. (SWL,Hs) DEPE TDEÎCE 141
Figure 7.15: SimS: Diagnostictest for the form of the dependence in the extremesof f15-
SWL. The solid line is the test statistic and the dotted line its pointwise95% confidence
interval.
logz
Parameters Siml Sim2 Sim3 DI]nJ Sim4 Sim5

PM 1.00 0.99 1 . 0 0 1.00 0.85 0.90
Pt 0.24 - 0 . 2 1 0.034 0.14 0 . 1 5 0.00
P2 -0.78 0.70 0.76
Lr21 0.0 0.9 0.5
ltzz t.2 0.8 0.3
o2r 0.5 1 . 0 0.9
o22 0.1 1.1 0.8
Table 7.3: Parameterestimatesfor the mixture of bivariate normalsappliedto the simu-

Iated (11s,SW L) data and for Sim3- to (Hs, Surge).
142 CHAPTER 7- DEPENDENCEESTIMATION
Figure7.16:Correlationfunction,p,,,between(Hs, SW L) versusthe threshold(in prob-

ability of non-exceedance)
for the differentsimulateddata sets.
Simulatsd
dara:Rho(Hs.SWL)
Figure7.17:Sim3: Correlationfunction,p,,,between(Hs, SW L) and(Hs, Surge)versus

the threshold (in probability of non-exceedance)
sina (Nodhwalesl:Fho(Hs.swl)andFho(Hs-surge)
a;
E
7.1. (SWL,Hs) DEPEI{DE TCE 143
Figure 7.18: Sim4: Correlationfunction,p,,, between(Hs,SWL) versusthe threshold

for the fitted mixture of bivariate normals and the
thresholdbivariate norma,l
Sim4 (Chrisrchurch):Bho{HS-S!VL)
Figure 7.19: SimS: Correlationfunction,p,,, between(Hs,SWL) versusthe threshold

for the fitted mixture of bivariate normals and the
thresholdbivariate normal
Sim5 (Dowsing)iBho(HS-SWL)
r44 CHAPTER7. DEPENDENCE
ESTIM,|TION
Figure 7.20: SimT Hs-SWL joint density contours: the contours of joint density value
0.01,0.001,0.0001,0.00001ale shown. The estimatedmodelis the bivariatenormal model.
45478
Figure 7.21: S1m2H1-SWL joint density contours: the contoursof joint density value
0.01,0.001,0.0001,0.00001
areshown.The estimatedmodelis the biva,riate normalmodel.
7.1. (SW L, H S) DEPENDENCE L45
Figure 7.22: Sirn3 I{5-Surge joint density contours: the contours of joint density value
0.01,0.001,0.0001,0.00001 are shown. The estimatedmodel is the bivariatenormal model.
I
Surgo
Figure 7.23: Sirr'J Hs-SWL joint density contours: the contours of joint density value
0.01,0.001,0.0001,0.00001are shown. The estimatedmodel is the bivariatenormal model.
11 12 13 14 15
146 CHAPTER7, DEPENDENCE
ESTIMATION
Figure 7.24: Sim4 Hs-SWL joint density contours: the contoursof joint density value
0.01,0.001,0.0001,0.00001
are shown. The estimatedmodel is the thresholdbivariate
normal model.
IF
0.0 0.5 1.0 1.5 2.O

WL
Figure 7.25: Sim4 Hs-SWL joint density contours: the contoursof joint density value
0.01,0.001,0.0001,
0.00001are shown. The estimatedmodel is the mixture of bivariate
normalsmodel.
IF
0.0 0.5 1.0 1.5

WL
7.1. (SWL, Hs) DEPEI\TDENCE 147
Figure 7.26: Sim5fIs-SWl, joint densitycontours:the contoursof joint densityvalue

0.01,0.001,0.0001,
0.00001are shown. The estimatedmodel is the thresholdbivariate
normal model.
joint densitycontours:the contoursof joint densityvalue

Figure 7.27: Sim5 11s-.9trV2
0.01,0.001,0.000f,0.00001are shown. The estimatedmodel is the mixture of bivariate
normalsmodel.
7.2 (5, Hs) Dependence for Simulated Data

Scatter plots of the (S,IIs) data for the five simulateddata sets are shown in Figures
7.28-7.32.From theseplots we seethat there is a strong positivedependence betweenthe
variablesfor Siml (Figure7.28)andweakdependence/independence betweenthe variables
for Sim2-Sim5. This is consistentwith the simulation modelsfor these variables(see
Chapter 5), for which Sim2 and Sim3havesmall degreesof negative/positivedependence
respectively,while Sim4 and Sim5 haveindependentvariables.
To estimatethe dependence betweenthe variableswe fit the linear regressionmodel,
given by equation (4.2.4),after transformationof eachvariableto a Gaussianscale,i.e.
(S.,113)with the variablesdefinedby equations(4.2.2)and (4.2.3). First we plot the
(S,.F15)data after transformationto the Gaussianscalein Figures 7.33-7.37. In the
Gaussianscale (onceties in the data and modified Ils valuesfor small ,9 levelsare ac-
countedfor) we seea clearlinear relationshipbetweenthe variablesfor Siml and little or
no evidenceof linear relationshipsin the other cases.
To assesswhat thresholdlevelto usefor lls (or equivalentlyfor IIi) in the regressron
modelwe now considerplots of E(S.l/1j > z) and Var(S-lHi > z) againstthe threshold
u in Figures7.38-7.42.Theseplots showthat for
SimL the rising mean showsa strong positive relationshipbetweenthe variables. The
variancestabilisesabovea thresholdz:0.5, so this level is taken for subsequent
analvsis.
Sim2 the falling mean with thresholdshowsa weak negativerelationshipbetweenthe

variables.The varianceis stableabovea thresholdlevel of u : -0.5, so this level is
usedin the subsequentanalysis.
SimB the rising mean showsa weak positive relationship,with stability of the variance
obtainedabovea thresholdof u : 0.5.
Sim4 and SimS There is no evidenceof non-constantmeanor varia,ncehere,indicating

independence
of the variables.It is thereforeunnecessary
to adopt a threshold.
Now we fit the linear regressionthresholdmodel to ,S* given I1}. Estimatesof the
intercept, a, gradient,b, and residualvariance,o2, for Siml-Sim5 are given in Table 7.4.
The estimatesconfirm the empiricalevidencefound above:highly significantdependence
for Siml, significant(but small) dependence for Sim2- Sim4,and independence for Sim5.
Sim4 is surprising as the data were generatedusing an independencestatistical model,
while the estimateddependence is non-zerobut still very weak.
7.2. (S,Hs) DEPE TDETCEFORSIMULA'.|EDDATA L49
We now examinethe quality of the fitted regressionmodel by plotting the estimated

regressionfunction (transformedback to the original spaceshownin Figures7.28-7.32),
and by comparinga samplesimulatedfrom the fitted model with the original sample
(shownin Figures7.43-7.47).
The fitted meanfunction is difficult to assess

for Sim2-Sim5due to the small varia-
tions in the meanrelativeto the high variability ofthe data. By contrastthe mean
function for Siml providesa very good modelfor the substantivevariationsof ,9 on
HS.
For Siml the data simulatedfrom the fitted model reproducethe main featuresof
the original data very well. The correspondingplots for Sim2-Sim5are lesseasyto
assess,but in eachcasethe original data structure appearsto be well replicated.
Data Thresholdz a b 02
Siml 0.5 0.43(0.02) 0.67(0.03) 0.43(0.03)
Sim2 -0.5 0.16(0.02) -0.11(0.02) 0.80(0.06)
Sim3 0.5 -0.02(0.03) 0.11(0.04) 0.8e(0.12)
Sim4 -oo 0.00(0.01) 0.05(0.01) 0.93(0.10)
Sim5 -oo 0.00(0.01) 0.00(0.01) 1.oo(0.10)
Table 7.4: Parameterestimatesfor the linear regressionof S* on I{ applied above a
thresholdz (on the Gaussianscale).Standarderrorsare given in parentheses.For Sim4
and Sim5 the valueof u usedin the resressionmodel is u : 0.
150 CHAPTER7. DEPENDENCE
ESTIMATION
Figure 7.28: Siml: S versusff5 and regressioncurve (plotted on the original scale).
0.0 0.5 1.0 1.5

HS
Figure 7.29: Sim2: S versus115and regressioncurve (plotted on the original scale).
01234
(5, HS)DEPENDENCEFORSIMUL,+TEDDATA 151
Figure7.30:SimS: ,9 versus fIs and regressioncurve (plotted on the original scale).
Figure 7.31: Sim4: 5 versusIls and regressioncurve (plotted on the original scale).
r52 CHAPTER7. DEPENDENCE
ESTIMATION
Figure 7.32: Sim5: S versus .t15and regressioncurve (plotted on the original scale).
Figure7.33: 5 versus 1/5 (plotted on the Gaussian

7,2. (S,HS)DEPENDENCE
FORSIMULATEDDATA IJJ
Figure 7.34: Sim2: S versus fI5 (plotted on the Gaussian scale).
.t a :r:i1-riJ:i.-::'-:-.i:::i,:jj:i.lii:i..r;.:l.l
j,
-2 -1 012 3
Figure7.35:SimS:,9 versusf1s (plottedon the Gaussianscale).
-202
1i)4 CHAPTER7. DEPENDENCE
ESTIMATION
Figure 7.36: Sim4: S versusf/5 (plotted on the Gaussianscale).
-2 -1 012 3
HS
Figure 7.37: Sim5; S versusl{5 (plotted on the Gaussianscale).
-202
HS
7,2, (S,11S)DEPENDENCEFORSIMULATEDDATA -Li)i)
Figure 7,38: Siml: Mean and Va,ria,nceof S (on the Gaussian scale) versus threshold for
I/s (on the Gaussian scale). Dotted lines give 95% confidence interval.
MeanversusThreshold VarianceversusThreshold
A9
tt
uJ
-1 0 1 2 -1 0 1 2
u (gaussian
scal€) u (gaussian
scale)
Figure 7.39: Sim2: Mean and Variance of .9 (plotted on the Gaussian scaie) versus thresh-
old for 1y'5(on the Gaussian scale). Dotted lines give 95% confidenceinterval.
i
E
\)) .
-'f012 -1 0 1 2
u (gaussianscale) u (gaussian
scal€)
Figure 7.40: Sim3: Mean and Varianceof ^9(plotted on the Gaussianscale)versusthresh-

old for lls (on the Gaussia;rscale). Dotted lines give 95% confidenceinterval.
ao !
I
uJ Eo
-1 0 1 2 -1 0 | 2
u (gaussian
scale) gcale)
u (gaussian
Figure 7.41: Sim4: Mean and Varianceof .9 (plotted on the Gaussianscale)versusthresh-

old for IIs (on the Gaussian scale). Dotted lines give 95% confidenceinterval.
I 4 .,i
ul
'''''' ''''''''''
''''-
'i'"^,..''.i
.1 0 1 2 .1 0 1 2
u (gaussian
scale) u (gausslan
scale)
7,2, (S,HS)DEPENDENCE
FORSIMULATEDDATA 157
Figure 7.42: Sim5: Mea,nand Variance of S (plotted on the Gaussian scale) versus thresh-
old for 1{s (on the Gaussian scale). Dotted lines give 95% confidenceinterval.
ol
I
E 6o.
I!
.1 0 1 2 -l 0 1 2
u (gaussian
6cale) u (gaussian
scale)
Figure 7.43: Siml: ,9 versus /1s for the original data and the data simulated from the
fitted resression.
Originaldata Dataslnulatedtrem
lhe lrned modet
ct
d
1.0 1.5
Figure 7.44: Sim2: S versus f1s for the original data and the data simulated from the
fitted reEression.
Originaldata DataI
tne
2 2
Hs t"1s
Figure 7.45: Sim3: S versus 115 for the original data and the data simulated from the
fitted resression.
Originaldata Datasjmulatedtram
tne flned model
8
7.2. (5, HS) DEPENDENCEFORSIMULATEDDATA 159
Figure 7.46: Sim4: .9 versusI1e for the original data and the data simulated from the
fitted regression.
Originaldata D"ASif|?{{etfi%fr3r
HS t-is
Figure 7.47: Sim5: S versus Ils for the original data and the data simulated from the
fitted resression.
Originaldata D+asifluFtedJr?m
10 I 10 14
Fts
CHAPTER7. DEPENDENCEESTIMATION
Chapter 8
Simultaneous Extremes of (SW L,,H S)
In this chapterwe focuson the estimationof the joint probability
Pr{SWL>r,Hs>gl:
Probabilitiesof this form, thought of as a function of c and g, are termed the joi,nt sur-
uiuorJunctionof the joint distribution ol (SW L, H s). Herewe are particularly interested
in finding all the possiblecombinationsof (r,y) suchthat the joint survivor function is a
small fixed probability, i.e. extrapolationfrom the sampleis required. We presentthese
combinationsas curves,which correspondto contoursof equaljoint survivor function.
Thesecontoursare the main output of the standardapproachof the current implemen-
tation of JPM (seeSection1.2) and are often used for initial designcalculations. The
refinedversionof current implementationusesthis information to derivethe probability
of failure.
In Sections8.1 and 8.2 we presentthe contoursof the estirnatedjoint survivorfunction
for Siml-Sim5 for caseswhere
r both the marginal distributionsare knou'n,and
e no knowledgeof the marginal distributionsis accountedfor,
respectively.Throughoutthesesections,contoursareshowncorresponding to 5, 10,50,100
and 1000year return periods (i.e. for eachpoint (2, g) on the associatedcontour curve,
the probability of the event{SW L > n, Hs } g} is suchthat this eventoccurson average
oncein the specifiedreturn period). Suchcontoursare shownfor
1 . the fitted joint distribution mode1,and
2. the joint distribution of the actual simulation model used to generate the Sim data
sets.
161
162 EXTREMESOF (SWL,HS)
CHAPTERB. SIMULTANEOUS
Beforewe considerthe estimatedjoint survivor function contourplots, we outline the

possibleforms that thesecontourcurvesare to be expectedto take. For generality,here
we considerarbitraryvariables(X, Y).
Limit values
The limits of the joint survivor function as either r -+ -cc with g fixed, or y -+ -cc with
r fixed, are important specialcasesas
Pr{X > r,Y > g} : Pr{Y > y}

,lim
Pr{X > n,Y > a) - P4X > r},
nlim
i.e. the limit valuesare the marginal exceedance
probabilities.
Complete positive dependence

the joint survivor function sim-
When the variablesfollow completepositivedependence,
plifies as
Pr{X > r,Y > s}: min (Pr{X > r}, Pr{Y > s}) .
In addition,ifX andY havethe samemarginaldistribution,i.e. Pr{X > c} : Pr{Y > r}
for all o, then X: Y always.
Independence
When the variablesare independent,the joint survivor function simplifiesas
Pr{X > r,Y > a} : Pr{X > z}Pr{Y > s}.
These three features are helpful in assessingthe contour curvesof the joint survivor
function as:
r limit valuesgive marginal return level/return period combinations;
o for completepositivedependence,
the 7 year contourconsistsof the valuesof (r, g)
which give
I
min(Pr{X > z}, Pr{r > s}) Ar. .r'
where No6, is the number of observationsper year. Defining r" and g/" by
I
Pr{X > or} : Pr{}'> g"}
B.l. MARGI iAL DISTHIBUTIOAISK rOWI\r 163
i.e. they are the marginal ? year return levelsfor the two variables,we have that
the contour takesthe form:
{@,A), y,= E7 for r < :tr,, and z : ry for y < y7'}.
This is just two perpendicularlines at the marginalT year return levelsfor X and
Y. For practical purposes,it is suficient to consideronly the intersectionpoint for
thesetwo lines,i.e. (rr,Ur).
o for independencethe contour of equaljoint survivor function is simple to obtain.

