Biometrics By The Border...11. An Approach To Poisson Mixed Models For -Omics Expression Data 12....
Transcript of Biometrics By The Border...11. An Approach To Poisson Mixed Models For -Omics Expression Data 12....
BiometricsByTheBorder
InternationalBiometricsSocietyAustralasianRegionConferenceKingscliff,NSW
26-30thNovember2017
BiometricsByTheBorder1. Welcome2. ProgrammeAtAGlance
1. Monday2. Tuesday3. Wednesday4. Thursday
3. ProgrammeAndAbstractsForMonday27thOfNovember1. AgriculturalAndAgri-EnvironmentalStatisticsWithSupportOf
GeospatialInformation,MethodologicalIssues2. DevelopingARegulatoryDefinitionForTheAuthenticationOf
ManukaHoney3. OnTestingRandomEffectsInLinearMixedModels4. AnalysingDigestionData5. APermutationTestForComparingPredictiveValuesInClinical
Trials6. ChallengesAndOpportunitiesWorkingAsAConsultingStatistician
WithAFoodScienceResearchGroup7. RobustSemiparametricInferenceInRandomEffectsModels8. RobustPenalizedLogisticRegressionThroughMaximumTrimmed
LikelihoodEstimator9. AMulti-StepClassifierAddressingCohortHeterogeneityImproves
PerformanceOfPrognosticBiomarkersInComplexDisease10. HowToAnalyseFiveDataPointsWithFun11. AnApproachToPoissonMixedModelsFor-OmicsExpressionData12. BayesianSpatialEstimationWhenAreasAreFew13. Knowledge-GuidedGeneralizedBiclusteringAnalysisForIntegrative
–OmicsAnalysis14. UnderstandingTheVariationInHarvesterYieldMapDataFor
EstimatingCropTraits15. CitizenScienceToSurveillance:EstimatingReportingProbabilities
OfExoticInsectPests16. IntroductionTo“Deltagen”-AComprehensiveDecisionSupportTool
ForPlantBreedersUsingRAndShiny17. EstimatingNitrousOxideEmissionFactors
4. ProgrammeAndAbstractsForTuesday28thOfNovember1. ClusterCapture-Recapture:ANewFrameworkForEstimating
PopulationSize2. PropensityScoreApproachesInThePresenceOfMissingData:
ComparisonOfBalanceAndTreatmentEffectEstimates3. VisualisingModelSelectionStabilityInHigh-DimensionalRegression
Models4. TheMissingLink:AnEquivalenceResultForLikelihoodBased
MethodsInMissingDataProblems5. DimensionalityReductionOfLIBSDataForBayesianAnalysis6. AnalysisOfMelanomaDataWithAMixtureOfSurvivalModels
UtilisingMulti-ClassDLDAToInformMixtureClass7. ForecastingHotspotsOfPotentiallyPreventableHospitalisationsWith
SpatiallyAggregatedLongitudinalHealthData:AllSubsetModelSelectionWithANovelImplementationOfRepeatedK-FoldCross-Validation
8. IdentifyingClustersOfPatientsWithDiabetesUsingAMarkovBirth-DeathProcess
9. ChallengesAnalysingCombinedAgriculturalFieldTrialsWithPartiallyOverlappingTreatments
10. AHiddenMarkovModelForSleepStageDetectionUsingRawTri-AxialWristActigraphy
11. SparsePhenotypingDesignsForEarlyStageSelectionExperimentsInPlantBreedingPrograms
12. ComparisonsOfTwoLargeLong-TermStudiesInAlzheimer'sDisease
13. AOne-StageMixedModelAnalysisOfCanolaChemistryTrials14. ASemi-ParametricLinearMixedModelsForLongitudinally
MeasuredFastingBloodSugarLevelOfAdultDiabeticPatients15. IndividualAndJointAnalysesOfSugarcaneExperimentsToSelect
TestLines5. ProgrammeAndAbstractsForWednesday29thOfNovember
1. StatisticsOnStreetCorners2. EstimatingOverdispersionInSparseMultinomialData3. AssessingMudCrabMeatFullnessUsingNon-InvasiveTechnologies4. AnalysisOfMultivariateBinaryLongitudinalData:Metabolic
SyndromeDuringMenopausalTransition5. StatisticalAnalysisOfCoastalAndOceanographicInfluencesOnThe
QueenslandScallopFishery
6. SavedByTheExperimentalDesign:TestingBycatchReductionAndTurtleExclusionDevicesInThePngPrawnTrawlFishery
7. RethinkingBiosecurityInspections:ACaseStudyOfTheAsianGypsyMoth(AGM)InAustralia
8. SubtractiveStabilityMeasuresForImprovedVariableSelection9. AComparisonOfMultipleImputationMethodsForMissingDataIn
LongitudinalStudies10. SpeciesDistributionModellingForCombinedDataSources11. AFactorAnalyticMixedModelApproachForTheAnalysisOf
GenotypeByTreatmentByEnvironmentData12. TheImpactOfCohortSubstanceUseUponLikelihoodOf
TransitioningThroughStagesOfAlcoholAndCannabisUseAndUseDisorder:FindingsFromTheAustralianNationalSurveyOnMentalHealthAndWell-Being
13. TheLASSOOnLatentIndicesForOrdinalPredictorsInRegression14. Whole-GenomeQTLAnalysisForNestedAssociationMapping
Populations15. AnAsymmetricMeasureOfPopulationDifferentiationBasedOnThe
SaddlepointApproximationMethod16. FastAndApproximateExhaustiveVariableSelectionForGLMsWith
APES17. OrderSelectionOfFactorAnalyticModelsForGenotypeX
EnvironmentInteraction18. MultipleSampleHypothesisTestingOfTheHumanMicrobiome
ThroughEvolutionaryTrees19. StatisticalStrategiesForTheAnalysisOfLargeAndComplexData20. OptimalExperimentalDesignForFunctionalResponseExperiments21. ToPCAOrNotToPCA22. AnEvaluationOfErrorVarianceBiasInSpatialDesigns23. NewModel-BasedOrdinationDataExplorationToolsFor
MicrobiomeStudies24. AlwaysRandomize?25. ComparingClassicalCriteriaForSelectingIntra-ClassCorrelated
FeaturesForThree-ModeThree-WayData26. ExploringTheSocialRelationshipsOfDairyGoats27. BayesianSemi-ParametricSpectralDensityEstimationWith
ApplicationsToTheSouthernOscillationIndex28. EfficientMultivariateSensitivityAnalysisOfAgriculturalSimulators29. BayesianHypothesisTestsWithDiffusePriors:CanWeHaveOur
CakeAndEatItToo?6. ProgrammeAndAbstractsForThursday30thOfNovember
1. AGeneralFrameworkForFunctionalRegressionModelling2. HockeySticksAndBrokenSticks–ADesignForASingle-Arm,
Placebo-Controlled,Double-Blind,RandomizedClinicalTrialSuitableForChronicDiseases
3. BoundingIVEstimatesUsingMediationAnalysisThinking4. CorrelatedBivariateNormalCompetingRisks---Structuring
EstimationInAnIll-PosedProblem5. BayesianRegressionWithFunctionalInequalityConstraints6. GeneticAnalysisOfRenalFunctionInAnIsolatedAustralian
IndigenousCommunity7. ThePerformanceOfModelAveragedTailAreaConfidenceIntervals8. BiasCorrectionInEstimatingProportionsByPooledTesting9. TheParametricCureFractionModelOfOvarianCancer10. TheSkillings-MackStatisticForRanksDataInBlocks
7. PosterAbstracts1. Multi-EnvironmentTrialAnalysisOfAgronomyTrialsUsing
EstablishedPlantPopulationDensity2. AdventuresInDigitalAgricultureInNewZealand3. ModellingCanopyGreennessOverTimeUsingSplinesAndNon-
LinearRegression4. OnTestingMarginalHomogeneityForSquareContingencyTables
WithOrdinalCategories5. AFactorAnalyticApproachToModellingDiseaseProgression
AcrossLeafLayersAndTime6. AssociatingStrawStrengthWithLikelihoodOfHeadLossInBarley
ForWesternAustralia7. ADeepLearningNeuralNetworkModelToIdentifyTheImportant
GenesInMetastaticBreastCancerFromCensoredMicroarrayData8. UsingRAndShinyToDevelopWebBasedSamplingApplications
ForTheAgriculturalAndEducationSectors9. GraphicalNetworkAnalysesInformsBiomarkerDiscoveryVia
QuantificationOfKeyDiseaseRelatedConnections10. Spatio-TemporalCorticalBrainAtrophyPatternsOfAlzheimer's
Disease11. EmpiricalModellingOfFruitFirmnessChangeDuringColour
Conditioning
WelcomeWelcometothe2017BiometricsbytheBorderconferenceinKingscliff,NewSouthWales.AswemeethereonthelandoftheBundjalungpeople,weacknowledgethetraditionalcustodiansofthelandandpayourrespecttotheElderspastandpresent.
IbidaspecialwelcometoallwhotravelledfarfromcountriessuchasBelgium,Brazil,China,Denmark,Germany,Italy,Japan,Nigeria,SouthAfrica,USAandVietnamtocomplementthemajorityfromAustraliaandNewZealand.
TheconferencewillbeopenedbythecurrentPresidentofIBS,ElizabethThompsonandclosedbytheincomingPresident,LouiseRyan.In-between,wehavesixoutstandingkeynotespeakerswhosettheframeworkforanexcitinganddiversescientificprogram.Iinviteyoualltoattendpresentationsontopicsfamiliartoyourresearchandtoexplorewhatisdoneinotherareas.
Wehaveaveryhealthynumberofstudentsandearlycareerresearchersamongthemorethan120delegatesattending,indicatinghowpopular,importantandattractivedatarichresearchcontinuestobeintheyearstocome.
Thisconferencewillbeasuccessbecauseofexcellentworkbymany,inparticularbyRossDarnell,MelissaDobbie,ClaytonForknall,AlisonKellyandBethanyMacdonald,whoformthelocalorganizingcommittee;byJamesCurran(ChairoftheScientificProgram),GarthTarrandEstherMeenkenwhoassistedJames;byWarrenMuller(Treasurer),DavidBaird(Secretary)andbyHansHockeywhoisourofficialphotographer,soalwayssmilewhenheisaround.
LastbutnotleastIthankourPlatinumsponsorsACEMS,GRDCandtheUniversityofSydneyfortheirgenerousfinancialsupport,manyaspectsoftheconferenceareonlypossiblebecauseoftheirhelp.
SamuelMueller,PresidentIBS-AR
ProgrammeAtAGlance
Monday
Time Monday27/11850 Housekeeping900 Welcome
SamuelMueller,PresidentTraditonalwelcometoAustraliabyarepresentativeofthe
BundjalungElizabethThompson,PresidentoftheIBS
940 AgriculturalAndAgri-EnvironmentalStatisticsWithSupportOfGeospatialInformation,MethodologicalIssues
ElisabettaCarfagnaUniversityofBologna,Italy
1030 MorningTea(30minutes)
Modelling Sampling,ComplexSurveys,AndMixedModels
1100 DevelopingARegulatoryDefinitionForTheAuthenticationOfManukaHoneyClaireMcDonald,MinistryforPrimaryIndustries,NZ
OnTestingRandomEffectsInLinearMixedModelsAlanWelsh,ANU
1120 AnalysingDigestionDataMaryannStaincliffe,AgResearchLtd.
APermutationTestForComparingPredictiveValuesInClinicalTrialsKoujiYamamoto,OsakaCityUniversity
1140 ChallengesAndOpportunitiesWorkingAsAConsultingStatisticianWithAFoodScienceResearchGroupM.GabrielaBorgognone,QueenslandDepartmentof
RobustSemiparametricInferenceInRandomEffectsModelsMichaelStewart,UniversityofSydney
AgricultureandFisheries1200 RobustPenalizedLogistic
RegressionThroughMaximumTrimmedLikelihoodEstimatorTongWang,ShanxiMedicalUniversty,China
1220 Lunch(1hour10minutes)1330 AMulti-StepClassifierAddressingCohortHeterogeneity
ImprovesPerformanceOfPrognosticBiomarkersInComplexDisease
JeanYangUniversityofSydney
General/Education Genetics1420 HowToAnalyseFiveDataPoints
WithFunPaulineO'Shaughnessy,UniversityofWollongong
AnApproachToPoissonMixedModelsFor-omicsExpressionDataIreneZeng,UniversityofAuckland
1440 BayesianSpatialEstimationWhenAreasAreFewAswiAswi,QueenslandUniversityofTechnology
Knowledge-guidedGeneralizedBiclusteringAnalysisforintegrative–omicsAnalysisQiLong,UniversityofPennsylvania
1500 UnderstandingTheVariationInHarvesterYieldMapDataForEstimatingCropTraitsDeanDiepeveen,CurtinUniversity
CitizenScienceToSurveillance:EstimatingReportingProbabilitiesOfExoticInsectPestsPeterCaley,CSIRO
1520 Afternoontea(30minutes)Computing/BigData EnvironmentalStatistics
1550 IntroductionTo“Deltagen”-AComprehensiveDecisionSupportToolForPlantBreedersUsingRAndshinyDongwenLuo,AgResearchLtd.
EstimatingNitrousOxideEmissionFactorsAlasdairNoble,AgResearchLtd.
1630 LightningTalksfollowedbyPosterSessionKindlySponsoredbytheGrainsResearchandDevelopment
Corporation
Tuesday
Time Tuesday28/11850 Housekeeping900 ClusterCapture-Recapture:ANewFrameworkForEstimating
PopulationSizeRachelFewster
UniversityofAuckland950 MorningTea(30minutes)
MissingData HighDimensionalData1030 PropensityScore
ApproachesInThePresenceOfMissingData:ComparisonOfBalanceAndTreatmentEffectEstimatesJannahBaker,TheGeorgeInsituteforGlobalHealth
VisualisingModelSelectionStabilityinHigh-DimensionalRegressionModelsGarthTarr,UniversityofSydney
1050 TheMissingLink:AnEquivalenceResultForLikelihoodBasedMethodsInMissingDataProblemsFirouzehNoghrehchi,UniversityofNewSouthWales
DimensionalityReductionofLIBSDataforBayesianAnalysisAnjaliGupta,UniversityofAuckland
Medical Modelling1110 AnalysisofMelanoma
DatawithaMixtureofSurvivalModelsUtilisingMulti-class
ForecastingHotspotsOfPotentiallyPreventableHospitalisationsWithSpatiallyAggregatedLongitudinalHealthData:AllSubsetModelSelectionWithANovel
DLDAtoInformMixtureClassSarahRomanes,UniversityofSydney
ImplementationOfRepeatedk-FoldCross-ValidationMatthewTuson,UniversityofWesternAustralia
1130 IdentifyingClustersOfPatientsWithDiabetesUsingaMarkovBirth-DeathProcessMugdhaManda,UniversityofAuckland
ChallengesAnalysingCombinedAgriculturalFieldTrialsWithPartiallyOverlappingTreatmentsKerryBell,QueenslandDepartmentofAgricultureandFisheries
1150 AHiddenMarkovModelForSleepStageDetectionUsingRawTri-AxialWristActigraphyMichelleTrevenen,UniversityofWesternAustralia
SparsePhenotypingDesignsForEarlyStageSelectionExperimentsInPlantBreedingProgramsNicoleCocks,UniversityofWollongong
1210 ComparisonsOfTwoLargeLong-TermStudiesInAlzheimer'sDiseaseCharleyBudgeon,UniversityofWesternAustralia
AOne-StageMixedModelAnalysisOfCanolaChemistryTrialsDanielTolhurst,UniversityofWollongong
1230 ASemi-ParametricLinearMixedModelsForLongitudinallyMeasuredFastingBloodSugarLevelOfAdultDiabeticPatientsLegesseKassaDebusho,UniversityofSouthAfrica
IndividualAndJointAnalysesOfSugarcaneExperimentsToSelectTestLinesRenataAlcardeSermarini,UniversityofSaoPaulo,Brazil
1250 Lunch(1hour10minutes)
Wednesday
Time Wednesday29/11850 Housekeeping900 StatisticsOnStreetCorners
DianneCookMonashUniversity
950 MorningTea(30minutes)Modelling Fisheries BiometricsI
1030 EstimatingOverdispersioninSparseMultinomialDataFarzanaAfroz,UniverityofOtago
AssessingMudCrabMeatFullnessUsingNon-InvasiveTechnologiesCaroleWright,QueenslandDepartmentofAgricultureandFisheries
AnalysisOfMultivariateBinaryLongitudinalData:MetabolicSyndromeDuringMenopausalTransitionGeoffJones,MasseyUniversity
1050 StatisticalAnalysisOfCoastalAndOceanographicInfluencesOnTheQueenslandScallopFisheryWen-HsiYang,UniversityofQueensland
SavedByTheExperimentalDesign:TestingBycatchReductionAndTurtleExclusionDevicesInThePngPrawnTrawlFisheryEmmaLawrence,CSIRO
RethinkingBiosecurityInspections:ACaseStudyOfTheAsianGypsyMoth(AGM)InAustraliaPetraKuhnert,CSIRO
ModelSelection Agriculture/Horticulture BiometricsII1110 SubtractiveStability
MeasuresForImprovedVariableSelectionConnorSmith,UniversityofSydney
GxEforGenomicPredictionModelsColleenHunt,QueenslandDepartmentofFisheriesandAgriculture
AComparisonOfMultipleImputationMethodsForMissingDataInLongitudinalStudiesMdHamidul
Huque,MurdochChildrensResearchInstitute
1130 SpeciesDistributionModellingForCombinedDataSourcesIanRenner,UniversityofNewcastle
AFactorAnalyticMixedModelApproachForTheAnalysisOfGenotypeByTreatmentByEnvironmentDataLaurenBorg,UniversityofWollongong
TheImpactOfCohortSubstanceUseUponLikelihoodOfTransitioningThroughStagesOfAlcoholAndCannabisUseAndUseDisorder:FindingsFromTheAustralianNationalSurveyOnMentalHealthAndWell-BeingChriannaBharat,NationalDrugandAlcoholResearchCentre,Australia
1150 TheLASSOOnLatentIndicesForOrdinalPredictorsinRegressionFrancisHui,ANU
Whole-GenomeQTLAnalysisForNestedAssociationMappingPopulationsMariaValeriaPaccapelo,QueenslandDepartmentofAgricultureandFisheries
AnAsymmetricMeasureofPopulationDifferentiationBasedontheSaddlepointApproximationMethodLouise
McMillan,UniversityofAuckland
1210 FastAndApproximateExhaustiveVariableSelectionForGLMsWithAPESKevinWang,UniversityofSydney
OrderSelectionOfFactorAnalyticModelsForGenotypeXEnvironmentInteractionEmiTanaka,UniversityofSydney
MultipleSampleHypothesisTestingOfTheHumanMicrobiomeThroughEvolutionaryTreesMartinaMincheva,TempleUniversity,USA
1230 Lunch(1hour10minutes)
AGM1340 StatisticalStrategiesfortheAnalysisofLargeAndComplexData
LouiseRyanUniversityofTechnologySydneyand
HarvardT.H.ChanSchoolofPublicHealthDesign Multivariate
1430 OptimalExperimentalDesignforFunctionalResponseExperimentsChristopherDrovandi,QueenslandUniversityofTechnology
ToPCAOrNotToPCACatherineMcKenzie,AgResearchLtd.
