P1:SFK/UKS P2:SFK/UKS QC:SFK/UKS T1:SFK CUUK244-Claeskens 978-0-521-85225-8 January13,2008 8:41 Model Selection and Model Averaging Givenadataset,youcanfitthousandsofmodelsatthepushofabutton,buthowdo youchoosethebest?Withsomanycandidatemodels,overfittingisarealdanger. IsthemonkeywhotypedHamletactuallyagoodwriter? Choosing a suitable model is central to all statistical work with data. Selecting the variables for use in a regression model is one important example. The past two decades have seen rapid advances both in our ability to fit models and in the theoreticalunderstandingofmodelselectionneededtoharnessthisability,yetthis book is the first to provide a synthesis of research from this active field, and it contains much material previously difficult or impossible to find. In addition, it givespracticaladvicetotheresearcherconfrontedwithconflictingresults. Modelchoicecriteriaareexplained,discussedandcompared,includingAkaike’s informationcriterionAIC,theBayesianinformationcriterionBICandthefocused informationcriterionFIC.Importantly,theuncertaintiesinvolvedwithmodelselec- tionareaddressed,withdiscussionsoffrequentistandBayesianmethods.Finally, modelaveragingschemes,whichcombinethestrengthofseveralcandidatemodels, arepresented. Worked examples on real data are complemented by derivations that provide deeper insight into the methodology. Exercises, both theoretical and data-based, guide the reader to familiarity with the methods. All data analyses are compati- ble with open-source R software, and data sets and R code are available from a companionwebsite. Gerda Claeskens is Professor in the OR & Business Statistics and Leuven StatisticsResearchCenterattheCatholicUniversityofLeuven,Belgium. Nils Lid Hjort is Professor of Mathematical Statistics in the Department of MathematicsattheUniversityofOslo,Norway. i P1:SFK/UKS P2:SFK/UKS QC:SFK/UKS T1:SFK CUUK244-Claeskens 978-0-521-85225-8 January13,2008 8:41 CAMBRIDGE SERIES IN STATISTICAL AND PROBABILISTIC MATHEMATICS EditorialBoard R.Gill(DepartmentofMathematics,UtrechtUniversity) B.D.Ripley(DepartmentofStatistics,UniversityofOxford) S.Ross(DepartmentofIndustrialandSystemsEngineering,UniversityofSouthernCalifornia) B.W.Silverman(St.Peter’sCollege,Oxford) M.Stein(DepartmentofStatistics,UniversityofChicago) Thisseriesofhigh-qualityupper-divisiontextbooksandexpositorymonographscoversallaspectsof stochasticapplicablemathematics.Thetopicsrangefrompureandappliedstatisticstoprobability theory,operationsresearch,optimization,andmathematicalprogramming.Thebookscontainclear presentationsofnewdevelopmentsinthefieldandalsoofthestateoftheartinclassicalmethods. While emphasizing rigorous treatment of theoretical methods, the books also contain applications anddiscussionsofnewtechniquesmadepossiblebyadvancesincomputationalpractice. Alreadypublished 1. BootstrapMethodsandTheirApplication,byA.C.DavisonandD.V.Hinkley 2. MarkovChains,byJ.Norris 3. AsymptoticStatistics,byA.W.vanderVaart 4. WaveletMethodsforTimeSeriesAnalysis,byDonaldB.PercivalandAndrewT.Walden 5. BayesianMethods,byThomasLeonardandJohnS.J.Hsu 6. EmpiricalProcessesinM-Estimation,bySaravandeGeer 7. NumericalMethodsofStatistics,byJohnF.Monahan 8. AUser’sGuidetoMeasureTheoreticProbability,byDavidPollard 9. TheEstimationandTrackingofFrequency,byB.G.QuinnandE.J.Hannan 10. DataAnalysisandGraphicsusingR,byJohnMaindonaldandJohnBraun 11. StatisticalModels,byA.C.Davison 