BRIEFING

1210 StatisticalToolsforProcedureValidation.TheUSPStatisticsExpertCommitteepresentsanewgeneral
informationchapter.ThischapterisproposedasacompanionchaptertoValidationofCompendialProcedures 1225 with
thepurposeofprovidingstatisticalmethodsthatcanbeusedinthevalidationofanalyticalprocedures.Specifically,this
chapterdiscussesallofthefollowinganalyticalperformancecharacteristicsfromastatisticalperspective:accuracy,precision,
range,detectionlimit,quantitationlimit,andlinearity.Additionalrelatedtopicsthatarediscussedinthisproposednewchapter
includestatisticalpower,twoonesidedtest(TOST)ofstatisticalequivalence,toleranceintervals,predictionintervals,
correctedAkaikeInformationCriterion(AICc),Bayesiananalysis,experimentaldesign,calibration,andvariancepooling
strategy.
(STAT:H.Pappa.)
CorrespondenceNumberC135740

Commentdeadline:November30,2014
Addthefollowing:
1210

STATISTICALTOOLSFORPROCEDUREVALIDATION

1.Introduction
2.WorkDoneBeforeValidation
2.1ExperimentalDesign
2.2OtherConsiderations
3.AccuracyandPrecision
3.1IntroductionandDefinitions
3.2ExperimentalDesignforAccuracyandPrecision
3.3RecommendedMethodsforVerifyingAccuracyandPrecision
3.3.1TestsofEquivalenceforSeparateAssessmentofAccuracyandPrecision
3.3.2CombinedValidationofAccuracyandPrecision
3.3.3NumericalExample
3.3.4PowerConsiderations
4.RangeDefinitions
5.LimitsofDetectionandQuantitation
5.1IntroductionandDefinitions
5.2EstimationofLOD
5.3LODExample
5.4EstimationofLOQ
5.5FinalCautionsandAssumptions
6.ModellingtheCalibrationRelationship(Linearity)
6.1IntroductionandDefinitions
6.2ReasonsforStudyingtheCalibrationRelationship
6.3CurrentPractice
6.4CalibrationModelDefinition

6.5MethodsfortheAssessmentofLinearity
6.6MethodologicalAssumptions
6.7TwoOneSidedTestsofEquivalencetoEvaluateBiasinReportedConcentrations
6.7.1TOSTforBiasWhenApproximatingaQuadraticwithaStraightLineModel
6.7.2TOSTforBiasWhenApproximatingaStraightLineModelwithaProportionalModel
6.8CorrectedAkaikeInformationCriterionforModelSelection
6.9Examples
6.9.1TOSTandAICcComparisonofQuadraticandStraightLineModels
6.9.2TOSTandAICcComparisonofStraightLineandProportionalModels
7.Appendix
8.References

1.INTRODUCTION
Validationisoneofthemilestonesinananalyticalprocedure'slifecycle.Itisaconfirmatorysteptodemonstrate,through
systematicexperimentationandformaldocumentation,thattheprocedureisfitforitsintendeduse.Inaddition,resultsof
validationareusefulfortheestablishmentofsystemsuitabilitycriteriatomonitorthelongtermperformanceoftheanalytical
procedure.Ananalyticalprocedureisdeemedvalidatedifitsperformancecharacteristicsareshowntobewithinrequiredlimits
withstatedconfidence.
Fromalifecycleperspective,aholisticapproachtoassessingandcontrollingthevariabilityofaprocedureincludesthree
stages:1)proceduredesign,2)performancequalification,and3)ongoingperformanceverification.Eachofthesestagescan
besupportedbyavarietyofstatisticaltoolsandapproachestoensuregooddecisionmaking.Theperformancequalification
stagegenerallyincludesexperimentsintendedtoconfirmthattheprocedureiscapableofmeetingitsdesignedintent.This
chapterisconcernedwithstatisticaltoolsthatsupporttheperformancequalificationstage.Forthepurposeofthischapter,the
term"validation"referstotheperformancequalificationstage,althoughtheimportanceofalifecycleperspectivethatincludes
gooddecisionmakinginallthreestagesisrecognized.
AlthoughintendedtoserveasacompaniontoValidationofCompendialProcedures 1225 ,thestatisticalmethods
presentedinthischapterarebroadlyapplicabletoprocedurevalidation,notonlyforcompendialpurposes.Thestatistical
methodscanbeappliedtoanalyticalproceduresforbothsmallandlargemoleculeproducts(seeBiologicalAssayValidation
1033 ).Eachofthefollowinganalyticalperformancecharacteristicsisdiscussedfromastatisticalperspectiveinthe
sectionsthatfollow:
Accuracyandprecision
Range
Limitsofdetectionandquantitation
Linearity
Forquantitativeprocedures,validationofaccuracyandprecisionprovidesthemostessentialevidencethattheprocedure
meetstherequirementsfortheintendedanalyticalapplication.Accordingly,thestatisticalhypothesistestingparadigmis
adoptedtoprovethattheanalyticalproceduregeneratesdata(reportableresult)thatissufficientlyaccurateandprecise.The
samplesizeofthevalidationexperimentshouldbedrivenmostlybythepowerconsiderationinprovingbothaccuracyand
precision.Otherfactorsthatarealsocharacterizedbythevalidationexperimentaremoredescriptiveinnature(e.g.,range,
detectionlimit,andquantitationlimit)oraremoreinternaltotheanalyticalprocedure(e.g.,linearity).
Theinformationprovidedinthischapterisnecessarilystatistical.Statisticalproceduresarepresentedassimplyaspossible
withoutlosingscientificrigor.Mostoftherecommendedcalculationscanbeperformedinaspreadsheetpackage,andsome
Excelcommandsareprovidedforthispurpose.Thetargetaudienceconsistsofmembersofaprocedurevalidationteamwho
areresponsibleforplanning,designing,andperformingthemostappropriateandscientificallyvalidanalysisofthedata.
Allofthestatisticalmethodsdescribedrequiretheestablishmentofprespecifiedacceptancecriteria.Theestablishmentof

numericalacceptancecriteriaonwhichtobaseavalidationtestcanbechallenging.Itrequiresconsiderationofmanyfactors,
including:1)knowledgeoftheprocessthatwillbemonitoredwiththeanalyticalprocedure2)pastperformanceofsimilar
proceduresandhistoricalnorms3)thelifecycleoftheanalyticalprocedure4)performanceoftheprocedureduringpre
validationworkand5)futuremaintenanceandcontroloftheanalyticalprocedure.Therequiredperformancecharacteristicsof
ananalyticalprocedurearesometimesreferredtoastheanalyticaltargetprofile(ATP).HowtodeveloptheATPisnotwithin
thescopeofthischapter.Inthischapter,itisassumedthattheATPhasbeenalreadyestablished.
Finally,althoughsomeofthestatisticalmethodsmayappearnew,theyarecurrentlyusedasstandardpracticeinmany
industriesoutsideofthepharmaceuticalindustry.Theyprovidebestpracticestatisticalproceduresforanalyzingvariationand
biasofmeasurementsystems.

2.WORKDONEBEFOREVALIDATION
Procedurevalidationisacornerstoneintheprocessofdevelopingananalyticalprocedure.Theaimofproceduredevelopment
istostudyallareasrelevanttothequalityandapplicabilityoftheprocedure,aswellastocollecttheinformationrequiredfor
optimallydesigningtheformalvalidationexperiments.Itisimportanttorealizethattheobjectiveofthevalidationexperiments
istoverifytheselectedproceduresettings,inputs,operatingconditions,equipment,limits,andranges,andpossiblyother
factorsthatcaninfluenceoutcomeinspecialcircumstances.Theseoperationaldetailsshouldhavebeendeterminedduring
proceduredevelopmentusingappropriatelydesignedexperimentsandthendocumentedinwrittenreports.Issuessuchas
identificationofruggednessfactorsthataffectintermediateprecision,appropriatenessofthenormalprobabilitymodel,needfor
transformation,variancepoolingstrategy,andsimilarquestionsshouldbeansweredbeforevalidation.Surprisingdiscoveries
(whethergoodorbad)abouttheprocedureduringvalidationshouldbeseenasafailureandshouldpromptareturntothe
proceduredevelopmentstage.Thegeneralprinciplesandtheplansforsamplepreparation,experimentaldesign,data
collection,statisticalevaluation,andchoiceofacceptancecriteriashouldbedocumentedinaformalvalidationexperimental
protocolthatissignedbeforeinitiationoftheformalvalidation.
Itisequallyimportanttorealizethatprocedurevalidationisnotaoneoffexperiment.Becausetheprocedurevalidation
acceptancecriteria(suchastherequiredprecision)shouldberelatedtotheuseoftheprocedure,theneedtorevalidate
shouldbeconsideredwheneveruseoftheprocedurechanges.Examplesofsuchchangesinuseoftheprocedureinclude:1)
introductionofanewstrengthoftheproduct,2)transferoftheproceduretoanewlab,3)testingofsampleswithanewtypeof
stresstest,and4)achangeinspecifications.Inanyofthesescenarios,arevalidation(possiblypartial)ismostlikely
appropriate.Sometimesareassessmentofexistingdatatorevisedacceptancelimitsissufficient.
Finally,althoughnotpartofprocedurevalidation,itisrecommendedthatsometypeofstatisticalprocesscontrolbeintroduced
tomonitortheperformanceoftheprocedure.Thiscanbeveryusefulbyprovidingearlywarningofdifferenttypesofdriftina
procedureperformanceparametersuchasprecision.Suchchangesarenotuncommon,andoftenoccurasaresultofwornout
equipment,changeofroutines,oragingreagents.
Althoughthischapterfocusesontheactualvalidationexperiment,someoftheimportantconsiderationsrelatedtothepre
validationworkarediscussedinthenexttwosections.
2.1ExperimentalDesign
Toapproachproceduredevelopment,validation,andmaintenancefromriskbasedandsciencebasedstandpoints,oneneeds
anunderstandingoftheprocedurebasedonaseriesofexperiments.Statisticaldesignofexperiments(DOE),togetherwith
goodpracticalunderstandingofthetaskathand,allowsonetoachievethisobjectivebyminimizingbiasandreducing
measurementvariability.Thisresultsingainsofefficiencyandanimprovedabilitytomakevalidconclusions.
Beforetheinitiationofvalidation,itisrequiredthatoperationaldetailsoftheprocedurearedescribedinawrittenprotocol.An
importantaspectoftheprotocolistoprescribetheallowablerangesforoperationalparameters,suchastemperatureandtime,
thataffecttheperformanceofananalyticalprocedure.Examinationusingexperimentaldesignstoestablishtherobustnessof
suchparameterrangesshouldbeperformedbeforestartingthevalidation.Awrittenreportsummarizingsuchrobustness
experimentsshouldbeincludedinthevalidationpackage.
Efficientexperimentaldesignisalsoneededtodemonstratethattheanalyticalerrorattributabletoprocedureperformance
(measurementerror)reliablyfallswithinacceptablelimits.Correctapplicationofstatisticalconceptsduringprocedure
developmentcanreducebiasandvarianceandhelpensuresuccessfulvalidationofbothaccuracyandprecision.Moreover,
carefulprevalidationworkcanrevealsuitableapproachesforreducingthetotalsizeoftheformalvalidationstudywithout