1450 AnEvaluationofErrorVarianceBiasinSpatialDesignsEmlynWilliams,ANU
NewModel-BasedOrdinationDataExplorationToolsForMicrobiomeStudiesOlivierThas,GhentUniversity,Belgium
1510 AlwaysRandomize?ChrisBrien,UniversityofSouthAustraliaand
ComparingClassicalCriteriaForSelectingIntra-ClassCorrelated
UniversityofAdelaide FeaturesForThree-ModeThree-WayDataLynetteHunt,UniversityofWaikato
1530 Afternoontea(20minutes)Agriculture/Horticulture BayesianStatistics
1550 ExploringthesocialrelationshipsofdairygoatsVanessaCave,AgResearchLtd.
BayesianSemi-ParametricSpectralDensityEstimationWithApplicationsToTheSouthernOscillationIndexClaudiaKirch,UniversityofMagdeburg,Germany
1610 EfficientMultivariateSensitivityAnalysisOfAgriculturalSimulatorsDanielGladish,CSIRO
BayesianHypothesisTestsWithDiffusePriors:CanWeHaveOurCakeAndEatItToo?JohnOrmerod,UniversityofSydney
Thursday
Time Thursday30/11850 Housekeeping900 AGeneralFrameworkForFunctionalRegressionModelling
SonjaGrevenLMUniversity,Munich
950 MorningTea(30minutes)
Biostatistics TheoryAndMethods
1030 HockeySticksAndBrokenSticks–ADesignForASingle-Arm,Placebo-Controlled,Double-Blind,RandomizedClinicalTrialSuitableForChronicDiseases
BoundingIVEstimatesUsingMediationAnalysisThinking
HansHockey,BiometricsMatters TheisLange,UniversityofCopenhagen
1050 CorrelatedBivariateNormalCompetingRisks---StructuringEstimationInAnIll-PosedProblemMalcolmHudson,MacquarieUniversity
BayesianRegressionwithFunctionalInequalityConstraintsJoshuaBon,UniversityofWesternAustralia
Genetics/Medical TheoryAndMethods
1110 GeneticAnalysisOfRenalFunctionInAnIsolatedAustralianIndigenousCommunityRussellThomson,WesternSydneyUniversity
ThePerformanceOfModelAveragedTailAreaConfidenceIntervalsPaulKabaila,LaTrobeUniversity
1130 DeconstructingTheInnateImmuneComponentOfAMolecularNetworkOfTheAgingFrontalCortexEllisPatrick,UniversityofSydney
BiasCorrectionInEstimatingProportionsbyPooledTestingGrahamHepworth,UniversityofMelbourne
1150 TheParametricCureFractionModelofOvarianCancerSerifatFolorunso,UniversityofIbadan,Nigeria
TheSkillings-MackStatisticForRanksDataInBlocksJohnBest,UniversityofNewcastle
1210 ConferenceCloseLouiseRyan
1230 Lunch(1hour10minutes)
ProgrammeAndAbstractsForMonday27thOfNovember
Keynote:Monday27th9:40Mantra
AgriculturalAndAgri-EnvironmentalStatisticsWithSupportOfGeospatialInformation,MethodologicalIssues
ElisabettaCarfagnaUniversityofBologna
Agri-environmentaltrade-offsareissuescriticalforpolicymakerschargedwithmanagingbothfoodsupplyandthesustainableuseoftheland.Reliabledataarecrucialfordevelopingeffectivepoliciesandforevaluatingtheirimpact.However,oftenthereliabilityofagriculturalandagro-environmentalstatisticsislow.
Duetothetechnologicaldevelopment,inthelastdecades,differentkindsofgeospatialdatahavebecomeeasilyaccessibleatdecreasingpricesandhavestartedtobeanimportantsupporttostatisticsproductionprocess.
Inthispaper,wefocusonmethodologicalissuesrelatedtotheuseofgeospatialinformationforsamplingframeconstruction,sampledesign,stratification,grounddatacollectionandestimationofagriculturalandagri-environmentalparameters.Particularattentionisdevotedtotheimpactofspatialresolutionofdata,changeofsupportaggregationanddisaggregationofspatialdata,whenremotesensingdata,GlobalPositioningSystemsandGeographicInformationSystems(GIS)areusedforproducingagriculturalandagro-environmentalstatistics.
Monday27th11:00Narrabeen
DevelopingARegulatoryDefinitionForTheAuthenticationOfManukaHoney
ClaireMcDonald1,SuzanneKeeling1,MarkBrewer2,andSteveHathaway11MinistryforPrimaryIndustries
2BioSS
ManukahoneyisapremiumexportproductfromNewZealandthathasbeenunderscrutinyduetoclaimsoffraud,adulterationandmislabelling.Althoughthereareseveralindustryapproachesfordefiningmanukahoney,thereiscurrentlynoscientificallyrobustdefinitionsuitableforuseinaregulatorysetting.Assuch,ensuringtheauthenticityofmanukahoneyischallenging.
HerewepresenttheresultsofathreeyearscienceprogrammewhichdevelopedscientificallyrobustdefinitionsformonofloralandmultifloralmanukahoneyproducedinNewZealand.Theprogrammeinvolved:selectingappropriatemarkerstoidentifyhoneysourcedfromLeptospermumscoparium(manuka),establishingplantandhoneyreferencecollections,developingtestmethodstodeterminethelevelsofthemarkersandanalysingthedatageneratedtodevelopthedefinitions.
Thesuitabilityof16markers(chemicalandDNA-based)wereevaluatedforuseinaregulatorydefinitionformanukahoney.Plantsampleswerecollectedfromtwofloweringseasonsrepresentingbothmanukaandnon-manukaspeciesfrombothNewZealandandAustralia.Honeysamples,alsorepresentingmanukaandnon-manukafloraltypes,weresourcedfromsevenNewZealandproductionseasons.Additionally,honeysamplesweresourcedfromanother12countriestoenablecomparison.Allsamplesweretestedforthemarkersbeingevaluatedusingthedevelopedtestmethods.
ThemethodofCART(ClassificationandRegressionTrees)wasusedtodevelopthemonofloralandmultifloralmanukahoneydefinitions.TheCARToutputswerefurtherprocessedusingasimulationapproachtodeterminethesensitivityandtherobustnessofthedefinitions.Thedefinitionsuseacombinationof5markers(4chemicaland1DNA)atsetthresholdstoclassifyasampleasmanukahoneyorotherwise.Wediscussthepracticalitiesofusingthescience-baseddefinitionswithinaregulatorycontext.
Monday27th11:00Gunnamatta
OnTestingRandomEffectsInLinearMixedModels
AlanWelsh1,FrancisHui1,andSamuelMueller21ANU
2UniversityofSydney
Wecanapproachproblemsinvolvingrandomeffectsinlinearmixedmodelsdirectlythroughtherandomeffectsorthroughparameterssuchasthatthevariancecomponentsthatdescribethedistributionsoftherandomeffects.Bothapproachesareusefulbutleadtodifferentissues.Forexample,workingwiththerandomeffectsraisesquestionsofhowtoestimatethemandcanmeandealingwithalargenumberofrandomeffects,whileworkingwithvariancecomponentsleadstotestinghypothesesontheboundaryoftheparameterspace.Inbothcases,findinggoodapproximationstothenulldistributionoftheteststatisticcanbechallengingsomodernapproachesoftenrelyonsimulation.Inthistalk,were-examinetheF-testbasedonlinearcombinationsoftheresponses,fortestingrandomeffectsinlinearmixedmodels.Wepresentageneralderivationofthetest,highlightitscomputationspeed,itsgenerality,anditsexactnessasatest,andreportempiricalstudiesintothefinitesampleperformanceofthetest.Weconcludethepresentationbyreportingourlatestresultsfromongoingresearchthatinvestigatesconnectionsbetweentestingandmodelselectionbydiscussingsometestsofsignificanceofrandomeffectsandexploringtheirrelationshiptomodelselectionprocedures.
Monday27th11:20Narrabeen
AnalysingDigestionData
MaryannStaincliffe,DebbieFrost,MustafaFarouk,andGuojieWuAgResearch
Foodtechnologistareinterestedinunderstandingifaddinggrainand/orvegetablestobeefincreasesthedigestibilityovermeatalone.Digestibilityof
meatwasmeasuredatupto4hoursusingapepsinandpancreatininvitromodel.Thismethodinvolvesusinggelsinlanes,wherethedensityofthecolourofthegelisanindicatorofthepresenceofaproteinorpeptideatthatlevelofkilodalton(kDa).TypicallykDaabove12kDaareconsideredtobetheproteinsandlowerthanthatarethepeptides.Inthepastwehaveanalysedthesedatausingtwoapproaches.Thefirstapproachistoselect4or5bandsthatexpectedtobeimportantandthenuseamixedeffectsmodeltocomparethemeanTraceQuantityofeachofthebands,whereMeattype,Additivetype(vegetableand/orgrain)andTimepointsarefixedfactorsandthegelisarandomeffect.ThesecondapproachistofitacurvetothechangeintheproportionofTraceQuantityabove12kDafromTimezero.Theproblemwiththesetwoapproachesisthatwestruggletoprovideaconsistentinterpretationoftheresults.Therefore,wewillexplorealternativemethodsforanalysingthistypeofdata.
Monday27th11:20Gunnamatta
APermutationTestForComparingPredictiveValuesInClinicalTrials
KoujiYamamotoandKanaeTakahashiOsakaCityUniversity
Screeningtestsordiagnostictestsareimportantforearlydetectionandtreatmentofdisease.Therearefourwell-knownmeasurements,sensitivity(SE),specificity(SP),positivepredictivevalue(PPV)andnegativepredictivevalue(NPV)indiagnosticstudies.ForcomparingSEs/SPs,McNemartestiswidelyused,butthereareonlyfewmethodsforthecomparisonofPPVs/NPVs.Moreover,allofthesemethodsarebasedonlarge-sampletheory.
So,inthistalk,firstly,weinvestigatetheperformanceofthosemethodswhenthesamplesizeissmall.Inaddition,weproposeapermutationtestforcomparingtwoPPVs/NPVswecanapplyevenifthesamplesizeissmall.Finally,weshowtheperformanceoftheproposedmethodwithsomeexistingmethodsviasimulationstudies.
Monday27th11:40Narrabeen
ChallengesAndOpportunitiesWorkingAsAConsultingStatisticianWithAFoodScienceResearchGroup
M.GabrielaBorgognoneQueenslandDepartmentofAgricultureandFisheries
Whenanestablishedresearchgrouphasbeenfunctioningformanyyearswithoutastatisticianasanintegralpartoftheteam,welcomingoneintothegroupcanpresentchallengesaswellasopportunitiesforallinvolved.
Challengesfortheresearchgroupinclude,forexample,involvingthestatisticianatthebeginningofthestudyinsteadofoncetheexperimentshavebeencompletedandthedatacollected;acquiringorincreasingknowledgeofexperimentaldesignprinciples;understandingthelimitationsofsomestatisticalanalyses,expandingtherangeofmethodstheyfeelfamiliarwith,andlearningwhen/howtoapplyeachone;andimprovingthepresentationofresultsinthiserawherepoorpresentationisperpetuatedbythegenerallackofsoundstatisticalmethodsintheliteratureoftheresearcharea.Challengesforthestatisticianinclude,forexample,overcominghis/herlackofgeneralknowledgeoftheunderlyingscientificareaanditsspecificvocabulary;determiningwhatexperimentaldesignswouldworkfromapracticalpointofview;developingunderstandingoftheirscientificquestions,datamanagementpractices,andtypesofdatacollected;navigatingthevarioussoftwaretheyuseandcheckingtheiradequaciesandlimitations;and,aboveall,communicatingwithpatienceandperseverance.
Correspondingly,allchallengespresentopportunitiesforimprovementandcollaborationbetweenscientistsandstatisticians.Workingasateamsupportsadecisionmakingprocessthatisrelevanttoindustryandthatisbasedongoodstatisticalpractices.Additionally,ithelpsscientistsbecomemorestatisticallyawareandempowered.AbitmorethanayearagoIstartedworkingasaconsultingstatisticianwithafoodscienceresearchgroup.InthispresentationIwillsharesomeofthechallengesandtheopportunitiestoincorporategoodstatisticalpracticeIhaveidentified,aswellassomeoftheimprovementswehavemadesofarworkingtogetherinthispartnership.
Monday27th11:40Gunnamatta
RobustSemiparametricInferenceInRandomEffectsModels
MichaelStewart1andAlanWelsh21UniversityofSydney
2ANU
Wereportonrecentworkusingsemiparametrictheorytoderiveprocedureswithdesirablerobustnessandefficiencypropertiesinthecontextofinferenceconcerningscaleparametersforrandomeffectmodels.
Monday27th12:00Gunnamatta
RobustPenalizedLogisticRegressionThroughMaximumTrimmedLikelihoodEstimator
HongweiSun,YuehuaCui,andTongWangShanxiMedicalUniversity
Penalizedlogisticregressionisusedtoidentifygeneticmarkersformanyhigh-dimensionaldatasetssuchasingeneexpression,GWAS,DNAmethylationstudiesandsoon.Butoutlierssometimesoccurduetomisseddiagnosisormisdiagnosisofsubjects,heterogeneityofsamples,technicalproblemsinexperimentsorotherproblems.Theycangreatlyinfluencetheestimationofpenalizedlogisticregression.Fewstudiesfocusontherobustnessofpenalizedmethodswhentheresponsevariableiscategorical,whichisstandardinmedicalresearch.ThisstudyproposedarobustLASSO-typepenalizedlogisticregressionbasedonmaximumtrimmedlikelihood(MTL-LASSO).Thedefinitionofbreakdownpoint(BDP)forpenalizedlogisticregressionwasgivenanditspropertyfortheproposedmethodwasproved.AmodificationofFAST-LTSalgorithmswasusedtoimplementtheestimation.Thereweightedstepwas
addedtoimproveperformancewhileguaranteeingrobustness.Thesimulationstudyshowstheproposedmethodcanresistagainstoutliers.Arealdatasetaboutgeneexpressionprofilesofmultiplesclerosispatientsandhealthycontrolswasanalysed.OutliersinthecontrolgroupidentifiedbyreweightedMTL-LASSObehavedifferentlyfromothers.Itunveilstheremaybeheterogeneityproblemincontrolgroup.Amuchbetterfitisobtainedafterremovingoutliers.
Keynote:Monday27th13:30Mantra
AMulti-StepClassifierAddressingCohortHeterogeneityImprovesPerformanceOfPrognosticBiomarkersInComplexDisease
JeanYangUniversityofSydney
Recentstudiesincancerandothercomplexdiseasescontinuetohighlighttheextensivegeneticdiversitybetweenandwithincohorts.Thisintrinsicheterogeneityposesoneofthecentralchallengestopredictingpatientclinicaloutcomeandthepersonalizationoftreatments.Here,wewilldiscusstheconceptofclassifiabilityobservedinmulti-omicsstudieswhereindividualpatients’samplesmaybeconsideredaseitherhardoreasytoclassifybydifferentplatforms,reflectedinmoderateerrorrateswithlargeranges.Wedemonstrateinacohortof45AJCCstageIIImelanomapatientsthatclinico-pathologicbiomarkerscanidentifythosepatientsthataremostlikelytobemisclassifiedbyamolecularbiomarker.Theprocessofmodellingtheclassifiabilityofpatientswasthenreplicatedinindependentdatafromotherdiseases.
Amulti-stepprocedureincorporatingthisinformationnotonlyimprovedclassificationaccuracyoverallbutalsoindicatedthespecificclinicalattributesthathadmadeclassificationproblematicineachcohort.Instatisticalterms,ourstrategymodelscohortheterogeneityviatheidentificationofinteractioneffectsinahighdimensionalsetting.Atthetranslationallevel,thesefindingsshowthatevenwhencohortsareofmoderatesize,includingfeaturesthatexplainthepatient-specificperformanceofaprognosticbiomarkerinaclassificationframeworkcansignificantlyimprovethemodellingandestimationofsurvival,
aswellasincreaseunderstanding.
Monday27th14:20Narrabeen
HowToAnalyseFiveDataPointsWithFun
PaulineO'Shaughnessy1,StephenRobson2,andLouiseRawlings21UniversityofWollongong
2ANU
While“bigdata”isoneofthebiggestbuzzwords,weoccasionallycomecrossdatawithveryfewdatapoints.ThisdatasetisfromastudyoftheincidenceratesofcommonlyperformedmedicalproceduresInAustralia,whichisavailableonlyatthestatelevel.WehavefivestatesinAustraliathusfivedatapoints.Sowhatcanwedowhenweonlyhavefivedatapoints?Oneofthenoveltyapproachesistofitthedatawitharegressionmodel.Howevergiventhechallengingnatureofthedatawithsmallsize,noguaranteecanbeplacedonthesatisfactoryofthelinearityandhomoscedasticityassumptionofthelinearregression,inturns,theinferencefromthestandardlinearmodeltheoryisnolongervalid.Doublebootstrapisusedtoprovidesolutiontothevalidstatisticalinferenceforthebestlinearapproximationfortherelationshipbetweenvariablesinanassumption-leanregressionsetting.
Monday27th14:20Gunnamatta
AnApproachToPoissonMixedModelsFor-OmicsExpressionData
IreneSuilanZengandThomasLumleyUniversityofAuckland
Weareinterestedinregressionmodelsformultivariatedatafromhigh-throughputbiologicalassays('omic'data).Thesedatahavecorrelationsbetweenvariables,andmayalsocomefromstructuredexperiments,soageneralisedlinearmixedmodelisappropriatetofittheexperimentalvariablesanddifferenttypesofomicsdata.However,thenumberofvariablesisoftenlargerthanthe
numberofobservations:astructuredcovariancemodelisnecessaryandsparsityinductionisbiologicallyappropriate.InthispresentationwedescribeanapproachtoPoissonmixedmodels,suitableforRNAseqgeneexpressiondata,basedontranscript-specificrandomeffectswithasparseprecisionmatrix.Weshowbysimulationsthattheoptimalsparsenesspenaltyforregressionmodellingisnotthesameasintheusualgraphestimationproblemandcomparesomeestimationstrategiesinsimulations.