It is not possibleto give a generaldescriptionof the form of contour shapeas it is
influencedby the form of the marginaldistributionsof the variables.
8.1 Marginal distributions known

In this sectionwe treat the marginal distributions as known. The situation is not un-
realistic as often we are able to incorporateadditional knowledgeabout the marginal
variablesinto the JPM. A consequence of includingthis extra information is that, in this
section,the only unknownfeatureof the ioint distribution of the seacondition va,riables
is the dependence structure. This allowsa better comparisonof our proposeddependence
modelswith the currentmethods.
We incorporatethe extra information through useof the joint densityft using the
1. known ma,rginaldistribution for eachelementof X,
2. estimateddependence
structure(i.e. as estimatedin Chapter7).
This joint densityestimateis modifiedfrom the full fitted form to havethe samemarginal
densitiesas thoseusedto simulatethe data. From this estimatedjoint densitythe associ-
ated joint survivor function, PI{SWL ) r,Hs > gr} is evaluatedby integrationover the
region{,SW-L> r,Hs > y}.
In practice,had contoursof the joint survivor function alreadybeenestimatedbefore
incorporating the knowledgeof the marginal distribution then the modification to the
joint survivor function contoursis simply obtainedby re-calibratingthe marginal scales
of the previousestimate.For example,changingthe 100year marginal retutn level from
the estimatedvalue to the known value. Consequently, estimatedcontoursof equaljoint
survivor function should agreeexactly with the contoursfrom the simulation model at
the limits (c -+ -oo and alsog -+ -oo) of the curves.
Unfortunately,in this study someinconsistencyarosein the evaluationof theseknown
marginalsso that the valuesusedin the current implementationwereslightly incorrectin
164 CHAPTER8. SIMULTANEOUS
EXTREMESOF (SWL,HS)
a few cases.The reasonfor this wasa changein the handlingof small waveheightswhich
the removalof resultedin extra observationsbeing generatedin somecases(lengthening
the tail of .I{5. Neverthelessthis error does not substantiallydetract from the inter-
comparisons.
Figures8.1-8.8showcontoursof equalioint survivorfunction estimatesfrom the fitted
model and derivedfrom the simulationmodel. Any significantdiscrepancybetweenthe
two setsof contourswould showthat either
1 . the statistical estimation,using the correctdependence

mode1,is poor, or
2. the fitted familiesof dependence

modelsare not sufficientlygeneralto fit thesedata
well.
We discussthe fits for eachdata set in turn beforedrawinggeneralconclusions.
Siml The two setsof curves,shownin Figure 8.1, exhibit excellentagreement.
Sim2 The two setsof curves,shownin Figure 8.2, alsoexhibit an excellentagreement.
SimS Here we have plots for (Surge,Hs) and (SWL,11s) in FiguresB.3 and 8.4 re-
spectively. The contoursagain seemwell reproducedby the fitted model. Some
discrepanciesoccur for (SWL,H1\, due primarily to the fact that as we have not
modified lhe SWL marginal distribution to follow the simulationmodel (which is
complexas it arisesfrom the combinationof independenttide and surge).
Sim4 Here we have plots for two different fitted dependencestructures: in Figure 8.5
the thresholdbivariate normal model is used,in Figure 8.6 the mixture of bivariate
normalsmodel is considered.In both casesthe Iitted statisticaldependence models
under-estimatethe contourcurves.For the thresholdmodelthe contoursare further
from the contoursgiven by the simulationmodel than for the mixture of bivariate
normals model. This is particularly obvious at the more extremecontours. For
thesedata the diagnosticplots suggestedthe variablescould reasonablybe taken
as completelydependent.From Figures8.5 and 8.6 we seethat a very closeagree-
ment with the simulationmodel is given by perpendicularlines - correspondingto
completedependence, and that completedependence is a better model than either
of the fitted models.
Simb Here we have the sametwo figrres as given for Sim4. We seethat the threshold
model averagesthe two forms of dependencethat occur in the simulation model,
the compromiseleadingto a poor fit (seeFigure 8.7). In contrast, the mixture of
bivariate normals givesan excellentfit. Note that here, although the dependence
8.1. MARGINAL DISTHIBU'ilOI\IS KI\IOWI\I 165
is strong, taking the variables to be completely dependent would be a substantial

over-estimate.
Once the marginal features have been corrected for, the main sourceof error in estimat-
ing contours of equal joint survivor function arises from not having a sufficiently flexible
parametric dependencemodel. The bivariate normal (and threshold fitted version) were
fine for Siml-Sim3 when this was the correct family of dependencestructures, but were
poor for both Sim4 and Sim5. On the other hand, the mixture of two bivariate normals
(which includes the bivariate normal as a special case) gives good fits in Sim4, when it
is not the correct form, as well as for Sim5, when it is the correct family. Note a.lsothat
differencesthat have been observedbetween the contours of the estimated and simulation
models occurred for combinations (r,g) where r > maxSWL and g > max Hs, and so
were due to long range joint extrapolations.
1()r) CHAPTER 8. SIMULTANEOUSEXTREMES OF (SWL,HS)
Figure 8.1: Siml: H1-SWL curvesof equaljoint survivor probability expressedin return
periods (in years)when the marginal distributionsare known
Figure 8.2: Sim2: Hs-SWL curvesof equaljoint survivor probability expressedin return
Iv
2345
8.1. MARGINALDISTRIBUTIOIISKI\rOm\r
Figure 8.3: Sim3: fI5 Surge curvesof equaljoint survivorprobability expressed

in return
Surge
Figure 8.4: Sim3: 115-SW,Lcurvesof equaljoint survivorprobability expressed

in return
periods(in years)when the marginaldistributions are known
168 CHAPTER 8. SIMUL'IANEOUSEXTREMES OF (SWL,HS)
Figure 8.5: Sim4: H;-SWL curvesof equaljoint survivorprobability expressedin return

periods(in years)whenthe marginaldistributionsa,reknown. Herethe dependence model
is the fitted bivariate normal thresholdmodel
-- 5flfi31"${P"dfJ,
Figure 8.6: Sim|: H s-SW L curves of equal joint survivor probability expressedin return
periods (in years) when the marginal distributions are known. Here the dependencemodel
is the fitted mixture of biva,riate normals model
-._Eliii[]EBnY,r*o-Jr
M ARGINAL DISTRIBUTIONS
KNOWN 169
Figure B.7: Sirn5: -[{5 SWL crtwesof equal joint survivor probability expressedin return
periods (in years) when the marginal distributions are known. Here the dependencemodel
is the fitted bivariate normal threshold model
Figure 8.8: SimS: fls SWL strves of equal joint survivor probability expressedin return
periods (in years) when the marginal distributions a,reknown. Here the dependencemodel
is the fitted mixture of bivariate normals model
CHAPTER8, SIMULTANEOUS
8.1.1 Comparison with current implementation

HR havemodified their estimatesof the joint survivor function contoursto incorporate
the extra marginal knowledgeby manual modification of the marginal scales. Results
havebeengiven for return periodsof 1, 10,100and 1000years. For comparisonwith the
plots in Figures 8.1-8.8,we show only the 10 and 1000year contour values;these are
tabulatedin Tables8.1 8.5 and givenin Figures8.9 8.12.
The comparisonwith the simulationmodelvaluesand the fitted modelaregivenbelow:
Siml Comparisonis not directly possiblehere as HR have separatelyestimated the

joint survivor function for wavesfrom two direction sectors. For direction sector
0' < e < 110", the 10 year and 1000year contourslie just abovethe respective
contoursfor the simulation mode1,whereasfor direction sector 110" < d < 360"
they lie much below. As the first sectoris slightly more likely, and the one most
likely to producethe extremecombinations,this providesa good estimate.
Sim2 The HR contour estimatesagreevery well with the simulation and fitted model
contours.
SimS The HR contourssignificantlyover-estimatethe (SW L, Hs) contoursas given by

both the simulationand fitted model. Someelementof this over-estimationis due
to the inconsistencyin the known marginal distributions for IIs, with the same
marginal being usedfor the simulationand fitted modelsbut slightly larger values
providedto HR for use.
Sim4 The HR contoursseempoor in this case. The strong form of dependencein the
simulated data has not been captured, and the contourshave been badly under-
estimated.For the 1000yea.rcontour,it is worrying that the error is so largein the
HR estimatedcontourwithin the rangeof the SW L data. Specifically,the contour
decreases rapidly in the Il5 value. This feature was not observedin either of the
statistically fitted models,or for the other four Sim data analyses.
SimS The HR estimatedcontoursappearto be good for low return periodsbut signifi-

cantly under-estimatefor high return periods.
In two casesthe HR implementationof current methodshas produceda good estimate

of the dependencestructure of extremeSWL and I1s combinations.In three other cases
the estimatesare poor) one in which there is a slight over-estimation(which would lead
to conservativedesign),and two casesof substantialunder-estimation(leadingto under-
design).
8.1. MARGINALDISTRIBUTIOI{SKt\rOWt\r 17l
Figure 8.9: Sim2: H1-SW L curvesof equaljoint sur-vivorprobability expressed

in return
periods (in years)when the marginal distributions are known. Curvesshownare 10 and
1000year return period for Estimatedmodel,Simulationmodel and HR estimates
-
5Pilfl3lfJ"{f$"J
Hti cslrmates
f,,
23456
Figure 8.10: Sim3:fls SWZ curvesof equaljoint survivorprobability expressed

in return
periods (in years)when the marginal distributionsare known. Curvesshownare 10 and
1000year return period for Estimatedmodel,Simulationmodel and HR estimates
sftqffiiBq{#$
172 CHAPTER 8. STMUT,iTANEOUS
EXTHEMES OF (SWL,Hs)
Figure B.11:Sim4: H1-SWL curvesof equaljoint survivorprobability expressedin return

periods (in years)when the marginal distributions are known. Curvesshownare 10 and
1000year return period for Estimatedmodel (Mixture Model), Simulationmodel and HR
estimates
Figure8.12: SimS:Hs-SWL curvesof equaljoint suryivorprobability expressed in return

periods (in years)wher the marginal distributions are known. Curvesshownare 10 and
1000year return period for Estimatedmodel (Mixture Model), Simulationmodel and HR
estimates
- EstimatedModel
- -- Simulationlvlodel
--- HR Estimates
8.1. MARGINAL DISTRIBUTIOIVSKl\tOWl\r 173
SWL 10 year 1{5 1000year .FIs

(0' < p s 110")| (110"< p < 360.) (0. < d < 110")| (110"< d < 360')
6.00 1.11 7.49
6.25 1.52 1.08 2.04 r.46
6.50 r.43 1.01 2.02 1.43
6.75 1.32 0.95 2.00 1.41
7.00 1.16 0.86 L.94 t.37
7.25 0.s4 0.67 1.82 r.28
7.50 1.48 1.09
Table 8.1: HR estimatesof points on the 10 and 1000year contoursof the joint survivor
function of (SW L, Hs) for Siml: separateestimatesconditionalon the wavedirection,d,
satisfying0" < 0 < 110' and 110"< 0 < 360'.
SWL 10 year I/s 1000year Il5

2.25 4.22 5.13
2.50 4.0B 5.04
3.93 4.83
3.00 3.50 4.65
3.25 3.00 4.39
3.50 2.06 3.93
A ?E
3.15
4.00 2.45
Table 8.2: HR estimates of points on the 10 and 1000 year contours of the joint survivor
function of (SW L,Ils) for Sim2, when the marginal distributions are known.
174 0HAPTER 8. STMULTANEOUSEXTREMES OF (SWL,Hi)
SWL 10 year Ils 1000year Ils

12.00 4.56 J. Oi)
L2.25 4-48 5.65

12.50 4.3r J.OJ
12.75 4.t7 J.OI
13.00 3.50 5.59

13.25 2.97 5.52
13.50 2.10 5.40
14.00 5.35
t4.25 5.03
14.50 4.69
r4.75 2.50
Table 8.3: HR estimatesof points on the 10 and 1000year contoursof the joint survivor
function of (SW L,.FIg) for Sim3,when the marginaldistributionsare known.
SWL 10 yea.rlls 1000year I1s

0.2 79.2
0.3 19.0
0.4 18.8
tt. i) 18.1
0.6 8.90 r t.o
0.7 B.B5 lo.1
0.8 8.47 t6.2

0.9 7.86 15.3
1.0 6.98 14.0
1.1 5.92 12.1
1.2 10.1
Table 8.4: HR estimatesof points on the 10 and 1000year contoursof the joiut survivor
functionof (SWL,Ils) for Sim4,whenthe marginaldistributionsare known.
8.2. MARGINAL DISTRIBUTIONSUI\rIfl{OWl{
SWL 10 year II5 1000year /Is

5.50 9.51
9.50
6.00 9.48 18.1
o.Jt, 9.18 r7.8
7.00 8.66 16.9
7.50 6.68 13.6
8.00 3.60 9.5
8.25 4.3
Table 8.5: HR estimatesof points on the 10 and 1000year contcursof the joint survrvor
function of (SI4zL,fls) for Sim5,when the marginaldistributions are known.
8.2 Marginal distributions unknown

Figures8.13-8.20showcontoursof equaljoint survivorfunction from the fitted model and
derivedfrom the simulationmodel. If the statistical model we haveusedto describethe
seaconditionsfits well then thesecontoursshouldbe in good agreement.Discrepancies
betweenthe two setsof contoursshowthat at least one asDectof the statistical estimate
is poor. Possiblesourcesof inadequacies
of the fit are;
1 . Poor marginalfits: the statisticalfamily is correcteachtime (exceptfor Sim3SW.L),

but a poor fit resultsfrom a poor fit of the GPD to the upper tail (detailsof these
fits are discussedin Chapter6).
2. Poor dependencestructure fits: the statistical family of fitted dependencemodels

is not alwaysthat from which the data were simulated,but even when it is, the
estimatedparametersmay be poor. This featurewasstudiedin isolationin Section
B.1 where the marginal distributions were treated as known, but contributeshere
when the marsina,ldistributionsare alsounknown.
We discussthe fits for eachdata set in turn beforedrawinggeneralconclusions.
Siml The two setsof curves(shownin Figure 8.13) exhibit good agreement.The only
obvioussmall differenceis in the extrapolationof the f/5 variable,with the fitted
model over-estimatingmarginal return levels.
Sim2 The two sets of curves(shownin Figure B.i4) exhibit moderateagreement.The

most obviousdifferencesstem from poor marginal estimates(for eachvariablethe
176 CHAPTER B. SIMUATANEOUSEXTREMES OF (SWL,HS)
fitted model under-estimatesreturn levels). The basicshapeof the contour curves

seemsreasonable,correspondingto the good fit of the dependencestructure found
in Section8.1. For thesedata the fitted model is from the samedependencemodel
family as usedto generatethe data.
Sim3 Herewe haveplots for (Surge,Hs) and (SWL,Hs) in Figures8.15and 8.16re-

spectively. The basic shapeof the contoursagain seemsfine, following from the
good estimation of dependencein Section8.1, but the return levelsfor Surge and
SW L are under-estimated,whereasthe ffs return levelsare over-estimated.
Sim4 Herewe haveplots for the two differentfitted dependence structures:in Figure8.17
the bivariatenormal thresholdmodelis used,whereasin Figure 8.18the mixture of
bivariate normalsmodel is considered.In eachcasethe samemarginal distribution
fits are used,so differencesbetweenthesefigulesare due to differencesin the fitted
dependencestructure. In Section8.1 we found that the mixture model fitted much
the better. That is still apparent here despitethe marginal return levels being
under-estimatedfor each margin (most notably for I1s which has the contrived
longer tail here). The dependencefeaturesare best consideredby comparingthe
shape of the contour curvesin the region of greatestcurvature. Most notable is
that in Figure 8.17this part of the curveis further from the true form derivedfrom
the simulationmodel than in Figure 8.18. This is particularly obviousat the most
extremecontours. The generalshapeof the contour of the fitted model in Figure
8.18seemsto representthe dependercewell (recallthat the simulationmodel here
had a strongerform of dependence in the extremevaluesthan any of the modelswe
are fitting).
SimS Herewe havethe sametwo figuresas givenfor Sim4. The most striking featureof
thesefiguresis the very poor estimateof return levelsfor "I1s,with the tail being
substantiallyover-estimated(althoughwithin the uncertaintyof the estimatesit is
not statistically significantlydifferent). The likely reasonfor this over-estimation
can be seenin the figures,as the data showthat the 300 year return level occurred
by chancein the sample (correspondingto a 10 year sample). For ,9Wtr, return
levelsare slightly under-estimated.This large distortion of the ma.rginalfeatures
makesdependenceaspectsof the fit hard to assess.
In conclusionwe have found that the main sourceof error in estimating contoursof
equaljoint survivor function is from marginal estimation (a feature largely outside the
remit of this current project). If no additional information were available,then relative
to the marginal estimation uncertainties,the dependencemodelling proposedhere for
8.2. M ARGINAL DIST EIBUTIOATSUI{Ifl\rOWl\r t77
(SW L, Hs) seemsreasonablygood. Qualitatively, what this meansis that to estimate

suchcontoursit seemsreasonableto recommendthat we need
r More than 10 yea,rsof data for eachof the marginalvariables;
r Simultaneous data for (SWL,Hs) coveringa 10 year period,

observational
to be able to estimate100year contoursand beyondwith any confidence.