increasingtheriskofdrawingthewrongconclusion.
ThesystematicDOEapproachismoreefficientthanonefactoratatimeexperimentationandisgenerallymoreefficientwhen
approachedsequentially.Forexample,ascreeningexperimentthatisperformedprevalidationcanprovidevalidation
acceptancecriteria.Thisapproachalsopermitsestimationofprocedurevariationandidentificationofcriticalfactorsthataffect
performance.Factorstobeassessedincludethephysical,chemical,andenvironmentalfactorslikelytoaffecttheprocedure's
response.Fractionalfactorialscreeningexperimentsallowforidentificationoffactorswiththegreatestinfluenceonthe
response.Ofthemanyfactorsexaminedduringscreening,itiscommontofindthatonlyoneortwofactorsarecriticaltothe
controloftheprocedure.Thus,asdescribedinlatersections,theconfirmatoryvalidationneednotinvolveacomplex
experimentaldesign.Inthisway,priorknowledgegainedduringproceduredevelopmentisleveragedtomakeformalvalidation
moresimple,targeted,andefficient.Acombinationofexpertandtheoreticalknowledgewithpreliminaryexperimentationis
usedtoidentifydisturbingfactorsandsuitablerangesforsuccessfulprocedurevalidation.
2.2OtherConsiderations
Prevalidationworkshouldideallyinvestigatethefollowingquestions:
Whatrangeoftrue(sometimesreferredtoasmeasurand)valuesmustbevalidated?
Determinationoftherangetobevalidatedisaniterativeprocessthatcouldpossiblyinvolveprocedure
optimizationtoachievethedesiredrange.
Shouldthetestresultsbecomparedtotruevalues,ortotestresultsfromanestablishedprocedure?
Todefinebias,itisnecessarytodefineatruevalue.Insomeexperiments,thesamplesaremadefrom
referencematerialandtargetedatparticularconcentrations.Inthesecases,thetruevaluesshouldbethe
targetedconcentrations.Inotherexperiments,thesampleswereincurredsampleswithfundamentally
unknownconcentration.However,ifthereisexternalinformationaboutthesesamples,e.g.,measured
concentrationsfromadifferentwellcharacterizedanalyticalprocedure,suchexternalinformationcanbe
usedasthetruevalues.
Itmustbedecidediftheabsoluteorrelativebiasistheprimaryendpoint.
Isbiasconstant,ordoesitvarywiththetruevalue?
Ifthebiasisconstantacrossmultipletruevalues,onecancombine(pool)dataacrosstruevalueswhen
estimatingthebias.Poolingisadvantageousbecauseitincreasesstatisticalpowerwhentestinghypotheses
toestablishaccuracy.Thismeanstheprobabilityofconcludingthataproceduremeetstheacceptability
criterionincreaseswhentheprocedureisindeedfitforpurpose.Thesuitabilityofpoolingcanbeinvestigated
duringprevalidationusingstandardanalysisofvarianceor,ifassumptionsofnormalityandvariance
homogeneityarenotfulfilled,usingnonparametricmethodssuchastheKruskalWallistest(1).
Whatisthemaximalacceptablebias?
Isvariabilityconstant,ordoesitchangeacrosstheexperimentalrange?
Statisticalpowercanalsobeincreasedbypoolingdatawheneitherthestandarddeviationorthenormalized
standarddeviation(NSD)isconstantacrosstheexperimentalrange.(TheNSDisdescribedmorefullyin
section3.1.)DifferencesinstandarddeviationsacrosstherangecanbeinvestigatedusingeitherLevene's
test(2)orBartlett'stest(3).
Isthenormalprobabilitymodelreasonablefordescribingthedata?
Theassumptionofnormalityisimportantforproperapplicationoftheformulasprovidedinthischapter.
Atypicalandoutliertestresultsprovideobjectiveevidenceofpotentialnonnormalbehaviorrequiring
investigationtodeterminetheacceptabilityoftheresults.Ifthevaluesareexpectedtobeslightly
asymmetricalorskewedwithalongtailtolargervalues,alogtransformationmaybeusedtoprovidedata
thataremoreconsistentwiththeunderlyingnormalityassumption.Thenormalityassumptioncanbe
investigatedusingtheShapiroWilk'stest(4)orbyvisualinspectionofnormalquantileplots.Ifdataappear
tobenonnormal,possibleremediesincludedatatransformation,changeofrangeendpoint,increased
samplesize,orconfidenceintervalsbasedonmoreappropriateprobabilitymodels.
Arepreliminaryestimatesofthevariancecomponentsavailable?
Howshouldoutliersbemanaged?
Thepresenceofanoutlierinthedatacouldcauseafalsefailureofaccuracyandprecisionvalidationcriteria.
Asuspectoutliercouldbeinvestigatedusinganyofseveraloutliertests(5).Outliertestsaredescribedin

chapterAnalyticalDataInterpretationandTreatment 1010 .
Whichvariablesshouldbecontrolledintheexperiment,andwhataretheirexperimentalranges?
Whichruggednessfactorsimpactintermediateprecision(asdefinedinsection3.2)andneedtobeincludedinthe
validationexperiment?
Whatistherequiredtargetmeasurementuncertaintyfortheanalyticalprocedure?
Adeterminationofmeasurementuncertaintyhelpsdefinethelikelyvaluesofthetrueanalyteconcentration,
giventheresultobtainedfromtheanalyticalprocedure.ProceduresfordoingsoaregivenintheGuidetothe
ExpressionofUncertaintyinMeasurements(6).Definingatargetmeasurementuncertaintyissometimesa
requirementofstandardsorganizationssuchasISO(InternationalStandardsOrganization).
Basedontheanswerstotheseandsimilarquestions,onecandesignasuitablevalidationexperimentalprotocol.
Statisticaltestsweresuggestedinthelistabovetohelpanswerprevalidationquestionsconcerningmodelassumptions.
However,itisnotrecommendedtomakedecisionsbasedexclusivelyonstatisticaltests.Thisisbecauserejectionofa
statisticalhypothesisisgreatlyimpactedbythesamplesize.Thesmallerthesamplesize,thelesslikelyoneistoconclude
thatthetestedassumptionisnotappropriate.Similarly,largesamplesoftenleadtorejectionofanassumptionbasedon
statisticalsignificance,evenwhenthereisnopracticalimpactofthedetecteddifference.Forthisreason,itisrecommended
touseavisualrepresentationofthedata,andpossiblyasimulationstudyofthesensitivityofresultstotheassumption
deviations,inconjunctionwithastatisticaltest.Bysupplementingastatisticaltestinthisway,itiseasiertoidentifysituations
whereassumptionsareobviouslynotreasonable,aswellassituationswheretheycanbereasonablyapplied.

3.ACCURACYANDPRECISION
3.1IntroductionandDefinitions
Amodelthatisusefulforrepresentingameasuredtestresultis:
MeasuredTestResult=TrueValue+SystematicBias+RandomError

[1]

whereboththeTrueValueandtheSystematicBiasareconstants,andtheRandomErrorisanormalrandomvariablewitha
meanofzero.
Accuracyistheclosenessbetweentheestimatedanalytelevelobtainedwiththetestprocedure(MeasuredTestResult)and
thecorrespondingtrueanalytelevel(TrueValue)oftestsamples.Closeness(orbias)isexpressedasthelongrunaverageof
thetestresultsminustheTrueValue.
Asdiscussedinsection2.2,biasmustbedefinedrelativetoatruevalue.Determinationoftruevaluerequiresexternal
informationandwillvarybythemethodandtheavailableinformation.Forexample,chapter 1225 notesthatareference
standardorawellcharacterizedprocedurecanbeusedtorepresentthetruevalue.Accuracyshouldbeestablishedacrossthe
procedure'srequiredanalyteconcentrationrange.Accuracyisdemonstratedbyestimatingthemagnitudeofbiasand
comparingittoaprespecifiedacceptancecriterion.
Theprecisionofananalyticalprocedureisthedegreeofagreementamongindividualtestresultswhentheprocedureisapplied
repeatedlytomultiplesamplings(possiblyunderdifferentconditions)ofahomogeneoussample.Imprecisionorvariabilityis
thedegreeofdisagreement.Precisionofatestproceduremaybeinfluencedbyvariousruggednessfactors,includinganalyst,
day,instrument,andwithininstrumentvariation.Asnotedearlier,identificationofsuchfactorsisrequiredintheprevalidation
work.Precisionshouldmeetpredefinedacceptancecriteria.
Themostcommonprecisionmetricisthestandarddeviation(SD).TheSDsquarediscalledthevariance.Precisionimproves
astheSDdecreases.Manycommonlyusedstatisticalproceduresrelyontheassumptionofthenormaldistribution,forwhich
theSDisanaturaldescriptorofvariability.
Forassaysthatarebasedonchemicalorbiologicalprinciples,themeasuredresults,whenexpressedinmassor
concentration(mass/volume)unitstendtovarymoreasthelevelincreases.Suchasituationmakesitcumbersometo
combineinformationacrossthevalidationrangeandmayrequireseveraldifferentexperimentstofullyvalidatetheprocedure
overtheentirerange.However,itmightbepossibletostabilizethevarianceovertheentirerangeanddecreasetheamountof
experimentationbyeithernormalizingthedataorapplyingalogtransformationtothedata.
Aformulafornormalizingvalidationdataexpressesthedifferencebetweenameasuredvalueandthetruevalueasa
percentageofanormalizingconstant.Thatis,

wherethenormalizingconstantisproportionaltotheSDofthemeasuredvalue.Inmanycases,thisnormalizingconstantis
thesamenumberthatisusedasthetruevalue.Insomeapplications,thelabelclaimisusedasthenormalizingconstant,and
themeasurementsarereportedas%labelclaim.Iftheanalyticalprocedureisintendedforadrugsubstance(bulkmaterial),
themassofthereferencestandardmaybeanappropriatenormalizingconstant.
TheSDcomputedwithnormalizeddataiscalledtheNSD.Althoughsimilarinconcept,theNSDshouldnotbeconfusedwith
thepercentrelativestandarddeviation(%RSD)asdefinedinbothchapters 1010 andUniformityofDosageUnits 905 .
The%RSDisdefinedasthesampleSD(ofnonnormalizeddata)expressedasapercentageofthesamplemean.The%RSD
iscalledthepercentagecoefficientofvariation(%CV)inthestatisticalliterature.TheessentialdifferencebetweentheNSD
and%RSDisthatNSDhasaconstantinthedenominatorthatisexternaltotheexperiment,andthedenominatorof%RSDis
asamplevaluecomputedwiththeexperimentaldata(i.e.,thedenominatorisarandomvariable).Thus,thereisgreater
uncertaintyassociatedwiththetruevalueof%RSD,anditismathematicallydifficulttoquantifythisuncertaintyinthe
validationexperiment.ByperformingvalidationwiththeNSD,onecanbettercontrolthedegreeofuncertaintyandminimizethe
probabilityofanincorrectvalidationdecision.
Properidentificationofthenormalizingconstantwillallowtheuseofonlyasinglevalidationexperiment,andthestatistical
power(probabilityofmeetingacceptancecriteria)willbeincreased.
Logarithmictransformationsarealsousefulforenablingthepoolingprocess.Forbiologicandvaccineproducts,potencymay
varycontinuouslyacrossseveralmagnitudes.USPchaptersBiologicalAssayChaptersOverviewandGlossary 1030
and 1010 defineotherversionsofprecisionforthisapplicationcalledthegeometricstandarddeviationandthegeometric
coefficientofvariation.
Thetotalvarianceofananalyticalprocedureoftenispartitionedintocomponentsattributabletothedifferentsourcesof
variability.Forthepurposesofthischapter,weconsidertwosourcesofvariability.Thefirstsourceistheobservedvariation
whenananalyticalprocedureisusedrepeatedlytoassessthesamesampleoverashortperiodoftimebyasingleanalyst
usingthesameequipment(whereeachreplicationinvolvestheentireprocessincludingthesamplepreparation).Thisis
referredtoastherepeatabilitycomponentandisdenotedbytheGreeksymbol
,whereEdenoteserrorassociatedwith
repeatability.Thesecondsourceisvariationinadditiontorepeatabilitythatoccurswhenananalyticalprocedureisusedinthe
samelaboratoryunderrandomconditionssuchasdifferentanalysts,equipment,ordays.Theserandomconditionsareknown
asruggednessfactors.ThissourceofvariationisdenotedbytheGreeksymbol