Monday27th14:40Narrabeen
BayesianSpatialEstimationWhenAreasAreFew
AswiAswi,SusannaCramb,EarlDuncan,andKerrieMengersenQueenslandUniversityofTechnlogy
Spatialmodellingwhentherearefew(<20)smallareascanbechallenging.Bayesianmethodscanbebeneficialinthissituationduetotheeaseofspecifyingstructureandadditionalinformationthroughpriors.However,careisneededasthereareoftenfewerneighboursandmoreedges,whichmayinfluenceresults.HereweinvestigateBayesianspatialmodelspecificationwhentherearefewareas,firstthroughasimulationstudy(numberofareasrangingfrom4to2500)andthenapplytoacasestudyondenguefeverin2015inMakassar,Indonesia(14areas).FourdifferentBayesianspatialmodelsnamely,anindependentmodeland3modelsbasedonaCAR(ConditionalAutoregressive)prior:theBesag,York&Mollié,Leroux,andalocalisedmodel(augmentstheCARpriorwithaclustermodelusingpiecewiseconstantintercepts)wereapplied.Dataweregeneratedforthesimulationstudyconsideringlowandhighspatialautocorrelationandlowandhighdiseaseincidence.ModelgoodnessoffitwascomparedusingDevianceInformationCriteria.AnalysisofvarianceandBonferroni'smethodwerealsousedtodeterminewhichmodelsweresignificantlydifferent.Thesimulationstudyshowedmodelsdifferedintheirperformancemainlyintwosituations:1.Whentherewereatleast25areasandboththediseaserateandspatialautocorrelationwaslow,and2.Forallareasizeswhentherewaslowspatialautocorrelationbutahighoveralldiseaserate.Likewise,resultsfromthecasestudyshowedthatallfourmodelsperformedsimilarly.Thisisprobablyduetothelownumberofareasandalowdisease
incidence.
Monday27th14:40Gunnamatta
Knowledge-GuidedGeneralizedBiclusteringAnalysisForIntegrative–OmicsAnalysis
ChanggeeChang1,YizeZhao2,MingyaoLi1,andQiLong11UniversityofPennsylvania
2CornellUniversity
Advancesintechnologyhaveenabledgenerationofmultipletypesof-omicsdatainmanybiomedicalandclinicalstudies,anditisdesirabletopoolsuchdatainordertoimprovethepowerofidentifyingimportantmolecularsignaturesandpatterns.However,suchintegrativeanalysespresentnewanalyticalandcomputationalchallenges.Toaddresssomeofthesechallenges,weproposeaBayesiansparsegeneralizedbiclusteringanalysis(GBC)whichenablesintegratingmultipleomicsmodalitieswithincorporationofbiologicalknowledgethroughtheuseofadaptivestructuredshrinkagepriors.Theproposedmethodscanaccommodatebothcontinuousanddiscretedata.MCMCandEMalgorithmsaredevelopedforestimation.Numericalstudiesareconductedtodemonstratethatourmethodsachieveimprovedfeatureselectionandpredictioninidentifyingdiseasesubtypesandlatentdrivers,comparedtoexistingmethods.
Monday27th15:00Narrabeen
UnderstandingTheVariationInHarvesterYieldMapDataForEstimatingCropTraits
DeanDiepeveen1,KarynReeves1,AdrianBaddeley1,andFionaEvans21CurtinUniversity
2MurdochUniversity
Ourrecentexploratoryresearchinvolvesextractingtangiblecroptraitsfromimagessuchasyield-maps.ResearchbyDiepeveenetal(2012)demonstrated
thatgeneticinformationcanbeextractedfromnear-infrared(NIR)imagesusingboththeimplicitknowledgeofdata,environmentaldataandusingamultivariateapproach.Yieldmapdataisgeo-referenceddataofgrain-yieldthatisgeneratedbyaharvestercuttingthecrop,threshingthestrawandextractingthegrain.Ourresultsshowthattheyield-mapdatahassignificantissuesassociatedwithit.Oneissueisthedelayfromtimeofcuttingthecroptoenteringintothestorage-binformeasurement.Thisisdependentonspeedoftheharvesterandiscompoundedwithmaintainingacriticalvolumegoingthroughtheharvestertooperateefficiency.Therearealsoissueswithvariationfromthedensityandplantsizeofthecropwithinthepaddockbeingharvested.Ourpreliminaryresultsjusthighlightthesignificantchallengesinextractingprecisecroptraitsfromyieldmapdata.
Monday27th15:00Gunnamatta
CitizenScienceToSurveillance:EstimatingReportingProbabilitiesOfExoticInsectPests
PeterCaley1,MarijkeWelvaert2,andSimonBarry11CSIRO
2UniversityofCanberra
Upuntilmid-2016,citizenscienceuploadstotheAtlasofLivingAustraliaincludedc.400bugspecies,andc.1,000beetlespecies.Giventheshorttimeperiod(c.3years)overwhichmostoftheserecordshaveaccumulated,thisrepresentsaconsiderablereportingeffort.Thekeyappliedquestionfromabiosecuritycontextishowthislevelofreportingtranslatestothedetectionandreportingofexoticinsectpestsintheeventofanincursion.
Weuseacase-controldesigntomodeltheprobabilityofexistinginsectspeciesbeingreportedviacitizensciencechannelsfeedingintotheAtlasofLivingAustralia.Theeffectofinsectfeatures(size,colour,pattern,morphology)andgeographicdistributiononreportingratesareexploredasexplanatoryvariables.Wethenapplythemodeltoexotichighprioritypestspeciestopredicttheirreportingratesintheeventoftheirintroduction.
Monday27th15:50Narrabeen
IntroductionTo“Deltagen”-AComprehensiveDecisionSupportToolForPlantBreedersUsingRAndShiny
DongwenLuoandZulfiJahuferAgResearch
Theobjectiveofthispresentationistointroduceauniquenewplantbreedingdecisionsupportsoftwaretool"DeltaGen",implementedinRanditspackageShiny.DeltaGenprovidesplantbreederswithasingleintegratedsolutionforexperimentaldesigngeneration,dataqualitycontrol,statisticalandquantitativegeneticanalyses,breedingstrategyevaluation/simulationandcostanalysis,patternanalysis,indexselectionandunderlyingbasictheoryonquantitativegenetics.Thissoftwaretoolcouldalsobeusedasateachingresourceinplantbreedingcourses.DeltaGenisavailableasFreewareonthelink:http://agrubuntu.cloudapp.net/shiny-apps/PlantBreedingTool/
Monday27th15:50Gunnamatta
EstimatingNitrousOxideEmissionFactors
AlasdairNobleandTonyVanDerWeerdenAgResearch
NitrousOxide(N2O)isanimportantgreenhousegaswithaglobalwarmingpotentialnearly300timesthatofcarbondioxide.UndertheKyotoProtocolNewZealandisrequiredtoreportagreenhousegasinventoryannuallywhichincludesN2O.InNewZealand,95%ofN2Oemissionsarederivedfromnitrogen(N)inputstoagriculturalsoils(e.g.animalexcretaandfertiliser).FieldexperimentsareconductedtoestimatetheseN2Oemissions,wheredataiscollatedandanalysedfollowingastandardmethodologytodetermineemissionfactors,whichestimatetheamountofN2OlostperunitofNappliedtosoil.Howeverforindividualdatasetstherearesomeaspectsofthedataareincompatiblewiththeproposedmodelsosomeadhocadjustmentsaremade.AmorerigorousBayesianapproachisproposedandsomeresultswillbediscussed.
ProgrammeAndAbstractsForTuesday28thOfNovember
Keynote:Tuesday28th9:00Mantra
ClusterCapture-Recapture:ANewFrameworkForEstimatingPopulationSize
RachelFewsterUniversityofAuckland
Askanywildlifemanager:theirfirstburningquestionis"Howmanyarethere?",andtheirsecondis"Aretheytrendingupwardsordownwards?"Capture-recaptureisoneofthemostpopularmethodsforestimatingpopulationsizeandtrends.Asthenamesuggests,itreliesonbeingabletoidentifythesameanimaluponmultiplecaptureoccasions.Thepatternofcapturesandrecapturesamongidentifiedanimalsisusedtoestimatethenumberofanimalsnevercaptured.
Physicallycapturingandtagginganimalscanbeadangerousandstressfulexperienceforboththeanimalsandtheirhumaninvestigators-orifittranspiresthattheanimalsactuallyenjoyit,biasedinferencemayresult.Consequently,researchersincreasinglyfavournon-invasivesamplingusingnaturaltagsthatallowanimalstobeidentifiedbyfeaturessuchascoatmarkings,droppedDNAsamples,acousticprofiles,orspatiallocations.Theseinnovationsgreatlybroadenthescopeofcapture-recaptureestimationandthenumberofcapturesamplesachievable.However,theyareimperfectmeasuresofidentity,effectivelysacrificingsamplequalityforquantityandaccessibility.Asaresult,capture-recapturesamplesnolongergeneratecapturehistoriesinwhichthematchingofrepeatedsamplestoasingleidentityiscertain.Instead,theygeneratedatathatareinformative---butnotdefinitive---aboutanimalidentity.
Iwilldescribeanewframeworkfordrawinginferencefromcapture-recapturestudieswhenthereisuncertaintyinanimalidentity.Intheclustercapture-
recaptureframework,weassumethatrepeatedsamplesfromthesameanimalwillbesimilar,butnotnecessarilyidentical,toeachother.Overlapisalsopossiblebetweenclustersofsamplesgeneratedbydifferentanimals.Wetreatthesampledataasaclusteredpointprocess,andderivethenecessaryprobabilisticpropertiesoftheprocesstoestimateabundanceandotherparametersusingaPalmlikelihoodapproach.
Becauseitavoidsanyattemptsatexplicitsample-matching,theclustercapture-recapturemethodcanbeveryfast,takingmuchthesametimetoanalysemillionsofsample-comparisonsasitdoestoanalysehundreds.Iwilldescribeapreliminaryframeworkforabundanceestimationfromacousticmonitoring.Clustercapture-recapturecanalsobeusedforbehaviouralstudies,andIwillshowanexampleusingcamera-trapdatafromapartially-markedpopulationofforestshiprats.
Tuesday28th10:30Narrabeen
PropensityScoreApproachesInThePresenceOfMissingData:ComparisonOfBalanceAndTreatmentEffectEstimates
JannahBaker1,TimWatkins1,2,andLaurentBillot1,31TheGeorgeInsituteforGlobalHealth
2UniversityofSydney3UniversityofNewSouthWales
Theuseofpropensityscoremethodscanpotentiallyimprovebalancebetweengroupsinobservationalstudydata,thusminimisingconfounding.However,afrequentproblemwithsuchstudiesisthepresenceofmissingdata.Wecomparethreeapproachestogeneratingthepropensityscoreaccountingformissingdatawithinthecontextofaclinicalcasestudyexaminingtheeffectoftelemonitoringonhypertension.Overall,4,642patientsdiagnosedwithhypertensionreceivingonlinehealthsupportinMyHealthGuardian–atelephonechronicdiseasesupportprogramofferedbyprivatehealthinsurerHCF–wereofferedatelemonitoringintervention.Ofthese,2,729acceptedandstartedtreatmentbetweenJuly2014toApril2015(designated“cases”),and1,913declined
(designated“controls”).Datawereavailablefromcasesandcontrolsonseveralbaselinevariablesincludingdemographic,lifestyleandclinicalcharacteristics.Outcomeswerethenumberofhospitalisations,totallengthofstayandtotalcostofhospitalisationbetween1Januaryto31December2016.Propensityscoremethodswereusedtobalancebaselinevariablesbetweengroups.Threeapproacheswereusedtogeneratepropensityscoresfromalogisticregressionmodelaccountingformissingdata:1)categorisationofallvariableswith“Missing”asacategory,2)multipleimputationoftreatmenteffect(MIte)wheretreatmenteffectestimatesarecombinedover20imputeddatasets,and3)multipleimputationofpropensityscore(MIps)wherepropensityscoresarefirstaveragedoverimputeddatasetspriortoestimationoftreatmenteffects.Thepropensityscorefromeachapproachwasthenusedintwoways:a)matchingcasestocontrols,andb)inverseprobabilityoftreatmentweighting(IPTW).Thebalanceachievedbyeachapproachwascomparedusingstandardiseddifferencesofmeansandproportionsinbaselinecharacteristicsbetweengroups.Thetreatmenteffectestimatesfromeachapproachwerealsocompared.Thediscussionwillcanvasourfindingsandrecommendationsforhandlingmissingdatawhenusingpropensityscoreapproaches.
Tuesday28th10:30Gunnamatta
VisualisingModelSelectionStabilityInHigh-DimensionalRegressionModels
GarthTarrUniversityofSydney
ThemplotRpackageprovidesaprovidesanimplementationofmodelstabilityandvariableinclusionplotsforresearcherstousetobetterinformthevariableselectionprocess.Theinitialfocuswasonexhaustivesearchesthroughthemodelspace,however,thisquicklybecomesinfeasibleforhighdimensionalmodels.Analternativeapproachforhighdimensionalmodelsistocombinebootstrapmodelselectionwithregularisationprocedures.Thereexistanumberoffastandefficientmethodregularisationmethodsforvariableselectioninhighdimensionalregressionsettings.Wehaveimplementedvariableinclusionplotsandmodelstabilityplotsusingtheglmnetpackage.Wedemonstratetheutilityofthemplotpackageinidentifyingstableregularisedmodelselectionchoices
withrespecttotwomainsourcesofuncertainty.Firstly,byresamplingthedataweareabletodeterminehowoftenvariousmodelsarechosenwhenthedatachanges.Secondly,weareabletoevaluatehowoftencompetingmodelsarechosenacrossarangeofvaluesforthetuningparameter.Exploringthesetwosourcesofuncertaintyinmodelselectiongeneratesalargeamountofrawdatathatneedstobeprocessed.Themplotpackageprovidesavarietyofmethodstovisualisethisrawdatatohelpinformaresearcher'smodelselectionchoice.
Tuesday28th10:50Narrabeen
TheMissingLink:AnEquivalenceResultForLikelihoodBasedMethodsInMissingDataProblemsFirouzehNoghrehchi,JakubStoklosa,SpiridonPenev,andDavidI.Warton
UniversityofNewSouthWales
Multipleimputationandmaximumlikelihoodestimation(viatheexpectation-maximizationalgorithm)aretwowell-knownmethodsreadilyusedforanalysingdatawithmissingvalues.Thesetwomethodsareoftenconsideredasbeingdistinctfromoneanother,duetotheirconstructionforestimationandtheirtheoreticalproperties.Weshowthatthereisacloserelationshipbetweenthetwomethods.Specifically,weshowthatatypeofmultipleimputationcanbeunderstoodasastochasticexpectation-maximisationapproximationtomaximumlikelihood.Asaresult,wecanexploretheapplicationofarangeoflikelihood-basedtoolsinthemultipleimputationcontextinordertoimproveitsperformance.Inparticular,wedevelopinformationcriteriaforselectinganimputationmodelgivenasetofcompetingmodels,andaflexiblelikelihoodratiotestwhenmodelsarefittedbymultipleimputation.Wedemonstrateourmethodsonrealandsimulateddatasets.
Tuesday28th10:50Gunnamatta
DimensionalityReductionOfLIBSDataForBayesianAnalysis
AnjaliGupta1,JamesCurran1,SallyCoulson2,andChristopherTriggs11UniversityofAuckland
2ESR
In2004,AitkenandLucypublishedanarticledetailingatwo-levellikelihoodratioformultivariatetraceevidence.Thismodelhasbeenadoptedinanumberofforensicdisciplinessuchastheinterpretationofglass,drugs(MDMA),andink.Moderninstrumentationiscapableofmeasuringmanyelementsinverylowquantitiesand,notsurprisingly,forensicscientistswishtoexploitthepotentialofthisextrainformationtoincreasetheweightofthisevidence.Theissue,fromastatisticalpointofview,isthattheincreaseinthenumberofvariables(dimension)intheproblemleadstoincreaseddatademandtounderstandboththevariabilitywithinasource,andinbetweensources.Suchinformationwillcomeintime,butusuallywedon’thaveenough.Onesolutiontothisproblemistoattempttoreducethedimensionalitythroughmethodssuchasprincipalcomponentanalysis.Thispracticeisquitecommoninhighdimensionalmachinelearningproblems.Inthistalk,IwilldescribeastudywhereweattempttoquantifytheeffectsofthisthisapproachontheresultinglikelihoodratiosusingdataobtainedfromaLaserInducedBreakdownSpectroscopy(LIBS)instrument.
Tuesday28th11:10Narrabeen
AnalysisOfMelanomaDataWithAMixtureOfSurvivalModelsUtilisingMulti-ClassDLDAToInformMixtureClass
SarahRomanes,JohnOrmerod,andJeanYangUniversityofSydney
MelanomaisaprevalentskincancerinAustralia,withcloseto14000newcasesestimatedtobediagnosedin2017.Survivaltimesaremarkedlydifferentfromoneindividualtothenext.Inparticular,thereappearstobethreeclassesofsurvivaloutcome.Thistalkconsidersintegratingsurvivaltimedatawithmicroarraygeneexpressiondata.Weconstructahybridmodelthatseamlesslyintegratesathree-classlineardiscriminantanalysismodel,mixtureofparametric
survivalmodels,andmodelselectioncomponents.Wefitthismodelusingavariationalexpectationmaximization(VEM)approach.Ourmodelselectioncomponentnaturallysimplifiesasafunctionoflikelihoodratiostatisticsallowingnaturalcomparisonswithtraditionalhypothesistestingmethods.Wecompareourmethodwithseveralnaïveapproacheswhichonlyaddressestheclassificationaspectorsurvivalmodelaspectinisolation.
Tuesday28th11:10Gunnamatta
ForecastingHotspotsOfPotentiallyPreventableHospitalisationsWithSpatiallyAggregatedLongitudinalHealthData:AllSubsetModelSelectionWithANovelImplementationOfRepeatedK-FoldCross-Validation
MatthewTuson,BerwinTurlach,KevinMurray,MeiRuuKok,AlistairVickery,andDavidWhyatt
UniversityofWesternAustralia
Itissometimesdifficulttotargetindividualsforhealthinterventionduetolimitedinformationontheirbehaviourandriskfactors.Insuchcasesplace-basedinterventionstargetinggeographical‘hotspots’withhigherthanaverageratesofhealthserviceutilisationmaybeeffective.Manystudiesexistexaminingpredictorsofhotspots,butoftendonotconsiderthatplace-basedinterventionsaretypicallycostlyandtaketimetodevelopandimplement,andhotspotsoftenregresstothemeanintheshort-term.Long-termgeographicalforecastingofhotspotsusingvalidatedstatisticalmodelsisessentialineffectivelyprioritisingplace-basedhealthinterventions.
Existingmethodsforecastinghotspotstendtoprioritisepositivepredictedvalue(i.e.correctpredictions)attheexpenseofsensitivity.Thisworkintroducesmethodstodevelopmodelsoptimisingbothpositivepredictedvalueandsensitivityconcurrently.Thesemethodsutilisespatiallyaggregated
administrativehealthdata,WAcensuspopulationdata,andABSgeographicboundaries,combiningallsubsetmodelselectionwithanovelimplementationofrepeatedcross-validationforlongitudinaldata.Resultsfrommodelsforecasting3-yearhotspotsforfourpotentiallypreventablehospitalisationsarepresented,namely:typeIIdiabetesmellitus,heartfailure,highriskfoot,andchronicobstructivepulmonarydisease(COPD).