In practice,simultaneousobservationaldata are difficult to obtain for longerthan 10
year periods,but separateinformationon SW L and IIs is availablebasedon longertime
periods of data. As the marginal and dependenceestimationhas beenseparatedinto a
two stageprocedure,additional marginal data can easilybe incorporatedinto the JPM.
This marginal information may take the form of
r extra data for the site: for examplewhen .FIsis hindcastfor a longerperiod,
r an estimateofthe distribution for the variableobtainedfrom a largerscaleanalysis:

for exampleestimatesof SW L extremesproducedby a spatial analysisas in Dixon
and Tawn (1994,1995).
This correctionfor ma,rginalfeaturesis particularly important for SW L as a direct ana.l-

ysis,usinga period of lessthan 18.61years,is likely to producepoor extrapolationsas
featuresof the nodal cycleof the tide are ignored.Incorporatingextra marginal informa-
tion overcomes suchpotentialweaknesses. In Section8.1we showedhow suchinformation
can be incorporatedinto the analysisto give improvedcontourcurves.
Finally we comparethe estimatesgiven by the approachdevelopedhere with those
obtainedby the current methodsimplementedby HR. There are two differencesbetween
the HR methodsand ours:
. marginal modelsand estimationtechniques,
r dependence
estimation.
Thus in addition to the differencesfound in Section8.1, wherethe marginal distribution

weretreated as known, here we haveadditionaldifferencesdue to marginal uncertainty.
178 EXTREMESOF (SWL, HS)
CHAPTERB, SIMULTANEOUS
Figure8.13: Siml: .fls-.9W.Lcurvesof equaljoint survivorprobability expressed

in return
periods(in years)
Figure 8.14: Sim2: Hs-SWL curves of equal joint survivor probability expressedin return
periods (in years)
I
8,2. MARGINAL DISTRIBUTIOIVS
UA'-I(NOI,4/'N 179
Figure 8.15: Sim3: fls-Szrgre curves of equal joint survivor probability expressed in
return periods (in years)
1
Surge
Figure 8.16: Sim3: 115-SWI curvesof equaljoint survivor probability expressedin return
periods (in years)
180 CHAPTER 8. SIMULTANEOUSEXTREMES OF (SWL,HS)
FigureB.lT: Sim4: ,[1s SIU.L curvesofequaljoint survivor probability expressedin return

periods (in years) for the fitted bivariate normal threshold model
- Esllmatedlrodel
--- SimulationModel
0.5
Figure 8.18: Sin4: Hs-SWL curves of equal joint survivor probability expressedin return
periods (in years) for the fitted mixture of biva.riate nolmals model
- 5f'lllEif,n{&da"J'
0.5
8.2. M ARGINAL DISTRIBUTIOATSUI\rI{IVOWN 181
Figure 8.19: Sim5: fI5-,9WL curves of equal joint survivor probability expressedin return
periods (in years) for the fitted bivariate normal threshold model
Figure8.20: Sim5: ,I15 ,914/-Lcurvesof equaljoint survivorprobability expressed

in return
periods(in years)for the fitted mixture of bivariatenormalsmodel
CHAPTER8. SIMULTANEOUS
8 . 2 . 1 Comparison with current implementation

HR haveimplementedthe current methodologyfor estimatingthe joint survivorfunction,
and producedcontoursof equalvalue of this function for return periodsof 1,10, and 20
years.For comparisonwith the plots in Figures8.13-8.20we only givethe 10 year contour
values,thesea,retabulatedin Tables8.6-8.10.For larger return periodsthe differencesin
the marginalextrapolationmethodsdominate,so suchcomparisonsa,reoutsidethe remit
of this project. The estimation of marginal distributions and dependencestructure by
HR usesdifferent methodsto thosedescribedin this report (for details seeHarvkesand
Hague, 1994). The comparisonwith the simulation model valuesand the fitted model
describedabovefor this contourcurve are givenbelow:
Siml Compa.rison is not directly possiblehereas HR haveseparatelyestimatedthe joint
survivorfunction for wavesfrom two directionsectors(thesedirectionsectorsdo not
exactlytie in with thoseusedto simulatethe data). Despitethe distribution of Str4r.L
being in a conditionalform, dependenton 0, the estimatedmarginal distribution of
SW L \s found to agreevery well with the simulationmodel.
Sim2 The HR contour estimatesgive return levelsfor .FIswhich over-estimateto the

samedegreeas our estimatedmodel under-estimates in Figure 8.14. For SWL lhe
HR estimatesare similar to our marginal return level estimates.The dependence
featuresare well captured.
SimS Herethe HR contour estimatesgive return levelsfor I15 which substantiallyover-

estimate (our estimates,given in Figure 8.15, also over-estimatebut to a lesser
extent), and are similar to our estimatesfor SWL. The dependence featuresseem
to be adequately captured.
Sim4 Here the HR contouris good at this return period. In Section8.1 the strong
form of dependencefor thesedata was not captured by the HR approach,so we
know that this good estimateis not sustainedfor larger return periods. Also the
marginal return levelsfor I19 at higher return periodsare under-estimated.The
SWl marginal distribution seemsto fit well.
Sim5 Again the HR contouris adequateat this return period. However,from Section8.1
we know that the strong form of dependence has not beencaptured (the curvature
of the contour is not sufficientlypeaked). The Ils return levelshave been under-
estimated (with the estimatedmarginal 10 year level the same as the simulation
model 5 year level) but is still - better than our estimate (given in Figure 8.19)
which is an over-estimate(correspondingto the simulation model 50 year level).
The SW L marginaldistribution seemsto be well fitted.
8.2. MARGINAL DISTRIBUTIOATSUI\rI(NOWI{
SWL Hg
(0"sd<1e0") (1e0'<d<270")
4.0 1.82 r.25
4.5 r.82 r.24
5.0 1.80 !.zo
D.O 1..77 1-22

6.0 1.69 r.77
D.J 7.52 1.09
6.75 1.41
7.0 L.23 0.97
7.25 0.88
0.68 0.73
Table 8.6: HR estimatesof points on the 10year contourof the joint survivor function for
Siml: separateestimatesconditionalon the wavedirection, d, satisfying0" < d < 190'
and190'<0<270".
si4/, | 1.50 r.75 2.00 2.25 2.50 2.75 3.00 3.25 3.50 3.75
Hs | +.or 4.bT 4.4s 4.28 4.06 z.zg 3.48 8.01 2.r7 o.sg
Table 8.7: HR estimatesof points on the 10 year contour of the joint survivor function
for Sim2.
s w , | 1 1 . 0 01 1 . 5 0 12.00 12.25 12.50 12.75 13.00

H5 | 5.25 5.19 5.00 4.87 4.70 4.38 3.61
Table B.B: HR estimatesof points on the 10 year contour of the joint survivor function
for Sim3.
184 CHAPTER 8. SIMULTANEOUSEXTREMES OF (SWL,HS)
SWL I 0.4 0 . 5 0 . 6 0 . 7 0 . 8 0 . 9 i . 0 1 . 1 1 . 2
H5 | 7.42 7.34 7.27 7.23 7.t5 6.88 6.48 5.86 4 . L 4
for Sim4.
sw L | 6.00 6.50 6.75 7.00 7.25 7.50 8.00

H5 18.52 8.20 8.03 7.62 7.r3 6.56 3.77
for Sim5.
Part III
Applications to the Estimation of

the Probability of Failure for an
Existing Design
Chapter I
Analysis of Overtopping Discharge

Rates
The methodologydevelopedin the first two parts of this report can be applied to any
designproblem. In this chapterwe examinean applicationwherethe mode of failure is
an extremeovertoppingdischargerate, and the analysisis requiredfor an existingdesign.
We fbcus on the problem of estimating return levelsof extremeovertoppingdischarge
rates,which can be rephrasedin terms of estimatingthe probability that the overtopping
dischargerate exceedssomecritical level.
In this chapterwe considerfive applicationsites (correspondingto the five simulated
data sets,Siml Sim5). For eachhypotheticalsife, the overtoppingdischargerate, Q6:,is
givenby equation(1.1.1),with the sea-r,r'all
designcharacteristics(a1,a2,o) givenin Table
9.1. This specificationof the problemmay be unrealistic,as the seaconditiondata are
offshoredata (as the waveshavebeenhindcast),whereasthe overtoppingdischargerate,
equation(1.1.1),presumes the wavesto be asmeasured at the sea-wall.Asaresultasmall
number of high waveheights,with very low steepness, tend to dominatethe overtopping
results.On the otherhand,the analysisis primarilyto
illustrate the application of the JPM to a design problem;
enable a comparison between the SVM and JPM;
enable comparison between the methods developed in this report and methods cur-
rently implemented, such as the two different levels of analysis described in Section
In this chapterwe first describespecialfeaturesof the methodsas appliedto overtop-

ping dischargerates (Section9.1), beforeapplying thesemethodsto the simulateddata.
A benefit of usine simulateddata for the sea condition data is that we ca^nderive the
187
188 CHAPTER 9, AATALYffS OF OVERTOPPING DISCHARGE RATES
associateddistribution of extremeovertoppingrates exactly,thus simplifying the com-

parisonsbetweenthe methods. To clarify our presentationof results, we focus on the
estimationof the 100year overtoppingdischargerate, giving estimatesand compa"risons
for eachdata set,from Siml in Section9.2 to Sim5in Section9.6. Finally in Sectiong.B,
for cornpleteness,
we giveestimatesfor the 10 and 50 year overtoppingdischargeratesfor
eachof thesesimulateddata sets.
Simulated data Ul A2 u
Siml 0.019246.96 8.5
Sim2 0.019246.96 5.0
SimS 0.019246.96 14.5
Sim4 0.019246.96 2.5
Sim5 0.019246.96 9.0
Table 9.1: Information on the structural designfor the simulateddata. Herethe sea-wall
is a simpleslopingwall with a 1:4 slope,n1 and a2 dimensionless constantsand u is the
crestelevationin metres(relativeto ODN exceptfor SimBand Sim5which are relativeto
Chart Datum). Seeequation(1.1.1)for detailsof how thesedesignparameters influence
overtoppingdischargerates.
9.1 Background Methods

We apply five differentmethodsof analysisof extremeovertoppingdischargerates,
two applicationsof the structure variablemethod,
three levelsof applicationof the joint probability method.
First we outline special featuresof these methods in their application to the specific
problem of estimatingextremesof overtoppingdischargerates, Q6:.
SVM applied to Qs
The standard SVM, as describedin Section2.1, is applieddirectly to the structure vari
able, Q" the overtoppingdischargerate. The resultswerederivedby applying the GPD
thresholdmodel to a rangeof thresholds(correspondingto upperquantilesover the range
e0- ee.5%).
9.1. BACKGROUNDMETHODS
SVM applied to )ogQg

The SVM is appliedindirectly,by statisticallyanalysinga transformedstructure variable,
log Q5r,and extrapolation.The return levelestimateson the Qc,scaleare givenby taking
the exponentialofthe estimateson the logQc' scale.The resultswerederivedby applying
the GPD thresholdmodelto a rangeof thresholds(corresponding to upper quantilesover
the range90 - 99.5%).
Of the two SVM, this is the closestin spirit to the generalapproachHR use for
univariate extrapolation to estimatereturn levels. We have applied our methods in a
routine approach,whereasin practicalstudiessubjectivechoice,suchas thresholdchoice,
and assessment of fit, would be carefully taken. The HR approachrequiresa greater
numberof cho'ices to be made,whichprotectsagainstphysicallyunrealisticextrapolations,
but potentially may force the extrapolationto be inconsistentwith the extremal data.
The HR estimatesfor theseSim data setswill be givenin the HR report which describes
the practical implementationof the methods.
This all acts as a note of warning not to extrapolatewithout at least assessingthe
sensitivity to the aspectsof the model that are subjectivelychosen,and evaluatingthe
estimationuncertaintv.
JPM applied using statistical model

Thejoint distributionof (SWL, Hs;,7.7) is givenby thejoint distributionof (SWL, Hs, S)
through the joint distribution ot (SWL, Hs) and the conditional distribution of SlHs
(due to the conditionalindependence of ,Sand SWL givenI15). The estimateof the joint
distributionof (SWL,11s) is givenin Chapter6 and Section7.1,whereasthe estimated
conditionaldistribution of S l,FIsis given in Section7.2. We haveappliedtwo versionsof
this method,basedon
r the fully estimatedjoint distribution of (SWI, I/s, S),
r the joint distribution of (SW L, ils,,9) estimatedwith the marginaldistributionsof

SWL and I1s taken to be known.
In eachcasethe resultsare obtainedusing the generalmethodologydescribedin Section

2.2.
Current implementafion of JPM: basic method

This methodcorresponds to the basicmethoddescribedin Section1.2. The estimated
joint survivorfunctionvaluesfor (SWL,H1) are givenin ChapterB. HR repeatedthe
analysisin eachapplicationusingtwo differentconstantwavesteepness
values:one being
190 CHAPTER 9. AI{ATYSIS OF OVERTOPPINGDISCHARGERATES
S: 0.06 (commonto all analyses),and the other a valuederivedfor eachindividual data

set. Two versionsof this method havebeenapplied;thesea,re
r estimatesof 1, 10,100 and 1000year return levelsof overtoppingdischargerates
basedon the estimatedjoint survivorfunction of (SWL,H1) with the marginal
distributions of SW L and ffe taken to be known,
o estimatesof 1, 10 and 20 year return levelsfor overtoppingdischargeratesbasedon

the fully estimatedjoint survivorfunctionof (SWL,Hs).
Current implementation of JPM: refined method

This methodcorresponds to the refinedmethoddescribedin Section1.2, in which the
exactfailure region(with respectto Hs and.SWL, but not with respectto 72 as steepness
is treated as fixed) is used. Here the joint densityof (SWL,H;) is derivedfrom the
estimatedjoint survivor function givenin the basicmethod (i.e. from the estimatesgiven
in Chapter 8) by suitable differencingover a grid (interval lengths of 0.5m). In each
applicationa value of constantwavesteepness is used(hereS: 0.06and a valuederived
from the data). This choicewas madefor consistencywith the basicmethod, but is not
always made in current 'risk analysis'studies. This method has only been applied to
Sim2, wheretwo levelsof implementationwereundertaken:
r estimatesbasedon the estimatedjoint densityof (SW L, fle) with known marginal
distribution,
o estimatesbasedon the exactjoint distributionof (SWL,Hs).
9.2 Application to Siml

For this hypotheticalsite, the five largestovertoppingdischargerates,and the associated
combinationsof (SWL,H;,T7), from a hypothetical10 year period are givenin Table
9.2. It is seenthat the largestobservedrate is an order of magnitudegreaterthan the
next largestvalues.This has a signilicantimpact on extrapolationsapplieddirectly to the
Qg data, producingmassivelyover-estimatedreturn levels. The estimateis not shown
in Figure 9.1 which illustratesthe other estimatesof the 100year overtoppingdischarge
rate. The target return levelsfor Q5r,from the simulationmodel,are given in Table 9.3.
The other estimatesare discussedbelow:
SVM the estimatedlevel is shownagainstthe thresholdfor the SVM appliedto log 8c.
The level is under-estimatedfor thresholdsbelowthe 96% quantile,but abovethis
thresholdlevel the return levelis well estimated.
APPLICATION'TO SIML 191
JPM statistical model the estimateis a slight over-estimateof the level (giving the
actual 140 year level). Here the statistical model is basedon estimatedma.rginal
distributions. Known marsinal distributionsare not usedbecauseof the direction-
ality.
JPM current implementation estimatesof the 100year level are givenfor two drrec-
tion sectorsonly, so direct comparisonis not possiblefor thesedata. Herewe report
only estimatesobtainedusing known marginal distribution for fls and SWL. For
the 0' < d < 110"directionsector,HR take ,9 : 0.06which givesthe estimated100
year overtoppingdischargerate of 0.016(corresponding to the actual 30 yea.rrate).
Also for the 110' < d < 360"directionsectorHR take S:0.025 obtainingthe 100
year overtoppingrate of 0.019(corresponding to the actual 40 year rate).
For this site eachof the methods appearto produceequally acceptableestimates(with

the exclusionof the SVM to Qc'). This is not too surprisingas
1. for this site, the estimatedjoint survivorfunction givenby the statisticalmodel and
current implementationseemsadequatelyestimated,
2. the dependence between(SWL,Hs) is moderately. strong,so the errorin the failure

region of the current implementationis not large,and
3. the variation of ^9given I1s is large,is relativelysmall, so taking this value as fixed
is a reasonableapproximation.
rn
SWL HS alc
7.28 7.47 3.66 0.01676
7.48 0.99 3 . 1 0 0.00397
6.59 l.D4 J./O 0.00237
6.98 1 . 2 5 3.45 0.00221
o.J1 0 . 5 9 6.82 0.00142
Table9.2: The largestfive seaconditionvaluesfor Siml in termsof the structurefunction,

overtoppingdischargerate Qs givenby equaticn(1.1.1).HereSWL.is in terms of metres
relative to ODN, I1s is in metres,arrdT2 in secs.Overtoppingdischargerate, Q6:,is in
metre3/sec/metre.
192 CHAPTER 9. AAIAIYSIS OF OVERTOPPING DISCHARGE RATES
Returnperiod Return level Return period Return level

(yearsJ Qc (yearsJ Qc
5 0.0040 100 0.0300
10 0.0075 200 0.0455
20 0.0127 300 0.0533
30 0.0161 400 0.0632
40 0.0192 500 0.0694
50 0.0213 600 0.0727
60 0.0232 700 0.0782
70 0.0251 B00 0.0829
80 0.0270 900 0.0856
90 0.0288 1000 0.0874
Table9.3: Siml: Returnlevelsfor Q6 (in metre3/sec/metre)

from simulationmodel.
Figure9.1: Siml Q6 100year return valuefor the SVM appliedto logQ5' and the JPM
basedon the bivariate normal dependencemodel for (SWL,Hs). The thresholdaxis is
relatedto applications
of the SVM
E
9.3. APPLICATION'IO SIM2
9.3 Application to Sim2

combinations of (SWL,Hs,Ts), are givenin Table9.4. As the largestobserved 8c rates
are consistentwith the other largelevels,we can apply the SVM to Q6rdirectly, but also
apply the SVM to the log Qc, data. The target return levelsfor Q6 for the simulation
model are given in Table 9.5. Estimatesof the 100 yea,rovertoppingdischargerate are
shownin Figure 9.2. Additionally, we haveestimatesfrom current implementationat the
basicand refinedlevelsof application.All theseestimatesare discussedbelow:
SVM the estimatedlevelis shownagainstthe thresholdfor the SVM appliedto Q6 and

to log Q5,. For Qs, the level is grosslyover-estimatedfor all thresholdsup to the
99% quantile. Alternatively,for logQ6, levelsare consistentlyunder-estimatedfor
all thresholds(giving the actual 25 year rate).
JPM statistical model the estimategiven for the statistical model very slightly over-
estimatesthe rate (giving the actual 140year rate). Here, the statistical model is
basedon estimatedmarginal distributions.
JPM current implementation: basic method If the standardvalue of S : 0.06 is

taken then the estimated100year valueis 0.56 (correspondingto a value lessthan
the actual 5 year rate). In contrast, taking ,9 : 0.03 (determinedfrom the Sim2
data) ihe estimateis increasedto 1.5 (the actual 35 year rate). Here, the known
marginal distributions are used.
JPM current implementation: refined method Using,9 : 0.06the refinedmethod

gaveequallypoor estimates.Howevertaking S - 0.03leadsto a slight over-estimate
(correspondingapproximatelyto the actual 200 year rate).
Of all the methods applied to the data, only the JPM statistical model and the JPM
current implementation(with ,9 speciallychosen)appearto produceacceptableestimates.
Possiblereasonsfor this poor performanceof the other estimatorsare that
1. the negativecorrelationof (SWL,H1) meansthat the basic rnethodof current

implementationis likely to under-estimate.
2. the high variability of S for largevaluesof /{s meansthat estimatesbasedon fixing

S at somevalue, a feature of all currently implementedmethods,will give biased
estimat€s.
For this hypotheticalsite we comparethe importanceof eachof the abovefeatures.