,whereCdenotescondition.Thesumof

thesetwocomponents,
,iscalledintermediateprecision(orruggedness).Itisdenotedbythesymbol
,and
representsthetotalvarianceofananalyticalprocedureusedinaparticularlabundervaryingconditionswithintheexpected
ranges.TheintermediateprecisionSDisdenotedby

3.2ExperimentalDesignforAccuracyandPrecision
Anappropriateexperimentaldesignisneededtoestimatebothaccuracyandprecision.Asnotedearlier,selectionofthis
designshouldbebasedoninformationgainedduringtheprevalidationstage.
AcommondesignusedtoestablishaccuracyforconcentrationlevelsintherangefromCmin toCmax istodefineatleastthree
truevaluesbetweenCmin andCmax ,inclusive.Typically,thethirdvalueistheexpectedresult(e.g.,labelclaim)forasample.
TheabilitytodefineCmin andCmax appropriatelydependsontheextentandqualityoftheworkdoneduringprocedure
development.Ifthelimitsoftheprocedurearenotfirmlydeterminedbeforevalidation,thenitisstronglyrecommendedthat
accuracyandprecisionareevaluatedatmorethanthreeconcentrationlevels.
Toestablishintermediateprecisionwhenthereareidentifiableruggednessfactorssuchasanalyst,equipment,ordays,one
mustcreateanumber(c)ofindependentexperimentalconditionsbasedonthesefactors.Theobjectivewhendesigningthe
experimentistoexplorethefulldomainofoperatingconditionsunderwhichtheprocedureisexpectedtooperate.Thevariation
fromconditiontoconditionisusedtoestimate

.Additionally,anumberofreplicates(r)ofeachconditionisneededto

estimate
.
Considerasituationwheretheobjectiveofthevalidationistoshowaccuracywithintherange75%to125%oflabelclaim.
Evidencecollectedbeforevalidationsuggeststhatafterthedatahavebeennormalized,boththebiasandtheSDareconstant
acrossthisrange.Thus,asingleexperimentaldatasetcanbeusedwithdatacollectedacrosstherangeoflabelclaimwithout
thenecessityofperformingaseparateexperimentforeachlevel.Onthebasisofpowerconsiderationstobediscussedlaterin
thischapter,itisdeterminedthattheexperimentaldesignshouldconsistofnineindependentexperimentalconditions.These
experimentalconditionsarecreatedbycombinationsofruggednessfactorsidentifiedduringtheprevalidationwork(e.g.,
analysts,days,orequipment).AnexampleofsuchanexperimentaldesignisshowninTable1.(Notethatthisdesignisnot
intendedforestimationofvariancecomponentsrelatedtotheruggednessfactors).
Table1.ExperimentalDesign

%ofLabelClaim ExperimentalCondition Prep1 Prep2 Prep3

75%
1
X
X
X
75%
2
X
X
X
75%
3
X
X
X
100%
4
X
X
X
100%
5
X
X
X
100%
6
X
X
X
125%
7
X
X
X
125%
8
X
X
X
125%
9
X
X
X
Itisimportantthatexperimentalconditionsbeasindependentaspossible.Forexample,ifprevalidationworkhasshownthat
thereissignificantanalysttoanalystvariationforaprocedure,thenideally,onewouldhaveadifferentanalystassociatedwith
eachexperimentalcondition.Otherwise,measurementsmadebythesameanalystwillbecorrelated.Practicalconsiderations
maymaketheidealunattainable,buttotheextentpossible,oneshouldtrytomakeeachconditionasindependentfromthe
othersaspossibleovertheentirerangeofexpectedlaboratoryconditions.
3.3RecommendedMethodsforVerifyingAccuracyandPrecision
Thissectionprovidesconfidenceintervalformulastouseinestimatingbiasandintermediateprecision.Confidenceintervals
canbeusedtoperformastatisticaltestofequivalenceagainstpredefinedacceptancecriteria.Toprovidethenecessary
formulas,thestatisticalmodelusedtorepresentthenormalizeddatainTable1is
Yij=+Ci+Eij i=1,...,cj=1,...,r

[3]

whereYijisthenormalizedvalueforthejthreplicateofexperimentalconditioni,isthemeanbias,Ciistheprocedureerror
duetotheithexperimentalcondition,Eijrepresentstheprocedureerrorassociatedwiththejthreplicationfromconditioni,cis
thenumberofexperimentalconditions(c=9inTable1),andristhenumberofreplicatesforeachcondition(r=3inTable1).
TheCiandEijareassumedtobeindependentrandomnormalvariables,eachwithmean0andwithvariances

and

respectively.
TheconsequenceofincludingCiinEquation[3]isthatobservationswithinthesameconditionarecorrelated.Thatis,
observationswithineachrandomconditionaremoresimilarthanareobservationsacrosstheexperimentalconditions.Interms
ofthemodelparameters,thecorrelationbetweentwoobservationswithinthesameconditionisdefinedas
Theratio iscalledtheintraclasscorrelation.Ifthiscorrelationisnotaccountedforinthestatisticalvalidationtest,
uncertaintyisunderestimated,andonewillinappropriatelypassvalidationmoreoftenthandesired.
3.3.1TESTSOFEQUIVALENCEFORSEPARATEASSESSMENTOFACCURACYANDPRECISION

Onegoalofprocedurevalidationistoprovideestimatesofand

.ThestatisticsneededtodothisforEquation[3]are

where

ThestatisticsinEquation[4]canbeobtainedusinganystatisticalpackageorspreadsheetthatcomputesaonewayanalysis
ofvariance.ThetermY(Ybar)istheobservedgrandmean.ThetermS12commonlyisreferredtoastheamonggroupmean
sumofsquares,andthetermS22commonlyisreferredtoasthewithingroupmeansumofsquaresorthemeansquared
error.Thepointestimatorsfortheparametersofinterestare

ThehatsymbolisplacedovertheGreeksymbolsinEquations[5]and[6]tosignifythatthecomputednumberisasample
estimateratherthanthetruevalue.Statisticalconfidenceintervalsprovideaninformativesummaryofthevalidation
experiment.Aconfidenceintervalcontainstheunknowntruevalueoftheparameterwithanassociatedconfidence(e.g.,95%).
Theconfidencelevelof95%definesthequalityofthestatisticalexperimentandmeasurestheabilityoftheconfidenceinterval
tocorrectlycapturethetruevalueoftheparameter.Theconfidenceintervalalsocanbeusedtoperformastatistical
equivalencetestagainstpredefinedacceptancecriteria.
A100(1 2 )%twosidedconfidenceintervalforthebias()isgivenby

wheret1

:c

1representsthepercentileofacentraltdistributionwitharea1

totheleftandc 1degreesoffreedom.

Forexample,with =0.05andc 1=8,t0.95:8=1.860.ThegeneralstatementinExcel2007toobtaint1

:c

1is=

TINV(2* ,c 1).Inthisexample,thestatement=TINV(0.10,8)returnsthevalue1.860.Theselectionof =0.05in

Equation[7]providesa100(1 20.05)%=90%twosidedconfidenceintervalfor.Theapplicationoftheaboveformulasto
anexampledatasetispresentedinsection3.3.3.
Forintermediateprecision,oneisconcernedwithonlythe100(1 )%upperconfidenceboundbecausetheacceptance

criteriaisonesidedtoguardagainstsituationswherethevariationistoolarge.Anupper100(1 )%confidenceboundUGW
for
isbasedonamethodfromGraybillandWang(7).Thismethodiscalledthemodifiedlargesampleconfidenceinterval
andhasbeenrecommendedforbiopharmaceuticalapplicationsbyNijhuisandVandenHeuvel(8).Thisformulais

where

representsthepercentileofacentralchisquareddistributionwithc 1degreesoffreedomandarea tothe

left.Forexample,ifc 1=8,c(r 1)=18and =0.05,

=2.73,

=9.39,H1=1.928,andH2=0.9168.The

generalstatementinExcel2007toobtain
is=CHIINV(1 ,c1).Intheaboveexample,thestatement=
CHIINV(0.95,8)returnsthevalue2.73.Section3.3.3containsaworkedexampleforthisformula.
Theconfidenceintervalsinthissectioncanbeusedtotestwhethertheanalyticalprocedureisfitforpurposebyperforminga
twoonesidedtest(TOST)ofstatisticalequivalence(9).Mosttypically,theTOSTusesatestsizeof5%.Thetestsizeisthe
maximumriskofdeclaringthattheacceptancecriterionissatisfied,whenintruthitisnotfulfilled.Forexample,supposethe
datahavebeennormalizedandthepredefinedacceptancecriterionrequiresthebiastobebetween 6.0%and+6.0%.Ifthe
entire100(1 2 )%twosidedconfidenceintervalfallswithintherangefrom 6.0%to+6.0%,thenithasbeendemonstrated
thatthetruebiasislessthan6%withatypeIerrorrateof .Thus,ifthedesiredtestsizeis5%, =0.05andthetwosided
confidencecoefficientis90%.
ThesameTOSTequivalenceapproachcanbeusedwiththeconfidenceintervalinEquation[8]tovalidateprecision,except
thatthistestisone(upper)sided.Forexample,supposethepredefinedacceptancecriterionrequiresthenormalized
intermediateprecisionSDtobelessthan3%.Ifthesquarerootoftheupper100(1
)%confidenceboundonthevariance
showninEquation[8]islessthan3%,thentheprecisionhasbeensuccessfullyvalidated.
3.3.2COMBINEDVALIDATIONOFACCURACYANDPRECISION

Whenassessingwhetherananalyticalprocedureisfitforitsintendedpurpose,itisimportanttounderstandtherelationship
betweenbiasandprecision.Thedegreetowhichthebiasaffectstheusefulnessofananalyticalproceduredependsinparton
theprecision.Thatis,aprocedurewitharelativelysmallintermediateprecisioncanacceptagreaterbiasthanaprocedure
withalargerintermediateprecision.Forthisreason,itisusefultoestablishasinglecriterionthatcanbeusedto
simultaneouslyvalidatebothaccuracyandprecision.Furthermore,becausetheintendedpurposeofananalyticalprocedureis
toprovideaccurateandprecisemeasurementsofsamples,onemayconsiderthattheprocedureisvalidatedifitisshownto
provideahighdegreeofassurancethatthetestresultsofthefuturesampleswillbeclosetotheirtruevalues.Onesuch
criterionproposedinaseriesofarticlesbyHubertetal.(1012)seekstoensure
Pr(

Y )

[9]

whereYisthenormalizedvalueofafuturesample, >0isanacceptablelimitdefinedaprioritobeconsistentwiththe
purposeoftheprocedure,and isthedesiredprobabilityforafuturemeasurementtohaveanerrorwithinthedefined
acceptablelimit(e.g., =0.99).
ThetestingstrategyforEquation[9]isbasedona expectationtoleranceinterval.Inparticular,ifa expectationtolerance
intervalfallscompletelywithintherangefrom to+ ,thenonecanclaimthatEquation[9]issatisfied.Aformulato

computethe100 %expectationtoleranceintervalforthemodelinEquation[3]ispresentedbyMee(13)as

where

HahnandMeeker(14)notethe expectationtoleranceintervaliscommonlyreferredtoasa100 %predictionintervalfora

futureobservation.Thus,itisinterpretedasarangethatwithagivenlevelofconfidence( ),willincludethenextobserved
normalizedvalue.
Anothertoleranceinterval,the contenttoleranceinterval,isusedtoprovidearangethat,withagivenlevelofconfidence,
includes100 %ofallfuturenormalizedvalues.HoffmanandKringle(15)recommendusingthe contenttoleranceinterval
tosimultaneouslyvalidatebothaccuracyandprecision.Atwosided contenttoleranceintervalthatcanbeusedwiththe
modelinEquation[3]is

where
[8],and

representsastandardnormalquantilewitharea

totheleft,UGWisthe100(1 )%upperboundinEquation

iscomputedusingEquation[6].Forexample,with =0.99,thenZ0.995=2.576.Thisvalueisobtainedusing

theExcelfunction=NORMINV(0.995,0,1).
ItispossibletoestimatethelefthandsideofEquation[9]directlyusingaBayesianapproach.Inparticular,onecanusethe
Bayesianapproachtoestimate
Pr(