Tuesday28th11:30Narrabeen
IdentifyingClustersOfPatientsWithDiabetesUsingAMarkovBirth-DeathProcess
MugdhaManda,ThomasLumley,andSusanWellsUniversityofAuckland
Estimatingdiseasetrajectorieshasincreasinglybecomemoreessentialtoclinicalpractitionerstoadministereffectivetreatmenttotheirpatients.Apartofdescribingdiseasetrajectoriesinvolvestakingpatients'medicalhistoriesandsociodemographicfactorsintoaccountandgroupingthemintosimilargroups,orclusters.Advancesincomputerisedpatientdatabaseshavepavedawayforidentifyingsuchtrajectoriesinpatientsbyrecordingapatient'smedicalhistoryoveralongperiodoftime(longitudinaldata):westudieddatafromthePREDICT-CVDdataset,anationalprimary-carecohortfromwhichpeoplewithdiabetesfrom2002-2015wereidentifiedthroughroutineclinicalpractice.WefittedaBayesianhierarchicallinearmodelwithlatentclusterstotherepeatedmeasurementsofHbA1candeGFR,usingtheMarkovbirth-deathprocessproposedbyStephens(2000)tohandlethechangesindimensionalityasclusterswereaddedorremoved.
Tuesday28th11:30Gunnamatta
ChallengesAnalysingCombinedAgriculturalFieldTrialsWithPartiallyOverlapping
Treatments
KerryBellandMichaelMumfordQueenslandDepartmentofAgricultureandFisheries
Tomakerecommendationsonwhichmanagementpracticeshavethepotentialtoincreasecropyieldthereneedstobeaconsistentpatterndemonstratedacrosstrialsfrommanyenvironments.Thispresentationconsidersacasestudylookingat31mungbeantrialsinnorthernAustralianfrom2014to2016.Thetrialsdidnotalwayshaveconsistentfactors(e.g.variety,rowspacingortargetplantdensity)orevenconsistentfactorlevels.Toovercometheissueofinconsistentfactors,environmentsweredefinedasthecombinationofsite,yearandanymanagementfactorsnotcommonacrosstrials(e.g.timeofsowing,irrigation,fertiliser).
Therewerenumerousfullfactorialcombinationswithinsubsetsofthedatathatcouldbeconsideredforinvestigationsothefirstchallengewastodeterminewhichfactorialcombinationstofocusontobestaddresstheresearchquestionsandreportingrequirements.Oncethiswasdetermined,allthedatafromthetrialsthatcontributedtothefactorialwereincludedinacombinedanalysisusinglinearmixedmodels.Inthismodel,thefactorialofinterestwaspartitionedinthetestoffixedeffectswhileeachtrials’designparametersandresidualvarianceswereestimatedusingallthedatafromeachtrial.Anexampleoftheabovementionedfactorialcombinationsisenvironmentbyrowspacingforoneparticularvariety.Thenextchallengewasthatwithsomanyenvironmentstherewasusuallyanenvironmentbyrowspacinginteractionwhichwasnotusefulformakingrecommendationsaboutrowspacing.
Clusteringofenvironmentsallowedforminggroupsthatdidnothaveasignificantinteractionbetweenrowspacingandenvironment.Thesegroupswerethengeneralisedtotypesofenvironmentswithcertainresponsestorowspacing.
Tuesday28th11:50Narrabeen
AHiddenMarkovModelForSleepStageDetectionUsingRawTri-AxialWrist
Actigraphy
MichelleTrevenen1,KevinMurray1,BerwinTurlach1,LeonStraker2,andPeterEastwood1
1UniversityofWesternAustralia2CurtinUniversity
Sleepisacomplexyetorganisedprocessconsistingofregularcyclesofsleepstages.Thesestagesarerapid-eyemovementandnon-rapideyemovement(lightsleepandslow-wavesleep).Sleepstagingisofgreatimportanceinthephysiologicalworldassleepdisordersoccurinaround20%ofthepopulationandareassociatedwithamultitudeofserioushealthimplicationsandconsiderableeconomicburden.HiddenMarkovmodelshavebeensuccessfullyusedintheclassificationofindividualsleepstagesmeasuredbypolysomnography,whichisconsideredthe'goldstandard'inassessingsleep,however,itisintrusiveandcostly.
Actigraphyisincreasinglybeingconsideredasanon-intrusiveandcosteffectivealternativemethodtoobjectivelymeasuresleeppatterns.However,thereislimitedresearchontheabilityofactigraphytodetectindividualsleepstages,furthermore,thisresearchindicatesthatthesecurrentmethodologiesdonothavetheabilitytodoso.Currentactigraphicapproachestosleepdetectionusefiltereduni-axialdatameasuredwithalowsamplingrate,whereas,rawtri-axialdatameasuredathighsamplingratesisfrequentlyusedintheassessmentofday-timeactivities.
Usingsimultaneouslymeasuredactigraphyandpolysomnographydatafrom100healthyyoungadultsintheWesternAustralianPregnancyCohortStudywecreatedandvalidatedanalgorithmtodeterminesleepstagesutilisingraw,tri-axialaccelerationdatafromwristactigraphy.Tenfeaturevariableswerecreatedfromeach30-secondblockofdataand50subjectswereusedtotrainthehiddenMarkovmodelwiththefeaturevariablesusedasinputparameters.Theremaining50subjectswereusedtovalidatethetrainedhiddenMarkovmodelagainstpolysomnography.
Validationsuggestedthatourmodelisabletoclassifysleepstagesusingrawtri-axialactigraphydata.Theseresultsdemonstratethatactigraphy-basedhiddenMarkovmodelscanfeasiblybeusedforautomaticsleepstaging.
Tuesday28th11:50Gunnamatta
SparsePhenotypingDesignsForEarlyStageSelectionExperimentsInPlantBreedingPrograms
NicoleCocks,AlisonSmith,DavidButler,andBrianCullisUniversityofWollongong
Theearlystagesofcerealandpulsebreedingprogramstypicallyinvolveinexcessof500testlines.Thetestlinesarepromotedthroughaseriesoftrialsbasedontheirperformance(yield)andotherdesirabletraitssuchasheat/droughttolerance,diseaseresistance,etc.Itisthereforeimportanttoensurethedesign(andanalysis)ofthesetrialsareefficientinordertoappropriatelyandaccuratelyguidethebreedersthroughtheirselectiondecisions,untilonlyasmallnumberofelitelinesremain.
ThedesignofearlystagevarietytrialsinAustraliaprovidedthemotivationfordevelopinganewdesignstrategy.Thepreliminarystagesoftheseprogramshavelimitedseedsupply,whichlimitsthenumberoftrialsandreplicatesoftestlinesthatcanbesown.Traditionally,completelybalancedblockdesignsorgridplotdesignsweresownatasmallnumberofenvironmentsinordertoselectthehighestperforminglinesforpromotiontothelaterstagesoftheprogram.Givenourunderstandingofvariety(i.e.line)byenvironmentinteraction,thisapproachisnotasensibleoroptimaluseofthelimitedresourcesavailable.
Anewmethodtoallowforalargernumberofenvironmentstobesampledforsituationswhereseedsupplyislimitedandnumberoftestlinesislargewillbediscussed.Thisstrategywillbereferredtoassparsephenotyping,whichisdevelopedwithinthelinearmixedmodelframeworkasamodel-baseddesignapproachtogeneratingoptimaltrialdesignsforearlystageselectionexperiments.
Tuesday28th12:10Narrabeen
ComparisonsOfTwoLargeLong-Term
StudiesInAlzheimer'sDisease
CharleyBudgeonUniversityofWesternAustralia
TheincidenceofAlzheimer’sdisease(AD),theleadingcauseofdementia,ispredictedtoincreaseatleastthreefoldby2050.Curingthisdiseaseisaglobalpriority.Currently,twomajorstudiesareattemptingtogainfurtherunderstandingofthisdisease;theAlzheimer'sDiseaseNeuroimagingInitiative(ADNI)andtheAustralianImaging,BiomarkerandLifestyleStudy(AIBL).Wedescribethesetwocohortstoassesstheimpactofcombiningthemtoprovidealargercohortforanalyses.
Aninitialcomparisonoftheprotocolswascarriedoutandrecruitmentstrategieswereshowntobemarginallydifferentbetweenthestudies.Inclusioncriteriaspecifiedagesbetween55and90yearsinADNIand>65yearsinAIBL.Marginallydifferentspecificationsfordiseasestageclassificationsofhealthycontrols(HC),mildcognitivelyimpaired(MCI)andADindividualswereobserved,forexample,differentMini-MentalStateExam(MMSE)cut-offs.However,bothstudieshadADdiagnosissupportedbytheNINDS/ARDAcriteria.BaselinecharacteristicswerecomparedbetweenADNIandAIBLcohorts.Overall,AIBLhadmoreHCscomparedtoADNI(69%vs30%),butfewerMCIindividuals(12%vs50%).TheADNIcohorthadahigherlevelofeducationandgenerally,withinadiseaseclassification,therewereminimaldifferencesinbaselineage,sex,MMSE,andPreclinicalAlzheimerCognitiveComposite(PACC)scores.
LongitudinalanalysescomparedthechangeovertimeforthetwocohortsanddiseaseclassificationsforPACCandMMSE.TherewerenosignificantdifferencesincohortswithintheHCandMCIgroups,butwithintheADgroup,subjectsintheADNIcohorthadgenerallyhigherpredictedPACCandMMSEscoresovertimethanthoseinAIBL.
OurresultssuggestthereisthepotentialtocombinetheADNIandAIBLcohortsforanalysispurposestoprovideonemorepowerfuldataset;however,considerationshouldbetakenforsomemeasures.
Tuesday28th12:10Gunnamatta
AOne-StageMixedModelAnalysisOfCanolaChemistryTrials
DanielTolhurst,KyMathews,AlisonSmith,andBrianCullisUniversityofWollongong
TheNationalVarietyTrials(NVT)programisusedbyplantbreedingcompaniestoevaluatetheyieldpotentialofnewcropvarietiesindependentlyacrossalargerangeofAustraliangrowingconditions.BycomparisonwiththeremainingNVTcrops,growerdecisionsarefurthercomplicatedincanolabecauseofitsvulnerabilitytotheinfestationofweeds.Ameasurehistoricallyusedbyfarmersforthemanagementofweedsistheapplicationofaherbicide(chemistry)treatment.Thechoiceofchemistryisimportantasitrestrictsvarietyselectiontothosebredwiththespecifictolerance.ThesetofvarietiescurrentlyevaluatedinNVTaretoleranttooneofthreechemistries,namelyimidazolinone(I;butmarketedasClearfield),glyphosate(RoundupReady;R)ortriazine(T),orhavenospecifictolerance(i.e.conventionalcanola;C).Consequently,canolahasamorecomplextestingregimethantheremainingNVTcropsaseachtrialhasanestedtreatmentstructureinvolvingbothchemistriesandvarieties.
CanolatrialsareconductedinlocationsacrosstheAustraliangrainbeltandreflectbestfarmerpracticeforeachdistrict.Everysiteispartitionedintoseveralfieldblocksandplotsareallocatedtothetreatmentsaccordingtoorthogonalblockdesigns.Asprayboomisusedtoadministereachchemistrybutispragmaticinthesensethatlargeareasaretreatedsimultaneously.Thisprecludestheapplicationofdifferentspraystoplotsinthesameblock.Randomisationisthereforerestrictedsovarietiesinasingleblockaretoleranttothesamechemistry.However,asthenumberofchemistriesandblocksareexactlyequal,thereisnoinformationtoestimatetheexperimentalerrorvariationandbotharestatisticallyconfounded.Consequently,growersarelimitedtoevaluatingvarietieswiththesametoleranceascomparisonsacrosschemistriesareinvalid.Thisalsohasimportantimplicationsonthestatisticalanalysis,whichisdiscussedinthistalk.
Tuesday28th12:30Narrabeen
ASemi-ParametricLinearMixedModelsForLongitudinallyMeasuredFastingBloodSugarLevelOfAdultDiabeticPatients
TafereTilahun1,BelayBirlie1,andLegesseKassaDebusho21JimmaUniversity
2UniversityofSouthAfrica
Thispaperfocusedonlongitudinaldataanalysisoffastingbloodsugar(FBS)levelofadultdiabeticpatientsatJimmaUniversitySpecializedHospitaldiabeticclinicusinganapplicationofsemi-parametricmixedmodel.ThestudyrevealedthattherateofchangeinFBSlevelindiabeticpatients,duetotheclinicinterventions,doesnotcontinueasasteadypacebutchangeswithtimeandweightofpatients.Furthermore,itclarifiedassociationsbetweenFBSlevelandsomecharacteristicsofadultdiabeticpatientsthatweightofadiabetespatienthasasignificantnegativeeffectwhereaspatientgender,age,typeofdiabetesandfamilyhistoryofdiabetesdidnothaveasignificanteffectonthechangeofFBSlevel.Undervariousvariancestructuresofsubject-specificrandomeffects,thesemi-parametricmixedmodelshadbetterfitthanlinearmixedmodel.Thiswaslikelyduetothelocalizedsplines,whichcapturedmorevariabilityinFBSlevelthanthelinearmixedmodel.
Tuesday28th12:30Gunnamatta
IndividualAndJointAnalysesOfSugarcaneExperimentsToSelectTestLines
AlessandraDosSantos1,ChrisBrien2,4,ClariceG.B.Demétrio1,RenataAlcardeSermarini1,5,GuilhermeA.P.Silva3,andSandroR.Fuzatto3
1UniversityofSãoPaulo2UniversityofSouthAustralia
3CTC-Piracicaba4UniverstiyofAdelaide5UniversityofAdelaide
Intheearlystagesofabreedingprogrammanyfieldtrialsareconducted,consideringseveralsoilandweatherconditions.Inthecaseofsugarcane,theseexperimentsareinstalledusingalargenumberoftestlines,butlimitationsinthefieldandoftheamountofgeneticmaterialdonotallowthereplicationofmany.Thisworkevaluated21trialsfromdifferentregions,fromaBraziliansugarcanebreedingprogram.Eachoftheseexperimentsoccupiedarectangulararrayofaround20rowsby25columnsinmostinstance,theplotswere12mlonger,double-furrowswith0.9mbetweenfurrowswithintheplotand1.5mspacingbetweendifferentplotsand1mbetweencolumns.Allthetrialshadatleast79%oftheareaplantedwithunreplicatedtestlines,theatmost21%oftheplotswereoccupiedbyfourcommercialvarieties,check.Aspecialcheck,interspersedalongdiagonals,wasplantedsystematicallyonadiagonalgrid,theotherthreewereequallyreplicatedandeachreplicatewasspreadoutinthreeneighbouringrowplots.Sevenofthe21experimentshadnosignificantdirectgeneticeffects,11presentedsignificantcompetitionattheresiduallevelandonlyonehadsignificantcompetitionatthegeneticlevel.Thecorrelationbetweentheselectedtestlinesforthedifferentexperimentsinasameregionwaslessthan0.54.However,thegeneticcorrelationwassignificantinthejointanalysesandstrongerthanthatfromtheindividualanalyses.Twosimulationstudieswereperformed:thefirstinvestigatedtheanalysisforasingleexperimentandtheresultsshowthatitisdifficulttofitamodelwhenthereisgeneticcompetitionwithorwithoutresidualcompetition.Thesamedifficultywasobservedinthesecondstudy,whichcomparedtheresultsfromindividualandjointanalyses.Itshowedthat,evenforthejointanalyses,onlyaround45to55%ofthetruebesttestlineswereselected.
ProgrammeAndAbstractsForWednesday29thOfNovember
Keynote:Wednesday29th9:00Mantra
StatisticsOnStreetCorners
DianneCookMonashUniversity
Perceptualresearchisoftenconductedonthestreet,withconveniencesamplingofpedestrianswhohappentobepassingby.Itisthroughexperimentsconductedusingpasser-bysthatwehavelearnedabouttheeffectofchange-blindness(https://www.youtube.com/watch?v=FWSxSQsspiQ)isinplayoutsidethelaboratory.
Indatascience,plotsofdatabecomeimportanttoolsforobservingpatterns,makingdecisions,andcommunicatingfindings.Butplotsofdatacanbevieweddifferentlybydifferentobservers,andoftenprovokeskepticismaboutwhetherwhatyousee"isreallythere?"Withtheavailabilityoftechnologythatharnessesstatisticalrandomisationtechniquesandinputfromcrowdswecanprovideobjectiveevaluationofstructurereadfromplotsofdata.
Thistalkdescribesaninferentialframeworkfordatavisualisation,andtheprotocolsthatcanbeusedtoprovideestimatesofp-values,andpower.Iwilldiscusstheexperimentsthatwehaveconductedthat(1)showthatthecrowd-sourcingdoesprovideresultssimilartostatisticalhypothesistesting,(2)howthiscanbeusedtoimproveplotdesign,(3)p-valuesinsituationswherenoclassicaltestsexist.Examplesfromecologyandagriculturewillbeshown.
JointworkwithHeikeHofmann,AndreasBuja,DeborahSwayne,HadleyWickham,Eun-kyungLee,MahbubulMajumder,NiladriRoyChowdhury,LendieFollett,SusanVanderplas,AdamLoy,YifanZhao,NathanielTomasetti
Wednesday29th10:30Narrabeen
EstimatingOverdispersionInSparseMultinomialData
FarzanaAfrozUniversityofOtago
Thephenomenonofoverdispersionariseswhenthedataaremorevariablethanweexpectfromthefittedmodel.ThisissueoftenariseswhenfittingaPoissonorabinomialmodel.Whenoverdispersionispresent,ignoringitmayleadtomisleadingconclusions,withstandarderrorsbeingunderestimatedandoverly-complexmodelsbeingselected.Inourresearchweconsideredoverdispersedmultinomialdata,whichcanariseinmanyresearchareas.Twoapproachescanbeusedtoanalyseoverdispersedmultinomialdata;theuseofthequasilikelihoodmethodorexplicitmodellingoftheoverdispersionusing,forexample,aDirichlet-multinomialorfinite-mixturedistribution.Useofquasilikelihoodhastheadvantageofonlyrequiringspecificationofthefirsttwomomentsoftheresponsevariable.Forsparsedata,suchasinacontingencytablewithmanylowexpectedcounts,useofquasilikelihoodtoestimatetheamountofoverdispersionwillbeparticularlyuseful,asitmaybedifficulttoobtainreliableestimatesoftheparametersinaDirichlet-multinomialorfinite-mixturemodel.Iconsiderfourestimatorsoftheamountofoverdispersioninsparsemultinomialdata,discusstheirtheoreticalpropertiesandprovidesimulationresultsshowingtheirperformanceintermsofbias,varianceandmeansquarederror.