12. SemiparametricRegression,byD.Ruppert,M.P.Wand,R.J.Carroll 13. ExercisesinProbability,byLoicChaumontandMarcYor 14. StatisticalAnalysisofStochasticProcessesinTime,byJ.K.Lindsey 15. MeasureTheoryandFiltering,byLakhdarAggounandRobertElliott 16. EssentialsofStatisticalInference,byG.A.YoungandR.L.Smith 17. ElementsofDistributionTheory,byThomasA.Severini 18. StatisticalMechanicsofDisorderedSystems,byAntonBovier 20. RandomGraphDynamics,byRickDurrett 21. Networks,byPeterWhittle 22. SaddlepointApproximationswithApplications,byRonaldW.Butler 23. AppliedAsymptotics,byA.R.Brazzale,A.C.DavisonandN.Reid 24. RandomNetworksforCommunication,byMassimoFranceschettiandRonaldMeester 25. DesignofComparativeExperiments,byR.A.Bailey ii P1:SFK/UKS P2:SFK/UKS QC:SFK/UKS T1:SFK CUUK244-Claeskens 978-0-521-85225-8 January13,2008 8:41 Model Selection and Model Averaging Gerda Claeskens K.U.Leuven Nils Lid Hjort UniversityofOslo iii P1:SFK/UKS P2:SFK/UKS QC:SFK/UKS T1:SFK CUUK244-Claeskens 978-0-521-85225-8 January13,2008 8:41 cambridgeuniversitypress Cambridge,NewYork,Melbourne,Madrid,CapeTown,Singapore,Sa˜oPaulo,Delhi CambridgeUniversityPress TheEdinburghBuilding,CambridgeCB28RU,UK PublishedintheUnitedStatesofAmericabyCambridgeUniversityPress,NewYork www.cambridge.org Informationonthistitle:www.cambridge.org/9780521852258 (cid:2)C G.ClaeskensandN.L.Hjort2008 Thispublicationisincopyright.Subjecttostatutoryexception andtotheprovisionsofrelevantcollectivelicensingagreements, noreproductionofanypartmaytakeplacewithout thewrittenpermissionofCambridgeUniversityPress. Firstpublished2008 PrintedintheUnitedKingdomattheUniversityPress,Cambridge AcataloguerecordforthispublicationisavailablefromtheBritishLibrary LibraryofCongressCataloguinginPublicationdata ISBN 978-0-521-85225-8hardback CambridgeUniversityPresshasnoresponsibilityforthepersistenceor accuracyofURLsforexternalorthird-partyinternetwebsitesreferredto inthispublication,anddoesnotguaranteethatanycontentonsuch websitesis,orwillremain,accurateorappropriate. iv P1:SFK/UKS P2:SFK/UKS QC:SFK/UKS T1:SFK CUUK244-Claeskens 978-0-521-85225-8 January13,2008 8:41 ToMaartenandHanne-Sara –G.C. ToJens,AudunandStefan –N.L.H. v P1:SFK/UKS P2:SFK/UKS QC:SFK/UKS T1:SFK CUUK244-Claeskens 978-0-521-85225-8 January13,2008 8:41 vi P1:SFK/UKS P2:SFK/UKS QC:SFK/UKS T1:SFK CUUK244-Claeskens 978-0-521-85225-8 January13,2008 8:41 Contents Preface pagexi Aguidetonotation xiv 1 Modelselection:dataexamplesandintroduction 1 1.1 Introduction 1 1.2 Egyptianskulldevelopment 3 1.3 Whowrote‘TheQuietDon’? 7 1.4 Survivaldataonprimarybiliarycirrhosis 10 1.5 Lowbirthweightdata 13 1.6 Footballmatchprediction 15 1.7 Speedskating 17 1.8 Previewofthefollowingchapters 19 1.9 Notesontheliterature 20 2 Akaike’sinformationcriterion 22 2.1 Informationcriteriaforbalancingfitwithcomplexity 22 2.2 MaximumlikelihoodandtheKullback–Leiblerdistance 23 2.3 AICandtheKullback–Leiblerdistance 28 2.4 Examplesandillustrations 32 2.5 Takeuchi’smodel-robustinformationcriterion 42 2.6 CorrectedAICforlinearregressionandautoregressivetimeseries 44 2.7 AIC,correctedAICandbootstrap-AICforgeneralised ∗ linearmodels 46 ∗ 2.8 BehaviourofAICformoderatelymisspecifiedmodels 49 2.9 Cross-validation 51 2.10 Outlier-robustmethods 55 2.11 Notesontheliterature 64 Exercises 66 vii P1:SFK/UKS P2:SFK/UKS QC:SFK/UKS T1:SFK CUUK244-Claeskens 978-0-521-85225-8 January13,2008 8:41 viii Contents 3 TheBayesianinformationcriterion 70 3.1 ExamplesandillustrationsoftheBIC 70 3.2 DerivationoftheBIC 78 3.3 Whowrote‘TheQuietDon’? 