Y )

andthencomparethisprobabilitydirectlytothedesiredvalueof .Theprocedureisvalidatedifthecomputedprobability
exceeds .
Ingeneral,Bayesiananalysisprovidesaframeworkformakinganinferenceabouteitherafutureobservationormodel
parametersbasedonnewdataandpriorbeliefs.ItbeginswithaparametricmodelsuchasEquation[3]fromwhichalikelihood
functionisderivedandassignmentofpriorprobabilitydistributionstoallfactorsthataccountforuncertaintiesinthe

parameters.Suchprobabilitydistributions,representingpriorbeliefsofthemodelparameters,usuallycanbeestimatedfrom
historicaldata.UsingBayes'rule,theposteriordistributionofmodelparameterscanbeobtainedbymultiplyingthelikelihood
functionofthenewdataandthepriordistributionoftheparameters.Thedistributionoffuturemeasurementscanbederived
fromthisposteriordistribution.
Bayesiananalysisappliedtotheprocedurevalidationprocesscombinesbothknowledgeandunderstandingoftheprocedure,
intermsofpriorbeliefs,withnewdatageneratedfromthevalidationstudy.Thepriorandnewdataarecombinedtopredictthe
behavioroffuturemeasurementswithregardtobiasandprecision.Theposteriordistributionfromoneexperimentcanserveas
thestartingpointforthepriordistributionforasubsequentexperiment.Bayesianmethodstherebyprovideacontinuous
learning,lifecycle,andriskbasedapproachandcanbeusefulforqualityriskassessment.
ABayesiantoleranceintervaloffuturemeasurementsprovidesaninterpretationthatcanbeusedtoevaluateEquation[9]
directly.ABayesiantoleranceintervalconsistentwithEquation[3]isprovidedinWolfinger(16)andcanbecomputedusingthe
statisticalsoftwarepackageWinBUGS(17,18).Bayesiananalysescanbechallenging,andtheaidofanexperienced
statisticianisrecommended.Moreinformationisprovidedin(19).
3.3.3NUMERICALEXAMPLE

Table2presentsadatasetconsistentwiththeexperimentaldesigninTable1.ThedatainTable2areintheoriginalformwith
unitofmeasurement%ofLabelClaim.Valuesshowninthefirstcolumnareassumedtobethetruevalues.
Table2.ExampleDataSet(RawData)

%ofLabelClaim ExperimentalCondition Prep1 Prep2 Prep3

75%
1
76.050 73.950 76.500
75%
2
75.900 73.650 75.450
75%
3
76.350 74.400 74.325
100%
4
100.600 99.700 99.600
100%
5
99.400 99.200 98.200
100%
6
102.500 102.500 102.800
125%
7
123.375 125.875 123.875
125%
8
128.875 127.250 127.125
125%
9
125.750 123.250 124.000
Figure1presentsaplotofthedataaftersubtractingthetruevaluefromeachresponse.Thenumberaboveeachcolumnof
circlesisthecolumnrange(maximumtominimum).Notefromthisfigurethatthespreadofthevaluesincreasesacrossthe
rangeof%labelclaim.Thissuggeststhat%labelclaimisanappropriatenormalizingconstant.

Figure1.PlotofmeasuredminustrueinTable2by%LabelClaim(withrangeofvalues).
Tostabilizethevariancesandallowpoolingacrossthelabelclaimrange,Equation[2]isusedtonormalizethedatausingthe
truevalueasthenormalizingconstant.Forexample,thevalue76.050forthefirstprepofexperimentalcondition1is
normalizedas

ThecompletesetofnormalizeddataisprovidedinTable3.
Table3.ExampleDataSet(NormalizedDatain%ofTarget)

Target%ofLabelClaim ExperimentalCondition Prep1

75%
1
1.4
75%
2
1.2
75%
3
1.8
100%
4
0.6
100%
5
0.6
100%
6
2.5
125%
7
1.3
125%
8
3.1
125%
9
0.6

Prep2
1.4
1.8
0.8
0.3
0.8
2.5
0.7
1.8
1.4

Prep3
2.0
0.6
0.9
0.4
1.8
2.8
0.9
1.7
0.8

Figure2presentsaplotofthedatainTable3demonstratingamoreconsistentspreadacrosstherangeoflabelclaimthanthe
plotinFigure1.

Figure2.Plotofnormalizeddatain%ofTarget(withrangeofvalues).
ThedatainTable3arenowusedtoconstructconfidenceintervalsandtoperformvalidationacceptancetests.Thesedataare
assumedtobebasedonreportablevaluesconsistentwiththeintendeduseoftheanalyticalprocedure.Therequiredstatistics
neededtocomputethedesiredconfidenceintervalsareY=0.374,S12=4.672,S22=1.299,c=9,andr=3.
AssumethatthepreselectedcriteriainTable4havebeenestablishedbeforevalidation.(Notethatthevaluesinthistableare
forillustrativepurposesonly.Theyarenotintendedtoserveasgenerallyrecommendedcriteria).Theselectedtestsize( )for
alltestsofequivalenceis0.05inthisillustration.Thecriterionforthe contenttoleranceintervalisnecessarilywiderthanfor
the expectationtoleranceintervalbecausethefocusoftheinferenceisonalargersetoffuturevalues.
Table4.PreselectedAcceptanceCriteria

Test

AcceptanceCriterion
Bias
Between 5%and+5%
IntermediatePrecisionSD
Lessthan3%
expectationToleranceInterval Between 10%and+10%
contentToleranceInterval

Between 15%and+15%

FromEquation[7]with =0.05,the100(1 2 )=90%twosidedconfidenceintervalforthebias(%)is

wheret0.95:8=1.860.Theconfidencecoefficientis90%becausetheTOSTtestofequivalenceconsidersatwosided
acceptancecriterionforbias.Becausethecomputedinterval[ 0.4%1.1%]fallswithintherangeof 5%to+5%specifiedin
Table4,theprocedureisvalidatedforaccuracy.
TheestimateoftheintermediateprecisionSDfromEquation[6]is

FromEquation[8]the95%upperboundontheintermediateprecisionvariancewithH1=1.928andH2=0.9168is

ThesquarerootoftheboundinEquation[16]providesthe95%upperboundon

IPof

=2.4%

Because2.4%islessthan3%asrequiredinTable4,theprocedureisvalidatedforprecision.
Nowconsiderthecombinedcriteriadescribedinsection3.3.2.FromEquation[10],the expectationtoleranceintervalwith
=0.99is

Becausetheboundsfallintherangefrom 10%to+10%asrequiredinTable4,theprocedurehasbeenvalidated.

FromEquation[12],thetoleranceintervalthatcontains100 %=99%ofthefuturemeasurementswith95%confidenceis

Thisintervalalsovalidatestheprocedurebasedonthecriteriaof 15%to15%showninTable4.
3.3.4POWERCONSIDERATIONS

Oncetheacceptancecriterionisselected,validationexperimentsshouldbeproperlypoweredtoensurethatthereare
sufficientdatatoconcludethataccuracyandprecisioncanmeetprespecifiedacceptancecriteriawithconfidence.Todothis,
astatisticalpowercalculationshouldbeperformedtodetermineappropriatevaluesforcandr.Statisticalpowerisdefinedas
theprobabilityofpassingthestatisticaltestasafunctionofthetruevalueoftheparameterofinterest.
Forexample,considerthevalidationofprecisionthatrequiresthesquarerootofthe95%upperboundinEquation[8]tobe
lessthan3%.Figure3presentspowercurvesfortwostatisticaldesignswithtwovaluesoftheintraclasscorrelation,

Figure4providespowercurvesfordifferentcombinationsofcandrwherecr=24.Thecurvesarebasedonacomputer
simulationof100,000valuesforeachvalueof

Figure3.Powercurvesforvariousvaluesofc,r,and .
.

Figure4.Powercurvesfortotalsamplesizeof24with =0.5anddifferentnumberofconditions.
Notethatpowerisafunctionofthesamplesizeusedinthedesign(candr),thetruevalueof

IP,andtheintraclass

correlation, (0.1and0.9inFigure3and0.5inFigure4).AsshowninFigure3,forafixedsamplesizeandfixedvalueof
IP,thepowerdecreasesas

increases.Thus,if isrelativelylarge,itisimportanttoensurethatthenumberofrandom

conditions,c,issufficientlylarge.AsshowninFigure4,greaterpowerisobtainedbyincreasingcratherthanrforafixedtotal
samplesize.

4.RANGEDEFINITIONS
Therangeofananalyticalprocedureistheintervalbetweentheupperandlowerlevelsofanalyte(includingtheselevels)that
havebeendemonstratedtobedeterminedwithasuitablelevelofprecisionandaccuracyusingtheprocedureaswritten.
Accuracyandprecisionrefertotheuncertaintyinthereportedresultsobtainedfromtestedsamples.Ingeneral,rangeisa
summarydescriptionofwhereprecisionandaccuracycriteriaaremet.Noseparateanalysesareneeded.Seesection3for
accuracyandprecision.

5.LIMITSOFDETECTIONANDQUANTITATION
5.1IntroductionandDefinitions
Thelimitofdetection(LOD)andlimitofquantitation(LOQ)aretworelatedquantitiesthataredeterminedinthevalidationof
CategoryIIprocedures(seechapter 1225 ).Theseareproceduresfordeterminationofimpuritiesordegradationproductsin
drugsubstancesandfinishedpharmaceuticalproducts.Onlyoneisneededforeachuse,namelyLOQforquantitativetests
andLODforqualitativelimittests.LOQiscalledforwhenevertheanalyticalprocedureyieldsaquantitativereportablevalue,
regardlessoftheformoftheacceptancecriterion.Forexample,foratestoftheformNMTxx%,theLOQneedstobelower
thanthestatedlimitofthetest(xx%)sothatthetestmaysubstantiatewhethertheamountofanalyteisaboveorbelowthe
allowedmaximumlimit.LODiscalledforwhenthereisnoquantitativereportablevalue.Forexample,iftheacceptance
criterionisbasedonacomparisonofpeakareas(withoutdeterminingaconcentration),theLODneedstobelessthanthe
concentrationofthecomparisonstandard.Theselimitsarealsoknownunderothernames,includingdetectionlimit(DL)for
LODandlowerlimitofquantitation(LLOQ)forLOQ.
Thefollowingdefinitionsareconsistentwithchapter 1225 andICHQ2(20)

Thelimitofdetectionisthelowestamountofanalyteinasamplethatcanbedetected,butnotnecessarilyquantitated,
underthestatedexperimentalconditions.
Thelimitofquantitationisthelowestamountofanalyteinasamplethatcanbedeterminedwithacceptableprecisionand
accuracyunderthestatedexperimentalconditions.NotethatagivenproceduremayhavemultiplevaluesofLOQ,depending
onitsapplication,asacceptableprecisionandaccuracymayvarybetweenapplications.
Thegeneralapproachistofirstestimatecandidatevalues(s)forLODorLOQ.Thecandidatevaluemustthenbeverified.This
isparticularlyimportantforLOQ,astheformulasfordeterminingcandidatevaluesdonotaddresstheacceptableaccuracyand
precisionrequirement.Thedeterminationofcandidatevaluesshouldbedoneprevalidation,withonlytheverificationstepas
partofvalidation.
5.2EstimationofLOD
ThebasicapproachtoestimatingLODisbasedonanalternativedefinitionadoptedbyInternationalUnionofPureandApplied
Chemistry(IUPAC)andISO.Thisdefinitionintroducestheconceptsoffalsepositiveandfalsenegativedecisions,thus
recognizingtheriskelementsinusingtheLODfordecisionmaking,andmakesclearthatthesevaluesaredependenton
laboratorycapability.
TheIUPAC/ISOdefinitionofLODisbasedontheunderlyingconceptofacriticalvalue(RC),definedasthesignal(readout,R)
thatisexceededwithprobability whennoanalyteispresentthatis,
RC=B+Z1
whereBistheestimatedmeanreadoutforblanks,Z1