Wednesday29th10:30Gunnamatta
AssessingMudCrabMeatFullnessUsingNon-InvasiveTechnologies
CaroleWright,SteveGrauf,BrettWedding,PaulExley,JohnMayze,andSuePoole
QueenslandDepartmentofAgricultureandFisheries
Thedecisionofwhetheramudcrabshouldberetainedatharvesthastraditionallybeenbasedonshellhardness.Thisismostcommonlyassessedbyusingthumbpressureappliedtothecarapaceofthemudcrab.Thecarapaceofarecentlymoultedmudcrabwillflexconsiderablyandisthereforereturnedtothewater.Thisassessmenthasalsobeenusedtodividemudcrabsintothreemeatfullnessgrades(A,BandC).ThehighermeatfullnessgradeAmudcrabsfetchagreaterpriceatmarketcomparedtothelowerBandCgrades.Thesubjectivenatureofthisassessmentwillalwaysresultindisputesattheboundariesofthegrades.Bydevelopingamoreobjectivescience-basedmethoddowngradesatthemarketwillbereducedwhileconsumersatisfactionandtheoverallindustryprofitabilitywillincrease.
Ascopingstudywasconductedthatevaluatedinnovativenon-invasivetechnologiestoassessmudcrabmeatfullnessbasedonpercentageyieldrecoveryofcookedmeatfromthedominantindividualmudcrabclaws.Thenon-invasivetechnologiesassessedincludednearinfraredspectroscopy(NIRS),candlingusingvisiblelight,andacousticvelocity.NIRSshowedthemostpotentialandwasreassessedinasecondstudywithslightimprovementstospectracapturemethodsandNIRlightsources.
94livemudcrabsfromtheMoretonBayareawereusedinthesecondstudy.Partialleastsquaresregression(PLS-R)wasperformedtobuildacalibrationmodeltopredictthepercentageyieldrecoveryofcookedmeatbasedonthespectraldata.ThePLS-Rhad and .
Aprincipalcomponentslineardiscriminantanalysis(PC-LDA)wasalsoconductedtodiscriminatebetweenthestandardthreegradesofmudcrabmeatfullness.Thiswascomparedtotheindustrystandardshellhardnessmethod.TheNIRSPC-LDAachievedaminimumof76%correctclassificationforeachofthethreegrades,comparedto24%fortheshellhardnessmethod.
Thenon-invasivetechnologiestrialledalongwiththeresultswillbediscussedinthistalk.
Wednesday29th10:30Bundeena
AnalysisOfMultivariateBinaryLongitudinalData:MetabolicSyndromeDuring
MenopausalTransition
GeoffJonesMasseyUniversity
Metabolicsyndrome(MetS)isamajormultifactorialconditionthatpredisposesadultstotype2diabetesandCVD.Itisdefinedashavingatleastthreeoffivecardiometabolicriskcomponents:1)highfastingtriglyceridelevel,2)lowhigh-densitylipoproteincholesterol,3)elevatedfastingplasmaglucose,4)largewaistcircumference(abdominalobesity),and5)hypertension.IntheUSStudyofWomen’sHealthAcrosstheNation(SWAN),a15-yearmulti-centreprospectivecohortstudyofwomenfromfiveracial/ethnicgroups,theincidenceofMetSincreasedasmidlifewomenunderwentthemenopausaltransition(MT).AmodelissoughttoexaminetheinterdependentprogressionofthefiveMetScomponentsandtheinfluenceofdemographiccovariates.
Wednesday29th10:50Narrabeen
StatisticalAnalysisOfCoastalAndOceanographicInfluencesOnTheQueenslandScallopFishery
Wen-HsiYang1,AnthonyJ.Courtney2,MichaelF.O’neill2,MatthewJ.Campbell2,GeorgeM.Leigh2,andJerzyA.Filar1
1UniverstiyofQueensland2QueenslandDepartmentofAgricultureandFisheries
Thesaucerscallop(Ylistrumballoti)otter-trawlfisheryusedtobethemostvaluablecommercially-fishedspeciesinQueenslandoceanwaters.Overthelastfewyears,therehasbeengrowingconcernamongfishers,fisherymanagersandscientistsoverthedeclineincatchratesandannualharvest.Aquantitativeassessmentconductedin2016showedthatscallopabundancewasatanhistoriclowlevel.Theassessmentuseddatasourcedfromthefisheryandindependentsurveys.Furtherinformationoncoastalandoceanographicinfluencesareavailableandmayrevealnewfactorsthatinfluencepopulationabundanceof
scallopsandimprovemanagementofthefishery.Inthisstudy,scallopcatchrateabundancedataandcoastalandphysicaloceanographicvariables(e.g.seasurfacetemperatureanomalies,coastalfreshwaterflowandChlorophyll-a)weremodelledtoidentifyspatialandtemporalenvironmentalprocessesimportantforconsiderationinfisherymanagementprocedures.
Wednesday29th10:50Gunnamatta
SavedByTheExperimentalDesign:TestingBycatchReductionAndTurtleExclusionDevicesInThePngPrawnTrawlFishery
EmmaLawrenceandBillVenablesCSIRO
Intrawlingforprawns,theprawncatchisoftenonlyasmallpartoftheresultsofanyonetrawl,withtheremaindercalled"bycatch".Reducingthebycatchcomponent,whilemaintainingtheprawncatch,isanimportantindustrygoal,primarilyforenvironmentalaccreditationpurposes,butalsoforeconomicreasons.
Wedesignedanat-seatrialfortheGulfofPapuaPrawnFishery,involvingfourvesselseachtowing"quadgear"(thatis,4separate,butlinkedtrawlnets)ineachtrawlshot,over18days.Theexperimentwasdesignedtoassesstheeffectivenessof27combinationsofTurtleExcluderDevices(TEDs)andBycatchReductionDevices,(BRDs),withacontrolnet,withoutanyattacheddeviceasoneofthenetsineachquad.AtBiometrics2015wediscussedhowweusedsimulatedannealingtogenerateahighlyefficientdesign,inseveralstages,tomeetthelargenumberofhighlyspecificlogisticalconstraints.
Thefocusofthistalkwillbetheanalysis,whichalsoprovedsomewhatchallenging.Wewillpresenttheresultsofouranalysisanddemonstratewhyputtingthetimeintothinkingaboutandgeneratinganon-standardexperimentaldesignallowedustoaccommodatethevariousglitchesandmisfortunesthatalwaysseemtohappenatsea.
Wednesday29th10:50Bundeena
RethinkingBiosecurityInspections:ACaseStudyOfTheAsianGypsyMoth(AGM)InAustralia
PetraKuhnert1,DeanPaini1,PaulMwebaze1,andJohnNielsen21CSIRO
2DepartmentofAgricultureandWaterResources
TheAsiangypsymoth(AGM)(Lymantriadisparasiatica)isaseriousbiosecurityrisktoAustralia’sforestryandhorticulturalindustries.WhilesimilarinappearancetotheEuropeangypsymoth(Lymantriadispardispar),theAsiangypsymothiscapableofflyingupto40kilometresandthereforehasthepotentialtoestablishandspreadinotherareaslikeAustralia.Inaddition,femalesareattractedtolightandwilloviposit(layeggs)indiscriminately.Asaresult,femalesareattractedtoshippingportsatnightandwillovipositonships.Theseshipsthereforehavethepotentialtospreadthismotharoundtheworld.
Thelife-cycleofthemothhasbeenwelldocumentedandisheavilydependentontemperature,witheggsundergoingthreephasesofdiapausebeforehatching.CurrentinspectionsofvesselsarrivingintoAustralianportsfromwhatisdeemedan“atrisk”portisalengthyandcostlyprocess.
ToassisttheDepartmentofAgriculturewiththeirprioritisationofships,wedevelopedanAGMToolintheformofanRShinyAppthat(1)showstheshortestmaritimepathfromanatriskporttoanAustralianportforavesselofinterestand(2)predictstheprobabilityofapotentialhatchandit’sreliabilityusingaclassificationtreemodelthatwasdevelopedtoemulatethelifecycleofthemothfromsimulateddata.Inthistalkwewilldiscussthemethodologythat(1)simulatestheAGMbiologyandpotentialhatchesofeggs,alongwithhowweextractedrelevanttemperaturedatathatwastheprimarydriverofthelifecycleforAGM,and(2)emulatesthissimulateddatausingastatisticalmodel,namelyaclassificationtreetopredicttheprobabilityofapotentialhatch.Wewillalsodiscussabootstrapapproachtoexplorethereliabilityofthepotentialhatchpredicted.
Wednesday29th11:10Narrabeen
SubtractiveStabilityMeasuresForImprovedVariableSelection
ConnorSmith1,SamuelMüller1,andBorisGuennewig1,21UniversityofSydney
2UniversityofNewSouthWales
ThistalkbuildsupontheInvisibleFence(Jiangetal.,2011)apromisingmodelselectionmethod.Utilizingacombinationofcoeffcient,scaleanddevianceestimatesweareabletoimprovethisresamplingbasedmodelselectionmethodforregressionmodels,bothlinearandgenerallinearmodels.Theintroductionofavariableinclusionplotallowsforavisualrepresentationforthestabilityofthemodelselectionmethodaswellasthevariablesbootstrappedrank.Thesuggestedmethodswillbeappliedtobothsimulatedandrealexampleswithcomparisonsaboutbothcomputationaltimeandeffectivenessmadetoselectionsthroughalternativeselectionprocedures.Wewillreportonourlatestresultsfromongoingworkinscalingupsubtractivestabilitymeasureswhenthenumbersoffeaturesislarge.
References:Jiang,J.,Nguyen,T.,&Rao,J.S.(2011).Invisiblefencemethodsandtheidentificationofdifferentiallyexpressedgenesets.StatisticsandItsInterface,4(3),403-415.
Wednesday29th11:10Bundeena
AComparisonOfMultipleImputationMethodsForMissingDataInLongitudinalStudies
MdHamidulHuque1,KatherineLee1,JulieSimpson2,andJohnCarlin11MurdochChildrensResearchInstitute
2UniversityofMelbourne
Multipleimputation(MI)forimputingmissingdataareincreasinglyusedin
longitudinalstudieswheredataaremissingduetonon-responseandlosttofollow-up.Standardmultivariatenormalimputation(MVNI)andfullyconditionalspecifications(FCS)aretheprincipleimputationframeworkavailableforimputingcross-sectionalmissingdata.Anumberofmethodshasbeensuggestedintheliteraturetoimputelongitudinaldataincluding(i)useofstandardFCSandMVNIwithrepeatedmeasurementsasseparatedistinctvariables(ii)useofimputationmethodsbasedongeneralizedlinearmixedmodels.NoclearevaluationoftherelativeperformanceofavailableMImethodsinthecontextoflongitudinaldata.Wepresentacomprehensivecomparisonofthealltheavailablemethodsforimputationlongitudinaldatainthecontextofestimatingcoefficientforbothlinearregressionmodelandlinearmixedeffectmodel.Wealsocomparedtheperformanceofthemethodstoimputebothbinaryandcontinuousdata.Atotalof10differentmethods(MVNI,JM-pan,JM-jomo,standardFCS,FCS-twofold,FCS-MTW,FCS-2lnorm,FCS-2lglm,FCS-2ljomoandFCS-Blimp)arecomparedintermsofbias,standarderrorandcoverageprobabilityoftheestimatedregressioncoefficients.Thesemethodsarecomparedusingasimulationstudybasedonapreviouslyconductedanalysisexploringtheassociationbetweentheburdenofoverweightandqualityoflife(QoL)usingdatafromtheLongitudinalStudyofAustralianChildren(LSAC).WefoundthatbothstandardFCSandMVNIprovidereliableestimatesandcoverageoftheregressionparameters.Amongothermethodslinearmixedmodelsbasedmethods,JM-jomoandFCS-Blimpapproachesholdgreatpromise.
Wednesday29th11:30Narrabeen
SpeciesDistributionModellingForCombinedDataSources
IanRenner1andOlivierGimenez21UniversityofNewcastle
2Centred’EcologieFonctionnelleetEvolutive
Increasingly,multiplesourcesofspeciesoccurrencedataareavailableforaparticularspecies,collectedthroughdifferentprotocols.Forsingle-sourcemodels,avarietyofmethodshavebeendeveloped:pointprocessmodelsforpresence-onlydata,logisticregressionforpresence-absencedataobtainedthroughsingle-visitsystematicsurveys,andoccupancymodellingfor
detection/non-detectiondataobtainedthroughrepeat-visitsurveys.Insituationsforwhichmultiplesourcesofdataareavailabletomodelaspecies,thesesourcesmaybecombinedviaajointlikelihoodexpression.Nonetheless,therearequestionsabouthowtointerprettheoutputfromsuchacombinedmodelandhowtodiagnosepotentialviolationsofmodelassumptionssuchastheassumptionofspatialindependenceamongpoints.
Inthispresentation,Iwillexplorequestionsofinterpretationoftheoutputfromthesecombinedapproaches,aswellasproposeextensionstocurrentpracticethroughtheintroductionofaLASSOpenalty,sourceweightstoaccountfordifferingqualityofdata,andmodelswhichaccountforspatialdependenceamongpoints.ThisapproachwillbedemonstratedbymodellingthedistributionoftheEurasianlynxineasternFrance.
Wednesday29th11:30Gunnamatta
AFactorAnalyticMixedModelApproachForTheAnalysisOfGenotypeByTreatmentByEnvironmentData
LaurenBorg,BrianCullis,andAlisonSmithUniversityofWollongong
Theaccurateevaluationofgenotypeperformanceforarangeoftraits,includingdiseaseresistance,isofgreatimportancetotheproductivityandsustainabilityofmajorAustraliancommercialcrops.Typically,thedatageneratedfromcropevaluationprogrammesarisefromaseriesoffieldtrialsknownasmulti-environmenttrials(METs),whichinvestigategenotypeperformanceoverarangeofenvironments.
Inevaluationtrialsfordiseaseresistance,itisnotuncommonforsomegenotypestobechemicallytreatedagainsttheafflictingdisease.AnimportantexampleinAustraliaistheassessmentofgenotypesforresistancetoblacklegdiseaseincanolacropswhereitiscommonpracticetotreatcanolaseedswithafungicide.Genotypesareeithergrownintrialsastreated,untreatedorasboth.
ThereareanumberofmethodsfortheanalysisofMETdata.Thesemethods,
however,donotspecificallyaddresstheanalysisofdatawithanunderlyingthree-waystructureofgenotypebytreatmentbyenvironment(GxTxE).Here,weproposeanextensionofthefactoranalyticmixedmodelapproachforMETdata,usingthecanolablacklegdataasthemotivatingexample.
Historicallyintheanalysisofblacklegdata,thefactorialgenotypebytreatmentstructureofthedatawasnotaccountedfor.Entries,whicharethecombinationsofgenotypesandfungicidetreatmentspresentintrials,wereregardedas`genotypes’andatwo-wayanalysisof‘genotypes’byenvironmentswasconducted.
Theanalysisofourexampleshowedthattheaccuracyofgenotypepredictions,andthenceinformationforgrowers,wassubstantiallyimprovedwiththeuseofthethree-wayGxTxEapproachcomparedwiththehistoricalapproach.
Wednesday29th11:30Bundeena
TheImpactOfCohortSubstanceUseUponLikelihoodOfTransitioningThroughStagesOfAlcoholAndCannabisUseAndUseDisorder:FindingsFromTheAustralianNationalSurveyOnMentalHealthAndWell-Being
LouisaDegenhardt1,MeyerGlantz2,ChriannaBharat1,AmyPeacock1,LuiseLago1,NancySampson3,andRonaldKessler31NationalDrugandAlcoholResearchCentre
2NationalInstituteonDrugAbuse3HarvardUniversity
Theaimsofthepresentstudyweretousepopulation-levelAustraliandatatoestimateprevalenceandspeedoftransitionsacrossstagesofalcoholandcannabisuse,abuseanddependence,andremissionfromdisorder,andconsiderthepotentialimpactsthatanindividual’sageandsexcohort’slevelofsubstance
usepredictedtransitionsintoandoutofsubstanceuse.Dataonlifetimehistoryofuse,DSM-IVusedisorders,andremissionfromthesedisorderswerecollectedfromparticipants(n=8,463)inthe2007AustralianNationalSurveyofMentalHealthandWellbeingusingtheCompositeInternationalDiagnosticInterview.
Lifetimeprevalenceofalcoholuse,regularuse,abuse,dependence,andremissionfromabuseanddependencewere94.1%,64.5%,22.1%,4.0%,16.1%and2.1%,respectively.Unconditionallifetimeprevalenceofcannabisuse,abuse,dependence,andremissionfromabuseanddependencewere19.8%,6.1%,1.9%,4.0%and1.5%.Increasesintheestimatedproportionofpeopleintherespondent’ssexandagecohortwhousedalcohol/cannabisasofagivenageweresignificantlyassociatedwithmosttransitionsfromusethroughtoremissionbeginningatthesameage.Clearassociationsweredocumentedbetweencohort-levelprevalenceofsubstanceuseandpersonalriskofsubsequenttransitionsofindividualsinthecohortfromusetogreatersubstanceinvolvement.Thisrelationshipremainedsignificantoverandaboveassociationsinvolvingtheindividual’sageofinitiation.Thesefindingshaveimportantimplicationsforourunderstandingofthecausalpathwaysintoandoutofproblematicsubstanceuse.
Wednesday29th11:50Narrabeen
TheLASSOOnLatentIndicesForOrdinalPredictorsInRegression
FrancisHui1,SamuelMueller2,andAlanWelsh11ANU
2UniversityofSydney
Manyapplicationsofregressionmodelsinvolveordinalcategoricalpredictors.AmotivatingexampleweconsiderisordinalratingsfromindividualsrespondingtoquestionnairesregardingtheirworkplaceintheHouseholdIncomeandLabourDynamicsinAustralia(HILDA)survey,withtheaimbeingtostudyhowworkplaceconditions(mainandpossibleinteractioneffects)affecttheiroverallmentalwellbeing.Acommonapproachtohandlingordinalpredictorsistotreateachpredictorasafactorvariable.Thiscanleadtoaveryhigh-dimensionalproblem,andhasspurredmuchresearchintopenalizedlikelihoodmethodsfor
handlingcategoricalpredictorswhilerespectingthemarginalityprinciple.Ontheotherhand,giventheordinalratingsareoftenregardedasmanifestationsofsomelatentindicesconcerningdifferentaspectsofjobquality,thenamoresensibleapproachwouldbetofirstperformsomesortofdimensionreductionbeforeenteringthepredictedindicesintoaregressionmodel.Inappliedresearchthisisoftenperformedasatwo-stageprocedure,andindoingsofailstoutilizetheresponseinordertobetterpredictthelatentindicesthemselves.
