110
8 H is to ric a l R e m a rk s a n d A lt e rn a tiv e A p p ro a c h e s
8.1
O v e rv ie w
(Thissection,2slides,hasbeenincludedintheIntroduction.)SeeHampeletal(1986):IntroductionbyF.HampelOutliers
Rejectionofoutliers:DanielBernoulli(1770),astronomers
Trimmedmean:1821Rejectionrules:Pierce(1852),...:Testextremeobservations(largest&smallest)forbeinganoutlier,Rejectiftestissignificant,averagetheremainingones.
111
RobustnessofTests
Fisher’sexacttheoryforthenormaldistr.−→“monopoly”.E.S.Pearson(1931):Non-robustnessofF-testforvariances−→Focusonrobustnessofthelevel,sometimesneglectingpower
Non-parametrictests:PowerofWilcoxontestisverygood!
Robustestimation
Non-robustnessofthemean:Tukey(1960)Minimaxtheory:Huber(1964)InfluenceFunctionandBreakdownpoint:Hampel(1968)
112
8.2
O th e r A p p ro a c h e s a n d R e su lt s
aRejectionrulesseeabove
bL-estimatorsPrototype:TrimmedmeanWeightedaverage,weightbasedonrank:
Tn= X
i aiX[i] ,X[1] ,...,X[n] ordered, X
i ai=1Functional:Defineaibyafunctionh:[0,1]→|R
ai= Z
[(i−1)/n,i/n] hhzidz . Z
[0,1] hhzidz
ThFi= RxhhFhxiidFhxiRhhzidz
113 .2
cR-estimatorsPrototype:Hodges-Lehmannestimator,derivedfromSignedRankTestTeststatistic−→p-value.Chooseµsuchthatp-value=1.
−→bµ=medh≤ih(Xh+Xi)/2i,medianof“Walshaverages”
Generalform:Foragivenskewsymmetricfunctionhh.i,
ThFiissolutionofZh 12 Fhyi+ 12 (1−Fh2t−yi) dFhyi=0fort.Theintegralgets0whenthesampleX1,..,Xnanditsmirrorimageaboutt,2t−X1,..,2t−Xnhavetheirranks“mixed”aswellaspossible.
1148.2
dAdaptiveestimatorsForn→∞,thedistributioncanbeestimatedfromthedata
−→asymptoticallyfullyefficientestimators!Needlargesamples.
eRobustnesstowardsdependencelongrangedependence:Hampel(1987),Beran(1992),...
fOtherAspectsofRobustnessRobustifiedLikelihoodRatioTest,ConfidenceInt.:Huber(1968)Capacities:Huber&Strassen(1973)
115.2
gFurtherDevelopmentsPrincetonRobustnessYear1970:Simulationformanylocationandscaleestimators
Firstresultsonregression:Huber(1971)CovarianceMatrix:Maronna(1976)TimeSeries:Martin(1979),K
¨unsch(1981)
Robusttests(infinitesimal):Ronchetti(1980)
Combinatorialestimators:Rousseeuwandhisstudents
116 8.2
hMainResearchGroups
•Tukey,BellLabs&Princeton
−→St.Morgenthaler(Lausanne),DavidTyler(Rudgers)
•HuberandHampelinZurich−→AlfioMarazzi(Lausanne)Ronchetti,Rousseeuw,Stahel,K
¨unsch,M Function” M-estimation,Hampel’sApproach“basedontheInfluence ¨achler(ETH)
•ElvezioRonchetti,Geneva−→Maria-PiaVictoria-Feser,EvaCantoni,StephaneHeritier
•PeterRousseeuw,Belgium−→MiaHubert,C.Croux
117
•Argentina:VictorYohai,RicardoMaronna(&DougMartin,Seattle)Highbreakdownest.,MinimaxBias,soundtheory!−→Book
•PennState:Hettmansperger−→Shrader&McKean(Aus)R-estimators
•Oja,Finland(connectedtoBerne):Multivariatecombinatorial,nonparametric
•Chicago:Portnoy,Koenker,X.HeRegressionQuantiles,soundtheory
•...
118 8.2
iBooksHuber(1981(2009))Hampel,Ronchetti,Rousseeuw,Stahel(1986)Rousseeuw&Leroy(1987):RegressionMaronna,Martin,Yohai(2006)
Conferences
RobustnessYearinPrinceton(1970)Minneapolis(1989)
L1norm(Neuchatel,1987ff)Int.ConferenceonRobustStatistics(ICORS)