[20]

isastandardnormalquantilewitharea1 totheleft,and

Eis

thetruerepeatabilitySD(seeFigure5).Forexampleif =0.05,1 =0.95,andZ0.95=1.645isobtainedusingtheExcel

function=NORMINV(0.95,0,1).
Thisdeterminationdependsonthedistributionofvaluesobtainedwhenanalyzingblanks.TheLODinthesignalspace(RD)is
definedasthatvalue,whichiftrue,issuchthatRCisexceededwithprobability1 ,namely
RD=RC+Z1

[21]

SolvingEquations[20]and[21]forRD,wehave
RD=B+(Z1

+Z1

[22]

Z1 isastandardnormalquantilewitharea totheleft.Notethatthisdefinitionallowsfortwovaluestobeselectedby
thelaboratory: and ,whichneednotbeequal.Thesymbol representsthetypeIorfalsepositiveerrorrate,andthe
symbol representsthetypeIIorfalsenegativeerrorrate.InFigure5,RCandRDareillustratedwith = =0.05for
normallydistributeddatasothatZ1

=Z1

toacommonruleforRD,namelyB+3.3

=1.645.Althoughthevaluesof and neednotbeequal,thischoiceleads

E(3.3

21.645.)

Figure5.DeterminationofRCandRD.
TheLODontheconcentrationscaleisthenfoundbyconvertingthevalueinthesignalscale,RD,tooneintheconcentration
scale,LOD,asshowninFigure6.Thissteprequiresthatthesignal(R)versusconcentration(X)line,R=B+mX,atlow
concentrationsaswellas

Ebeknownexactly.TheLODontheconcentrationscaleisthencalculatedas

Figure6.DeterminationofLODfromRD.
Asastatisticalprocedure,thisisincompleteintwoways.First,because
besttoestimatethisparameter.Thisiscomplicatedbecause

Eisgenerallyunknown,itmustbedeterminedhow

Eistypicallyconcentrationdependent.Twocommonchoices

are:1)theSDoftheblankresponses,and2)theSDobtainedfromdeviationsabouttheregressionlineofsignalon
concentration.ThechoiceneedstobethevaluethatbestrepresentstheSDintheneighborhoodoftheLOD.Laboratorieswill
oftenpickagreatestworstcasevaluefortheSD.IftheLODoftheprocedureisstillsuitableforitsintendeduse,the
laboratoriesareprotectedagainstpickingavaluethatistoosmallandunderstatingtheLOD,whichwouldresultinaninflated
typeIIerrorrate( )andadeflatedtypeIerrorrate( ).
Thesecondaspecttobeconsideredishowtoincorporatethefactthattheslopeoftheregressionlineofsignalon
concentrationandSDaboutthelineareestimatedandnotknownexactly.Becausethelineisestimated,theestimateof
usedtodetermineRDinEquation[22]istoosmall.Thisiscorrectedbyusingstatisticalpredictionintervalsforfuture
observations,i.e.,usingintervalsaboutthelineratherthanthelineitself.Thepredictionintervalstakeintoaccountthe
uncertaintyintheestimatedlineaswellasthevariabilityassociatedwithafutureobservation.
Theexpandedformulaforthecriticalvalue,RC,originallydefinedinEquation[20]thataccountsforthisuncertaintyis

whereBistheestimatedinterceptofthefittedcalibrationline,theXi'saretheconcentrationvaluesusedindeterminingthe
line,andt1

:N 2isthecentraltquantilewithdegreesoffreedomN 2andarea1 totheleft.Asdemonstratedin

Equation[7],t1

:N 2canbecomputedusingtheTINVfunctionofExcel.Equation[24]differsfromEquation[20]because

thetdistributionisusedinsteadofthenormaldistributionforthemultiplier,andtwoadditionaltermsappearinthesquareroot
tocapturetheuncertaintyoftheslopeandintercept.
AsecondequationforRC[25]answersthequestion,Abovewhichconcentrationcanwebeconfidentthatwewillobtain
signalsthataredistinguishablefrombackground?Thisquestionisansweredbyusingthelower100(1 )%prediction
boundofthecalibrationcurve(seeFigure7).Figure7issimilartoFigure6butusestwodashedcurvesinsteadofthesolid
calibrationline.Here

.ByequatingEquations[24]and[25]andcancellingtheBterms,wethenhaveanequationforLOD:

Equation[26]isaquadraticequationforLODthatcanbesolvedexactlyorbyusingiterativesearchtoolsavailablein
spreadsheets.Aslightlyconservative(overlylarge)approximationforLODthatdoesnotrequireaquadraticsolutionis
obtainedbyassumingthatLODisnegligiblecomparedtoX.Theresultingequationunderthisassumptionis

whichissimilarinformtoEquation[23].BothEquation[26]andEquation[27]recognizethegeneralcasethatthetwoerror
probabilities, and ,maybedifferent.Oftentheyarebothtakenasequalto0.05,asintheexamplethatfollowsinsection
5.3.

Figure7.DeterminationofLODusingpredictionbounds.
5.3LODExample
ThedatainTable5areusedtodemonstratecalculationoftheLOD.
Table5.DataforLODExample

Concentration(X) Area(signal)
(/mL)
(mAUmin)
0.01
0.00331
0.02
0.00602
0.05
0.01547
0.1
0.03078
0.15
0.04576
0.25
0.07592
Fittingastandard(unweighted)linearregressiontothesedatayieldstheregressionline:
Area=0.000235+0.3032Concentration
[27a]
ThevaluesneededtocomputeLOQasshowninEquation[27]with = =0.050areprovidedinTable6.
Table6.StatisticsNeededtoComputeLOQinConcentrationUnits

Statistic
N
m(slope)
S

Value
6
0.3032
0.0002

t1
t1

:N 2=t0.95:4=tinv(0.1,4) 2.132
:N 2=t0.95:4=tinv(0.1,4) 2.132

0.0967
0.0419

ThevalueofLODcomputedfromEquation[27]is

5.4EstimationofLOQ
TheimportantconsiderationindeterminingtheLOQistheacceptableprecisionandaccuracyportionofthedefinition
providedinsection5.1.Ideally,thelaboratoryknowswhatLOQisrequiredfortheprocedure,basedontheintended
application.Inthatcase,thevalidationproceedsbydocumentingtheaccuracyandprecisionintheneighborhoodofthe
requiredLOQ.Intheabsenceofsuchknowledge,orwhenthelaboratorywantstodeterminehowlowtheLOQmightbe(e.g.,
forpotentialotheruses),thenthelaboratorycanstartwithpotentialLOQvaluesgreaterthanbutneartheLOD.Alternatively,
methodsfordeterminingtheLODcanbeadaptedtotheLOQascandidatestartingvalues.Essentially,theformulausedto
computeLODinEquation[27]canbeusedtocomputeLOQbyreplacing(t1

:N 2+t1

:N 2with10.Valuesotherthan

10canbeusedifjustified.Oncecandidatevalueshavebeenobtained,oneshouldproceedtoverifywhethertheaccuracyand
precisionatthosevaluesmeetrequirements(seesection3).
5.5FinalCautionsandAssumptions
Ageneralcautionforallofthemethodspresentedinsection5isthattheyarebasedontwoassumptions:linearityand
homogeneityofvarianceacrosstherangeofconcentrationsusedindeterminingthecalibrationcurve.Neitherisanecessary
assumption.Thecalibrationcurvemaybenonlinear,andaweightedleastsquaresapproachcanbeusedtoallowfor
heterogeneityofvariance.IfthecurveisnonlinearortheconcentrationvariancesvarygreatlyintherangeoftheLODand
LOQ,itisbesttoseekexpertstatisticalhelpindefiningLODandLOQ.Ifvariabilityaboutastraightlineexistsbutisnotlarge,
anunweightedregressionofthecalibrationcurvewillprovideanaveragevariabilitythatcanbeusedintheLODandLOQ
formulas.
Otherproceduresthanthoseshownabove,suchassignaltonoiseratios,canbeusedtoestimateLODandLOQ.Ineither
case,analystsshouldconsiderthesevaluesaspreliminaryandproceedtoverifythem,particularlyiftheyfallbelowthe
concentrationvaluesusedindeterminingthecalibrationcurve.Verificationmeansanalyzingsampleswithconcentrationsnear
thepreliminaryLODandLOQ.ConsiderationshouldbegiventohowlowanLODandLOQarerequiredfortheproceduretobe
suitable.Forexample,ifdataarealreadyavailableatalevelbelowtherequiredLODandasignalwasdetectableatthatlower
value,thenthatlowervaluemaybetakenasaverifiedLOD.Thereislittlevalueinfurtherverification,giventhecurrent
requirement.However,therecouldstillbevalueinverificationofalowervalueincasetherequirementchanges.

6.MODELLINGTHECALIBRATIONRELATIONSHIP(LINEARITY)