194 CHAPTER 9, A]VALYSIS OF OVERTOPPING DISCHARGE RATES
For the current implementationthe refinedmethod givesa higher overtoppingdis-

chargerate than the basicmethod,asit includesa wider rangeof loadingconditions
which contributeto the failure region. There is a degreeof coarseness
in the refined
method due to the coarseness of the grid used for convertingthe estimatedjoint
survivor function into a joint densityfunction. More refinementwould be possible,
but is time-consuming,sinceit relieson manualinterpolationof return periodsfrom
the estimatedcontoursofjoint survivorfunctions. This stageofthe refinedmethod
is made more difficult when the known marginal distributions are used,sincethe
marginal axesare a non-lineartransformationof marginal axeswhen the marginal
distributionis estimated.
A feature of the coarseness is that the refined method will tend to overestimate
overtopping(if ,9 is really a constantvalue),becausecentral valuesof wave height
and water level are usedfor each'cell', whereasthe probability density function is
concentratedtowardsthe lowercornersof the cell. This featureis seenin the above
estimates.
IT
SWL tts rz Qc
2.38 3.34 10.31 1.1465
2.18 3.22 9.28 1.0920
2.46 3.42 10.44 0.939
1.88 3.89 11.00 0.9314
3.26 3.64 7.65 0.8798
Table9.4: The largestfive seaconditionvaluesfor Sim2in termsof the structurefunction,

overtoppingdischarge rateQc givenby equation(1.1.1).Here SWLisintermsof metres
relative to ODN, fIs is in metres,and 7s in secs.Overtoppingdischargerate, Q6r,is in
metre3/sec/metre.
9.3. APPLICA'IION TO SIM2 195
Return period Return level Return period Return level

(years) Qc (years) Qc
5 0.9476 100 1.8673
10 1.1516 200 2.L462
20 7.3640 300 2.3205
30 1.4881 400 2.4042
40 r.5745 500 2.4703
50 1.6308 600 2.5544
60 1.6991 700 2.6169
70 r.7613 800 2.7049
80 1.8036 900 2.7498
90 1.8357 1000 2.7517
Table 9.5: Sim2: Return levels for Q6 (in metre3/sec/metre) from simulation model
Figure 9.2: Sim2 Q6 100year return value for the SVM applied to Q6 and logQc, and
the JPM basedon the bivariatenormaldependence modelfor (SWL,Hs\. The threshold
axis is related to applicationsof the SVM
- vadabt€
Slructure on ec
.. gifri'nffi$ffi,liiiHl*
0.94 0.96
Throshold
196 CHAPTER 9. AÎALYSISOF OVERTOPPI]VGDISCHARGERATES
9.4 Application to SimB

For this hypotheticalsite the five largestovertoppingdischargerates,and the associated
combinations of (SWL, Hs,T2), aregivenin Table9.6,andthe targetreturnlevelsfor Qc,
from the simulationmodel,are givenin Table 9.7. The largesttwo observedovertopping
dischargerates are twice the sizeof the fourth and fifth largestobservations,which leads.
to the SVM based on the Q6 data producing a long estimatedtail resultiug in over-
estimatedreturn levels. The 100year overtoppingdischargerate estimatesare shownin
Figure9.3,theseare discussed below:
SVM the estimatedlevelis shownagainstthe thresholdfor the SVM appliedto Q5,and

logQ5'. The level is badly over-estimatedfor Q5' for all thresholds,and slightly
over-estimatedfor log Q6rdata.
JPM statistical model the estimateis very good (the estimatedvalueis the artual 90
year rate). Here,the statisticalmodelis basedon estimatedmarginaldistributions.
JPM current implementation For ,9 : 0.06 the estimateof the 100year rate is 1.86
(correspondingto the actual 25yearrate), whereasfor the derivedvalueofS : 0.048
the estimated100year rate is 2.3 (correspondingto the actual 70 year rate). Here,
the known marginaldistributionswereused.
With the exclusionof the SVM appliedto Qc,lhe methodsagreeto a reasonabledegree.

In comparingthesemethodswe shouldnote that:
1. For the currentimplementation,knownmarginaldistributionhavebeenusedwhereas

for the statistical model estimatedmarginaldistribution wereused;
2. The estimatedjoint survivor function is slightly over-estimatedin the current im-

plementationbut not in the statistical model. This should lead to over-estimated
overtoppingdischargerates using the current implementation;
3. The failure regionusedin the current implementation(basicmethod) alwaysleads

to under-estimationof overtoppingdischargerates;
4. The omission,in the current implementation,of the largevariation of S, giventhat

I15 is large, leadsto under-estimation.
Thesefeaturesseemto havebalancedeachother for thesedata to producean acceotable

estimatefor the current implementationof the JPM.
9.4. APPLICATION TO SIMT 107
SWL Hg Tg Qc
1 3 . 1 14-J I 8.57 2.20413
12.88 4.95 7.92 1,8589
12.45 4.30 8.70 r.2745
12.78 3.62 7.76 0.9231
1 1 . 8 3 5.05 7(:)0 0.8149
Table9.6: The largestfive seaconditionvaluesfor Sim3in termsof the structurefunction,

overtopping dischargerate Qc givenby equation(1.1.1).HereSWL is in termsof metres
relative to ODN, lls is in metres,and Tz in secs.Overtoppingdischargerate, Q6r,is in
metre3/sec/metre.

(years) Qc (yearsJ Qc
5 1.1171 100 2.5782
10 1.4094 200 2.9652
20 7.7253 300 3-3309
30 1.9158 400 3.6324
40 2.0449 500 4.7264
50 2.1698 600 4.5034
60 2.2977 700 4.5779
70 2.3685 800 5.4004
80 2.4291 900 5.5087
90 2.4986 1000 5.7691
Table 9.7: Sim3: Return levels for Q6 (in metre3/sec/metre) from simulation model.
198 CHAPTER 9. AAIALYSIS OF OVERTOPPING DISCHARGE RATES
Figure 9.3: Sim3 8c 100 year return value for the SVM applied to Qc' and log Q5rand
the JPM basedon the bivariatenormaldependence modelfor (SWL,Hs). The threshold
Et!ilffifi,[q'ffi[,9$&,""
.E
0.94
Threshold
9.5. APPLICATION TO SIM4
9.5 Application to Sim4

For this hypotheticalsite the five largestovertoppingdischargerates,and the associated
combinations of (SWL,H;,T7), are givenin Table9.8, and the target return levelsfor
Q6, from the simulationmodel,are givenin Table9.9. Relativeto the previoushypothet-
ical sites(Tables9.2,9.4 and 9.6)the overtoppingdischarge ratesare muchlarger.This
featureis due to the extremewaveheight and waveperiod data beingmuch larger- Here,
variationsin SWL arelargelyunimportant in affectingthe overtoppingrate. Correspond-
ingly,the dependence between(SWL,Hs) is alsounimportant.This is disappointingas
thesedata havestrongdependence for (SWL,Hs) whichis generallypoorly estimated).
This hypotheticalsite has exceptionallylarge waves,for two reasons
1. the data on which this hypotheticalsite wasbasedis Christchurch,which has large

offshorewaves,
2. we choseto lengthenthe tail of the wave height variable (for overtoppingstudies

this was clearly an inappropriatechoice).
The simulatedwavesard completelyunrealisticas inshorewaves,so the resultingovertop

pinggivenby equation(1.1.1)is quitefalse.Nevertheless,asanassessment of performance
under varying conditions,the comparisonstill has va.lue.The 100 year overtoppingdis-
chargerate estimatesare shownin Figures9.4 and 9.5, theseare discussedbelow:
SVM the estimatedlevelis shownagainstthe thresholdfor the SVM appliedto Qs and

logQ6. The level is badly over-estimatedfor 8c at thresholdslessthan the 98%
quantile and under-estimatedfor higher thresholds.No stability with thresholdis
seen,so this method is poor. For logQ6' data, the estimatesare stablewith respect
to thresholdbut under-estimatebadly, (the estimatedrate correspondsto actual 10
yearrate).
JPM statistical model Four estimatesare siven here:
o an estimate based on the threshold bivariate normal dependence model for

(SW L, H s) with estimated marginal distributions (shown in Figure 9.4),
. an estimate based on the threshold bivariate normal dependence model for
(SWL, Hs) with known marginal distributions (shownin Figure 9.4),
. an estimate based on the mixture of bivariate normals dependencemodel for
(SW L, Hs) with estimatedmarginal distributions (shownin Figure 9.5),
r an estimate based on the mixture of birariaie normals dependencemodel for
(SW L, Hs) with known marginal distributions (shownin Figure 9.5).
200 CHAPTER 9. AAIALYSF OF OVERTOPPI]VGDISCHARGERATES
Three featuresemergefrom consideringthesefigures:
1 . all four estimatesare reasonablebv comparisonwith the SVM. When the

marginaldistributionsare estimated,the actual return period of the estimated
rate is 200years,whereaswhenthe known marginaldistributionsare used,the
return period of the estimateis 400 years;
2 . the choiceof dependencemodel makes almost no differencehere (this was
explainedabovefrom an intuitive standpoint,the estimateshereconfirmthis);
2
the useof knownmarginaldistributionshasmadethe estimatesworsethan the
estimatesbasedon unknownmarginal distribution. This is counter-intuitive,
but resultsfrom the Q variabledominatingthis extrapolationwith this com-
ponent of the model unaffectedby treating the distributions of fls and 5l4/tr
as known. Hence,any error in its modelling (i.e. the modellingof ,9l}/s and
11s)outweighsother aspectsof the joint distribution model. Recall,this is the
site for which a falserelationshipbetweenS and fls wasfound in Section7.2.
JPM current implementation For ,5 : 0.06the esti.mateof the 100 year rate is 15.0
(correspondingto the actual 18 year rate), whereasfor ,5 - 0.04 (derivedfrom the
data) the estimateis 20.8 (corresponding to the 45 year rate). Here,the known
marginal distributions were used. Holding ,S constant has been beneficial,as the
sensitivity to Ts has beenreduced.
For these data, in Section 7.1.2 we identified that without being overly conservative
(SWL,Hs) couldbe takento be completelydependent.Howeverincreasingthe depen-
dencemakesthe JPM statistical model estimateworse.In contrastfor the JPM current
implementation, with S:0.04, with thesevariablestakento be completelydependent
the actual 100year rate is estimatedalmostperfectly.
Detailed comparisonof methodsis unjustifiedhere. Again the JPM (at a1llevelsof
implementation)seemsto be preferableto the SVM.
9.5. APPLICATION TO SIM4
SWL Hs IZ Qc
I.I7 6.64 1 2 . 1 3 8.003
0.50 6.38 t4-75 7.924
0.75 6.54 72.t5 6.444
0.89 6.89 7t.22 6.413
0.97 5.75 12.24 O . U O J
Table9.8: The largestfive seaconditionvaluesfor Sim4in termsofthe structurefunction,

overtoppingdischargerateQ6 givenby equation(1.1.1).Here SWLisintermsof metres
relative to ODN, 1/s is in metres,and 7z in secs.Overtoppingdischargerate, Q5r,is in
metre3 /sec/metre.
Returnperiod Return level Returnperiod Return level

(yearsJ Qc (years) Qc
5 8.6723 100 27.1627
10 11.7261 200 33.5968
20 15.5245 300 35.9757
30 17.8495 400 3B.0409
40 t9.9222 500 40.8586
50 21.6399 600 44-3648
60 23.0572 700 46.2059
70 24.3832 800 47.9264
80 25.6450 900 48.3332
90 26.3787 1000 48.8283
Table 9.9: Sim4: Return levelsfor Qs (in metre3

fsec/metre) from simulationmodel.
202 CHAPTER 9. AI\IA-LyS6 OF OVERTOPPING DISCHARGE RATES
Figure 9.4: Sim4 Q6 100 year return value for the SVM applied to Q6 and logQg and
the JPM based on the bivariate normal dependencemodel fot (SW L, Hs) with estimated
and known marginal parameters. The threshold axis is related to applications of the SVM
.?
g
0.94. 0.96
Threshold
Figure 9.5: Sim4 8c 100year return valuefor the SVM appliedto Qs and log Q6 and the
JPM basedon the mixture of biyariate normalsdepeldencemodel for (SWL,H1) with
estimatedand known marginal parameters.The thresholdaxis is related to applications
of the SVM
- Slructur€
Vadableon Oc
ilflYFl90#8{,"l,Xilis?.
ixtureof BVN
--- f;HiFI""#811ff
ffi3,lid.Pxtureol BvN+fixed
6
E
;V
0.94 0.96
Threshold
APPLICATION TO SIM\ 203
9.6 Application to SimS

combinations of (SWL, Hs,Tz), are givenin Table9.10,and the target return levelsfor
Qs, from the simulationmodel,are given in Table 9.11. The largeovertoppingdischarge
rates for Sim5 are similar to thosefor Sim4 (seeTabte 9.8). This featureis again due to
the long tail of the extremewaveheight distribution. This resultsin variationsin SWL
being largely unimportant in determiningthe overtoppingrate. The dependence between
(SW L, Hs) is alsounimportant. The simulatedwavedata, althoughconsistentwith data
at Dowsing,representoffshorewaves,and are unrealisticas inshorewaves,so the resulting
overtopping,givenby equation(1.1.1),is dominatedby their values.This reducesthe
value of the comparisonat this site, as the care given to modellingdependencebetween
(SW L, H s) in Chapters7 and 8 is not exploitedwhen (Hs,Ts) dominatethe variation in
8c values. The 100 year overtoppingdischargerate estimatesare shownin Figures9.6
and 9.7; theseare discussedbelorv:
SVM the estimatedlevelis shownagainstthe thresholdfor the SVM appliedto Q6:and
IogQ6'. The levelis badly over-estimatedfor Q6, with the degreeof over-estimation
reducedfor larger thresholds. No stability with thresholdis seen,so this method
is poor. For logQ6 data, the estimatesare stablewith respectto threshold,but
under-estimatethe rate (corresponding to the 50 year rate).
JPM statistical model Four estimatesare given here:
e an estimate based on the threshold bivariate normal dependencemodel for

(SWL, Hs) with estimatedmarginaldistributions(shownin Figure9.6),
o an estimate basedon the threshold bivariate normal dependencemodel for
(SWL,Hs) with knownmarginaldistributions(shownin Figure9.6),
r an estimatebasedon the mixture of bivariate normalsdependencemodel for
(SWL, Hs) with estimatedmarginaldistributions(shownin Figure9.7),
o an estimatebasedon the mixture of bivariate normalsdependencemodel for
(SWL, Hs) with knownmarginaldistributions(shownin Figure9.7).
Three featuresemergefrom consideringthesefigures:
1. The useof knownmarginaldistributionshasimprovedthe estimates,asthe I1s

upper tail was badly over-estimatedwhich led to over-estimatedovertopping
return levels;
2. The choiceof dependence model makessomediflerencehere,with the ml.<ture
of bivariate normalsproducingthe best estimates;
204 CHAPTER 9. AATA'YSISOF OVERTOPPINGDISCHARGERATES
3. All four estimatesare reasonableby comparisonwith the SVM appliedto Qc.