Inthistalk,weproposetheLASSOonLatentIndices(LoLI)forhandlingordinalcategoricalpredictorsinregression.TheLoLImodelsimultaneouslyconstructsacontinuouslatentindexforeachorgroupsofordinalpredictorsandmodelstheresponseasafunctionofthese(andotherpredictorsifappropriate)includingpotentialinteractions,withacompositeLASSOtypepenaltyaddedtoperformselectiononmainandinteractioneffectsbetweenthelatentindices.Asasingle-stageapproach,theLoLImodelisabletoborrowstrengthfromtheresponsetoimproveconstructionofthecontinuouslatentindices,whichinturnproducesbetterestimationofthecorrespondingregressioncoefficients.Furthermore,becauseoftheconstructionoflatentindices,thedimensionalityoftheproblemissubstantiallyreducedbeforeanyvariableselectionisperformed.Forestimation,weproposefirstestimatingthecutoffsrelatingtheobservedordinalpredictorstothelatentindices.Thenconditionalonthesecutoffs,weapplyapenalizedExpectationMaximizationalgorithmviaimportancesamplingtoestimatetheregressioncoefficients.AsimulationstudydemonstratestheimprovedpoweroftheLoLImodelatdetectingtrulyimportantordinalpredictorscomparedtobothtwo-stageapproachesandusingfactorvariables,andbetterpredictiveandestimationperformancecomparedtothecommonlyusedtwo-stageapproach.
Wednesday29th11:50Gunnamatta
Whole-GenomeQTLAnalysisForNestedAssociationMappingPopulations
MariaValeriaPaccapelo1,AlisonKelly1,JackChristopher2,andArunasVerbyla3,4
1QueenslandDepartmentofAgricultureandFisheries2QueenslandAllianceforAgricultureandFood
3Data614CSIRO
Geneticdissectionofquantitativetraitsinplantshasbecomeanimportanttoolinbreedingofimprovedvarieties.ThemostcommonlyusedmethodstomapQTLarelinkageanalysisinbi-parentalpopulationsandassociationmappingindiversitypanels.However,bi-parentalpopulationsarerestrictedintermsofallelicdiversityandrecombinationevents.Despitethefactthatassociationmappingovercomestheselimitations,ithaslowpowertodetectrareallelesassociatedwithatraitofinterest.Multi-parentpopulationssuchasmulti-parentadvancedgenerationinter-cross(MAGIC)andnestedassociationmapping(NAM)populationshavebeendevelopedtocombinestrengthsofbothmappingapproaches,capturingmorerecombinationeventsandallelicdiversitythanbi-parentalpopulationsandinagreaterfrequencythanadiversitypanel.NestedassociationmappingusesmultipleRILfamiliesconnectedbyasinglecommonparent.Suchapopulationstructurepresentssomeadditionalchallengescomparedtotraditionalmapping,inparticularthepopulationdesignandthelargenumberofmolecularmarkersthatneedtobeintegratedsimultaneouslyintotheanalysis.WepresentamethodforQTLmappingforNAMpopulationsadaptedfrommulti-parentwholegenomeaverageintervalmapping(MPWGAIM)wheretheNAMdesignisincorporatedthroughtheprobabilityofinheritingfounderallelesforeverymarkeracrossthegenome.Thismethodisbasedonamixedlinearmodelinaone-stageanalysisofrawphenotypestogetherwithmarkers.Itsimultaneouslyscansthewhole-genomethroughaniterativeprocessleadingtoamulti-locusmodel.TheapproachwasappliedtoawheatNAMpopulationinordertoperformQTLmappingforplantheight.ThemethodwasdevelopedinR,withmaindependenciesbeingtheRpackagesMPWGAIMandasreml.ThisapproachestablishesthebasisforfurtherstudiesandextensionssuchasthecombinationofmultipleNAMpopulations.
Wednesday29th11:50Bundeena
AnAsymmetricMeasureOfPopulationDifferentiationBasedOnTheSaddlepointApproximationMethod
LouiseMcMillanandRachelFewsterUniversityofAuckland
Inthefieldofpopulationgeneticstherearemanymeasuresofgeneticdiversityandpopulationdifferentiation.ThebestknownisWright'sFst,laterexpandedbyCockerhamandWeir,whichisverywidelyusedasameasureofseparationbetweenpopulations.Morerecentlyamultitudeofothermeasureshavebeendeveloped,fromGsttoD,allwithdifferentfeaturesanddisadvantages.Onethingthesemeasuresallhaveincommonisthattheyaresymmetric,whichistosaythattheFstbetweenpopulationAandpopulationBisthesameasthatbetweenpopulationBandpopulationA.FollowingmyworkonGenePlot,avisualizationtoolforgeneticassignment,Iamnowworkingonthedevelopmentofanasymmetricmeasure,wherethefitofAintoBmaynotbethesameasthefitofBintoA.Thismeasurewillenablethedetectionofscenariossuchas"subsetting",therelationshipbetweenalarge,diversepopulationAandasmallerpopulationBthathasexperiencedgeneticdriftsincebeingseparatedfromA.Themeasurehasseveralfeaturesthatdistinguishitfromexistingmeasures,andisconstructedusingthesamesaddlepointapproximationmethodunderlyingGenePlot,andwhichisusedtoapproximatethemulti-locusgeneticdistributionsofpopulations.
Wednesday29th12:10Narrabeen
FastAndApproximateExhaustiveVariableSelectionForGLMsWithAPES
KevinWang,SamuelMueller,GarthTarr,andJeanYangUniversityofSydney
Obtainingmaximumlikelihoodestimatesforgeneralisedlinearmodels(GLMs)iscomputationallyintensiveandremainsasthemajorobstacleforperformingallsubsetsvariableselection.Exhaustiveexplorationofthemodelspace,evenforamoderatelylargenumberofcovariates,remainsaformidablechallengeformoderncomputingcapabilities.Ontheotherhand,efficientalgorithmsforexhaustivesearchesdoexistforlinearmodels,mostnotablytheleapsandboundalgorithmand,morerecently,themixedintegeroptimisationalgorithm.Inthistalk,wepresentAPES(APproximatedExhaustiveSearch)anewmethodthat
approximatesallsubsetselectionforagivenGLMbyreformulatingtheproblemasalinearmodel.Themethodworksbylearningfromobservationalweightsinacorrect/saturatedgeneralisedlinearregressionmodel.APEScanbeusedinpartnershipwithanyotherstate-of-the-artlinearmodelselectionalgorithm,thusenabling(approximate)exhaustivemodelexplorationindimensionsmuchhigherthanpreviouslyfeasible.WewilldemonstratethatAPESmodelselectioniscompetitiveagainstgenuineexhaustivesearchviasimulationstudiesandapplicationstohealthdata.Extensionstoarobustsettingisalsopossible.
Wednesday29th12:10Gunnamatta
OrderSelectionOfFactorAnalyticModelsForGenotypeXEnvironmentInteraction
EmiTanaka1,FrancisHui2,andDavidWarton31UniversityofSydney
2ANU3UniversityofNewSouthWales
Factoranalytic(FA)modelsarewidelyusedacrossarangeofdisciplinesowingtocomputationaladvantagesfromdimensionreductionandpossibleabilitytointerpretthefactors.Inplantbreeding,FAmodelprovidesanaturalframeworktomodelthegenotypexenvironmentinteraction.AnFAmodelisdictatedbythenumberoffactors(orderofthemodel).Ahigherorderlendstomoreparametersinthemodelandthisnecessitatestheorderselectiontoachieveparsimony.Weintroduceanorderselectionmethodviatheorderedfactorlasso(OFAL).Weillustrateitsperformancebasedonasimulationonarealwheatyieldmulti-environmentaltrial.
Wednesday29th12:10Bundeena
MultipleSampleHypothesisTestingOfTheHumanMicrobiomeThroughEvolutionaryTrees
MartinaMincheva1,HongzheLi2,andJunChen31TempleUniversity
2UniversityofPennsylvania3MayoClinic
Nextgenerationsequencingtechnologiesmakeitpossibletosurveymicrobialcommunitiesbysequencingnucleicacidmaterialextractedfrommultiplesamples.Themetagenomicreadcountsaresummarizedasempiricaldistributionsonareferencephylogenetictree.ThedistancebetweenthemisevaluatedbytheKantorovich-Rubinstein(kr)metric,equivalenttothecommonlyusedweightedUniFracdistanceonatree.Thispaperproposesamethodtotestthehypothesisthattwosetsofsampleshavethesamemicrobialcomposition.TheasymptoticdistributionsofthekrdistancebetweenthetwoFrechetmeansandtheFrechetvariancesarederivedandareshowntobeindependent.TheteststatisticisdefinedastheratioofthosedistancesanditisshowntofollowanasymptoticF-distribution.Itsgeneralitystemsfromthefactthatthetestisnonparametricandrequiresnoassumptionsontheprobabilitydistributionsofthecountdata.Itisalsoapplicableforvaryingsetsizesandsamplesizes.ItisanextensionofEvansandMatsen(2012)whosuggestatesttoonlycomparetwosinglesamples.Thecomputationalefficiencyoftheproposedtestcomesfromtheexactasymptoticdistributionoftheproposedteststatistic.Extensivedataanalysisshowsthatthetestissignificantlyfasterthanthepermutation-basedmultivariateanalysisofvarianceusingdistancematrices(permanova)(McArdleandAnderson,2001).Atthesametime,ithascorrecttype1errorsandcomparablepower,whichmakesitpreferableintheanalysisoflargescalemicrobiomedata.
Keynote:Wednesday29th13:40Mantra
StatisticalStrategiesForTheAnalysisOfLargeAndComplexData
LouiseRyan1,2,StephenWright1,3,andHonHwang11UniversityofTechnologySydney
2HarvardT.H.ChanSchoolofPublicHealth3AustralianRedCross
Thistalkwillfocusonchallengesthatarisewhenfacedwiththeanalysisofdatasetsthataretoolargeforstandardstatisticalmethodstoworkproperly.Whileonecanalwaysgofortheexpensivesolutionofgettingaccesstoamorepowerfulcomputerorcluster,itturnsoutthattherearesomesimplestatisticalstrategiesthatcanbeused.Inparticular,we'lldiscusstheuseofsocalled"DivideandRecombine"strategiesthatrelegatesomeoftheworktobedoneinadistributedfashion,forexampleviaHadoop.Combiningthesestrategieswithcleversubsamplinganddatacoarseningideascanresultindatasetsthataresmallenoughtomanageonastandarddesktopmachine,withonlyminimalefficiencyloss.TheideasareillustratedwithdatafromtheAustralianRedCross.
Wednesday29th14:30Narrabeen
OptimalExperimentalDesignForFunctionalResponseExperiments
JeffZhangandChristopherDrovandiQueenslandUniversityofTechnology
Functionalresponsemodelsareimportantinunderstandingpredator-preyinteractions.Thedevelopmentoffunctionalresponsemethodologyhasprogressedfrommechanisticmodelstomorestatisticallymotivatedmodelsthatcanaccountforvarianceandtheover-dispersioncommonlyseeninthedatasetscollectedfromfunctionalresponseexperiments.However,littleinformationseemstobeavailabletothosewishingtoprepareoptimalparameterestimationdesignsforfunctionalresponseexperiments.Wedevelopaso-calledexchangedesignoptimisationalgorithmsuitableforinteger-valueddesignspaces,whichforthemotivatingfunctionalresponseexperimentinvolvesselectingthenumberofpreyusedforeachobservation.Further,wedevelopandcomparenewutilityfunctionsforperformingrobustoptimaldesigninthepresenceofparameteruncertainty,whicharegenerallyapplicable.Themethodsareillustratedusingapublishedbeta-binomialfunctionalresponsemodelforanexperimentinvolvingthefreshwaterpredatorNotonectaglauca(anaquaticinsect)preyingonAsellusaquaticus(asmallcrustacean)asacasestudy.
Wednesday29th14:30Gunnamatta
ToPCAOrNotToPCA
CatherineM.McKenzie1,WeiZhang1,StuartD.Card1,CoryMatthew2,WadeJ.Mace1,andSivaGanesh1
1AgResearch2MasseyUniversity
Whentherearegroupingsofobservationspresentinthedata,manyresearchersresorttoutilisingaPrincipalComponentsAnalysis(PCA)aprioriforidentifyingpatternsinthedata,andthenlooktomapthepatternsobtainedfromPCAtodifferencesamongthegroupings,orattributebiologicalsignaltothem.Isthisappropriate,giventhatPCA’sarenotdesignedtodiscriminatebetweenthegroupings?Isagroup-orientedmultivariatemethodologysuchasmultivariateanalysesofvariance(MANOVA)orCanonicalDiscriminantAnalysis(CDA)preferable?Whichmethodhasmorerelevancewheninvestigatingfactoreffectsofbiochemicalpathways?Weexplorethisquestionviaabiologicalexample.
Twobiologicalmaterials(B1&B2)wereanalysedforthesame19primarymetabolites,withthreefactorsofMethods(M1&M2),Treatment(T1,T2,T3andT4),andAge(A1&A2),withthreereplicatevaluesgivingatotalof48observationsforeachbiologicalmaterial.Univariateandmultivariateanalysesofvariance(ANOVAandMANOVA,respectively)werecarriedout,forwhichthereweremanystatisticallysignificantinteractioneffects.Inaddition,othermultivariatetechniquessuchasPCAandCDAwereusedtoexplorerelationshipsbetweenthevariables.ThequestionremainsastotheappropriatenessofcarryingoutPCAtoexplorebiochemicalpathways,thecomparisonbeingbetweentailoringthepatternextractionaprioritomatchtheknowngroupingswithinthedataversusstartingwithanunrestrainedpatternanalysisandseekingtoexplainthepatternsdetectedpost-analysis?
Wednesday29th14:50Narrabeen
AnEvaluationOfErrorVarianceBiasInSpatialDesigns
EmlynWilliams1andHans-PeterPiepho21ANU
2UniversityofHohenheim
Spatialdesignandanalysisarewidelyused,particularlyinfieldexperimentation.However,itisoftenthecasethatspatialanalysisdoesnotenhancemoretraditionalapproachessuchasrow-columnanalysis.Itisthenofinteresttogaugethedegreeoferrorvariancebiasthataccrueswhenaspatially-designedexperimentisanalysedasarow-columndesign.ThistalkbuildsontheworkofTedin(1931)who,withR.A.Fisherasadvisor,studiederrorvariancebiasinknight'smoveLatinsquares.
Wednesday29th14:50Gunnamatta
NewModel-BasedOrdinationDataExplorationToolsForMicrobiomeStudies
OlivierThas1,StijnHawinkel1,andLucBijnens21GhentUniversity
2JanssenPharmaceutics
High-throughputsequencingtechnologiesalloweasycharacterizationofthehumanmicrobiome,butthestatisticalmethodsforanalysingmicrobiomedataarestillintheirinfancy.DataexplorationoftenreliesonclassicaldimensionreductionmethodssuchasPrincipalCoordinateAnalysis(PCoA),whichisbasicallyaMultidimensionalScaling(MDS)methodstartingfromecologicallyrelevantdistancemeasuresbetweenthevectorsofrelativeabundancesofthemicroorganisms(e.gBray-Curtisdistance).
Wewilldemonstratethattheseclassicalvisualisationmethodsfailtodealwithmicrobiome-specificissuessuchasvariabilityduetolibrary-sizedifferencesandoverdispersion.Nextweproposeanewtechniquethatisbasedonanegativebinomialregressionmodelwithlog-link,andwhichreliesontheconnectionbetweencorrespondenceanalysisandthelog-linearRC(M)modelsofGoodman(AnnalsofStatistics,vol.13,1985);seealsoZhuetal.(EcologicalModelling,vol.187,2005).InsteadofassumingaPoissondistributionforthecounts,anegativebinomialdistributionisassumed.Tobetteraccountforlibrarysize
effects,weadoptadifferentweightingscheme,whichnaturallyarisesfromtheparameterisationofthemodel.AniterativeparameterestimationmethodisproposedandimplementedintoR.Thenewmethodisillustratedonseveralexampledatasets,anditisempiricallyevaluatedinasimulationstudy.Itisconcludedthatourmethodsucceedsbetterindiscoveringstructureinmicrobiomedatasetsthanwithotherconventionalmethods.
Inthesecondpartofthepresentationweextentthemodel-basedmethodtoaconstrainedordinationmethodbyusingsample-specificcovariatedata.Themethodlooksforatwo-dimensionalvisualisationthatoptimallydiscriminatesbetweenspecieswithrespecttotheirsensitivitytoenvironmentalconditions.AgainwebuilduponresultsofZhuetal.(2005)andZhangandThas(StatisticalModelling,vol.12,2012).Themethodisillustratedonrealdata.
AllmethodsareavailableasanRpackage.
Wednesday29th15:10Narrabeen
AlwaysRandomize?
ChrisBrien1,21UniversityofSouthAustralia
2UniverstiyofAdelaide
Fishergaveusthreefundamentalprinciplesfordesignedexperiments:replication,randomizationandlocalcontrol.Consonantwiththis,Brienetal.(2011)[Brien,C.J.,Harch,B.D.,Correll,R.L.,&Bailey,R.A.(2011)Multiphaseexperimentswithatleastonelaterlaboratoryphase.I.Orthogonaldesigns.JournalofAgricultural,Biological,andEnvironmentalStatistics,16,422-450.]exhorttheuseofrandomizationinmultiphaseexperimentsviatheirPrinciple7(Allocateandrandomizeinthelaboratory).Thisprincipleisqualifiedwith‘whereverpossible’,whichleadstothequestion‘whenisrandomizationnotpossible?’.
Situationswhererandomizationisnotapplicablewillbedescribedforbothsingle-phaseandmultiphaseexperiments.Thereasonsfornotrandomizingincludepracticallimitationsand,formultiphaseexperiments,difficultyinestimatingvarianceparameterswhenrandomizationisemployed.Forthelatter
case,simulationstudiescanvassinganumberofpotentialdifficultieswillbedescribed.ANonrandomizationPrinciple,andanaccompanyinganalysisstrategy,formultiphaseexperimentswillbeproposed.
Wednesday29th15:10Gunnamatta
ComparingClassicalCriteriaForSelectingIntra-ClassCorrelatedFeaturesForThree-ModeThree-WayData
LynetteHunt1andKayeBasford21UniversityofWaikato
2UniversityofQueensland
Manyunsupervisedlearningtasksinvolvedatasetswithbothcontinuousandcategoricalattributes.Onepossibleapproachtoclusteringsuchdataistoassumethatthedatatobeclusteredcomefromafinitemixtureofpopulations.Therehasbeenextensiveuseofmixtureswherethecomponentdistributionsaremultivariatenormalandwherethedatawouldbedescribedastwomodetwowaydata.Thefinitemixturemodelcanalsobeusedtoclusterthreewaydata.Themixturemodelapproachrequiresthespecificationofthenumberofcomponentstobefittedtothemodelandtheformofthedensityfunctionsoftheunderlyingcomponents.
Thistalkillustratestheperformanceofseveralcommonlyusedmodelselectioncriteriainselectingboththenumberofcomponentsandtheformofthecorrelationstructureamongsttheattributeswhenfittingamixturemodeltothefinitemixturemodeltoclusterthreewaydatacontainingmixedcategoricalandcontinuousattributes
Wednesday29th15:50Narrabeen
ExploringTheSocialRelationshipsOfDairyGoats
VanessaCave1,BenjaminFernoit2,JimWebster1,andGosiaZobel11AgResearch
2AgrosupDijon
Goatsaresentientbeingscapableofanemotionalresponsetotheirlives.Yetdespitethis,thesocialrelationshipsbetweenanimalsarelargelyoverlookedincommercialsystems.Goatshavebeenshowntorecognise,andmakedecisionsbasedon,thepresenceofotherspecificgoats.Thedegreetowhichtheychoosetoassociatewithindividualshasnotbeenestablished,butsocialbondsinotheranimalshavebeenshowntobufferagainststressfulsituations,andconverselycanbeasignificantsourceofstresswhensuchbondsaredisrupted.