6.1IntroductionandDefinitions
Thelinearityofananalyticalproceduresometimesreferstotherelationshipbetweenreportedandknownsample
concentrations.Thistypeoflinearityisdirectlyconcernedwithrelativeaccuracy(seesection3.0)andisnotthesubjectofthis
section.
Thissectionconcernsthecalibrationrelationshipbetweenthecalibrator(orstandard)concentrationandprocedureresponse
acrosstheanalyticalrange.Inmostapplications,thesampleandstandardmatricesarenotidentical.However,thetrue
calibrationrelationshipisapropertyofthestandardreferencematerialthatdoesnotinvolvesamplematrixconsiderationsand
isinternaltotheanalyticalprocedure.Furthermore,thissectionaddressesonlythesystematicbiasinducedbythechoiceof
mathematicalmodelforthestandardcurve.
6.2ReasonsforStudyingtheCalibrationRelationship
Thecriticalperformancemeasureforaprocedureisitsabilitytoprovideadequatelypreciseandaccuratereported
concentrationsacrosstheanalyticalrange(domain).Aswithchapter 1225 ,thetermaccuracyinthiscontextrefersto
unbiasedness.Acalibrationmodelisusuallyonlyanapproximationofthetruecalibrationrelationship.Acalibrationmodelthat
isapoorapproximationcanintroducebiasinreportedconcentrationsatcertainregionswithintheanalyticalrange.
Adequacyofthecalibrationmodelisideallyconfirmedduringprevalidation.Aspartoftheseprevalidationstudies,therelative
contributionofthecalibrationmodeltoprocedurebiasshouldbedetermined.Ifthiscontributionisimportant,itmaybe
appropriatetoincludeanassessmentofthecalibrationmodeltogetherwiththerangeaspartoftheformalprocedure
qualificationstudy.Whethertoincludesuchanassessment,andtheburdenofproofrequired,aredecisionsbestmadeaspart
ofanoverallriskassessment.
Beforeformalvalidation,itisdesirabletoexplorecalibrationoptionsandselectacalibrationmodelthatdoesnotcontribute
measurablytobias.Choosingtherightcalibrationmodelthussupportsgoodprocedureaccuracy.Forinstance,insomecases
itmaybeefficienttoassumeastraightlinerelationshipwheninfactthetruerelationshiphasslightcurvature.Inothercases,
itmaybeefficienttoassumeaproportionalrelationship(straightlinepassingthroughtheorigin),wheninfactthetrue
relationshipformsastraightlinebutdoesnotpassthroughtheorigin.Themethodsdescribedinthissectionareintendedto
aidinfindinganappropriatecompromisebetweenefficiencyandaccuracyinsuchcases.
Becauseoftheefficienciesprovidedwhenacalibrationrelationshipcanberepresentedbyastraightline(e.g.,onlyoneortwo
calibratorsrequired),modelingacalibrationrelationshipisoftenreferredtoasalinearitystudy.However,theformerterm
betterdescribestheintentandconductofsuchstudies.
Assessmentofaccuracyiscriticalbecauseitisatoplevelprocedureperformancecharacteristic.Properidentificationofthe
calibrationrelationshipsupportsthescientificfoundationoftheanalyticalprocedurebyprovidingdeeperinsightintothecauses
andmagnitudesofanyprocedureinaccuracies.However,calibrationrelationshipinadequacyisonlyonepossiblesourceof
procedureinaccuracy.Other,andperhapsmoreimportant,sourcesincludecontributionsfromsampletype/matricesbeing
tested.Dependingontherelativemagnitudeofinaccuracycontributedbyvarioussources,ensuringanadequatecalibration
relationshipmaynotrequirethesameburdenofproofasdemandedforensuringoverallaccuracy.
6.3CurrentPractice
Aninitialvisualexaminationofaplotofanalyticalsignalsasafunctionofanalyteconcentrationisrecommendedaccordingto
1225 .Amongcurrentpractices,visualexaminationremainsthefirststepforassessinglinearity.However,duringpre
validationitisimportanttodevelopobjectiveexperimentalevidencetojustifythechoiceofthecalibrationmodel.Validationis
aregulatedconfirmatoryexercise,anditismoreusefultohaveawelldefinedevidentialstandardwithinterpretablemetrics.
Thus,visualinspectionshouldbesupplementedwithstatisticalevidencethatthechosencalibrationmodelcontributes
negligiblytobiasinthereportedconcentrationsacrosstheclaimedrange.
ThePearsoncorrelationcoefficient(R)isawidelyusedlinearityindicator.However,itisnotanappropriatemetricforthis
purposebecauseverydifferentdatapatternscanhaveidenticalcorrelationcoefficients,andacorrelationcoefficientveryclose
toonecanresultfromanobviouslycurvilinearrelationship(21).Requiringthecorrelationcoefficienttoattainsomelevelof
statisticalsignificanceisalsonothelpfulbecauseevenaweakcorrelationcoefficient(e.g.,0.2)canattainstatistical
significancewithasufficientlylargesamplesize.Further,itisimpossibletosetathresholdforthecorrelationcoefficientorits
statisticalsignificancebeyondwhichthecalibrationrelationshipcanbeconvincinglyestablished.Statisticaltestsof

hypothesesformodelparameters,suchasslopeandinterceptsuffersimilardifficulties.
Rejectingacalibrationmodelbasedonalackoffit(LOF)Ftest(22,23)isalsoacommon,butproblematic,practice.TheLOF
testrequiresindependent,replicatetestingatmultiplecalibrationlevels.Intuitivelyitseemspreferabletoavoidreplicationin
favoroftestingmorelevelstoelicitthetruecalibrationrelationship.Finally,aswiththecorrelationcoefficient,unimportant
deviationsfromaproposedcalibrationmodelcanbefoundstatisticallysignificant,andimportantdeviationsmaynotbefound
statisticallysignificantdependingonsamplesize.
BoththeLOFFtestandthecorrelationcoefficientsignificancetestscompareaproposedcalibrationmodeltoarangeof
alternatives,includingunrealistic,nonmonotonicmodels(includingcircularpatterns)thatwouldhavebeenruledoutearlyin
methoddevelopment.Itwouldbescientificallypreferabletolimitalternativestoarealisticclassofmonotoniccurves.Mandel's
Ftest(24),forinstance,comparesastraightlinetoaquadraticalternative.Lesscommonapproachestocalibration
qualificationincludethequalitycoefficient(22)andtheMarkWorkmantest(23,25).However,theseapproacheshave
deficienciessimilartothoseofthemethodsdiscussedabove.
Moststatisticaltests,includingthosediscussedabove,givethebenefitofdoubttothesimplercalibrationmodel.Yettheycan
provideevidenceonlyagainst,butnotfor,thesimplermodel.Thisseemstobetooweakanevidentialstandardforvalidation
purposes.Analyticalprocedureswithgoodprecisionmaybepenalizedwithahighfalserejectionrate,whilepoorprecisionis
rewardedwithfalseconfirmationofthesimplerandmoreconvenientmodel.
Statisticaltestsgivingthebenefitofdoubttothemorecomplexcalibrationmodel,andthususingahigherevidentialstandard,
areavailable(26).Unfortunately,themetricsarecomplicatedandarebasedonthesignal,ratherthantheconcentrationscale.
Thismakesitchallengingtosetfitforpurposeacceptancelimitsthatlinktopracticaleffects.Inthenextsection,wepropose
twoapproaches.Oneofthese,theTOSTapproach,isbasedonahigherandmoreappropriateevidentialstandard,the
acceptancelimitsofwhichdirectlylinktobiasinreportedconcentration.TheotheroneisbasedonthecorrectedAkaike
informationcriterion(AICc),whichisacommonmetricinstatisticalmodelselection.
6.4CalibrationModelDefinition
Thetermcalibrationmodelreferstoastatisticalfunctionthatdescribesacalibrationrelationshipbetweenobservedresponse
(signal)andstandardconcentration.LetYirepresenttheithresponseatconcentrationx i,withi=1,...,N.Threecommon
calibrationmodelsthatdescribetherelationshipbetweenYiandx iare:
ProportionalModel:Yi=

propx i+E i

[29]

StraightLineModel:Yi= 0str+ 1strx i+Ei

[30]

QuadraticModel:Yi= 0quad+ 1quadx i+ 2quadx i2+Ei

[31]

TheGreeksymbol representsanunknownmodelparameterthatdeterminestheshapeofthecalibrationrelationship.Itis
nottobeconfusedwithitsearlierusageasaprobability.ThevariableEirepresentsarandomerrorcreatedbythe
measurementprocessinthedeterminationofY.Themethodsdescribedbelowassumetheerrortermsarenormallydistributed
withameanofzeroandwithmodelspecificSD.
Anassumptionofproportionalityshouldbeexploredduringproceduredevelopmenttoconsiderroutineuseofasinglecalibrator
level.Ifanassumptionofproportionalityisnotjustified,thenacalibrationmodelconsistingofastraightinewithnonzero
intercept(Equation[30]above)maybeconsidered.Inthiscase,thestraightlineassumptionshouldbeexploredduring
proceduredevelopment.Inthissectiontheconceptoflinearityisbroadenedtoincludebothproportionalandstraightline
relationships.Statisticaltoolsforexploringeachmodelarerecommended.
Moregenerally,acalibrationrelationshipmaybecurvilinearandthusmayrequiremorecomplexmodelsforcalibration.
However,modelsmorecomplexthanthequadraticarenotconsideredhere.Notethatnonlinearmodelssometimescanbe
madelinearbyasimpledatatransformation,suchasalogtransformation.
6.5MethodsfortheAssessmentofLinearity
Agivencalibrationmodelisexpectedtoholdoveraspecificrangeofanalyteconcentrations(CmintoCmax ).Weassumethat

sucharangehasbeenproposedduringproceduredevelopmentwork.Inthischapter,wefocusonstatisticalmethodsthat
demonstratethatthecalibrationrelationshipisrepresentedbyanappropriatemodelwithinsuchaprespecifiedrange.
Statisticalmethodsfordeterminingthisrangebasedondevelopmentdatacanbeadaptedfrommethodsdiscussedherebut
areoutsidethescopeofthisdiscussion(seealsosection3.0).
Asageneralprinciple,itispreferabletorepresentthetruecalibrationrelationshipwithasimplemodelasopposedtoa
complexmodel.Thatis,onewantsacalibrationmodelwithasfewparametersasnecessarytoadequatelyapproximatethe
truecalibrationrelationship.Dependingontheriskassociatedwiththechoice,twoevidentialstandardsarerecommended.
ThefirstapproachusestheTOSTofequivalence(9)toevaluatethebiasthatcanresultfromapproximatingthequadratic
analyticalresponseinEquation[31]withthestraightlinefunctioninEquation[30]orapproximatingthestraightlinefunctionin
Equation[30]withtheproportionalfunctioninEquation[29].ATOSTforbiasprovidesademonstrationthatthebiasinreported
concentrationsduetotheuseofthesimplercalibrationmodelislessthanthatrequiredbytheprocedureacrossthereportable
range.Thegreaterthebias,thegreatertheriskthesimplermodelisinadequate.
Thesecondapproachisbasedonasimpleresultofinformationtheory,thecorrectedAkaikeInformationCriterion(AICc).The
AICcprovidesasimpleindexthatcanbeusedtoselectbetweencompetingmodels.ComparisonoftheAICcstatisticamong
competingmodelswillidentifythemostparsimoniousmodel.Themostparsimoniousmodelprovidestheleastnumberof
parameterswhilestillprovidinganadequaterepresentationofthedata(27,28).ThelowertheAICcstatistic,thelowertherisk
thatthesimplermodelisinadequate.
6.6MethodologicalAssumptions
Therecommendedmethodologydependsonthefollowingassumptions:
Responsenormalizedvaluesarenormallydistributed.Thisassumptionensurestheaccuracyoftheconfidencelevelofthe
TOSTandisanunderlyingassumptionoftheAICcstatistic.Itisimportanttoverifythisassumptionbeforeusingthemethods
describedherebecauseresponsedeviationsinsomeanalyticalproceduresarebetterdescribedbyotherdistributions.Insuch
cases,alogtransformationoftheresponsemayresultinnormallydistributederrors.
Responsenormalizedvaluesareindependent.Thisassumptionimpliesthatallresponsemeasurementsareuncorrelated.
Correlationamongresponsemeasurementsobtainedinthesameanalyticalruncouldoccur,forinstance,ifresponse
measurementswereobtainedfrommultipleanalyticalrunsandtheanalyticalprocedurewassuchthatruntorunvariance
contributedimportantlytoprocedurevariance.Ifresponsedatamustbeobtainedfrommultipleruns,itisimportanttoensure
thatruntorunvariancedoesnotcontributesignificantlytoprocedureimprecision.Similarcautionsapplywhenthereare
multipleinstruments,operators,days,andreagentlots.
TheSDofthenormalizedvaluesisthesameateachcalibratorlevel.Inmanycases,theNSD,ratherthantheSD,isconstant
acrosstherangeofcalibratorlevels.Thismightbetrue,forinstance,withanalyticalprocedureswithnormalizedvaluesthat
followalognormaldistribution.Inthiscase,alogtransformationoftheresponsemayresultinaSDthatisapproximately
constantacrosstheanalyticalrange.
Themorecomplexmodelaccuratelyrepresentstheshapeofthecalibrationrelationship.Thisassumptionpresumesthata
definitiveproceduredevelopmentplanortheoreticalknowledgeisavailableconcerningtheshapeofthecalibrationrelationship.
Additionally,thesimplermodelmustbeaspecialcaseofthemorecomplexmodel.Thecomparisonbetweenthetwomodels
isonlymeaningfulifthecomplexmodelis,forallpracticalpurposes,anunbiasedrepresentationoftheresponse
concentrationrelationship.Forexample,thecomparisonbetweenquadraticandstraightlinemodelsisnotusefulifthetrue
relationshipissigmoidal.
Alinearityacceptancecriterionhasbeenpredetermined.Beforeconductingalinearityqualificationexperiment,thedecision
criterionmustbespecified.InthecaseoftheTOSTforbias,thiswilltaketheformofamaximumallowablebias,M,in
concentrationunits.Todemonstratethatthebiasforthesimplermodelislessthanthismaximum,theconfidenceintervalfor
biasattributabletononlinearitymustresideinanintervalfrom Mto+Macrosstheanalyticalrange.Toensureatestsizeof
5%,atwosided90%confidenceintervaliscomputedforthemaximumbiasacrosstheanalyticalrange.Inthecaseofa
decisionbasedonAICc,theAICcforthesimplermodelmustbelessthanorequaltothatofthemorecomplexmodel.
Ananalyticalrange(CmintoCmax )hasbeenproposed.Suggestedrangesforvariousapplicationsareprovidedinchapter 1225
.Theselectedcalibrationmodelmustbeadequateacrosstheproposedanalyticalrange.Ifnot,eitheramorecomplex
calibrationmodelisrequiredortheanalyticalrangemustbenarrowedtoaccommodatetheselectedcalibrationmodel.
ThefirstthreeassumptionssimplifytheTOSTcalculationsandmustholdwhenevertheusualstatisticaltestsassociatedwith
regressionareused.Ifmultiplereplicateresponsemeasurementsareobtainedateachcalibratorlevel,theseassumptionscan