By comparisonwith the SVM applied to log Qg, only the known ma.rginal
distributions casesare better. For known marginal distribution, with either
dependencemodel,the JPM statisticalmodel estimateis very good.
JPM current implementation For S : 0.06the estimateof the 100year levelis 10.32
(correspondingto the actual 20 year level),whereasfor ^9- 0.05 (derivedfrom the
data) the estimateis 15.1 (corresponding to the actual 45 year rate). Here, the
known marginal distributionswereused.
The statisticalmodellingusedin the JPM givesthe bestestimateshere. As Ils dominates

the overtoppingdischargerate distribution, the choiceof dependencemodel has only a
small infuence, with the most critical componentof the JPM being accurateestimation
of the marginal distribution of f15. Again, the JPM performsbetter than the SVM.
SWL HS 41
Qc
5.89 10.48 1 1 . 3 1 6.236
7.64 7.02 9.87 5.982
5.90 9.83 1 1 . 3 4 5.680
6.01 8.63 10.34 3.845
6.47 7.64 o o o 3.650
Table 9.10: The largestfive seacondition valuesfor Simb in terms of the structure func-
tion, overtoppingdischargerate Qs givenby equation(1.1.1).HereSWL is in terms of
metresrelativeto ODN, Ifs is in metres,and Q in secs.Overtoppingdischargerale, Qs,
is in metre3/sec/metre.
9.6. APPLICATIONTO SIM\

(years) Qc (yearsJ Qc
5.9903 100 19.4855
10 8-2648 200 23.8109
20 11.0693 300 26.7994
30 12.9900 400 29.5869
40 14.44B8 500 30.7396
50 15.6470 600 32.4526
60 16.6410 700 33.5968
70 17.4921 800 34.3186
80 18.2494 900 35.7950
90 18.9181 1000 37.4294
Table9.11:Sim5:Returnlevelsfor Q6 (in metre3/sec/metre)

from simulationmodel.
Figure 9.6: SimSQ5' 100 year return value for the SVM applied to Qs and log Q5 and
the JPM basedon the bivariatenormal dependence model for (SW L, Hs) with estimated
and known marginalparameters.The thresholdaxis is relatedto applicationsof the SVM
- S!rucluleVaiabtg.qn Qc
----- ilSY'Fl!"8.U'P'R"l,Ytr$?-"ur
-- slruclurevadableon loooc
JolntProbabllity
l\rethod+BvN+lixed
marglns
-l
0.94 0.96
Thfeshold
206 CHAPTER 9, ANA.LYffS OF OVEHTOPPINGDISCHARGERATES
Figure 9.7: Sim5 Q6 100year return valuefor the SVM appliedto Q6 and log Qc' aud the
JPM basedon the mixture of bivariate normalsdependencemodel for (SWI, f/s) with
of the SVM
- V?riabte
Struclut€ 0n Oc \
o,BvN
---- trfriFii{rli$&:bY$'d,?r.M*,ure \
JJiHiF'1"%XB'iifrl'&&llod+ffixt're
otevru*nxeomarsins \
\
\
6^ \
8- \
\
.
__:___-_-_:___:_-,-:---
0.94 0.96
Threshold
9-7. COMMEN? OAI CURREAIT IMPLEMENTA?/OT\TRESUL?S
9.7 Comment on Current Implementation Results

The estimatesof extreme overtoppingrate given by the basic method of the current
implementationof the JPM for the 100 year level were found to be under-estimatesin
Sections9.2 9.6. Here we briefly look to seeif this under-estimationis consistentover
data sets. In Table 9.12 we give the actual return period for the levelspredictedby the
basic method to be the 100 and 1000year levels.Two return periodsare given for each
data set, correspondingto estimateswith S: 0.06 and with a fixed value for ,9 derived
for eachdata set. For thesehypotheticalsites,the typical level of protectionofferedby a
designto the 100year rate would actually be approximately25 years(if S : 0.06is used),
or more realistically 45 years (when ,S is derivedfrom the data). This is a reasonably
consistentbias. The estimatesfor the 1000year level show that this degreeof under-
estimationis similar for other return levels,althoughis not as consistentlyobservedover
the data sets. This suggeststhat a rough re-calibrationrule is to double the estimated
return period ofreturn levelsestimatedusingthe currentimplementation(basicmethod).
To givesomemea,sure of the performanceof the currentimplementationrelativeto the
JPM statistical model we also summariseestimatesfor the JPM statistical model. For
Siml-Sim5 the actual return period of the overtoppinglevelestimatedto havea 100year
returnperiodare140,140,90,200(400)and 1000(85)respectively (wherethe numbersin
brackctsare estimatesobtainedusingthe knownmarginaldistributions). Theseestimates
arenot directly comparablewith thosefrom the currentimplementationasherethe known
marginal distributions havenot beenused. For Siml-Sim3 the performanceof the JPM
statisticalmodelis better than the currentimplementation.Sim4and Simbarecompletely
unrealisticdata sets,particularly in relation to the structure function, hencecomparison
is essentiallymeaningless.For completeness the methodscompareas follows: for Sim4
the current implementationis best, yet this is in contra.stto the estimatesin Chapter 8,
so is totally due to the modellingof S; and for Sim5, when we are comparinglike with
like, i.e. 85 v 45 year return periodsthe JPM statistical model appearsbest.
9.8 Other Figures

For completeness, we alsogive the estimatedreturn levelsfor return periodsof 10 and 50
years. Generallythesevalues,shownin Figures9.8-9.21,showthe samefeaturesidentified
for the 100year level of overtoppingdischargerate.
The comparisonsso far haveexcludedestimatesfor the JPM applied using the basic
current implementationwith estimatedmarginals. Table 9.13 gives estimated 10 year
Qc valuesbasedon the current implementationof the JPM. We seethere is a reason-
208 CHAPTER 9. AI{AIYSIS OF OVERTOPPINGDISCHARGERATES
Site s: 0.06 Derived5

100yearI lOO0
y.u. 100yearI toooy.u.
Siml 4J 750 40 550
Sim2 <5 5 JD 250
Sim3 2000 70 2000
Sim4 25 150 45 480
Sim5 20 100 45 350
Table 9.12: Actual return period of designlevelsestimatedby the current implementation

(basiclevel)of the JPM: for S :0.06 and when S is derivedfrom the data. For Siml
these results correspondto the dominant direction. Owing to the dominanceof wave
period for Sim4 and Sim5 theseestimatesshouldbe largely ignored.
ably consistentpattern of improvementin the estimatedlevelswhen the known marginal

distribution are used.
site S : 0.06 Derived ,5 actual values

estimatedI kno*o estimated I koo*n
Siml 0.005 0.003 0.0035/0.0117 0.0021/0.0043 0.0075
Sim2 0.298 0.347 0.97 1.06 1.151
Sim3 0.902 0.763 1..22 1.05 1.48
Sim4 3.95 o.J/ 5.9 11..7
blmD 3.88 4.7 6.5 8.3
Table 9.13: Estimated 10 year return levelsfor Qc obtainedusing estimatedand known

marginalsvia the current implementation(basiclevel) of the JPM. For Siml (.9 : 0.06)
theseresultscorrespondto the dominant direction,whereasfor derived,9 eachdirection
sectoris eiven.
9.8. OTHERFIGURES 209
Figure9.8: Siml Qc 10 year return valuefor the SVM appliedto logQ6 and the JPM
basedon the bivariate normal dependencemodel for (SWL,11s). The thresholdaxis is
related to applicationsof the SVM
- Shuclurevafiableon loo
- tJtYFt#0Jltfveilv&l+e:
o
9-
0.92 0.s4 0.96 0.98 1.00

Threshold
Figure 9.9: Siml Q5' 50 year return value for the SVM applied to logQ6, and the JPM
basedon the bivariatenormaldependence modelfor (SWL,Hs). The thresholdaxis is
related to applicationsof the SVM
0.92 0.94 0.96 0.98 1.00

Threshold
210 CHAPTER 9. AAIAIYffS OF OVERTOPPIATGDISCHARGE RATES
Figure9.10: Sim2 Q6 10 year return valuefor the SVM appliedto Q6 and logQg and
the JPM basedon the bivariatenormaldependence modelfor (SWL,,11s).The threshold
c2
E
E c'l
E
ll
0.94 0,96
Threshold
Figure 9.11: Sim2 Qc 50 year return value for the SVM applied to QL- and 1ogQ6 and
the JPM basedon the bivariatenormaldependence modelfor (SWL,Hs). The threshold
-?^
9:
Eq
0.94 0.96
Thteshold
9.8, OTHERFIGURES 271
Figure 9.12: Sim3 Q5, 10 year return value for the SVM applied to Q6 and log Q6r and
the JPM based on the bivariate normal dependencemodel fot (SW L, H s). The threshold
axis is related to applications of the SVM
.:
0.94 0.96
Threshold
Figure 9.13: Sim3 Qc 50 year return value for the SVM applied to Qg and 1ogQ6, and
the JPM basedon the bivariatenormal dependence model fot (SWL,Hs). The threshold
0.94 0.96
Threshold
2t2 CHAPTER 9. AATAIYSIS OF OVEHTOPPING DISCHARGE RATES
Figure 9.14: Sim4 Q6 10 year return value for the SVM applied to Qg and log Q5, and
the JPM basedon the bivariatenormal dependence model for (SW L, /1s) with estimated
and known marginalparameters.The thresholdaxis is relatedto applicationsof the SVM
Structure
Vafiableon Oc
Simulation
ModelValue
JoinlProbabilitv
frethod+BVN
$ilHf
FISbYBiif,
9rffin$.gfN+rxed
marsins
3r
0.94 0.96
Threshold
Figure9.15:Sim4Q6' 10yearreturnvaluefor the SVM appliedto Qs and logQ6'and the

JPM basedon the mixture of bivariate normalsdependencemodel for (SWL,Hs) with
of the SVM
0.94 0.s6
Threshold
9.8. OTHERFIGURES 213
Figure 9.16: Sim4 Qg 50 year return value for the SVM applied to Q6 and logQ6 and
the JPM based on the bivariate normal dependencemodel for (SI4IL,.I15) with estimated
and known marginal parameters. The threshold axis is related to applications of the SVM
- StrudureVadableon Oc
-:
--
- !3ifftrtt'8uU,rl&"r[,rv"i[uo?.ev'r
structurevrnaDl€ on loo uc
- Joinl PfobabiljtyMethocf+BvN+ljxed
maeins
5H
E
G
0.90 0.94 0.96

Threshold
Figure 9.17: Sim4Qc 50 yearreturn valuefor the SVM appliedto Qg and 1ogQ6 and the
JPM basedon the mixture of bivariate normalsdependence model for (SWl, t15) with
of the SVM
E
i
E6
P
o
0.94 0.96
Threshold
2r4 CHAPTER 9. AAIAlYSff OF OVERTOPPING DISCHARGE RA.IES
Figure9.18: SimSQc 10 year return valuefor the SVM appliedto Q5' and logQ6 and
the JPM basedon the bivariatenormal dependencemodel for (SW L, Hs) with estimated
and known marginalparameters.The thresholdaxisis relatedto applicationsofthe SVM
:
E
E
0.94. 0.96
Thfeshold
Figure9.19:Sim58c 10yearreturnvaluefor the SVM appliedto Qg and.IogQsand the

JPM basedon the mixture of bivariate normalsdependencemodel for (SWL,H) wifh
estimatedand known marginal paramete$. The thresholdaxis is related to applicatrons
of the SVM
of BVN
margins
of BvN+fixed
.?
0.s4 0.96
Threshold
9.8. OTHER FIGURES
Figure 9.20: Simb 8c 50 year return value for the SVM applied to Qs and logQs and
the JPM basedon the bivariatenormal dependence model for (SW L, Hs) with estimated
and known marginalparameters.The thresholdaxis is relatedto applicationsofthe SVM
- Vadable
Slructure on Oc
-- IJH'tt$"H8$"l,rY?t'""a.ev',l
slruclurevaiableon loooc
JointPrcbability
MethodlBvN+fixed
marglns
Ei6
E
0.94 0.96
Threshold
Figure 9.21: Sim5Qc 50 year return valuefor the SVM appliedto Q6rand log Q6rand the
JPM basedon the mixture of bivariate normalsdependencemodel for (SI71,,t15) with
of the SVM
$ffifldlHfi{#fl{tu?1
ijiHfHISb:8r'lf'y'fi
sllid-
Xixtureol BVN
Mlxtureof BvN+lixadmarqins
0.94 0.s6
Threshold
216 CHAPTER 9. ANAIYSIS OF OVET{TOPPINGDISCHARGER.ATES
Chapter 10
Comparisons and Conclusions
In this report we have developednew statistical modelsand proceduresfor joint proba-

bility methodsappliedto designproblemswherethe environmentalvariableswhich cause
failuresare extremeseaconditions.
The new statistical methodsdevelopedin Part I involve:
o Extreme value modelsfor the marginalvariablesol Hs and SWL.
o A rangeof dependence modelsfor (SWL,H;), whichenableflexiblemodellingof

different dependence
structuresbetweenthesevariables.
r A techniquefor explainingvariationsin Q through wavesteepness

and associated
f15 values.
o An objectivemodellingprocedureand diagnostictestsallowingaspectsof the mod-

elling assumptionsto be checked.
o A movetowardsa model-basedmethod for extrapolations,rather than the current

methodswhich rely in part on subjectiveextrapolationby eye.
r Statistical modelsthat enableprobabilitiesfor eventsof interest to be evaluated.

The main techniquefor this is thiough simulationof.a seriesof pseudoseacondition
data. Two optionswere developed:a continuoustime simulationof the conditions
at high water levelsand direct simulationof extremeevents.
Together,thesetechniquesprovidea general/flexibleframeworkfor addressingthe prob-
lem of evaluatingthe probability of failure of a structure.
The applicationsof the joint distribution fits in Part II showthat:
r The statistical modelshavesulficientflexibility to capture the featuresof the joint
distribution of seaconditionswhich produceextremeloadingconditionsfor a range
of observationaland simulateddata sets.
277
218 CHAPTER 10. COMPARISOATS
AI\rD COI\rC, USIOArS
The proposedmethodsappearto performbetter than currentmethodsin estimating

the joint extremesof (SW L.,I{s) and the behaviourof S for extremewaves.
Dependencemodelling of (SW L,,fls) appearsto be lessimportant than marginal

modelling. Methods for including extra marginal distribution were discussedand
illustrated.
The illustration of the techniquesto the problem of estimatingextremeovertopping

dischargerates in Part III of the report revealed:
In general,the newly developedjoint probability methodsare better than structure

variablemethods.This wasnot the casewith existingmethodsas the JPM, unlike
the SVM, ignoredvariation in 72.
Currently implementedversionsof the JPM, aimedat calculatingthe joint survivor

functions with given return periods,tend to under-estimateovertoppingdischarge
rates. The main reasonfol this under-estimationis the exclusionof variation of ,9
in the analysis.By selectingan appropriatefixed value for ,5 the impact of this is
reduced.
The simplilicationof the failure regiontor (SW L, Hs) causesrelatively little contri-
bution to any under-estimation(typically the return period of an overtoppinglevel
is over-estimatedby a factor of 1.5-2.0,which may be important in someapplica-
tions). Conversely,the refinedversionof the current implementation,which re-uses
resultsfrom the basicmethod to producea better estimateof overtopping,has po-
tential for someover-estimationdue to the methodsusedto evaluatethe probability
of failing in a rectangularregion.
The importanceof carefulmarginalmodelling. This featureprovedparticularly im-

portant in this study, as the waveheightsand periodsdomlnatedthe extrapolations
of overtoppingdischargerates. This was probably due to the useof'olfshore' wave
conditionsas if they were 'inshore'waves,so that they were often unrealistically
large.
The need for future comparisonsof the methods, basedon a range of carefully
implementeddesignproblems.
Chapter 11
References
Alcock, G.A. (1984). Parameteri,sing EntremeStill Water Leuelsand Wauesi,n Desi.gn

LeuelStud,i,es.Report no. 183,Institute of Oceanographic Sciences.
Carter, D.J.T. and Challenor,P.G. (1981). Esti,mating Retum Valuesof WaueEei,ght.
Reportno 116.Institute of Oceanographic Sciences.
Coles,S.G. and Tawn, J.A. (1994). Statisticalmethodsfor multivariateextremes:an
applicationto structuraldesign(with discussion). Appl. Stati,st.,43,1 48.
Davison,A.C. and Smith,R.L. (1990).Modelsfor exceedances overhigh thresholds(with
discussion). J. Roy. Statist.Soc.,8,52,393 442.
Dixon,M.J. and Tawn,J.A. (i994). EstimatesoJEfiremeSea-Leuels at the UK A-Class
Sr.tes:Stte by Si,teanalgse.s.
Internal Documentno. 65, ProudmanOceanographicLab-
oratory.
Dixon, M.J. and Tawn, J.A. (1995).Ertreme Sea Leuelsat the UK A ClassSr.tes:Op-
ti,maLSite-bg Site Analysesfor the East Coast. Internal Document no. 72. Proudman
OceanographicLaboratory.
Flather,R. A. (1987).Estimatesof extremeconditionsoftide and surgeusinga numerical
'
modelof the north-westEuropeancontinentalshelf.Est. Coast.Shelf.1ci.,24,69-93.
Hawkes,P. and Hague, R. (1994). Valid,ati,on oJ J oint Probability Methodsfor Large
Wauesand.High Water Leuels,Report SR 347, HR Wallingford.
Hydraulics ResearchStation (1980). Designof seawallsallowing for wave overtopping.
HRS Report924,HR Wallingford.
Leadbetter,M.R., Lindgren,G. and Rootzen,H. (1983).Ertremesand,RelatedProperti,es
oJRandomSequences arul,Seri,es.New York: Springer.
Ledford, A.W. and Tawn, J.A. (1996). Statisticsfor near independencein multivariate
extremevallues.Biometri.ka.83,169 187.
Pickands,J. (1975).Statisticalinference usingextremeorderstatistics.Ann. Statist.,3,
119-131.
2t9
CHAPTERll. REFERENCES
Prandle,D. and Wolf, J. (1978).The interactionof surgeand tide in the North Seaand
River Thames.Geophgs. J. R. Astr. Soc.,55,203-216.
Smith, R.L. (1989).Extremevalueanalysisof environmental time series:an application
to trend detectionin ground-level
ozone(with discussion),Statist.gci,.,4,367-393.
Tawn, J.A. (1992). Estimatingprobabilitiesof extremesea levels,AppL Statist., 41,
77 93.
Appendix A
Background information
A.1 Joint Distributions

Considerthe situation where there is a d-dimensionalvector of random variables,X :
(Xr,...,Xa)- For continuousrandomvariablesthe extensionofthe probabilitydensity
function for a singlevariable to the vector variablecaseis the joint probability density
the followingproperties:
function, Jx(rt,,... ra), whichsatisfies
.
" f x ( r r ,. . . r o ) : / x ( x ) l 0 f o r a l l x : ( r L ,. . . , n a ) .
r The rlfold (d-dimensional)integral
E l:ft(x)do1...d,ra:
I: I:/x(x)dx:1
r The probabilityof someevent,E say,is givenby
Pr{Xe il: [/x(x)dx.