Toinvestigatewhethersocialrelationshipsexistamongdairygoats,4non-consecutivedaysofvideofocusingonagroupof12goatswasanalysed.Atoneminutescanintervals,theproximityofeverygoatrelativetoeveryothergoatwasrecordedasanordinalvariablewithfourlevels(incontact,withinaheadlength,withinabodylength,oralone).Eachgoat’slocationinthepenwasalsonoted(feeding,bedding,climbingplatform).
Avarietyofstatisticaltechniques,includingheatmapsandnetworkanalyses,wereusedtostudythesocialrelationshipsamonggoatsbasedonproximity.Socialrelationshipswerecharacterisedbyspecificpairsofgoatsreliablyspendingalotoftimeincloseproximity.
Resultsindicatethatwhilstsomegoatswere“sociable”(e.g.,spendingmorethan60%oftheirtimewithothergoats),otherstendedtobe“loners”(e.g.,spendingmorethan60%oftheirtimealone).Interestingly,therewasevidenceofbothpreferredandavoidedcompanionships.
Thissmall-scalestudyprovidesthefirstevidencetosuggestthatcommonmanagementpracticesresultingintheregroupingofdairygoatscouldhaveanimpactontheirwelfare.
Wednesday29th15:50Gunnamatta
BayesianSemi-ParametricSpectralDensityEstimationWithApplicationsToThe
SouthernOscillationIndex
ClaudiaKirch1,MattEdwards2,AlexanderMeier1,andRenateMeyer21UniversityofMagdeburg2UniversityofAuckland
StandardtimeseriesmodellingisdominatedbyparametricmodelslikeARMAandGARCHmodels.EventhoughnonparametricBayesianinferencehasbeenarapidlygrowingareaoverthelastdecade,onlyveryfewnonparametricBayesianapproachestotimeseriesanalysishavebeendeveloped.Mostnotably,CarterandKohn(1997),Gangopadhyay(1998),Choudhurietal.(2004),andRosenetal(2012)usedWhittle'slikelihoodforBayesianmodelingofthespectraldensityasthemainnonparametriccharacteristicofstationarytimeseries.Ontheotherhand,frequentisttimeseriesanalysesareoftenbasedonnonparametrictechniquesencompassingamultitudeofbootstrapmethods(KreissandLahiri,2011,KirchandPolitis,2011).
AsshowninContreras-Cristanetal.(2006),thelossofefficiencyofthenonparametricapproachusingWhittle'slikelihoodapproximationcanbesubstantial.Ontheotherhand,parametricmethodsaremorepowerfulthannonparametricmethodsiftheobservedtimeseriesisclosetotheconsideredmodelclassbutfailifthemodelismisspecified.Therefore,wesuggestanonparametriccorrectionofaparametriclikelihoodthattakesadvantageoftheefficiencyofparametricmodelswhilemitigatingsensitivitiesthroughanonparametricamendment.WeuseanonparametricBernsteinpolynomialprioronthespectraldensitywithweightsinducedbyaDirichletprocess.ContiguityandposteriorconsistencyforGaussianstationarytimeserieshavebeenshowninapreprintbyKirchetal(2017).BayesianposteriorcomputationsareimplementedviaaMH-within-Gibbssamplerandtheperformanceofthenonparametricallycorrectedlikelihoodisillustratedinasimulation.WeusethisapproachtoanalysethemonthlytimeseriesoftheSouthernOscillationIndex,oneofthekeyatmosphericindicesforgaugingthestrengthofElNinoeventsandtheirpotentialimpactsontheAustralianregion.
Wednesday29th16:10Narrabeen
EfficientMultivariateSensitivityAnalysisOf
AgriculturalSimulators
DanielGladishCSIRO
Complexmechanisticcomputermodelsoftenproducemultivariateoutput.Sensitivityanalysiscanbeusedtohelpunderstandsourcesofuncertaintyinthesystem.Muchoftheliteraturearoundsensitivityanalysishasfocusedonunivariateoutput,withsomerecentadvancesusingmultivariatecorrelatedoutput.Onepromisingmethodformultivariatesensitivityanalysisinvolvesdecompositionthroughbasisfunctionexpansion.However,thesemethodsoftenrequireseveralmodelrunsandmaystillbecomputationallyintensiveforpracticalpurposes.Emulatorshavebeenaprovenmethodforreducingcomputationaltimeforunivariatesensitivityanalysis,withsomerecentdevelopmentformultivariatecomputermodels.Weproposetheuseofgeneralizedadditivemodelsandrandomforestscombinedwithaprincipalcomponentanalysisforemulationforamultivariatesensitivityanalysis.Wedemonstrateourmethodusingacomplexagriculturalsimulator.
Wednesday29th16:10Gunnamatta
BayesianHypothesisTestsWithDiffusePriors:CanWeHaveOurCakeAndEatItToo?
JohnOrmerod,MichaelStewart,WeichangYu,andSarahRomanesUniversityofSydney
WeintroduceanewclassofpriorsforBayesianhypothesistesting,whichwename"cakepriors".ThesepriorscircumventBartlett'sparadox(alsocalledtheJeffreys-Lindleyparadox);theproblemassociatedwiththeuseofdiffusepriorsleadingtononsensicalstatisticalinferences.Cakepriorsallowtheuseofdiffusepriors(havingonescake)whileachievingtheoreticallyjustifiedinferences(eatingittoo).WedemonstratethismethodologyforBayesianhypothesestestsforscenariosunderwhichtheoneandtwosample -tests,andlinearmodelsaretypicallyderived.Theresultingteststatisticstaketheformofapenalized
likelihoodratioteststatistic.ByconsideringthesamplingdistributionunderthenullandalternativehypothesesweshowforindependentidenticallydistributedregularparametricmodelsthatBayesianhypothesistestsusingcakepriorsarestronglyChernoff-consistent,i.e.,achievezerotypeIandIIerrorsasymptotically.Lindley'sparadoxisalsodiscussed.
ProgrammeAndAbstractsForThursday30thOfNovember
Keynote:Thursday30th9:00Mantra
AGeneralFrameworkForFunctionalRegressionModelling
SonjaGrevenandFabianScheiplLMUMunich
Recenttechnologicaladvancesgenerateanincreasingamountoffunctionaldata,datawhereeachobservationrepresentsacurveoranimage(RamsayandSilverman,2005).Examplesoftechnologiesthatgeneratefunctionaldataincludeimagingtechniques,accelerometers,spectroscopyandspectrometry.Anykindofmeasurementcollectedovertime-datausuallyreferredtoaslongitudinal-canalsobeviewedaspotentiallysparselyobservedfunctionaldata.
Researchersareincreasinglyinterestedinregressionmodelsforfunctionaldatatorelatefunctionalobservationstoothervariablesofinterest.Wewilldiscussacomprehensiveframeworkforadditive(mixed)modelsforfunctionalresponsesand/orfunctionalcovariates.Theguidingprincipleistoreframefunctionalregressionintermsofcorrespondingmodelsforscalardata,allowingtheadaptationofalargebodyofexistingmethodsforthesenoveltasks.Theframeworkencompassesmanyexistingaswellasnewmodels.Itincludesregressionfor'generalized'functionaldata,meanregression,quantileregressionaswellasgeneralizedadditivemodelsforlocation,shapeandscale(GAMLSS)forfunctionaldata.Itadmitsmanyflexiblelinear,smoothorinteractiontermsofscalarandfunctionalcovariatesaswellas(functional)randomeffectsandallowsflexiblechoicesofbases-inparticularsplinesandfunctionalprincipalcomponents-andcorrespondingpenaltiesforeachterm.Itcoversfunctionaldataobservedoncommon(dense)orcurve-specific(sparse)grids.Penalized
likelihoodbasedandgradient-boostingbasedinferenceforthesemodelsareimplementedinRpackagesrefundandFDboost,respectively.Wealsodiscussidentifiabilityandcomputationalcomplexityforthefunctionalregressionmodelscovered.Arunningexampleonalongitudinalmultiplesclerosisimagingstudyservestoillustratetheflexibilityandutilityoftheproposedmodelclass.ReproduciblecodeforthiscasestudyisalsoavailableonlinewiththerecentdiscussionpaperGrevenandScheipl(2017)thistalkisbasedon.
Thursday30th10:30Narrabeen
HockeySticksAndBrokenSticks–ADesignForASingle-Arm,Placebo-Controlled,Double-Blind,RandomizedClinicalTrialSuitableForChronicDiseases
HansHockey1andKristianBrock21BiometricMattersLtd.
2CancerResearchUKClinicalTrialsUnit
Thisworkismotivatedandexemplifiedbyageneticdisordercausingearlyonsetdiabetes,blindnessanddeafness,whichisextremelyrare,inevitablyfatalandhasnocurrentdirecttreatment.Whilethestandardplacebo-controlledRCTisthegoldstandardrequiredbytheregulatoryagencyforanewproposeddrugstudy,itisconjecturedthatpotentialstudyparticipantswillpreferadesignwhichguaranteesthattheyarealwaysassignedtothedrugunderstudy.Asingle-armdesignisproposedwhichmeetsthispatientneedandhenceprobablyincreasesrecruitmentandcompliance.Atthesametime,itmeetstherequirementforfullrandomization.Analyseswhichfollownaturallyfromthisdesignarealsodescribedandwereusedintrialsimulationsforsamplesizingandforexaminationoftheeffectofunderlyingassumptions.
Thursday30th10:30Gunnamatta
BoundingIVEstimatesUsingMediation
AnalysisThinking
TheisLange1,21UniversityofCopenhagen
2PekingUniversity
Inthispaperweinitiallydemonstratethatthewell-knownassumptionsforconductingIV-analysescanequivalentlybeexpressedusingnaturaleffectsfrommediationanalysis.Viewingtheassumptions,anindeedthewholeIVanalysis,fromamediationanalysisperspectiveopensupanovelpossibilityforboundingthetruecausaleffectoftheexposurewhenthedistributionalassumptionsoftheIVanalysisfail;e.g.ifthereisaninteractionsbetweentheunmeasuredconfoundersandexposure.Theprocedureworksacrossalleffectscales(riskdifference,hazardratioetc.)andtypesofviolationsofthedistributionalassumptions.TheproposedmethodcanalsobeviewedasasensitivityanalysisfortheIV-analysiswherethereisonlyasingletuningparameter,whichcanbestraightforwardlyinterpreted.ForthepurelybinarycasetheproposedboundsconvergestotheBalkeandPearlboundswhenthetuningparametertendstoinfinity(i.e.wheneventhemostextrememisspecificationisconsidered).Astheproposedmethodiscomputationallydemandingsometimeisspentonimplementationconsiderations.
Thursday30th10:50Narrabeen
CorrelatedBivariateNormalCompetingRisks---StructuringEstimationInAnIll-PosedProblem
MalcolmHudsonMacquarieUniversity
Estimationofcorrelationbetweencompetingriskshaslongbeenknowntobeanill-posedproblem,duetolackofidentifiablity,termedtheidentifiabilitycrisis[Crowder,ScandJStat1991].Recently,manysemi-parametricmodelsandafullyparametricmodelhavebeenusedtopermitestimationandassess
sensitivityofresultstodegreeofcorrelation[JeongandFineBiostatistics2007;seealsoTaietal,SIM2008].Parametricmodelsarenon-standardandcalculationscomplex;JeongandFine'sparametricmodelemploysaGompertzdistribution.Forlog-Normaldata,complexityofcalculationsinabivariateNormal(BVN)modelisforbiddinganddirectoptimizationunstable.
WedescribeaparametricsolutionavailableintheBVNcase.OurapproachusesanEMalgorithmforcompetingrisks(generalisingtheAitkin'sEMforunivariatesurvival).ComplexcalculationsareevaluatedinclosedformusingalemmaofStein[Liu,StatistProbLetters1994].Thisprobabilisticandstatisticalplatformforexploringtheill-posednessisimplementedwithinthebncR-package(inpreparation).
Weintroducethesevariouscomponentsinthetalk.
Thursday30th10:50Gunnamatta
BayesianRegressionWithFunctionalInequalityConstraints
JoshuaBon1,BerwinTurlach1,KevinMurray1,andChristopherDrovandi21UniversityofWesternAustralia
2QueenslandUniversityofTechnology
WeinvestigatehowtoconductBayesianinferencewithfunctionalinequalityconstraintsovertheparameterspace.Thisproblemisanalogoustosemi-infiniteprogramminginoptimisation.AnovelmethodusingSequentialMonteCarlo(SMC)isgiven.TheSMCalgorithmapproximatesthedistributionofminimainordertosuccessivelyiteratetowardstheconstrainedparameterspace.Wedemonstrateonforensicmorphometricdataofhumanskullswheremonotonicityisrequiredinmorethanonedimension.Methodsarecomparedtothosecurrentlyavailablewithonedimensionalconstraintsandtheresultsdiscussed.
Thursday30th11:10Narrabeen
GeneticAnalysisOfRenalFunctionInAn
IsolatedAustralianIndigenousCommunity
RussellThomson1,BrendanMcMorran2,WendyHoy3,MatthewJose4,TimThornton5,GaétanBurgio2,andSimonFoote2
1WesternSydneyUniversity2ANU
3UniversityofQueensland4UniversityofTasmania
5UniversityofWashington
Incloseconsultationwiththelocallandcouncil,andwithethicalapprovalfrommanyethicscommittees,wehaveperformedagenome-wideassociationstudy(GWAS)onasamplecohortfromthe1990s.Wealsohaveafollowupdatasetof120studyparticipantsfrom2014,withDNAsequencedata.
Iwilldiscussthestatisticalanalysesofthesedatasets,withrespecttoissuesofdegradedDNA,higherrorratesandcorrelatedsamples.
Thursday30th11:10Gunnamatta
ThePerformanceOfModelAveragedTailAreaConfidenceIntervals
PaulKabailaLaTrobeUniversity
Commonlyinappliedstatistics,thereissomeuncertaintyastowhichexplanatoryvariablesshouldbeincludedinthemodel.Frequentistmodelaveraginghasbeenproposedasamethodforproperlyincorporatingthis"modeluncertainty"intoconfidenceintervalconstruction.Suchproposalshavebeenofparticularinterestinenvironmentalandecologicalstatistics.
Theearliestapproachtotheconstructionoffrequentistmodelaveragedconfidenceintervalswastofirstconstructamodelaveragedestimatoroftheparameterofinterestconsistingofadata-basedweightedaverageoftheestimatorsofthisparameterunderthevariousmodelsconsidered.Themodel
averagedconfidenceintervaliscenteredonthisestimatorandhaswidthproportionaltoanestimateofthestandarddeviationofthisestimator.However,thedistributionalassumptiononwhichthisconfidenceintervalisbasedhasbeenshowntobecompletelyincorrectinlargesamples.
AnimportantconceptualadvancewasmadebyFletcher&Turek(2011)andTurek&Fletcher(2012)whoputforwardtheideaofusingdata-basedweightedaveragesacrossthemodelsconsideredofproceduresforconstructingconfidenceintervals.Inthiswaythemodelaveragedconfidenceintervalisconstructedinasinglestep,ratherthanfirstconstructingamodelaveragedestimator.
WereviewtheworkofKabailaetal(2016,2017)whichevaluatestheperformanceofthemodelaveragedtailareaconfidenceintervalofTurek&Fletcher(2012)inthe“testscenario”oftwonestednormallinearregressionmodels.Ourassessmentofthisconfidenceintervalisthatitperformsquitewellinthisscenario,providedthatthedata-basedweightfunctioniscarefullychosen.
References:
1. Kabaila,P.,Welsh,A.H.,&Abeysekera,W.(2016)Model-averagedconfidenceintervals.ScandinavianJournalofStatistics.
2. Kabaila,P.,Welsh,A.H.andMainzer,R.(2017)Theperformanceofmodelaveragedtailareaconfidenceintervals.CommunicationsinStatistics-TheoryandMethods.
Thursday30th11:30Narrabeen
DeconstructingTheInnateImmuneComponentOfAMolecularNetworkOfTheAgingFrontalCortex
EllisPatrick1,MarikoTaga2,MartaOlah2,Hans-UlrichKlein2,CharlesWhite3,JulieSchneider4,LoriChibnik5,DavidBennett4,SaraMostafavi6,
ElizabethBradshaw2,andPhilipDeJager21UniversityofSydney2ColumbiaUniversity
3TheBroadInstitute4RushUniversity
5HarvardUniversity6UniverstiyofBritishColumbia
Alzheimer'sdiseaseispathologicallycharacterizedbytheaccumulationofneuritic-amyloidplaquesandneurofibrillarytanglesinthebrainandclinicallyassociatedwithalossofcognitivefunction.ThedysfunctionofmicrogliacellshasbeenproposedasoneofthemanycellularmechanismsthatcanleadtoanincreaseinAlzheimer'sdiseasepathology.InvestigatingthemolecularunderpinningsofmicrogliafunctioncouldhelpisolatethecausesofdysfunctionwhilealsoprovidingcontextforbroadergeneexpressionchangesalreadyobservedinmRNAprofilesofthehumancortex.
WehaveusedmRNAsequencingtoconstructgeneexpressionprofilesofmicrogliapurifiedfromthecortexof11subjectsfromalongitudinalcohortofaging,RushMemoryAgingProject(MAP).Bystudyingthesemicrogliageneexpressionprofilesinthecontextoftissue-levelprofilesofthecortexof542subjectsfromtheMAPandReligiousOrdersStudy(ROS)weaddressdualproblems.ByusinginformationfromthelargeROSMAPcohort,weareabletoisolatethegeneswhicharestronglyassociatedwithimmuneresponse.Conversely,weillustratethatthemicrogliasignaturecanbeusedtohighlightpredefinedsetsofcoexpressedgenesinROSMAPthatarehighlyenrichedformicrogliagenes.AddressingthesetwoquestionsallowsustoidentifysetsofmicrogliaspecificgeneswhichareassociatedwithvariousAlzheimer'sdiseasetraitsfurtheremphasizingthemolecularconsequencesofmicrogliadysfunctioninthisdisease.Specifically,weareabletoidentifyasetofmicrogliarelatedgenesassociatedwithtauandamyloidpathologyaswellasanactivatedmicroglialmorphology.
Thursday30th11:30Gunnamatta
BiasCorrectionInEstimatingProportionsByPooledTesting
GrahamHepworth1andBradBiggerstaff2
1UniversityofMelbourne2CentersforDiseaseControlandPrevention
Pooledtesting(orgrouptesting)ariseswhenunitsarepooledtogetherandtestedasagroupforthepresenceofanattribute,suchasadisease.Wehaveencounteredpooledtestingproblemsinplantdiseaseassessmentandprevalenceestimationofmosquito-borneviruses.