beexaminedfromthelinearitydatathemselves.Ideally,theseassumptionswouldbejustifiedbyknowledgeacquiredduring
theearlydevelopmentstagesofananalyticalprocedure.
Itisrecognizedthatoneormoreoftheassumptionsnotedabovemaynotbejustified.Forinstance,itiscommontouse
multiplecalibrationrunsincalibrationstudies.Ifruntorunvarianceisappreciable,thengeneralizedregressionmethodsare
neededforanalysis(see,e.g.,(29)and(30)).Insuchcases,theresultingasymptoticvariancesandcovarianceestimatesstill
maybeappropriateforuseinmakingtheTOSTcomparisons.Equivalenceteststocomparethequalityofastraightlinefitto
thatofahigherorderpolynomial,undermoregeneralexperimentalconditionsthanthosedescribedhere,areavailable(31).
Thesemethodsuseorthogonalpolynomialsandgeneralizedpivotalquantitiestoestimatetheprobabilityofequivalence
betweenastraightlineandapolynomialmodelwithrespecttoeitherassaysignalorreportedconcentration.However,
implementationofthesemethodsrequiresspecialiststatisticalassistance.
6.7TwoOneSidedTestsofEquivalencetoEvaluateBiasinReportedConcentrations
Althoughitisalwaysdesirabletoeliminatebiasinreportedconcentrationsdeterminedbyananalyticalprocedure,sometimes
thisissimplynotpractical.Thebiasmaybetoosmalltobedetectedinareasonablylargeexperiment,ortheeffectofbiason
decisionrisksassociatedwiththereportableresultmaybenegligible,renderingitunnecessarytoreducebiasfurther.Insuch
cases,itmaybeacceptabletouseacalibrationmodelthatissimplerthanacomplexmodelthatcontributesnobiasbutmay
requiremorecalibrationlevelsandmorecomplexdatareduction.Whensomelevelofbiascanbetolerated,theobjectivesofa
linearityassessmentaretoestimatethemagnitudeofthebiasacrosstheconcentrationrangeandprovideevidencethatthe
magnitudeofthebiasisbelowthemaximumthatcanbetolerated.
TheTOSTapproachassumesthatthebiasassociatedwiththesimplermodelisunacceptable,unlesscontradictedbydata
(32).Thisisareversaloftheusualhypothesistestingschemeandrepresentsahighevidentialstandard,placingtheburdenof
proofonthelaboratorytodemonstratingthatthebiasisacceptable.Thisapproachfavorsexperimentswithadequatesample
sizesandlowervariability,incontrasttostandardsignificancetesting,whereimprovedprecisionandincreasedtestingcanbe
penalized.
Someguidanceonselectingthenumberofcalibratorlevelsandthenumberofreplicateswithineachcalibratorlevelisprovided
inLeblondetal.(30).Therequiredsamplesizesdependonthemagnitudeofbiasandanticipatedlevelofprocedureprecision,
aswellastherequiredlimitsforbiasacrosstheanalyticalrange.
6.7.1TOSTFORBIASWHENAPPROXIMATINGAQUADRATICWITHASTRAIGHTLINEMODEL

TheTOSTforbiasrecommendedherereliesonFieller'smethod(33).Tosimplifycalculations,orthogonalpolynomial(OP)
transformationsofEquations[30]and[31]areusedbyYangetal.(34).TheformulaeintheAppendixprovidetheintermediate
quantities

,andgQfromexperimentaldata.

Apointestimateforthebiasattheconcentration(x)atwhichthemaximumabsolutebiasisgreatestis

A90%confidenceintervalforbias xbasedonFieller'stheorem,is

Ifthecomputed90%confidenceintervalforthebias xiscontainedwithintheallowedbiasfortheprocedure,thenthesimpler

(straightline)modelistakenasanadequatecalibrationmodel.Otherwise,thequadraticmodelisselected.
6.7.2TOSTFORBIASWHENAPPROXIMATINGASTRAIGHTLINEMODELWITHAPROPORTIONALMODEL

ATOSTforbiaswhenapproximatingastraightlinemodelwithaproportionalmodelcanbeobtainedusingFieller'smethodand
orthogonalpolynomialsinamannersimilartothatintheprevioussection.TheformulaeintheAppendixprovidethe
intermediatequantitiesjmax ,wjmax ,L,R,U,B(wjmax ,L),B(wjmax ,R),andB(wjmax ,U)fromexperimentaldata.
Apointestimateforthebiasattheconcentration(x)atwhichthemaximumabsolutebiasisgreatestis

A90%confidenceinterval(LowerUpper)forthebias x,basedonFieller'stheoremis

Ifthe90%computedconfidenceintervalforbiasiscontainedwithintheallowedbiasfortheprocedure,thenthesimpler
(proportional)modelistakenasanadequatecalibrationmodel.Otherwise,thestraightlinemodelisselected.
6.8CorrectedAkaikeInformationCriterionforModelSelection
ThecorrectedAkaikeInformationCriterion(AICc)wasdevelopedforuseinmodelselectionbasedoninformationtheory.
Unlikeapproachesbasedonhypothesistestingwhichgivethebenefitofdoubttothesimplermodel,theAICcdirectly
evaluatestheparsimoniesofallcandidatemodelsgiventhedataathand.TheAICcisappropriateforthesmallersamplesizes
usuallyavailableforcalibrationstudies.TheAICcforagivencalibrationmodeliseasilycalculated.

whereNisthetotalnumberofdatapoints,Kisthetotalnumberofestimatedregressionparametersinthemodel(includingthe
populationpureerrorvariance),ln()indicatesthenaturallogfunction(LOGinExcel),SSEistheresidualsumofsquares
obtainedfromtheleastsquaresregression,and isthemodelpredictedresponseattheithconcentration.Themodelwith
thesmallerAICcisthepreferredmodelbrcauseitismoreparsimonious.
Theinformationcriteriaapproachexplicitlyacknowledgesthefactthatgivenmoredata,amorecomplicatedmodelmaywellbe
moreappropriate.TheAICccriterionisbasedonmodelparsimony,giventheavailabledata.Incontrast,theTOSTprocedure
incorporatestwoadditionalconsiderations:1)theamountofbiasinreportedresultsthatcanbetoleratedbyalesscomplicated
calibrationmodel,and2)theconfidencelevelatwhichsuchbiascanbeestimated.BecauseAICcandTOSTidentifythe
bestmodelbasedondifferentprinciples,theymayleadtodifferentconclusions.Theteamresponsibleforprocedure
validationmustdecidewhethertheadditionalconsiderationsaffordedbytheTOSTareneeded.
6.9Examples
ThissectionpresentssimpleexamplestoillustratetheTOSTandAICcapproachesusingonlyasinglecalibrationrun.Inthis

case,allreplicatemeasurementsaretakentobeindependent.Withthisassumption,simpleregressionmethodsaresufficient.
Thecalculationsbelowcanbeimplementedeasilyinaspreadsheet(30).Thesameexampledataareusedforbothmodel
comparisons.TheTOSTcriterionassumesthattheproceduremustmaintainabsolutebias,inconcentration(x)units,below8
overtherangeofconcentrationsfrom0to100.
6.9.1TOSTANDAICcCOMPARISONOFQUADRATICANDSTRAIGHTLINEMODELS
Inthefirstexample,weconsidertestingforconcentrationbiasduetoapproximatingaquadraticmodelwithastraightline
modelthisiscomparisonofEquation[30]toEquation[31].Considerthecalibrationdataandassociatedcalculated
intermediateresultsshowninTable7andTable8.FormulasareshownintheAppendix.
Table7.ExampleCalibrationDataandIntermediateCalculationstoCompareQuadraticandStraightLineModels

i
1
2
3
4
5
6
7
8
9
10
11
12
13
14
N=15

xi

Yi

0
0
0
20
20
20
50
50
50
70
70
70
100
100
100

12.08
1.47
3.23
28.43
31.75
28.57
56.00
46.74
48.89
73.80
74.69
72.86
97.95
117.79
108.59

f0i
[38]
0.2582
0.2582
0.2582
0.2582
0.2582
0.2582
0.2582
0.2582
0.2582
0.2582
0.2582
0.2582
0.2582
0.2582
0.2582

h1i
[40]
48
48
48
28
28
28
2
2
2
22
22
22
52
52
52

f1i
[42]
0.3497
0.3497
0.3497
0.2040
0.2040
0.2040
0.0146
0.0146
0.0146
0.1603
0.1603
0.1603
0.3788
0.3788
0.3788

h2i
[45]
1191.08
1191.08
1191.08
388.54
388.54
388.54
1257.96
1257.96
1257.96
837.58
837.58
837.58
1292.99
1292.99
1292.99

f2i
[47]
0.2925
0.2925
0.2925
0.0954
0.0954
0.0954
0.3090
0.3090
0.3090
0.2057
0.2057
0.2057
0.3176
0.3176
0.3176

[51]
7.82
7.82
7.82
24.99
24.99
24.99
53.33
53.33
53.33
73.96
73.96
73.96
107.51
107.51
107.51

Table8.IntermediateCalculationswithCmin=0andCmax =100

j
1
2
3
4
5
...
997
998
999
1000
1001

wj
[54]
0.0000
0.1
0.2
0.3
0.4
...
99.6
99.7
99.8
99.9
100.05904

StatisticsfromthisdatasetaregiveninTable9.

d1j
[55]
0.3497
0.3490
0.3482
0.3475
0.3468
...
0.3759
0.3767
0.3774
0.3781
0.3788

d2j
[56]
0.2925
0.2901
0.2877
0.2853
0.2829
...
0.3077
0.3102
0.3126
0.3151
0.3176

0.2925
0.2901
0.2877
0.2853
0.2829
...
0.3077
0.3102
0.3126
0.3151
0.3176

[60]
5.76
5.76
5.76
25.66
25.66
25.66
55.51
55.51
55.51
75.41
75.41
75.41
105.27
105.27
105.27

Table9.IntermediateStatisticsCalculatedfortheTOSTTesttoCompareQuadraticandStraightLineModels

Statistic
N
s1

Equation

[39]

Result
15
185.9

c1

[41]

137.26

s2

[43]

13787.82

s3

[44]

13586.01

c2

[46]

4071.42

[48]

207.2924

[49]

136.5863

[50]

7.0534

[52]

5.9185

g0

[53]

0.0060

m0

[58]

43.5904

Cmax

100.0

[32]

2.2511

[33]