JE
The probability densityfunction of the separatevariables,known as the marginal vari-

ables, can be derivedfrom the joint probability density function by summingthe joint
densityover the other variables.For example,for variableX,, this is given by the (d- 1)-
fold integral
rm rco
( r 1 ): | | f 7 $ ) d 4 . . . r l n i - 1 d , r i 1. .; .1d r ; , f o t i : 7 , . . . , d .
"fx, J -co J -co
That is, marginal distributions are obtained by summing over all the possibilities of the
other random variables. Similarly, by summing over a set of the random variables, the
joint probability density function of the subset of the remaining variables is obtained, i.e.
the joint density function of (&+r, . . . , Xa) is given by
" f x ( r ; * ' ,. . . , r a ) : [*
J -c<r
. .[*
J-m
1;*)d.q...dni.
221
222 APPET{DIX A. BACKGROU ID il\TFORMATIO T
An important conceptfor joint probability problemsis that of conditional proba-

bility. The conditional probability of an event occurringgiven another event is known
to have occurredarisesfrequentlywhen working with dependentrandom variables.The
conditionalprobability of event E2 given .81is known to haveoccurredis definedby
Pr{X e E2lX e E'} -

Pr{X e E1}'
where -E : Et 1 Ez is the intersectionof eventsE1 and E2. Similarly, the conditional

probabilitydensityfunctionfor (X1,. . . , X;) given(X,+t = r;+t,. ..Xa: ra), is givenby
"f"(*)
. -.,,o)
f 6r*r,...,xo1(r;+t,
where the denominator is the joint probability density function of (X,;+r, . . . , Xa)
AII the information about the variation of the vector of random variables is contained
in the joint probability density function. Summariesof this function are helpful in compar-
ing different distributions, so summary statistics a.reoften evaluated. The most common
summa.riesare the marginal mean and variance of each individual variable and, as mea-
sures of dependencebetween the variables, the couariance and,correlation between each
pair of variables.
Correlation and covariance are measuresof linear dependencebetween two va,riables.
Let J (r;, r) denote the joint probability density function for the pair (X', X ) Let p; and
oo2denote the marginal mean and variance of variable e, then Cov(X;, Xi), the covariance
of (&, Xj), is given by
Cov(X,,X7) : I:I: (rn - t"o)(ri - p)f (t;,n)dr;dri,
and the correlationbetweenthesetwo variablesis givenby
y,1 = Co"(x'' xi)

correlation(x,,
" oio j
The correlation, p, is a non-dimensional quantity always lying in the range -1 < p < 1.
When p:0 there is no linear dependencebetweenthe variables;when p > 0 there is a
positive linear association between the variables, with large values in each variable (and
small values in each variable) coinciding more often than if there were no relationship
between the variables; and when p < 0 there is a negative linear association between the
variables, with large values in one variable coinciding with small values in the other van-
able, and vice-versa, more often than if there were no relationship between the variables.
4.2. THE NORMALDISTRIBUTION
4.2 The Normal Distribution

In this section we give the distribution and properties of a univariate normal distribution
(i.e. one dimensional variable) and a multivariate normal distribution (i.e. a vector of
univariate normal random variables).
4.2.1 lJnivariate Normal Distribution

The probability density function of a normal randomvariableY is given by
| (/,, -,r)?'l
pt',r)- ------;- exp { _i_f,| for -all -co<y<oo
J \y/ (2n1rtro_.., , },, -".
2o2 J
where o > 0. The expectation of the variable is p and the variance is o2, hence the
standard notation V - N(p,,oz). The distribution function for this random variable
cannot be evaluated analytically, but symbolically is expressedas
* : I:_#^*'{-#}*,
Pr{Y<
: o[(s- r")1"].
When p:0 and o: l then the distributionis said to be in standardform, i.e. Y -
1r/(0,1). To transformY - N (p,,o2)to Z - N(0,1) we usethe relationship
Y-p.
Z:
o
4.2.2 Multivariate Normal Distribution

First we give resultsfor Z : (2t,. ..,2;), a randomvariablewhichfollowsa multivari-
ate normal distribution with standardnormal marginal distributions,i.e. eachmarginal
distributionis Zo * ^719,1).It followsthat Z hasjoint densityfunction
1 - - ''zJ)
) : --. - - * - . . - * e * p r
f z ( z ) : d ( z"t L 1
en)d/\der(D))U2'-i'p\-iz-
wherethe rangeof zi, for i : L,. ..,d, is -m ( z;.--cr.).Heredet(D) is the determinant
of the d x d matrix X, termed the variance-covariancematrix,
lott
":l I
\ oar
224 APPENDIXA. BACKGROU\rDINFORMATIOI\r
with (z,j)th element corresponding to the pairwise covariance of variables Zi, and, 2,1.
Thus, as
p,; : Correlatio Cov(z;'zi) - --
ru r' -
n(2,. Z,\
(var(Z;)Yar(2,))'t'
we have
(' " " nra)
'-l^ 0"
s'-|"'"'Pij " I
;J
\ Par " " L )
where p;3 is the correlation of Z; and 27. Therefore, the diagonal of D is a seriesof entries
of ones and the matrix is symmetric.
To illustrate the notation, we now give this joint density in the bivariate case:
1 ( 1 )
rl'2'1)
fzr,z"(21,22):((21,22):
G$: t;+/,exr{-4f-a1tz!-2p2""+A}'
where-oo < (zr, zz) < m. Herep: prz: p21,wilh -l < p ( 1, is the correlation
between21 and 22, so in terms of D,
'-/t o\
"- t )
\,
and
t ( t -P\
F
' -r--(r-t")\-P
1)
Property l: An importantspecialcaseis whenp: 0, i.e. no correlation.Then
z2): *o
lz,,z"(zy, * =
# {-t t": "3)\ f ,,(",)f,,("r),
so (fi, 22) are independent.
Property 2: The marginal densityof Zt - 11719,1).
This result followsfrom
I z,kt) : l*,- I rr)d

",.r,lr,, "r.
Property 3: The density ol Z1lZ2 : 22 is
1r7,.,'l
'"P * Pz2)'J
@F4W t-41 - l)\zt
so that z1lz2: zt * N(pzt, (1 - pt)) The resultfollowsfrom
f- - (t. z \
: 4i#(!.
f z,lz,=,"(zrltr)
l 221r'21
4.2. THE NORMAL DISTRIBUTIAN
Now considerthe more generalsituationwhereY : (yr, ...,Yi, a randomvariable

which follows a multivariate normal distribution with marginal distributions being [ -
N (t"0, for i : 1, .. .,d. It followsthat Y hasjoint densityfunction
"?)
:
/Y(y) 6V+rE r*o {-}r" - p)rL tv- r)},
w i t h g q: - *
l At< oo for i:1,...,d. H e r ep : ( h , . . . , p a ) r a n d X i s t h e v a r i a n c e -
covariance matrix with (rt,j)th element corresponding to the pairwise covariance of vari-
ables )t and Yi. Thus, as
Cov({, Y3)
nri: Con(Yu,Yt): : oijl@ioi),
(Var(4)Var(r1))1/'?
we have
("t
D: I
t
\ olodqd|
wherep;7is the correlationofY; and Y7.To illustrate the notation we againgivethis joint
densityin the bivariatecase:
( t -/i,)'?)],
J Y t y ' r: r P x l l - 2 1 1 - p e 1 ( 1 y , , -u , 1 ' _ r o ( a r - p t \ ( ! 2 - t r , \ + f ? ,
qr ot / \ ot /\ 02 / \ 02 / /)
with
1
c: - p'Y
21To1oil1
and -co < (Ayyz) < cc. Herep: pr2: p21is the correlationbetweenYi and Y2. Now,
Y r l ( Y " :a t )- N ( t L , r + ! p f u-, t i , o ? ( t - p ' ) ) .
We often denotethe distribution of Y by
Y - MVNa(p,D)
wherepr : (U,. .. , p;)z is the vectorof meanvaluesof Y, and D is the variance-covariance

matrix of Y. Note MVN6 denotesmultivariatenormaldistribution of d-dimensions.When
d,: 2 the notation is often simplifiedto
Y - BVN(p,x).
226 APPE TDIX A. BACKGROU\ID ilIFORMATIOAT
A.3 The Probability Integral Tlansform

In studying dependencebetweenvariables,it helpsto haveall the variablesfollorvingthe
samemarginaldistribution, as then marginaland dependence propertiesof the statistical
modelscan be easily separated.However,in any problem the random variablescannot
be expectedto follow the samedistribution, so sometransformationis required before
suchan assumptioncan be made. In this sectionwe givethe transformationnecessaryto
transform a random variable,Y, with distribution function, F(y), to a random variable,
7, with distribution function G(l). The approachwe useis the probability integraltrans-
form, which is a standardgeneralapproachto transformationof random variables.The
transformationwe use is
T : G-l@(Y)), (A.3.1)
where G-1 denotesthe inverseof the function G. The transformedrandom variable ?
hasdistributionfunctionG, since:
Pt{T <t} : P|{G-1(F(1,))
<4
: Pr{F(Y)<c(4}
: pr{y < r-,[G(r)]]
: F(F-tlG(t)]) as the distributionfunctionof Y is -F
: G(r).
4.4 Likelihoods for Dependence Models

Here we give the likelihood function which is requiredin fitting the dependencemodels
for significantwave height and still water level. The likelihood is the probability of the
data for a givenset of parametervalues,regardedas a function of the parameters.As the
data are independentpairs, the likelihoodis a product of the likelihoodcontributionsfor
the separatedata pairs, hencewe give only the likelihoodcontribution of a typical pair.
Throughout,we usethe notation (Xt, Xr) to denotethe pair of still water leveland sig-
nificant waveheightvariablesin the originalspace,and (Xi, X;) to denotethesevariables
after eachvariablehasbeentransformedto followa standardNorrnalrandomvariable,i.e.
usingequation(A.3.1) in SectionA.3. We denotethe distribution function of a unilariate
standardNormalrandomvariableby O(z). i.e.
Pt(xi < z) :Pr(xi < z) - a(z)

Similarly we let O(21,e2) denotethe joint distribution of bivariate Normal variables,i.e.
if (Xi, X;) follow a bivariate Normal distribution, with joint density given by equation
(A.2.1),then
Pr(Xi ! zu Xi < zz): Q(n, zz).
4.4. LIKELIHOODS FOR DEPEAIDEIfCE MODELS
4.4.L Bivariate Normal Model

The likelihood contribution for the observation(rj, ri) is
"i##{: d@i,o)
A2d\1."r ."+ \
where /(ri,ri) is the joint density function of the bivariate Normal distribution with
standardmarginaldistributions.
4.4.2 Bivariate Normal Threshold Model

Underthe thresholdmodeldescribedin Section4.2.7,we havea full modelfor the joint
distribution of observations havingboth variablesabovesuitablethresholds,but otherwise
have an incompletemodel. As a consequence) points which have one or both marginal
variablevaluesbelowthe respectivethresholdare censoredat the threshold.Herewe con-
sidertwo cases,the first wherethe observations are of standardNormal randomvariables,
i.e. observations of (Xi,Xi), with thresholds("I,"i), and the secondwhere(X1,X2) are
observed,thesebeing taken to haveGPD margiualsabovethresholds(u1,u2).
Normal Marginals
For an observation(rj, ri), the likelihoodcan take oneoffour formsdependingon whether

the two thresholdsare exceededor not:
r If zj < ui and r\ ( zj then the likelihoodcontribution is
o(zl,uj);
r If oj > ui and.r\ ( u| then the liketihood contribution is
}a@\,u).
0'i
o If ri < ui and ri > ui then the likelihoodcontribution rs
r If ri > u\ and r) > z| then the likelihoodcontributionrs
d@i,"\).
228 APPET{DIXA, BACKGROU,ryD
IATFO.RMATION
Original Marginals with GPD Tails
ConsiderXi ('i : 1,2) to be modelledby a GPD oversuitablethresholdsui U : 1,2),

so that the distribution function of X is taken to be
FyJ.x;):1 - ,\c{1+ {,(rn - uu)loo}n'/tn

, for r; > u,6,
where .\; : Pr(d; > u;), and o,; ) 0. Then for ri > ui, the marginal variablesare
transformedto normality through the probability integral transform (A.3.1), giving
x;: o-'(r",(&))fori:r,2,
and in particular,
ui: Q 1 ( F x , ( u e )f)o r i : 1 , 2 ,
representthe thresholdsmapped onto the normal scale. For an observation(21,12) the

Iikelihood can take one of four forms depending on whether the two thresholds are ex-
ceededor not:
r If 11 { u7 and.fr2 ( z2 then the likelihood contribution is
o(ul,uj);
r If 11 ) u1 and 12 ( z2 then the likelihood contribution is
)a(ri,u)) d,rI.
)ri drt'
. If zr ( ur and 12 ) u2 then the likelihoodcontribution is
la(ui, r\) d,ri.

lxi dr2'
r lf 11 > ur and r,z ) u2 then the likelihoodcontribution is
,b@;,';)##
for i : 1, 2,
In the aboveexpressions,
: + &(zr- u;)I o,jl'-r/e';

H #rr
4.4. LIKELIHOODS FOR DEPENDEI\rCEMODELS
4.4.3 Mixture of Bivariate Normals

Here the variablesare taken to be modelledby standardNormal marginal random vari-
ables, but with dependencestructure which arisesfrom the mixture of two bivariate
normal random variables. For the construction,we take T - (71,72) to be a mixture
of two bivariate ncrmal random variables.We first constructthe joint distribution func-
tion of T and then, via marginal transformationof T, obtain the joint distribution of
T- : ("i,?lj), wherethe marginal distributionsare standardnormal random variables.
The vector randomvariableT is definedby:
T:eZt+(7-e)22
Le.
Tj : eZti + (I - e)221fot j :1,2
where
1. e :0 with probability 7 - pv, and e : l with probability py;
2. 21 * BVN(prr, D1) where Ft: }try, lt tz) and.
o?,Ptoton
,,-(
"'
\ Ptottotz o1z )
3. Zz - BVN(prr, D2) where Ft: Qrn, tL,zz)and
- ( pzoztotz\
03,
"':
\ Prorrorz oEz )
It followsthat the joint densityfunction, ft, of T is
fr(t) : o, tr'rt) + (1- Pv)f7"ft)

where /2, and fz2 arc the joint densities of 21 and 22 respectively.
We now derive the marginal distribution ol T1 (j : 1, 2). As
p,{4 s t} : Fr,Q): era (t--Jti) * tt - trto (t---'.z.) ,

' oti ozt
\ / \ )
it follows that
r; : a-r (Frj(]:j)) for j : 1,2
are standardnormal randomvariables.