Intheestimationofproportionsbypooledtesting,theMLEisbiased,andseveralmethodsofcorrectingthebiashavebeenpresentedinpreviousstudies.WeproposeanewestimatorbasedonthebiascorrectionmethodintroducedbyFirth(1993),whichusesamodificationofthescorefunction.Ourproposedestimatorisalmostunbiasedacrossarangeofproblems,andsuperiortoexistingmethods.WeshowthatforequalpoolsizesthenewestimatorisequivalenttotheestimatorproposedbyBurrows(1987),whichhasbeenusedbymanypractitioners.
Thursday30th11:50Narrabeen
TheParametricCureFractionModelOfOvarianCancer
SerifatFolorunso,AngelaChukwu,andAkintundeOdukogbeUniversityofIbadan
WeproposeincorporatingtheGammalinkfunctiontothegeneralizedGammausingamixturecurefractionmodel.Themathematicalpropertiesoftheproposedmodelwereexploredandtheinferencesforthemodelswereobtained.TheproposedmodelcalledGamma-generalizedgammamixturecuremodel(GGGMCM)willbevalidatedusingareallifedatasetonovariancancer.
Thursday30th11:50Gunnamatta
TheSkillings-MackStatisticForRanksDataInBlocks
JohnBestUniversityofNewcastle
SkillingsandMackgaveastatisticfortestingtreatmentdifferencesforranksdatainblocks-bothcompleteandincompleteblocksorblockswithmissingvalues.Thisgeneralstatisticisthusquiteuseful.ThereisanRpackageforitscalculation.Howeverweillustrateaproblemwithtiedranks.Sensoryevaluationdataisused.
PosterAbstracts
Multi-EnvironmentTrialAnalysisOfAgronomyTrialsUsingEstablishedPlantPopulationDensity
MichaelMumfordandKerryBellQueenslandDepartmentofAgricultureandFisheries
Thereisagrowinginterestwithinthegrainsindustrytodeterminetheyieldresponsetoplantpopulationdensity.Thiscanvarydependingonthetriallocation,thehybridthatisplantedandthemanagementpracticesthatareemployed.
Inthepast,theanalysisofplantpopulationdensityhasbeenperformedusingthetargetedplantpopulationdensityasdeterminedbythenumberofseedsinitiallyplanted.However,ithasbeenfoundthatthetargetedplantpopulationdensitymaydiffersignificantlyfromtheestablishedplantpopulationdensity,i.e.theplantpopulationdensityobservedinthefield.Insuchcircumstances,itbecomesimportanttousetheestablishedratherthantargetedplantpopulationdensitiesinordertoprovidethemostinformativemeasureofyieldresponsetoplantdensity.Thisisespeciallytruewhenperformingamulti-environmenttrial(MET)analysisduetovariationbetweenthetargetedandestablishedplantpopulationdensityateachtrial.
WeproposeaprocedureforperformingaMETanalysisusingestablishedplantpopulationdensityasanexplanatoryvariable.Usingalinearmixedmodelframework,separateregressionlines/curvesarefittoeachcombinationoftrialandmanagementpractice,whichwedefineas‘environment’.Environmentsarethengroupedaccordingtotheresultsofaclusteranalysisperformedusingtheestimatesoftheregressioncoefficients,ensuringthatenvironmentswithinaclustershareasimilarresponsetoplantdensity.Thelines/curvesfittedtoenvironmentswithinclustersarethenparallel,allowingustothenperformFisher’sleastsignificantdifferencetestingforyielddifferencesbetween
environmentswithineachcluster.
ThispresentationwillfocusontheapplicationoftheproposedmethodtoalargesetofmaizeandsorghumagronomytrialsconductedacrossNewSouthWalesandQueenslandoverthelastthreeyears.Forsimplicity,aseparateanalysiswasperformedforeachhybrid.
AdventuresInDigitalAgricultureInNewZealand
EstherMeenkenandVanessaCaveAgResearch
NewZealandisanationforwhichprimaryindustriesaretheeconomicbackbone.NewZealand'shighlyvariablelandscapemeansthereareawiderangeinclimaticandsoilconditions,withmultiplelandusesexistinginrelativelysmallspatialareas.ThismeanstherearemanypossibilitiestobeexploredinDigitalAgricultureasadvancesinsensorytechnologiesaremade.Thetechnicalchallengesknowntoberelatedtotheuseofsensorsincludeissuesaroundbigdata,missingdata,biasedandvariablesensors,andvarioustypesofstructuredandnon-normaldata.WewilldescribeasuiteofstudiesthathavebeenrecentlyundertakenthatcanbelooselygroupedundertheDigitalAgricultureumbrella.Casestudiesexplorearangeofscientificandstatisticalchallengesintheprimaryindustriesarena,andincludetheuseofclassicalandBayesianPrincipledExperimentalDesign,HierarchicalBayesianModels,neuralnetclassifiers,visualisation,animaltrackingviaGPS,anddatamanagement.
ModellingCanopyGreennessOverTimeUsingSplinesAndNon-LinearRegression
BethanyMacdonald1,JackChristopher2,andAlisonKelly11QueenslandDepartmentofAgricultureandFisheries
2QueenslandAllianceforAgricultureandFood
Theabilityofaplanttoretainleafgreennessforanextendedtimeafteranthesis,
knownasstay-green,hasbeenlinkedtogreateryieldinwheatandcanvaryamongstgenotypes.Manyaspectsofaplant'ssenescencecontributetostay-green;therefore,accuratelymodellingsenescenceandunderstandingthegeneticvariationinsenescencedynamicsisintegraltoselectingforthestay-greenphenotype.Normaliseddifferencevegetativeindex(NDVI)providesameasureofcanopygreennessandwhencollectedovertime,cangiveanindicationofwhensenescencebeginsandtherateatwhichitcontinues.Christopheretal.(2014)capturedgeneticvariationinsenescencedynamicsusingmultiplestay-greentraitsderivedfromalogisticregressionoflongitudinalNDVImeasurements.Thesetraitswereestimatedindependentlyforeachplotwithintheexperiment,ignoringanysourcesofcovariance,andthensubsequentlyanalysedusingalinearmixedmodel.However,noestimationerrorswerecarriedfromthefirsttothesecondstageoftheanalysis.
Wediscusstwoalternativeone-stagemethodsformodellingsenescencedynamicsbasedonlongitudinalNDVImeasurements.Bothofthesemethodsenablethevariabilityinsenescencepatternstobepartitionedintogeneticandnon-geneticcomponents,whilealsoincorporatingtheexperimentalstructure.Thefirstmethodinvolvestheuseofsplinesinalinearmixedmodelframework.Thismethodallowsanappropriatecovariancestructureforrepeatedmeasurestobeutilisedand,unlikeotherapproaches,enablesflexibilityinthesenescencepatterns.Thesecondmethodinvolvesfittinglogisticequationsinanon-linearmixedmodelframework.Theestimatedparametershavebiologicalinterpretations,aidinginthegeneticcomparisonofsenescencedynamics.Thesetwomethodsofferamorestatisticallyrobustapproachtomodellingcropsenescenceandtheunderlyinggeneticvariability.
OnTestingMarginalHomogeneityForSquareContingencyTablesWithOrdinalCategories
KoujiTahataTokyoUniversityofScience
Fortheanalysisofsquarecontingencytableswithorderedcategories,themarginalhomogeneitymodel,whichindicatestherowmarginaldistributionis
equaltothecolumnmarginaldistribution,wereconsidered.Stuart(1955)proposedtheteststatistics(denotedbyQ)forthemarginalhomogeneitymodel.Howeveritdoesnotusetheinformationofthecategoryordering.Agresti(1983)comparedbetweenQandtheMann-Whitneytestaboutthepowerbysimulations.Inthispaper,weconsidersomemeasurestorepresentthedegreeofdeparturefromthemarginalhomogeneity.Forexample,Tomizawa,MiyamotoandAshihara(2003)andTahata,IwashitaandTomizawa(2008).InthesituationgivenbyAgresti(1983),wecomparebetweenQandtestbasedonthesemeasuresaboutthepower.
AFactorAnalyticApproachToModellingDiseaseProgressionAcrossLeafLayersAndTime
ClaytonForknall,GregPlatz,LisleSnyman,andAlisonKellyQueenslandDepartmentofAgricultureandFisheries
ControllingandlimitingtheimpactoffoliardiseasesisachallengefacedbytheAustraliangrainsindustry.Foliardiseasescompromiseanddestroyphotosyntheticarea,thuslimitingplantresourcesandadverselyaffectingcropproductivity.Suchdiseasesofteninfectthelowerleaflayersofplantsearlyintheseasonand,dependentontheircompatibilitywiththehostandenvironment,progresstowardsthetopmostleaflayersovertime.
Synthesizingthecomplicateddynamicsofdiseaseprogression,overdifferentleaflayersandacrossagrowingseason,intoameasuretoestimatetheimpactoflossofleafareaonproductivityrequirestheassessmentoftheproportionofleafareacompromisedbydisease(LAD)atmultipletimesthroughouttheseason.AsimplemeasurethatcapturesbothLADanddiseasedurationonagivenleaflayeristheAreaUndertheDiseaseProgressCurve(AUDPC),formedbyapplyingthetrapezoidruletoLADassessmentsovertime.TheAUDPCisthenoftencorrelatedtoameasureofproductivityforestimatingtheimpactofdiseaseateitheravarietyorexperimentalunitlevel.
WeproposeanalternativeapproachtotheAUDPCtomodeldiseaseprogressionovertimeandleaflayersmoreefficiently.Usingafactoranalyticapproach,thecovariancebetweenleaflayers,bothwithinandacrossassessmenttimes,is
capturedonavarietybasisinalinearmixedmodelframework.Thisapproachwillnotonlyprovidemorereliableestimatesofthediseasepressureexertedonvarieties,butalsoinformofrelationshipsbetweenleaflayersandhowconsistenttheserelationshipsareacrosstime.Betterqualityinformationregardingtherelationshipbetweenlossofleafareaandproductivitycanbemadeavailabletoindustrythroughtheimprovedestimatesandunderstandingofdiseaseprogressionobtainedfromthismodellingapproach.
AssociatingStrawStrengthWithLikelihoodOfHeadLossInBarleyForWesternAustralia
DeanDiepeveen1,KefeiChen1,ChengdaoLi2,andDavidFarleigh31CurtinUniversity
2MurdochUniversity3DepartmentofPrimaryIndustriesandDevelopment
Barleyheadlossischaracterisedbystrawbreakagejustbelowtheheadatornearplantmaturity.Ourrecentresearchintobarleyheadlosshasenabledustoevaluatethevariouscomponentsofplantmaturityandhavefoundacleargeneticcomponenttothisindustryissue.Themorechallengingcomponentsofthisresearchistounderstandwhyourmoretolerantcultivarsalsoshowsubstantialheadlossincertainunseasonalconditionsduringharvest.Ourpreliminaryanalysesusesenvironmentaldatawithplantmeasurementdataandlooksattheincidencesofextremeweathereventsonbarleyheadloss.Whatconfoundstheanalysisisthenutrient/growthconditionsoftheplantandthepredispositionofthesebarleycropstoheadlossduringtheseweatherevents.Thispaperwillprovidesomelessonslearntandfuturedirections.
ADeepLearningNeuralNetworkModelToIdentifyTheImportantGenesInMetastaticBreastCancerFromCensoredMicroarrayData
Quoc-AnhTrinhVietnamNationalUniversity
Microarraydatahavebeenshowntocorrelatedwithsurvivalinbreastcancer.Themostdifficultinanalysisofthisdataconsistinasimultaneousmeasureofhugegeneexpressionlevelsoverafewpatientsavailable.TheconventionalsurvivalmodelssuchastheproportionalhazardmodelofCoxarenomoreappropriatedincludingthehazardproportionalhypothesisandthelinearcovariateseffecthypothesis.Prognosticstudiesinvolvetheidentificationofgenesthatcorrelatewithsurvivalinordertoprovidenewinformationonpathogenesis.Thisresultmayaidintheresearchofdrugdesignfornewtargets.Weintroduceinthispaperadeeplearningrecurrentneuralnetworkapproachtopredictsurvivaltimesfortheindividualpatientbasedonmicroarraymeasurement.Thisneuralnetworkdefinesthehazardfunctionofsurvivalnottomentiontheproportionalityandlinearityhypothesis.Wepresentadeeplearningapproachtotheidentificationofgenescriticalforthemetastasesofbreastcancer.Weidentifiedasetofhighlyinteractivegenesbyanalysingtheconnectivitymatrices.
UsingRAndShinyToDevelopWebBasedSamplingApplicationsForTheAgriculturalAndEducationSectors
PeterKasprzak,OlenaKravchuk,andAndyTimminsUniveristyofAdelaide
Samplingisanimportantaspectofagriculturalstatisticsandanimportantfeatureofindustrialreality.Enablinganefficientdialoguewithresearchersandagronomistsaboutsoundmethodsforfielddatacollectionwillincreasetheefficiencyandreliabilityofagriculturalprojects.TheShinypackageinRallowsinteractivewebapplicationstobecreatedbaseduponexistingRpackages.ThispresentationwilloutlinethestepsinvolvedincreatinganappthatimplementssamplingtechniquessuchasSRS,StratifiedSampling,andRatioSampling,withinavisuallyappealinggraphicalenvironmentforuseintheeducationandagriculturalsectors.
GraphicalNetworkAnalysesInformsBiomarkerDiscoveryViaQuantificationOfKeyDiseaseRelatedConnections
JamesDoeckeCSIRO
Determinationoftheoptimalsetofbiomarkersoftenreliesonlargescalefeatureselection,thatselectsaminimalsetofindependentbiomarkers,withoutconsiderationofanyinformationregardingthebiologicalnetwork.Thisresearchisaimedatdefiningbiologicalnetworkstructuresrelatedtodiseasephenotype,throughathree-levelstatisticalprocess.Briefly,wefirstimplementanunsupervisedcorrelationstructureacrossthesetofbiomarkers,usedtodefineclustersofbiomarkersviastrictthresholding.Next,usingthegraphicalnetworkanalysismethodDifferentialNetworkAnalysisinGenomics(DINGO),wedefinethedifferentialnetworkstructureforhealthyanddiseasestates.Lastlyfromthederiveddifferentialconnectivityscoreweclustertheindividualgroupsofbiomarkers.Fromthisprocess,weareabletoproduceanetworkofsimilarlygroupedsetsofbiomarkersthatcanseparatehealthyfromdiseasesubjects.Lastlyweperformarandomeffectsmodelforeachclustertoestimatetheassociationwithdiseasestatus.Usingthisdesign,wehavebeenabletoincorporateinformationfrommultiplecorrelatedbiomarkerstoincreasetheproportionofvarianceexplainedindiseasestatus.Wetestthisdesignontwodatasets,1)asetof~2000peptidebiomarkers,withtheexpressaimofdefiningsetsofbiomarkersassociatedwithAlzheimer’sDiseasepathology,and2)asetof~2000mRNAprobesetscombinedwithasetof~2000DNAmethylationprobeswiththeexpressaimofdefiningsetsofbiomarkersassociatedwithbreastcancer.Wepresentnetworkplotsandconnectivitystructuresforeachdataset,andhighlighttheincreasedperformanceofclusteredbiomarkersascomparedtoindividualbiomarkerstoseparatediseasestatus.Lastly,weshowthatbyutilizinganetworkapproach,wecanincorporatemorebiologicalinformationfromthedatatoexplainoutcome.
Spatio-TemporalCorticalBrainAtrophyPatternsOfAlzheimer'sDisease
MarcelaCespedes1,JamesMcGree1,ChristopherDrovandi1,KerrieMengersen2,JamesDoecke3,andJurgenFripp3
1QueenslandUniversityofTechnology2QueenslandUniversityofTechnlogy
3CSIRO
Thedegenerationofthecerebralcortexisacomplexprocesswhichoftenspansdecades.Thisdegenerationcanbeevaluatedonregionsofinterest(ROI)inthebrainthroughprobabilisticnetworkanalysis.However,currentapproachesforfindingsuchnetworkshavethefollowinglimitations:1)analysisatdiscreteagegroupscannotappropriatelyaccountforconnectivitydynamicsovertime;and2)morphologicaltissuechangesareseldomunifiedwithnetworks,despiteknowndependencies.Toovercometheselimitations,aprobabilisticdynamicwombledmodelisproposedtosimultaneouslyestimateROIcorticalthicknessandnetworkcontinuouslyoverage,andwascomparedtoanageaggregatedmodel.Theinclusionofageinthenetworkmodelwasmotivatedbytheinterestininvestigatingthepointintimewhenconnectionsalteraswellasthelengthoftimerequiredforchangestooccur.Ourmethodwasvalidatedviaasimulationstudy,andappliedtohealthycontrols(HC)andclinicallydiagnosedAlzheimer'sdisease(AD)groups.Theprobabilityofalinkbetweenthemiddletemporal(akeyADregion)andtheposteriorcingulategyrusdecreasedfromage55(posteriorprobability>0.9),andwasabsentbyage70intheADnetwork(posteriorprobability<0.12).ThesameconnectionintheHCnetworkremainedpresentthroughoutages55to95(posteriorprobability≥0.75).Theanalysespresentedinthisworkwillhelppractitionerschoosesuitablestatisticalmethodstoidentifykeypointsintimewhenbraincovarianceconnectionschange,inadditiontomorphologicaltissueestimates,whichcouldpotentiallyallowformoretargetedtherapeuticinterventions.
EmpiricalModellingOfFruitFirmnessChangeDuringColourConditioning
LindyGuoandRingoFengPlantandFood
Gold3(Actinidia.chinensisZesy002,marketedasZespri®SunGoldKiwifruit)isanewyellow-fleshedkiwifruitcultivarbeingusedtoreplaceHort16Athathas
beenalmostalldestroyedbytheoutbreakofPseudomonassyringaepv.actinidiae(Psa)in2010.TheSmartMonitoringprogrammestartedin2011,recordsfruitdevelopment,maturation,harvestingandstorageperformanceofGold3kiwifruitfrom12orchardsinafairlyconsistentmanneryearafteryear.
Theaimofthisprojectistosummarizefruitdevelopmentandstoragepotentialsoverthe5years,andhenceprovideabetterunderstandingoftherolesofseasonalweatherconditions,orchardmanagement,andfruitattributesatharvestaffectingstoragepotential.
Forthisproject,storagelifewasdefinedasstoragetimeforabatchoffruit(fruitharvestedfromanorchardonaparticulardate)tosoftento1kfg.Severalwaystocalculatedstoragelifebasedonfirmnessmonitoringdataarecomparedandthestorageliveswerecorrelatedwithweatherconditions,orchardmanagement,andfruitattributesatharvestusingdifferentalgorithms.Theresultswillbepresentedtohighlightexperienceslearntinthisdataanalysis.Recommendationswillbemadeondealingwithmultiyeardatawithcollinearvariables.