1.1165to5.6456

90%CIforbias

Becausethe90%CIforbiasinconcentrationiscontainedwithinthemaximumallowedbiasof8,thesimpler(straightline)
modelisadequateforcalibration.
ThedatainTable7canalsobeusedtocompareaquadraticandastraightlinemodelusingtheAICcobtainedfromthe
Equation[36].ForthequadraticandstraightlinemodelsK=4and3,respectively.Also,SSE=420.3and470.1forthe
quadraticandstraightlinemodels,respectively.ApplyingEquation[36],AICc=62.0and59.9,respectively.BytheAICc,the
straightlinemodelisamoreparsimoniouscalibrationmodelthanisthequadraticmodelbecauseithasalesserAICcvalue.
Accordingly,thestraightlinemodelwouldbeselectedforcalibrationbytheAICc.ThisconclusionisthesameastheTOST
procedureconclusion,althoughtheTOSTandAICcconclusionswillnotalwaysbeexpectedtoagree.
6.9.2TOSTANDAICcCOMPARISONOFSTRAIGHTLINEANDPROPORTIONALMODELS
Thesecondexampleseekstodeterminewhetheraproportionalmodel,(Equation[29])isanappropriateapproximationofa
straightlinemodel(Equation[30]).Assumethattheproceduremustmaintainabsolutebiasbelow8overtherangeof
concentrationsfrom0to100,andthatasinglepointcalibrationwithastandardofconcentration100(=Xstd)isdesired.Manyof
thecalculationsneededforthiscomparisonhavealreadybeengivenintheprevioussection.Additionalstatisticsneededare
showninTables10,11,and12.
Table10.IntermediateCalculations

i
1

0.00

2
3
4
5
6
7
8
9
10
11
12
13
14
15

0.00
0.00
21.45
21.45
21.45
53.64
53.64
53.64
75.09
75.09
75.09
107.27
107.27
107.27

Table11.CalculationWorksheet

j
1
2
3
4
5
...
997
998
999
1000
1001

wj
[54]
0.000
0.100
0.200
0.300
0.400
...
99.600
99.700
99.800
99.900
100.000

B(wj,L)
[67]
0.733
0.732
0.731
0.730
0.730
...
0.003
0.002
0.001
0.001
0.000

B(wj,R)
[67]
5.470
5.464
5.459
5.453
5.448
...
0.022
0.016
0.011
0.005
0.000

B(wj,U)
[67]
9.775
9.766
9.756
9.746
9.736
...
0.039
0.029
0.020
0.010
0.000

Bmaxj
[68]
9.775
9.766
9.756
9.746
9.736
...
0.039
0.029
0.020
0.010
0.000

Table12.FinalResults

Statistic
X

Equation
[59]
[62]

Result
48
6.013

gS

[63]

0.006

R
L
U
xstd

[64]
[65]
[66]

1.518
1.375
1.660
100

ValueofXatmaximumbias
biasx

[34]

0
5.470

90%CIofbiasx

[35]

0.733to9.775

Becausethe90%CIofbias,0.733to9.775,isnotcontainedwithinthemaximumbiaslimitsof8,thesimplermodel
(proportional)isnotadequateasacalibrationmodel,andthestraightlinemodelshouldbeselectedbasedonthiscriterion.

ThesedatacanalsobeusedtocomparestraightlineandproportionalmodelsusingtheAICcparsimonycriterion.Forthe
straightlineandproportionalmodelsK=3and2,respectively.Equation[37]yieldsSSE=470.1and645.5,respectively.
ApplyingEquation[36]yieldsAICc=59.9and61.4,respectively.BytheAICc,thestraightlinemodelisamoreparsimonious
calibrationmodelthanistheproportionalmodelbecauseithasthelesserAICcvalue.

7.APPENDIX
ThisappendixprovidesformulaeusedforthecalibrationlinearitybiasTOSTtests.TheTOSTforbiasrecommendedhere
reliesonFieller'smethod(33).Orthogonalpolynomial(OP)transformationsofEquations[30]and[31]areusedtosimplify
calculations.Thefollowingcalculationsprovidethe90%confidenceintervalforbiasattheconcentration(x),atwhichthe
maximumabsolutebiasisgreatest.Theintermediatestatisticsaremeanttobecalculatedintheordergivenbelow.

Equations[48]to[50]giveestimatesfortheintercept,linear,andquadraticOPcoefficients,respectively.

Thepredictedresponsebasedonthequadraticmodelisgivenby

AnestimateoftheSDoftherandomimprecisioncomponentforthequadraticmodelisgivenby

wheret0.95:N3representsthepercentileofacentraltdistributionwitharea0.95totheleftandN 3degreesoffreedom.
Equations[54]to[57]identifytheconcentrationatwhichtheabsolutebiasisgreatest,bysamplingacrosstheconcentration
rangefromCmintoCminoverafinegridof1001evenlyspacedpoints.

wherej=1,...1001,jmax isthevalueofjforwhichtheabsolutevalueofd2jismaximum,and
m0=c 1d2jmax

[58]

ATOSTforbiaswhenapproximatingastraightlinemodelwithaproportionalmodelcanbeobtainedusingFieller'smethodand
orthogonalpolynomialsinamannersimilartothatintheprevioussection.StatisticsaregeneratedusingEquations[38]to
[42],[48],[49],and[54]asintheprevioussection.Inaddition,

PredictedresponsevaluesbasedonthestraightlineandproportionalmodelsaregivenbyEquations[60]and[61]below.

AnestimateoftheSDoftherandomimprecisioncomponentforthestraightlinemodelisgivenby

Tosimplifypresentationoftheseformulae,wedefinethefunctionB(h,p),as

wherex STDisthefixedstandardconcentrationintendedtobeusedinsinglepointcalibration.Substitutingvaluesforhandp
intothisfunctionweobtainB(wj,L)forj=1,...1001,and

ThefollowingstepfindsthevalueofjatwhichBmax jismaximized.
Finally,thequantitiesB(wjmax ,L),B(wjmax ,R),andB(wjmax ,U)areobtainedbysubstitutingtheindicatedvaluesforhandpinto
Equation[67].

8.REFERENCES
1. KruskalW,WallisA.Useofranksinonecriterionvarianceanalysis.JAmStatAssoc.195247:583621.
2. LeveneH.Robusttestsforequalityofvariances.In:Contributionstoprobabilityandstatistics:essaysinhonorof
HaroldHotelling.PaloAlto,CA:StanfordUniversityPress1960.p278292.
3. BartlettMS.Propertiesofsufficiencyandstatisticaltests.ProceedRStatSocLondA.1937160(901):268282.
4. ShapiroSS,WilkMB.Ananalysisofvariancetestfornormality(completesamples).Biometrika.196552(34):591
611.
5. BarnettV,LewisT.Outliersinstatisticaldata.In:Wileyseriesinprobabilityandmathematicalstatistics.3rded.New
York:JohnWiley&Sons1994.
6. JCGMmemberorganizations(BIPM,IEC,IFCC,ILAC,ISO,IUPAC,IUPAP,andOIML).Internationalvocabularyof
metrologybasicandgeneralconceptsandassociatedterms(VIM).3rded.Geneva:JCGM200:2012.
7. GraybillFA,WangCM.Confidenceintervalsonnonnegativelinearcombinationsofvariances.JAmStatAssoc.
198075:869873.
8. NijhuisMB,VandenHeuvelER.Closedformconfidenceintervalsonmeasuresofprecisionforaninterlaboratory
study.JBiopharmStat.200717(1):123142.
9. SchuirmannDJ.Acomparisonofthetwoonesidedtestsprocedureandthepowerapproachforassessingthe
equivalenceofaveragebioavailability.JPharmacokinetiBiopharmaceut.198715(6):657680.
10. HubertP,NguyenHuuJJ,BoulangerB,etal.Harmonizationofstrategiesforthevalidationofquantitativeanalytical
procedures.ASFSTPproposalpartI.JPharmBiomedAnal.200436(3):579586.
11. HubertP,NguyenHuuJJ,BoulangerB,etal.Harmonizationofstrategiesforthevalidationofquantitativeanalytical
procedures.ASFSTPproposalpartII.JPharmBiomedAnal.200745(1):7081.
12. HubertP,NguyenHuuJJ,BoulangerB,etal.Harmonizationofstrategiesforthevalidationofquantitativeanalytical
procedures.ASFSTPproposalpartIII.JPharmBiomedAnal.200745(1):8296.
13. MeeRW.BetaexpectationandbetacontenttolerancelimitsforbalancedonewayANOVArandommodel.
Technometrics.198426(3):251254.
14. HahnGJ,MeekerWQ.Statisticalintervals:aguideforpractitioners.NewYork:JohnWiley&Sons1991.p.412.
15. HoffmanD,KringleR.Atotalerrorapproachforthevalidationofquantitativeanalyticalmethods.PharmRes.

16.
17.
18.
19.
20.

21.
22.
23.
24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
34.

200724(6):11571164.
WolfingerRD.ToleranceintervalsforvariancecomponentmodelsusingBayesiansimulation.JQualityTechnol.
199830(1):1832.
NtzoufrasI.BayesianmodelingusingWinBUGS:anintroduction.NewYork:JohnWiley&Sons2009.
SpiegelhalterD,ThomasA,BestNG,GilksWR.BUGS0.5examples,volume1,version1.1996..
BurdickRK,LeBlondDJ,SandellD,YangH.Statisticalmethodsforvalidationofprocedureaccuracyandprecision.
Stimulitotherevisionprocess.PharmacopeialForum.201339(3).
ICH.ICHharmonizedtripartiteguideline.Validationofanalyticalprocedures:textandmethodologyQ2(R1).2005.
http://www.ich.org/fileadmin/Public_Web_Site/ICH_Products/Guidelines/Quality/Q2_R1/Step4/Q2_R1__Guideline.pdf.
Accessed21April2014.
AnscombeFJ.Graphsinstatisticalanalysis.AmStatistician.197327(1):1721.
VanLocoJ,ElskensM,CrouxC,BeernaertH.Practitioner'sreport.Linearityofcalibrationcurves:useandmisuseof
thecorrelationcoefficient.AccredQualAssur.20027:281285.
BrggemannL,QuappW,WennrichR.Testfornonlinearityconcerninglinearcalibratedchemicalmeasurements.
AccredQualAssur.200611(12):625631.
MandelJ.Thestatisticalanalysisofexperimentaldata.In:Doverbooksonmathematics.NewYork:JohnWiley&
Sons1964.
MarkH,WorkmanJ.Chemometricsinspectroscopy.Chapter27:Linearityincalibration.London:ElsevierPress,
AcademicPressasanimprint2007.[OriginallypublishedinSpectroscopy.199813(6):1921.]
LiuJ,HsiehE.Evaluationoflinearityinassayvalidation.EncyclopediaofBiopharmaceuticalStatistics.s.l.:Informa
Healthcare,2010,p467474.{couldnotfind}
BurnhamKP,AndersonDR.Modelselectionandmultimodelinference:apracticalinformationtheoreticapproach.
2nded.NewYork:Springer,2002.
BurnhamKP,AndersonDR.Multimodelinference:understandingAICandBICinmodelselection.Sociological
MethodsRes.200433(2):261304.
LeBlondD,TanCY,YangH.Confirmationofanalyticalmethodcalibrationlinearity.Stimulitotherevisionprocess.
PharmacopeialForum.201339(3).
LeBlondD,TanCY,YangH.Confirmationofanalyticalmethodcalibrationlinearity:practicalapplication.Stimulitothe
revisionprocess.PharmacopeialForum.201339(5).
NovickS,YangH.Directlytestingthelinearityassumptionforassayvalidation.JChemometrics.2013:27(5):117
125.{couldnotfind}
BergerRL,HsuJC.Bioequivalencetrials,intersectionuniontestsandequivalenceconfidencesets.StatSci.
199611(4):283319.
FinneyDJ.Statisticalmethodinbiologicalassay.2nded.London:CharlesGriffin&Co.,Ltd1952.
YangH,NovickSJ,LeBlondD.Testinglinearityundergeneralexperimentalconditions.JBiopharmStat.Acceptedfor
publication. 2S(USP38)

AuxiliaryInformationPleasecheckforyourquestionintheFAQsbeforecontactingUSP.