82 3.4 TheBICandAICforhazardregressionmodels 85 3.5 Thedevianceinformationcriterion 90 3.6 Minimumdescriptionlength 94 3.7 Notesontheliterature 96 Exercises 97 4 Acomparisonofsomeselectionmethods 99 4.1 Comparingselectors:consistency,efficiencyandparsimony 99 4.2 Prototypeexample:choosingbetweentwonormalmodels 102 4.3 StrongconsistencyandtheHannan–Quinncriterion 106 4.4 Mallows’sC anditsoutlier-robustversions 107 p 4.5 Efficiencyofacriterion 108 4.6 EfficientorderselectioninanautoregressiveprocessandtheFPE 110 4.7 Efficientselectionofregressionvariables 111 ∗ 4.8 Ratesofconvergence 112 ∗ 4.9 Takingthebestofbothworlds? 113 4.10 Notesontheliterature 114 Exercises 115 5 Biggerisnotalwaysbetter 117 5.1 Someconcreteexamples 117 5.2 Large-sampleframeworkfortheproblem 119 5.3 Aprecisetolerancelimit 124 5.4 Toleranceregionsaroundparametricmodels 126 5.5 Computingtolerancethresholdsandradii 128 5.6 Howthe5000-mtimeinfluencesthe10,000-mtime 130 5.7 Large-samplecalculusforAIC 137 5.8 Notesontheliterature 140 Exercises 140 6 Thefocussedinformationcriterion 145 6.1 Estimatorsandnotationinsubmodels 145 6.2 Thefocussedinformationcriterion,FIC 146 6.3 Limitdistributionsandmeansquarederrorsinsubmodels 148 6.4 Abias-modifiedFIC 150 6.5 CalculationoftheFIC 153 6.6 Illustrationsandapplications 154 ∗ 6.7 Exactmeansquarederrorcalculationsforlinearregression 172 P1:SFK/UKS P2:SFK/UKS QC:SFK/UKS T1:SFK CUUK244-Claeskens 978-0-521-85225-8 January13,2008 8:41 Contents ix 6.8 TheFICforCoxproportionalhazardregressionmodels 174 6.9 Average-FIC 179 ∗ 6.10 ABayesianfocussedinformationcriterion 183 6.11 Notesontheliterature 188 Exercises 189 7 FrequentistandBayesianmodelaveraging 192 7.1 Estimators-post-selection 192 7.2 SmoothAIC,smoothBICandsmoothFICweights 193 7.3 Distributionofmodelaverageestimators 195 7.4 Whatgoeswrongwhenweignoremodelselection? 199 7.5 Betterconfidenceintervals 206 7.6 Shrinkage,ridgeestimationandthresholding 211 7.7 Bayesianmodelaveraging 216 ∗ 7.8 AfrequentistviewofBayesianmodelaveraging 220 ∗ 7.9 Bayesianmodelselectionwithcanonicalnormalpriors 222 7.10 Notesontheliterature 223 Exercises 224 8 Lack-of-fitandgoodness-of-fittests 227 8.1 Theprincipleoforderselection 227 8.2 Asymptoticdistributionoftheorderselectiontest 229 ∗ 8.3 Theprobabilityofoverfitting 232 8.4 Score-basedtests 236 8.5 Twoormorecovariates 238 8.6 Neyman’ssmoothtestsandgeneralisations 240 ∗ 8.7 AcomparisonbetweenAICandtheBICformodeltesting 242 ∗ 8.8 Goodness-of-fitmonitoringprocessesforregressionmodels 243 8.9 Notesontheliterature 245 Exercises 246 9 Modelselectionandaveragingschemesinaction 248 9.1 AICandBICselectionforEgyptianskulldevelopmentdata 248 9.2 Lowbirthweightdata:FICplotsandFICselectionperstratum 252 9.3 SurvivaldataonPBC:FICplotsandFICselection 256 9.4 Speedskatingdata:averagingovercovariancestructuremodels 258 Exercises 266 10 Furthertopics 269 10.1 Modelselectioninmixedmodels 269 10.2 Boundaryparameters 273 ∗ 10.3 Finite-samplecorrections 281 P1:SFK/UKS P2:SFK/UKS QC:SFK/UKS T1:SFK CUUK244-Claeskens 978-0-521-85225-8 January13,2008 8:41 x Contents 10.4 Modelselectionwithmissingdata 282 10.5 When p andq growwithn 284 10.6 Notesontheliterature 285 Overviewofdataexamples 287 References 293 Authorindex 306 Subjectindex 310