APPENDIXA. BACI<GROUI{DIAIFORMA?/OI\T
For likelihood calculationswe now give the joint densityof T- - (Ti,T)):
/'. ft') : {lz,(t1(ti),t2fti))p - oll , ,, 9jlil9(t?]

t,(t;))(1
+ fz,(tt(tj), '112 ,rr,,
7'1 | J
J \t/ \t/ | ) ) \12 \r'2 ) )
wheret3(ij): IE'(o({)).
In practicethis statisticalmodelis over-parametrised
for statisticalpurposes.
To avoid
parameter redundancv we take
! 4 = 0 , o n = 1 a n do n : 7 .
List of Figures
3.1 Joint densitycontoursfor the bivariatelogisticdependence

structure when
the marginal distributionsare standardnormal, i.e. Zi = O-1(exp(-1/X;))
for X; a unit Fr6chetvariableand i: parameter
1,2. The dependence
o : 1 ,0 . 9 , 0 . 8 , 00. 7. 5, , 0 . 3 .
3.2 Joint density contours for the bivariate normal dependencestructure when
the marginal distributionsare standardnormal, i.e. Zn : O-i (exp(- 1/X,))
for X1 a unit Fr6chet variable and i: 1,2. The dependenceparameter
p : - 0 . 5 ,- 0 . 2 , 0 . 00,. 2 , 0 . 5 , 0 . 8 .
4.1. Estimation of the distribution function of Ts at Christchurch: estimates

from the empirical distribution of Tz, constant ,9 and independenceof
(Hs,S). 55
5 . 1 Cardiff observationaldata and Siml data: .F15vs SWL 78

. K ? Cardiff observationaldata and Siml data: Hs vs T2 78
5 . 3 Cardiff observationaldata and Siml data: SWL vs 7z 70
5 . 4 Cardiff observationaldata and Siml data: Hs vs 0 79

5 . 5 Cardifi observationaldata and Siml data: t11 --^
tz \n v
O
80
5 . 6 Cardiff observationaldata: I{s vs Surge 80
Dover observationaldata and Sim2 data: 115vs SWL B1
5 . 8 Dover observationaldata and Sim2 data: Hs vs Tg Bi
5 . 9 Dover observationaldata a,ndSim3 data: SWL vs ?2 82
5 . 1 0Dover observational data: I1s vs Surge 82
5 . 1 1North Walesobservationaldata and Sim3 data: I{s vs SWL 83
5 . 1 2North Walesobservationaldata and Sim3 data: Hs vs T2 83
5 . 1 3North Walesobservationaldata and Sim3 data: SWL vs Q 84
5 . 1 4North Walesobservationaldata and Sim3 data: Fls vs Surge 84
Christchurchobservationaldata and Sim4 data: 11.<vs SWL B5
5 . 1 6Christchurch observational data and Sim4 data: Hs ls T2 85
5 . r 7Christchurch observational data and Sim4 data: SWL vs ?7 B6
231
232 LISI: OF FIGURES
5 . 1 8Christchurchobservationaldata: Ifs vs Surge 86

5 . 1 9Dowsingobservationaldata and Sim5 data: /1s vs SWL 87
5 . 2 0Dowsingobservationaldata and Simb data: Hs 'ts Ts . 87
D . ZI Dowsingobservationaldata and Sim5 data: SWL vs ?2 88
5 . 2 2Dowsingobservationaldata: Ils vs Surge BB
6 . 1 GPD shapeparameterestimatesversusthe empiricaldistribution function

calculatedin z for SW L data at Cardi{f and Christchurch. Dots denote
maximum likelihood estimates. Lines denotelower and upper bounds of
91
6.2 GPD shapeparameterestimatesversusthe empiricaldistribution function
calculatedin u for SWL data at Dowsinp and Shoreham. Dots denote
maximum likelihood est,imates.Lines denote lower and upper bounds of
92
6.3 GPD shapepa.rameterestimatesversusthe empiricaldistribution function
calculatedin u for SWL data at North Wales and Dover. Dots denote
maximum likelihood estimates. Lines denotelower and uooer bounds of
the 95% confidenceintervals.. 93
o . 4 Probability and quantile plots for the GPD fitted to SWL daha.
ai Cardiff
and North Wales.Also shownin eachplot is the r = E line. o,4.
6 . 5 GPD shapeparameterestimatesversusthe empiricaldistribution function
calculatedin u for Surgelevel data at Cardiff and Christchurch.Dots de-
note maximumlikelihoodestimates.Linesdenotelower and upper bounds
of the 95% confidenceintervals. 97
o.o GPD shapeparameterestimatesversusthe empiricaldistribution function
calculatedin z for Surgeleveldata at Dowsingand Shoreham.Dots denote
maximum likelihood estimates. Lines denote lower and upper bounds of
98
6.7 GPD shapeparameterestimatesversusthe empiricaldistribution function
calculatedin a for Surgeleveldata at North Walesand Dover. Dots denote
maximum likelihood estimates. Lines denote lower and upper boundsof
the 95% confidenceintervals. qq
6 . 8 Probability and quantile plots for the GPD fitted to Surgelevel data at
Cardiffand Christchurch.Also shownin eachplot is the z: g line. 100
6 . 9 Probability and quantile plots for the Weibull distribution fitted to I15
observationsat Cardiff and Christchurch. Also shownin eachplot is the
r:u line. . .
LIST OF FIGURES
6.10 Probability and quantile plots for the Weibull distribution fitted to flg
observationsat Dowsing and Shoreham. Also shown in each plot is the
s : y line.
6.11 Probability and quantile plots for the Weibull distribution fitted to Ils
observationsat North Wales and Dover. AIso shownin each plot is the
r - y line. 107
6.12 Probability and quantileplots for the TruncatedWeibull distribution fitted
to Fls exceedancesat Cardiff and Christchurch.Also shownin eachplot is
the z :3r line.. 108
to -FIsexceedancesat Dowsingand Shoreham.Also shownin eachplot is
the z: g line.. 109
to I{s exceedancesat North Whlesand Dover. Also shownin eachplot is
the r : 9 line. . 110
6.15 Probability and quantile plots for the GPD fitted to /1s exceedances
at
Cardiff and Christchurch.Also shownin eachplot is the r :9 line. 111
6.16 Probability and quantile plots for the GPD fitted to Ils exceedances
at
Dowsingand Shoreham.Also shownin eachplot is the r: grline. . t72
6.17 Probability and quantile plots for the GPD fitted to lls exceedances at
. l s os h o w ni n e a c hp l o ti s t h ec : y l i n e . . . . . . 1 1 3
N o r t hW a l e sa n dD o v e r A
6.18 Cardifi: Estimatesof two extreme cuantiles under the Weibull and the
GPD modelsversusthe empiricaldistribution function calculatedin u. 1r4
6.19 Christchurch: Estimatesof two extremequantilesunder the Weibull and
the GPD modelsversusthe empiricaldistribution function calculatedin t. 1 1 5
6.20 Dowsing: Estimatesof two extremequantilesunder the Weibull and the
GPD modelsversusthe empiricaldistribution function calculatedin u. 116
6.21 Shoreham:Estimatesof two extremequantilesunder the Weibull and the
GPD modelsversusthe empiricaldistribution function calculatedin z. tL7
6.22 North Wales: Estimatesof two extremequantilesunder the Weibull and
the GPD modelsversusthe empiricaldistribution function calculatedin u. 1 1 8
6.23 Dover: Estimatesof two extremequantilesunderthe Weibull and the GPD
modelsversusthe empiricaldistribution function calculatediu z. 119
6.24 Histogramof Steepness
at varioussites.
6.25 Histogramof measuredsteepness
at Morecambe,Boygrift and L1.me. 1t1
6.26 Histoeramof WaveDirection at varioussites. r25

234 LIST OF FIGURES
7.7 Correlationfunction,p.,,between(Hs, SW L) versusthe threshold(in prob-

ability of non-exceedance)for the differentsites . . . 130
7.2 Correlationfunction, p,, between(Hs,Surge) versusthe threshold(in
probability of non-exceedance)
for the differentsites . . . . . . . 130
JA
Cardiff:Correlationfunction,p,, between(Hs,SWL) versusthe threshold
separatelyfor g < 190" and d > 190" . . 131
'7^
Cardiff: Correlationfunction,p,,,between(Hs, Surge)versusthe threshold
(in probability of non-exceedance)separatelyfor d < 190" and d > 190' . . 131
Christchurch:Estimatedcorrelationfunction, p,,,between(Hs,SWL) and
(H s, Surge) versusthe threshold(in probability of non-exceedance)
for the
fitted mixture of bivariate normalsand the thresholdbivariate normal . 133
7.6 Dowsing: Estimatedcorrelationfunction, p,,, between(H1,SWL) and
(Hs,Surge) versusthe threshold (in probability of non-exceedance) for
the fitted mixture of biva;riatenormalsand the thresholdbivariate normal . 133
7.7 Dover: Correlationfunction,pu, between(Hs,SWL) and (-F15, Szrge) ver-
sus the threshold(in probability of non-exceedance)
of bivariate normalsand the thresholdbivariate normal . . . . . 134
7.8 North Wales:Estimatedcorrelationfunction, p,, between(Hs, SW L) and,
(Hs,Surge) versusthe threshold(in probability of non-exceedance)
for the
fitted mixture of bivariatenormalsand the thresholdbivariatenormal . 134
70 Cardiff: Correlation function, p,,, between(Hs, SW L) and (Hs, Surge)
versusthe threshold(in probability of non-exceedance)for the fitted mix-
ture of bivariate normalsand the.thresholdbivariatenormal . . . . . . . 135
7.10 Shoreham:Estimatedcorrelationfunction,p,, between(Hs,SWL) and.
(Hs,Surge) versusthe threshold(in probabilityof non-exceedance) for
the fitted mixture of bivariate normalsand the thresholdbivariate normal . 135
7 . r l Siml: Diagnostictest for th e form of the dependencein the extremes of //s-
SWL. The solid line is the test statistic and the dotted line its pointwise
95% confidence interval.
7.12 Sim2:Diagnostictestfor the form of the dependencein the extremes of IIs-
SH/.L. The solid line is the test statistic and the dotted line its oointwise
7.13 Sim3: Diagnostictest for the form of the dependence in the extremesof IIs-
SW,L. The solid line is the test statistic and the dotted line its pointwise
LIST OF F/GURES
7.14 Sim4: Diagnostictest for the form ofthe dependence in the extremesof 1{s-
SWL. The solid line is the test statistic and the dotted Iine its Dointwise
95% confidenceinterval.
7.15 Sim5:Diagnostictestfor the form ofthe dependence in the extremesof fls-
SW L. The solid line is the test statistic and the dotted line its pointwise
9 5 %c o n f i d e n icnet e r v a l . ...141
7.16 Correlationfunction,p,, between(Hs, SW L) versusthe threshold(in prob-
ability of non-exceedance) for the differentsimulateddata sets. r42
7.17 SimS: Correlationfunction, pu, between(Hs,SWL) and (Hs,Surge) ver
sus the threshold(in probability of non-exceedance) r42
7.18 Sim4: Correlationfunction,p,,,between(Hs,SWL) versusthe threshold
(in probability of non-exceedance) for the fitted mixture of bivariate nor-
mals and the thresholdbiva.riatenormal . . 143
7.19 SimS: Correlationfunction,p,,, between(Hs, SW L) versusthe threshold
for the fitted mixture of biva,riatenor-
mals and the thresholdbivariatenormal . 1,43
7.20 Siml Hs SWL joint densitycontours:the contoursof joint densityvalue
0.01,0.001,0.0001,0.00001
areshown.The estimatedmodelis the bivariate
n o r m am
l odel. ....I44
7.2I S\m2 Hy-SWL joint densitycontours:the contoursof joint density value
0.01,0.001,0.0001,
0.00001areshown.The estimatedmodelis the bivariate
n o r m am
l odel. ....144
7.22 Sim3 Ils Surgejoint densitycontours:the contoursof joint density value
0.01,0.001,0.0001,
n o r m am
l odel. ... . ...145
7.23 Sim3 H7-SWL joint density contours:the contoursof joint density value
0.01,0.001,0.0001,
n o r m am
l odel. ....145
7.24 Sim4 Hs SWL joint density contours:the contoursof joint densityvalue
0.01,0.001.0.0001.
0.00001are shown.The estimatedmodelis the thresh-
old bivariate normal model.
7.25 Sim4 Hs SW L joint densitycontours:the contoursof joint density value
0.01,0.001,
0.0001,
0.00001areshown.The estimatedmodel is the mixture
of bivariate normals model.
7.26 Sim5H;-SWL joint densitycontours:the contours of joint density value
0.01,0.001,
0.0001,
0.00001
areshown.The estimated model is the thresh-
old bivariate normal model.
236 LIST OF FIGURES
7.27 Sim5 H;-SWL joint densitycontours:the contoursof joint density value

0.01,0.001,0.0001,
0.00001areshown.The estirnated
modelis the mixture
of bivariate normals model. t47
7.28 Siml: S versusFIs and regressioncurve (plotted on the original scale). . . 150
7.29 Sim2: S versusI1s and regressioncurve (plotted on the original scale). . 150
7.30 Sim3: S versusI{s and regressioncurve (plotted on the original scale). . 151
7.31 Sim4: S versusffs and regressioncurve (plotted on the original scale). . 151
7.32 SimS:,5 yersusHs and regressioncune (plotted on the original scale). . 152
7 . 3 3S i m l : S v e r s uIsf s ( p l o t t e do n t h e G a u s s i asnc a l e ) . ......I52
7.34 Sim2:S versusI15 (plottedon the Gaussian scale). . . .. . . .153
7.35 Sim3: S versusIls (plottedon the Gaussian scale). . . . 153
7.36 Sim4: S versus/Is (plottedon the Gaussian scale). . . . 154
7.37 Sim5: ,9 versusIls (plottedon the Gaussian scale). . . . 154
7.38 Siml: Mean and Varianceof ,9 (on the Gaussianscale)versusthreshold
for I15 (on the Gaussianscale).Dotted lines give 95% confidenceinterval. . 155
7.39 Sim2: Mean and Varianceof S (plotted on the Gaussianscale) versus
thresholdfor fJs (on the Gaussianscale).Dotted linesgive95%confidence
interval. .......155
thresholdfor Ils (on the Gaussianscale).Dotted linesgive95%confidence
thresholdfor Ils (on the Gaussianscale).Dotted linesgive 95%confidence
7.42 Sim5: Mean and Varianceof ,9 (plotted on the Gaussianscale) versus
thresholdfor I1s (on the Gaussianscale).Dotted linesgive95%confidence
interval. ......157
7.43 Siml: S versusfls for the original data and the data simulatedfrom the
f i t t e dr e g r e s s i o n ...I57
7.44 Sim2: S versusIIs for the original data and the data simulatedfrom the
f i t t e dr e g r e s s i o n .......158
7.45 Sim3: ,9 versus.FI5for the original data and the data simulatedfrom the
f i t t e dr e g r e s s i o n ...158
7.46 Sim4: 5 versusIls for the original data and the data simulatedfrom the
fitted regression . .159
7.47 Sim5: ,9 versusf15 for the original data and the data simulatedfrom the
fitted resression . 159
LIST OF FIGURES
8 . 1 Siml: 11s-.9W1 curvesof equal joint survivor probability expressedin

return periods(in years)when the marginaldistributions are known 166
8 . 2 Sim2: Ils-SWl curvesof equal joint survivor probability expressedin
return periods(in years)when the marginaldistributions are known 166
8 . 3 Sim3: fls Szrge curvesof equaljoint survivor probability expressedin
return periods (in years)when the marginaldistributions are known 167
8 . 4 Sim3: fIs-StrU-L curvesof equal joint survivor probability expressedin
return periods (in years)when the marginal distributionsare known 1.67
B . 5 Sim4: H7-SWL curvesof equal joint survivor probability expressedin
return periods(in years)when the marginaldistributionsare known. Here
the dependence model is the fitted bivariate normal thresholdmodel . 168
8 . 6 Sim4: H 3-SW -L curvesof equal joint survivor probability expressedin
the dependence model is the fitted mixture of bivariatenormalsmodel 168
8 . 7 SimS: Ils-,9WL curvesof equal joint survivor probability expressedin
return periods(in years)whenthe marginaldistributionsare known. Here
the dependence model is the fitted bivariate normal thresholdmodel . 169
8 . 8 SimS: II5-Str4u-Lcurvesof equal joint survivor probability expressedin
the dependencemodel is the fitted mixture of bivariatenormalsmodel 169
8 . 9 Sim2: H1-SWL curvesof equaljoint survivor probability expressedin re-
turn periods(in years)whenthe marginaldistributionsare known. Curves
shownare 10 and 1000yearreturn period for Estimatedmodel,Simulation
t7r
8.10 SimS:IIs .9WL cwvesof equaljoint sun'ivor probability expressedin re-
shownare 10 and 1000year return period for Estimatedmodel,Simulation
model and HR eslimates t71"
8.11 Sim4: H;-SWL curvesof equaljoint survivor probability expressedin re-
shown are 10 and 1000year return period for Estimatedmodel (Mixture
Model),Simulationmodeland HR estimates 172
8.12 SimS: H7-SW L curvesof equaljoint survivor probability expressedin re-
shown are 10 and 1000year return period for Estimated model (Mixture
M o d e l )S
, i m u l a t i omn o d e al n dH R .e s t i m a t e s .......172
238 LIST OF FIGURES
8.i3 Siml: H7-SWL curvesof equal joint survivor probability expressedin

return periods (in yea.rs) . . . 178
8.14 Sim2: Hs SWL curvesof equal joint survivor probability expressedin
return periods(in years) 178
8.15 Sim3: Hs-Surge curvesof equal joint survivor probability expressedin
return periods (in years) . . 179
8.16 Sim3: H1-SWL curvesof equal joint survivor probability expressedin
return periods (in years) . . . 779
8.17 Sim4: HS-SWL curvesof equal joint survivor probability expressedin
return periods (in years)for the fitted bivariate normal thresholdmodel . . 180
8.18 Sim4: H7-SWL curvesof equal joint survivor probability expressedin
return periods(in years)for the fitted mixture of bivariatenormalsmodel . 180
8.19 SimS: Hy-SWL curvesof equal joint survivor probability expressedin
return periods(in yeats)for the fitted bivariatenormal thr-esholdmodel . . 181
8.20 Sim5: H;-SWL curvesof equal joint survivor probability expressedin
return periods(in years)for the fitted mixture of bivariatenormalsmodel . 181
9.1 Siml Qc 100 year return value for the SVM applied to log Qg and the
JPM based on the bivariate normal dependencemodel for (SW L, Hs).
The thresholdaxis is relatedto applicationsof the SVM . . . . . 192
9.2 Sim2 Q6' 100year return valuefor the SVM appliedto Qs and.IogQ6' and
the JPM basedon the bivariatenormal dependence model for (SW L, H s).
9.3 Sim38c 100yearreturnvaluefor the SVM appliedto Q6 and logQg and
the JPM basedon the bivariatenormaldependence modelfor (SWL,Hs).
The thresholdaxis is relatedto applicationsof the SVM . . . 198
9.4 Sim4 Q6' 100year return valuefor the SVM appliedto Qs and logQo and
the JPM basedon the bivariatenormaldependence modelfor (SWL,Hs)
with estimated and known marginal parameters. The threshold axis is
related t o a p p l i c a t i oo
n fst h e S V M .......202
9.5 Sim4 Qc 100year return valuefor the SVM appliedto Qc and log Q6 and
the JPM basedon the mixture of bivariate normalsdependencemodel for
(SW L, Hi with estimatedand knownmarginalparameters.The threshold
axis is relatedto applicationsof the SVM . . . . 202
9.6 Sim5 Qc 100year return valuefor the SVM appliedto Q6 and logQ6rand
the JPM basedon the bivariatenormal dependence model for (SW L., Hs)
with estimated and known marginal paraneters. The threshold axis is
n fst h e S V M ...... .205
LIST OF FIGURES 239
9.7 Simb Q6 100year return valuefor the SVM appliedto Qc andlog Qa and
(.SWL, Hs) with estimatedand knownmarginalparameters.The threshold
axis is related to applicationsof the SVM . . . . 206
9.8 Siml Qc 10 year return valuefor the SVM appliedto log Q6,and the JPM
based on the bivariate normal dependencemodel for (Str;tr/],Ifs). The
thresholdaxis is related to applicationsof the SVM . . . 209
9.9 Siml Q,?50 yearreturnvaluefor the SVM appliedto logQ6 and the JPM
based on the bivariate normal dependencemodel for (SW-L,I1s). The
thresholdaxis is relatedto applicationsof the SVM . . . 209
9.10 Sim2 Q6 l0 year return valuefor the SVM appliedto Qs ald logQo and
the JPM basedon the bivariatenormaldependence modellor (SWL,Hs).
The thresholdaxis is relatedto applicationsof the SVM . . . 210
9.11 Sim2 Qcr 50 year return valuefor the SVM appliedto Qs and log Q6: and
the JPM basedon the bivariatenormaldependence modelfor (SWL,Hs).
9.12 Sim3 Q6r10 year return valuefor the SVM appliedto Q6' and log Qc, and
the JPM basedon the bivariatenormal dependence model fot (SW L, H s).
9.13 Sim3 Qc 50 year return valuefor the SVM appliedto Q6 and 1ogQs and
the JPM basedon the bivariatenorrnaldependence model for (SWL, Hs).
The thresholdaxisis relatedto applications
of the SVM . . . . . 2I-
9.14 Sim4 Q6, 10 year return valuefor the SVM appliedto Q6 and log Q6 and
the JPM basedon the bivariatenormaldependence modelfor (SWL,Hs)
relatedt o a p p l i c a t i oo
n fst h e S V M .......212
9.15 Sim4 Q6 l0 yearreturn valuefor the SVM appliedto Qa1and logQ6 and
(SWL, Hs) with estimatedandknownmarginalparameters. The threshold
axis is related to applicationsof the SVM . . . . 2I2
9.16 Sim4 Q6 50 year return valuefor the SVM appliedto Q5rand log Q6: and
the JPM basedon the bivariatenormaldependence modelfot (SWL,Hs)
relatedto applicatioon fst h e S V M ...... .213
240 LIS'| AF FIGURES
9.17 Sim4 Q6: 50 year return valuefor the SVM appliedto Qs and logQ6 and
(SW L, Hs) with estimatedand knownmarginalparameters.The threshold
axis is relatedto applicationsof the SVM . . . . 2I3
9.18 Simb Qg 1-0year return valuefor the SVM appliedto Qs and log Q5' and
n fst h e S V M .......274
9.19 Sim5Qc 10 yearreturn valuefor the SVM appliedto Q6 and.logQg and
the JPM basedon the mixture of bivariate normalsdependence model for
(SW L, Hs) with estimatedand knownmarginalparameters.The threshold
axis is related to applicationsof the SVM . . . . 2I4
9.20 SimSQc,50 yearreturn valuefor the SVM appliedto Q6'and logQ6'and
related to applicatioo n fst h e S V M .......2I5
9.21 SimSQ6: 50 year return valuefor the SVM appliedto Q5rand log Q6rand
the JPM basedon the mixture of bivariate normalsdependence model for
(SWL, Hs) with estimatedandknownmarginalparameters. The threshold
axis is relatedto applicationsof the SVM . . . . 21,5
List of Tables
4.1 Resultsfrom simulating 10000samplesof size6000from a Weibull distri-

bution with knownparameters o,: 0, b : 1 and c: !.2.
4 . 2 Resultsfrom simulating 10000samplesof size6000from a Weibuli distri-
a : 0, b - 1 and c:7.5.
bution with knownparameters
Resultsfrom simulating 10000samplesof size6000from a Weibull distri-
butionwith knownparameters a: 0, b: 1 and c: 1.7.
4.4 Resultsfrom simulating 10000samplesof size6000from a Weibull distri-
a : 0, b : 1 and c:0.7.
bution with knownparameters 50
5.1 Information about the observationaldata sites: SWL - still water level.
Waves - significant wave height, wave period, wave direction. 70
5.2 Differencesin the statisticalmodelsfor the simulateddata setsSiml-Sim5.
6.1 Maximum likelihoodestimatesof the parametersof the GPD fitted.to SW L

data at observationsites (standarderrorsin pa,rentheses). 90
6.2 Maximum likelihoodestimatesof the parametersof the GPD fitted lo SW L
simulatedfor the synthetic data sets (standarderrorsin parentheses). 95
6.3 Maximum likelihoodestimatesof the parametersof the GPD fitted to Surge
level data at various sites (standard errors in parentheses). 96
6.4 Maximum likelihoodestimatesof the oarametersof the GPD fitted to sim-
ulated Surge level data (standard errors in parentheses). 101
6.5 Maximumlikelihoodestimatesof the parametersof the Weibull distribution
fitted to -FIsobservationsat varioussites (standarderrorsin parentheses).
. 103
6.6 Maximum likelihoodestimatesof the parametersof the TruncatedWeibull
at varioussites (standard errors in
distribution fitted to IIs exceedances
parentheses).
6.7 Maximum likelihood estimatesof the parametersof the GPD fitted to I1s
exceedancesat varioussites (standarderrorsin parentheses). . . 104
24r
.,,4n
LIST OF TABLES
6.B Maximum likelihoodestimatesof the parametersof the GPD fitted to simu-

lated I{s thresholdexceedances
for the syntheticdata sets(standarderrors
in parentheses). ...120
7.r Parameterestimatesfor the mixture of bivariatenormalsappliedto obser-

vational (Hs, SWL) data. IJI
7.2 Parameterestimatesfor the mixture of bivariatenormalsappliedto obser-

vational(Hs, Surge)data. 132
7.3 Parameterestimatesfor the mixture of bivariate normals applied to the
simulated(Hs, SW L) data and for Sim3.to (Hs,,Surge).. I4I
7.4 Parameterestimatesfor the linear regressionof S* on flj applied abovea
thresholdu (on the Gaussianscale). Standarderrors a"regiven in paren-
theses.For Sim4 and Sim5 the value of u usedin the reqressionmodel is
u:0.. ...........149
8 . 1 HR estimatesof points on the 10 and 1000year contoursof the joint sur-

vivor function of (SW L, Hs) for Siml: separateestimatesconditional on
the wavedirection, 0' < 0 < 110"and 110' < d < 360'. ...173
d, satisfying
vivor function ol (SW L, Hs) for Sim2,when the marginaldistributionsare
known.
qe
HR estimatesof points on the 10 and 1000year contoursof the joint sur-
vivor function of (SW L, Hs) for Sim3,when the marginaldistributionsare
known. .......774
vivor function of (SW L, H s) for Sim4,when the marginaldistributionsare
known.
5.5 HR estimatesof points on the 10 and 1000year contoursof the joint sur-
vivor function of (SW L,lJ5) for Sim5,whenthe marginaldistributionsare
known. . .....175
8 . 6 HR estimatesof points on the 10 yearcontourof the joint survivorfunction
for Siml: separateestimatesconditionalon the wavedirection,0, satisfying
0 " < P < 1 9 0 ' a n dL 9 0 ' < 0 < 2 7 0 " .
8.7 HR estimatesof points on the 10 yearcontourof the joint survivorfunction
for Sim2.
8.8 HR estimates of points on the 10 year contour of the joint survivor function
for Sim3. 183
LIST OF TABLES 243
8.9 HR estimatesof points on the 10 yearcontourof the joint survivorfunction

for Sim4. 184
8.10 HR estimatesof points on the 10 yearcontourof the ioint survivorfunction
for Sim5. 184
9.1 Information on the structural designfor the simulateddata. Here the sea-
wall is a simple sloping wall with a 1:4 slope, a4 a;nda2 dimensionless
constantsand u is the crestelevationin metres(relativeto ODN exceptfor
Sim3and Sim5which are relativeto Chart Datum). Seeequation(1.1.1)
for details of how thesedesignparametersinfluenceovertoppingdischarge
rates.. ......I88
9.2 The largest five sea condition valuesfor Siml in terms of the structure
function,overtoppingdischargerale Q6 grvenby equation(1.1.1). Here
SWL is in terms of metresrelative to ODN, I1s is in metres, and Ts in
s e c sO
. v e r t o p p i ndgi s c h a r gnet e , Q 6 , i s i n m e t r e 3 / s e c / m e t r.e. ... . . . 1 9 1
9.3 Siml: Returnlevelsfor Q6 (in metre3
/sec/metre)from simulationmodel.. 192
9.4 The largest five sea condition valuesfor Sim2 in terms of the structure
function,overtoppingdischargerate Q6 givenby equation(1.1.1). Here
SW L \s in terms of metresrelative to ODN, lls is in metres,and Tz in
s e c sO
. v e r t o p p i ndgi s c h a r gr ea l , eQ, s , i s i n m e t r e 3 / s e c / m e t r.e. ... . . . 1 9 4
9.5 Sim2:Returnlevelsfor Qc (in metre3/sec/metre)
from simulationmodel.. 195
function,overtoppingdischargerate Q5:givenby equation(1.1.1). Here
SWL is in terms of metresrelative to ODN, Ils is in metres, and.Ts in
secs.Overtoppingdischargente, Q6, is in metre3/sec/metre. . . . . . . . 197
9.7 Sim3: Return levelsfor Q6 (in metre3/sec/metre)from simulationmodel. . 197
function,overtoppingdisch-arge rate Q6 givenby equation(1.1.1). Here
SW L is in terms of metresrelative to ODN, J{s is in metres, and T2 in
s e c sO
. v e r t o p p i ndgi s c h a r gr a
e t e ,Q c , i s i n m e t r e 3 / s e c / m e t r.e. ... . . . 2 0 I
9.9 Sim4: Return levelsfor Qq (in metre3/sec/metre)from simulationmodel. . 201
9.10 The largest five sea condition valuesfor SimS in terms of the structure
function,overtoppingdischargerate Q6: givenby equation(1.1.1). Here
SWL \s in terms of metres relative to ODN, Hs is in metres, and Tz in
s e c sO
. v e r t o p p i ndgi s c h a r gnel e , Q 6 , i s i n m e t r e 3 / s e c / m e t r.e. ... . . . 2 0 4
from simulationmodel.. 205
9.11 SimS:Returnlevelsfor Q6 (in metres/sec/metre)
244 LISTOFTABLES
9.12 Actual return period of designlevelsestimatedby the current implementa-

tion (basiclevel) of the JPM: for S : 0.06and when S is derivedfrom the
data. For Siml theseresultscorrespondto the dominantdirection. Owing
to the dominanceof waveperiod for Sim4and Sim5theseestimatesshould
be largely ignored.
9.13 Estimated10year return levelsfor 8c obtainedusingestimatedand known
marginalsvia the current implementation(basic level) of the JPM. For
Siml (S : 0.06)theseresultscorrespond to the dominantdirection,whereas
for derivedS eachdirection sectoris Eiven. 208

SR537 JOINSEA Probablility Waves Water HRWallingford

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

SR537 JOINSEA Probablility Waves Water HRWallingford

Uploaded by

Copyright:

Available Formats

The Joint Probabilityof

F€grsr.r.d h Enghnd 6. 2s320€0,HRw.llifrglord i3 ! wholy dn.d .ubltdt.ry ot HRw.tlnqr6d c6up Lrd.

The HR WallingfordJobNumberwasCCS14W.Thework at HR Wallingford

A rigorous but practical new approach

A common application ofjoint probability in coastalengineenngis to the

- removesmost of the subjectivity from the presentmethods;

- allows the distribution of wave period to be included in calculations,

- allows long-term simulation of the combinedwave height, wave

This researchwas funded by the Ministry of Agriculture, Fisheriesand Food, and

Availabilitvof Drosramsand reDorts

Part I HR Wallingford Report

!"**,"o.* vl1 sR 53?04r'05/00

Part II Lancaster University Report

- Iargewave heightsand high water levels;

- large river flows and high sealevels;

- Iarge surgesand high astronomicaltidal levels.

Importanceof ioint nrobabilitv

Predictionof ioint Drobabilitv

Record interval and event duration

Joint structuralprobability and return pqtied

Joint exceedanceorobability and retum period

The discreoancybetweenjoint exceedance andjoint stxctural probabilities

Scatterdiagramsto illustratethe degre€of dependence

SURGERESIDUAL LEVEL (M)

No DATA/CALMS O 1 O 1212 211 11 I0 0 C o n t o u or sf

RECORDSANALYSED FROM DAY 1 OF MONTH 2 OF L9'17 PO DAY 31 OF MONTH 3 0F 1990

-0.?0 -0,40 -0.10 0.20 0.50 0-80 L-10 1-40

RECORDS ANALYSED FROM DAY 1 OF MONTH 1 OF 1978 To DAY 31 OF MONflj 3 0F 1990

Figure 3 Probability density contours showing positive dependencebetweensurge and wave

1.4 Reviewof alternativemethods

Full analvsisof dependenceandjoint orobabilities

Extrapolation of ioint orobabilitvdensitv

Direct hindcastingof designvanables

1.5 Outlineof the new method

1 . Preparationof input data"consistingof many independentrecordsofwave height,wave period and water

Fitting of statisticaldistributionsseparatelyto the wave heights,the water levelsandthe wave steepnesses,

5. Extremes analysis of a range of responsevariables basedon the simulated data.

- optionally,division of the datainto two or threeseparatepopulations(datasets)corresponding

(optionally)recordswith zerowaveheightandrecordswith very low

The TESTDATA diagnostic proeram

2.2 Fittingof statisticaldistributionsto the marginalvariables

The GeneralisedPareto Distribution (GPD)

Importanceof wave period

Incorporation of impleyeclplqeligliellfunarginal extremes

2.3 Statisticalmodelsfor dependence

!"**rroo,.o l0 sR5tt 04/r)5/{11)

2.5 Analysisof joint exceedanceextremesand structuralresponsefunctions

Direct analysisof structuralresponsevariables

- overtoppingrate on a smooth slope;

Analvsis of more comBlgl_ylli4blgs

!"**r,*'.0 11 sR 53? 04/{t5ll)t)

Fittine of statisticalmodels(BVN andMIX

Simulation of long data samoles(SIMBVN and S/MMID

Extremesfrom the simulateddata (ANALISIS)

Comparisons madebv Lancaster University

t"* \4hllingford t3 sR53? 0:V05Xr

Someconclusionsfrom the LancasterUniversity comparisons

["**,roo* ).4 sR 53?04/05/(x)

Resultsand conclusionsfrom direct hindcasting

Threshold(ie the PredictedextremeovenoDpingrateO" (mYs/m)

Table 2 Direct hindcastingand extrapolationof overtoppingrate, appliedto data set Sim2

Threshold(ie the Predicted extreme overtopping rate Q" (m'is/m)

t"* Wallingford 15 sR53?lM^)tu)

Threshold(ie the Predictedextremeovertopoinqrate O. (mls/m)

E"o"u'"'.0 22 sR5t7 04 rtrx)